Merge master.kernel.org:/pub/scm/linux/kernel/git/lethal/sh-2.6
[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 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 /*
286  * Check whether bios must be queued in the device-mapper core rather
287  * than here in the target.
288  *
289  * m->lock must be held on entry.
290  *
291  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
292  * same value then we are not between multipath_presuspend()
293  * and multipath_resume() calls and we have no need to check
294  * for the DMF_NOFLUSH_SUSPENDING flag.
295  */
296 static int __must_push_back(struct multipath *m)
297 {
298         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
299                 dm_noflush_suspending(m->ti));
300 }
301
302 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
303                   unsigned was_queued)
304 {
305         int r = DM_MAPIO_REMAPPED;
306         unsigned long flags;
307         struct pgpath *pgpath;
308
309         spin_lock_irqsave(&m->lock, flags);
310
311         /* Do we need to select a new pgpath? */
312         if (!m->current_pgpath ||
313             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
314                 __choose_pgpath(m);
315
316         pgpath = m->current_pgpath;
317
318         if (was_queued)
319                 m->queue_size--;
320
321         if ((pgpath && m->queue_io) ||
322             (!pgpath && m->queue_if_no_path)) {
323                 /* Queue for the daemon to resubmit */
324                 bio_list_add(&m->queued_ios, bio);
325                 m->queue_size++;
326                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
327                     !m->queue_io)
328                         queue_work(kmultipathd, &m->process_queued_ios);
329                 pgpath = NULL;
330                 r = DM_MAPIO_SUBMITTED;
331         } else if (pgpath)
332                 bio->bi_bdev = pgpath->path.dev->bdev;
333         else if (__must_push_back(m))
334                 r = DM_MAPIO_REQUEUE;
335         else
336                 r = -EIO;       /* Failed */
337
338         mpio->pgpath = pgpath;
339
340         spin_unlock_irqrestore(&m->lock, flags);
341
342         return r;
343 }
344
345 /*
346  * If we run out of usable paths, should we queue I/O or error it?
347  */
348 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
349                             unsigned save_old_value)
350 {
351         unsigned long flags;
352
353         spin_lock_irqsave(&m->lock, flags);
354
355         if (save_old_value)
356                 m->saved_queue_if_no_path = m->queue_if_no_path;
357         else
358                 m->saved_queue_if_no_path = queue_if_no_path;
359         m->queue_if_no_path = queue_if_no_path;
360         if (!m->queue_if_no_path && m->queue_size)
361                 queue_work(kmultipathd, &m->process_queued_ios);
362
363         spin_unlock_irqrestore(&m->lock, flags);
364
365         return 0;
366 }
367
368 /*-----------------------------------------------------------------
369  * The multipath daemon is responsible for resubmitting queued ios.
370  *---------------------------------------------------------------*/
371
372 static void dispatch_queued_ios(struct multipath *m)
373 {
374         int r;
375         unsigned long flags;
376         struct bio *bio = NULL, *next;
377         struct mpath_io *mpio;
378         union map_info *info;
379
380         spin_lock_irqsave(&m->lock, flags);
381         bio = bio_list_get(&m->queued_ios);
382         spin_unlock_irqrestore(&m->lock, flags);
383
384         while (bio) {
385                 next = bio->bi_next;
386                 bio->bi_next = NULL;
387
388                 info = dm_get_mapinfo(bio);
389                 mpio = info->ptr;
390
391                 r = map_io(m, bio, mpio, 1);
392                 if (r < 0)
393                         bio_endio(bio, bio->bi_size, r);
394                 else if (r == DM_MAPIO_REMAPPED)
395                         generic_make_request(bio);
396                 else if (r == DM_MAPIO_REQUEUE)
397                         bio_endio(bio, bio->bi_size, -EIO);
398
399                 bio = next;
400         }
401 }
402
403 static void process_queued_ios(struct work_struct *work)
404 {
405         struct multipath *m =
406                 container_of(work, struct multipath, process_queued_ios);
407         struct hw_handler *hwh = &m->hw_handler;
408         struct pgpath *pgpath = NULL;
409         unsigned init_required = 0, must_queue = 1;
410         unsigned long flags;
411
412         spin_lock_irqsave(&m->lock, flags);
413
414         if (!m->queue_size)
415                 goto out;
416
417         if (!m->current_pgpath)
418                 __choose_pgpath(m);
419
420         pgpath = m->current_pgpath;
421
422         if ((pgpath && !m->queue_io) ||
423             (!pgpath && !m->queue_if_no_path))
424                 must_queue = 0;
425
426         if (m->pg_init_required && !m->pg_init_in_progress) {
427                 m->pg_init_required = 0;
428                 m->pg_init_in_progress = 1;
429                 init_required = 1;
430         }
431
432 out:
433         spin_unlock_irqrestore(&m->lock, flags);
434
435         if (init_required)
436                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
437
438         if (!must_queue)
439                 dispatch_queued_ios(m);
440 }
441
442 /*
443  * An event is triggered whenever a path is taken out of use.
444  * Includes path failure and PG bypass.
445  */
446 static void trigger_event(struct work_struct *work)
447 {
448         struct multipath *m =
449                 container_of(work, struct multipath, trigger_event);
450
451         dm_table_event(m->ti->table);
452 }
453
454 /*-----------------------------------------------------------------
455  * Constructor/argument parsing:
456  * <#multipath feature args> [<arg>]*
457  * <#hw_handler args> [hw_handler [<arg>]*]
458  * <#priority groups>
459  * <initial priority group>
460  *     [<selector> <#selector args> [<arg>]*
461  *      <#paths> <#per-path selector args>
462  *         [<path> [<arg>]* ]+ ]+
463  *---------------------------------------------------------------*/
464 struct param {
465         unsigned min;
466         unsigned max;
467         char *error;
468 };
469
470 static int read_param(struct param *param, char *str, unsigned *v, char **error)
471 {
472         if (!str ||
473             (sscanf(str, "%u", v) != 1) ||
474             (*v < param->min) ||
475             (*v > param->max)) {
476                 *error = param->error;
477                 return -EINVAL;
478         }
479
480         return 0;
481 }
482
483 struct arg_set {
484         unsigned argc;
485         char **argv;
486 };
487
488 static char *shift(struct arg_set *as)
489 {
490         char *r;
491
492         if (as->argc) {
493                 as->argc--;
494                 r = *as->argv;
495                 as->argv++;
496                 return r;
497         }
498
499         return NULL;
500 }
501
502 static void consume(struct arg_set *as, unsigned n)
503 {
504         BUG_ON (as->argc < n);
505         as->argc -= n;
506         as->argv += n;
507 }
508
509 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
510                                struct dm_target *ti)
511 {
512         int r;
513         struct path_selector_type *pst;
514         unsigned ps_argc;
515
516         static struct param _params[] = {
517                 {0, 1024, "invalid number of path selector args"},
518         };
519
520         pst = dm_get_path_selector(shift(as));
521         if (!pst) {
522                 ti->error = "unknown path selector type";
523                 return -EINVAL;
524         }
525
526         r = read_param(_params, shift(as), &ps_argc, &ti->error);
527         if (r)
528                 return -EINVAL;
529
530         r = pst->create(&pg->ps, ps_argc, as->argv);
531         if (r) {
532                 dm_put_path_selector(pst);
533                 ti->error = "path selector constructor failed";
534                 return r;
535         }
536
537         pg->ps.type = pst;
538         consume(as, ps_argc);
539
540         return 0;
541 }
542
543 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
544                                struct dm_target *ti)
545 {
546         int r;
547         struct pgpath *p;
548
549         /* we need at least a path arg */
550         if (as->argc < 1) {
551                 ti->error = "no device given";
552                 return NULL;
553         }
554
555         p = alloc_pgpath();
556         if (!p)
557                 return NULL;
558
559         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
560                           dm_table_get_mode(ti->table), &p->path.dev);
561         if (r) {
562                 ti->error = "error getting device";
563                 goto bad;
564         }
565
566         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
567         if (r) {
568                 dm_put_device(ti, p->path.dev);
569                 goto bad;
570         }
571
572         return p;
573
574  bad:
575         free_pgpath(p);
576         return NULL;
577 }
578
579 static struct priority_group *parse_priority_group(struct arg_set *as,
580                                                    struct multipath *m)
581 {
582         static struct param _params[] = {
583                 {1, 1024, "invalid number of paths"},
584                 {0, 1024, "invalid number of selector args"}
585         };
586
587         int r;
588         unsigned i, nr_selector_args, nr_params;
589         struct priority_group *pg;
590         struct dm_target *ti = m->ti;
591
592         if (as->argc < 2) {
593                 as->argc = 0;
594                 ti->error = "not enough priority group aruments";
595                 return NULL;
596         }
597
598         pg = alloc_priority_group();
599         if (!pg) {
600                 ti->error = "couldn't allocate priority group";
601                 return NULL;
602         }
603         pg->m = m;
604
605         r = parse_path_selector(as, pg, ti);
606         if (r)
607                 goto bad;
608
609         /*
610          * read the paths
611          */
612         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
613         if (r)
614                 goto bad;
615
616         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
617         if (r)
618                 goto bad;
619
620         nr_params = 1 + nr_selector_args;
621         for (i = 0; i < pg->nr_pgpaths; i++) {
622                 struct pgpath *pgpath;
623                 struct arg_set path_args;
624
625                 if (as->argc < nr_params)
626                         goto bad;
627
628                 path_args.argc = nr_params;
629                 path_args.argv = as->argv;
630
631                 pgpath = parse_path(&path_args, &pg->ps, ti);
632                 if (!pgpath)
633                         goto bad;
634
635                 pgpath->pg = pg;
636                 list_add_tail(&pgpath->list, &pg->pgpaths);
637                 consume(as, nr_params);
638         }
639
640         return pg;
641
642  bad:
643         free_priority_group(pg, ti);
644         return NULL;
645 }
646
647 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
648 {
649         int r;
650         struct hw_handler_type *hwht;
651         unsigned hw_argc;
652         struct dm_target *ti = m->ti;
653
654         static struct param _params[] = {
655                 {0, 1024, "invalid number of hardware handler args"},
656         };
657
658         r = read_param(_params, shift(as), &hw_argc, &ti->error);
659         if (r)
660                 return -EINVAL;
661
662         if (!hw_argc)
663                 return 0;
664
665         hwht = dm_get_hw_handler(shift(as));
666         if (!hwht) {
667                 ti->error = "unknown hardware handler type";
668                 return -EINVAL;
669         }
670
671         m->hw_handler.md = dm_table_get_md(ti->table);
672         dm_put(m->hw_handler.md);
673
674         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
675         if (r) {
676                 dm_put_hw_handler(hwht);
677                 ti->error = "hardware handler constructor failed";
678                 return r;
679         }
680
681         m->hw_handler.type = hwht;
682         consume(as, hw_argc - 1);
683
684         return 0;
685 }
686
687 static int parse_features(struct arg_set *as, struct multipath *m)
688 {
689         int r;
690         unsigned argc;
691         struct dm_target *ti = m->ti;
692
693         static struct param _params[] = {
694                 {0, 1, "invalid number of feature args"},
695         };
696
697         r = read_param(_params, shift(as), &argc, &ti->error);
698         if (r)
699                 return -EINVAL;
700
701         if (!argc)
702                 return 0;
703
704         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
705                 return queue_if_no_path(m, 1, 0);
706         else {
707                 ti->error = "Unrecognised multipath feature request";
708                 return -EINVAL;
709         }
710 }
711
712 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
713                          char **argv)
714 {
715         /* target parameters */
716         static struct param _params[] = {
717                 {1, 1024, "invalid number of priority groups"},
718                 {1, 1024, "invalid initial priority group number"},
719         };
720
721         int r;
722         struct multipath *m;
723         struct arg_set as;
724         unsigned pg_count = 0;
725         unsigned next_pg_num;
726
727         as.argc = argc;
728         as.argv = argv;
729
730         m = alloc_multipath(ti);
731         if (!m) {
732                 ti->error = "can't allocate multipath";
733                 return -EINVAL;
734         }
735
736         r = parse_features(&as, m);
737         if (r)
738                 goto bad;
739
740         r = parse_hw_handler(&as, m);
741         if (r)
742                 goto bad;
743
744         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
745         if (r)
746                 goto bad;
747
748         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
749         if (r)
750                 goto bad;
751
752         /* parse the priority groups */
753         while (as.argc) {
754                 struct priority_group *pg;
755
756                 pg = parse_priority_group(&as, m);
757                 if (!pg) {
758                         r = -EINVAL;
759                         goto bad;
760                 }
761
762                 m->nr_valid_paths += pg->nr_pgpaths;
763                 list_add_tail(&pg->list, &m->priority_groups);
764                 pg_count++;
765                 pg->pg_num = pg_count;
766                 if (!--next_pg_num)
767                         m->next_pg = pg;
768         }
769
770         if (pg_count != m->nr_priority_groups) {
771                 ti->error = "priority group count mismatch";
772                 r = -EINVAL;
773                 goto bad;
774         }
775
776         return 0;
777
778  bad:
779         free_multipath(m);
780         return r;
781 }
782
783 static void multipath_dtr(struct dm_target *ti)
784 {
785         struct multipath *m = (struct multipath *) ti->private;
786
787         flush_workqueue(kmultipathd);
788         free_multipath(m);
789 }
790
791 /*
792  * Map bios, recording original fields for later in case we have to resubmit
793  */
794 static int multipath_map(struct dm_target *ti, struct bio *bio,
795                          union map_info *map_context)
796 {
797         int r;
798         struct mpath_io *mpio;
799         struct multipath *m = (struct multipath *) ti->private;
800
801         if (bio_barrier(bio))
802                 return -EOPNOTSUPP;
803
804         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
805         dm_bio_record(&mpio->details, bio);
806
807         map_context->ptr = mpio;
808         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
809         r = map_io(m, bio, mpio, 0);
810         if (r < 0 || r == DM_MAPIO_REQUEUE)
811                 mempool_free(mpio, m->mpio_pool);
812
813         return r;
814 }
815
816 /*
817  * Take a path out of use.
818  */
819 static int fail_path(struct pgpath *pgpath)
820 {
821         unsigned long flags;
822         struct multipath *m = pgpath->pg->m;
823
824         spin_lock_irqsave(&m->lock, flags);
825
826         if (!pgpath->path.is_active)
827                 goto out;
828
829         DMWARN("Failing path %s.", pgpath->path.dev->name);
830
831         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
832         pgpath->path.is_active = 0;
833         pgpath->fail_count++;
834
835         m->nr_valid_paths--;
836
837         if (pgpath == m->current_pgpath)
838                 m->current_pgpath = NULL;
839
840         queue_work(kmultipathd, &m->trigger_event);
841
842 out:
843         spin_unlock_irqrestore(&m->lock, flags);
844
845         return 0;
846 }
847
848 /*
849  * Reinstate a previously-failed path
850  */
851 static int reinstate_path(struct pgpath *pgpath)
852 {
853         int r = 0;
854         unsigned long flags;
855         struct multipath *m = pgpath->pg->m;
856
857         spin_lock_irqsave(&m->lock, flags);
858
859         if (pgpath->path.is_active)
860                 goto out;
861
862         if (!pgpath->pg->ps.type) {
863                 DMWARN("Reinstate path not supported by path selector %s",
864                        pgpath->pg->ps.type->name);
865                 r = -EINVAL;
866                 goto out;
867         }
868
869         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
870         if (r)
871                 goto out;
872
873         pgpath->path.is_active = 1;
874
875         m->current_pgpath = NULL;
876         if (!m->nr_valid_paths++ && m->queue_size)
877                 queue_work(kmultipathd, &m->process_queued_ios);
878
879         queue_work(kmultipathd, &m->trigger_event);
880
881 out:
882         spin_unlock_irqrestore(&m->lock, flags);
883
884         return r;
885 }
886
887 /*
888  * Fail or reinstate all paths that match the provided struct dm_dev.
889  */
890 static int action_dev(struct multipath *m, struct dm_dev *dev,
891                       action_fn action)
892 {
893         int r = 0;
894         struct pgpath *pgpath;
895         struct priority_group *pg;
896
897         list_for_each_entry(pg, &m->priority_groups, list) {
898                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
899                         if (pgpath->path.dev == dev)
900                                 r = action(pgpath);
901                 }
902         }
903
904         return r;
905 }
906
907 /*
908  * Temporarily try to avoid having to use the specified PG
909  */
910 static void bypass_pg(struct multipath *m, struct priority_group *pg,
911                       int bypassed)
912 {
913         unsigned long flags;
914
915         spin_lock_irqsave(&m->lock, flags);
916
917         pg->bypassed = bypassed;
918         m->current_pgpath = NULL;
919         m->current_pg = NULL;
920
921         spin_unlock_irqrestore(&m->lock, flags);
922
923         queue_work(kmultipathd, &m->trigger_event);
924 }
925
926 /*
927  * Switch to using the specified PG from the next I/O that gets mapped
928  */
929 static int switch_pg_num(struct multipath *m, const char *pgstr)
930 {
931         struct priority_group *pg;
932         unsigned pgnum;
933         unsigned long flags;
934
935         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
936             (pgnum > m->nr_priority_groups)) {
937                 DMWARN("invalid PG number supplied to switch_pg_num");
938                 return -EINVAL;
939         }
940
941         spin_lock_irqsave(&m->lock, flags);
942         list_for_each_entry(pg, &m->priority_groups, list) {
943                 pg->bypassed = 0;
944                 if (--pgnum)
945                         continue;
946
947                 m->current_pgpath = NULL;
948                 m->current_pg = NULL;
949                 m->next_pg = pg;
950         }
951         spin_unlock_irqrestore(&m->lock, flags);
952
953         queue_work(kmultipathd, &m->trigger_event);
954         return 0;
955 }
956
957 /*
958  * Set/clear bypassed status of a PG.
959  * PGs are numbered upwards from 1 in the order they were declared.
960  */
961 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
962 {
963         struct priority_group *pg;
964         unsigned pgnum;
965
966         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
967             (pgnum > m->nr_priority_groups)) {
968                 DMWARN("invalid PG number supplied to bypass_pg");
969                 return -EINVAL;
970         }
971
972         list_for_each_entry(pg, &m->priority_groups, list) {
973                 if (!--pgnum)
974                         break;
975         }
976
977         bypass_pg(m, pg, bypassed);
978         return 0;
979 }
980
981 /*
982  * pg_init must call this when it has completed its initialisation
983  */
984 void dm_pg_init_complete(struct dm_path *path, unsigned err_flags)
985 {
986         struct pgpath *pgpath = path_to_pgpath(path);
987         struct priority_group *pg = pgpath->pg;
988         struct multipath *m = pg->m;
989         unsigned long flags;
990
991         /* We insist on failing the path if the PG is already bypassed. */
992         if (err_flags && pg->bypassed)
993                 err_flags |= MP_FAIL_PATH;
994
995         if (err_flags & MP_FAIL_PATH)
996                 fail_path(pgpath);
997
998         if (err_flags & MP_BYPASS_PG)
999                 bypass_pg(m, pg, 1);
1000
1001         spin_lock_irqsave(&m->lock, flags);
1002         if (err_flags) {
1003                 m->current_pgpath = NULL;
1004                 m->current_pg = NULL;
1005         } else if (!m->pg_init_required)
1006                 m->queue_io = 0;
1007
1008         m->pg_init_in_progress = 0;
1009         queue_work(kmultipathd, &m->process_queued_ios);
1010         spin_unlock_irqrestore(&m->lock, flags);
1011 }
1012
1013 /*
1014  * end_io handling
1015  */
1016 static int do_end_io(struct multipath *m, struct bio *bio,
1017                      int error, struct mpath_io *mpio)
1018 {
1019         struct hw_handler *hwh = &m->hw_handler;
1020         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1021         unsigned long flags;
1022
1023         if (!error)
1024                 return 0;       /* I/O complete */
1025
1026         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1027                 return error;
1028
1029         if (error == -EOPNOTSUPP)
1030                 return error;
1031
1032         spin_lock_irqsave(&m->lock, flags);
1033         if (!m->nr_valid_paths) {
1034                 if (__must_push_back(m)) {
1035                         spin_unlock_irqrestore(&m->lock, flags);
1036                         return DM_ENDIO_REQUEUE;
1037                 } else if (!m->queue_if_no_path) {
1038                         spin_unlock_irqrestore(&m->lock, flags);
1039                         return -EIO;
1040                 } else {
1041                         spin_unlock_irqrestore(&m->lock, flags);
1042                         goto requeue;
1043                 }
1044         }
1045         spin_unlock_irqrestore(&m->lock, flags);
1046
1047         if (hwh->type && hwh->type->error)
1048                 err_flags = hwh->type->error(hwh, bio);
1049
1050         if (mpio->pgpath) {
1051                 if (err_flags & MP_FAIL_PATH)
1052                         fail_path(mpio->pgpath);
1053
1054                 if (err_flags & MP_BYPASS_PG)
1055                         bypass_pg(m, mpio->pgpath->pg, 1);
1056         }
1057
1058         if (err_flags & MP_ERROR_IO)
1059                 return -EIO;
1060
1061       requeue:
1062         dm_bio_restore(&mpio->details, bio);
1063
1064         /* queue for the daemon to resubmit or fail */
1065         spin_lock_irqsave(&m->lock, flags);
1066         bio_list_add(&m->queued_ios, bio);
1067         m->queue_size++;
1068         if (!m->queue_io)
1069                 queue_work(kmultipathd, &m->process_queued_ios);
1070         spin_unlock_irqrestore(&m->lock, flags);
1071
1072         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1073 }
1074
1075 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1076                             int error, union map_info *map_context)
1077 {
1078         struct multipath *m = (struct multipath *) ti->private;
1079         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1080         struct pgpath *pgpath = mpio->pgpath;
1081         struct path_selector *ps;
1082         int r;
1083
1084         r  = do_end_io(m, bio, error, mpio);
1085         if (pgpath) {
1086                 ps = &pgpath->pg->ps;
1087                 if (ps->type->end_io)
1088                         ps->type->end_io(ps, &pgpath->path);
1089         }
1090         if (r != DM_ENDIO_INCOMPLETE)
1091                 mempool_free(mpio, m->mpio_pool);
1092
1093         return r;
1094 }
1095
1096 /*
1097  * Suspend can't complete until all the I/O is processed so if
1098  * the last path fails we must error any remaining I/O.
1099  * Note that if the freeze_bdev fails while suspending, the
1100  * queue_if_no_path state is lost - userspace should reset it.
1101  */
1102 static void multipath_presuspend(struct dm_target *ti)
1103 {
1104         struct multipath *m = (struct multipath *) ti->private;
1105
1106         queue_if_no_path(m, 0, 1);
1107 }
1108
1109 /*
1110  * Restore the queue_if_no_path setting.
1111  */
1112 static void multipath_resume(struct dm_target *ti)
1113 {
1114         struct multipath *m = (struct multipath *) ti->private;
1115         unsigned long flags;
1116
1117         spin_lock_irqsave(&m->lock, flags);
1118         m->queue_if_no_path = m->saved_queue_if_no_path;
1119         spin_unlock_irqrestore(&m->lock, flags);
1120 }
1121
1122 /*
1123  * Info output has the following format:
1124  * num_multipath_feature_args [multipath_feature_args]*
1125  * num_handler_status_args [handler_status_args]*
1126  * num_groups init_group_number
1127  *            [A|D|E num_ps_status_args [ps_status_args]*
1128  *             num_paths num_selector_args
1129  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1130  *
1131  * Table output has the following format (identical to the constructor string):
1132  * num_feature_args [features_args]*
1133  * num_handler_args hw_handler [hw_handler_args]*
1134  * num_groups init_group_number
1135  *     [priority selector-name num_ps_args [ps_args]*
1136  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1137  */
1138 static int multipath_status(struct dm_target *ti, status_type_t type,
1139                             char *result, unsigned int maxlen)
1140 {
1141         int sz = 0;
1142         unsigned long flags;
1143         struct multipath *m = (struct multipath *) ti->private;
1144         struct hw_handler *hwh = &m->hw_handler;
1145         struct priority_group *pg;
1146         struct pgpath *p;
1147         unsigned pg_num;
1148         char state;
1149
1150         spin_lock_irqsave(&m->lock, flags);
1151
1152         /* Features */
1153         if (type == STATUSTYPE_INFO)
1154                 DMEMIT("1 %u ", m->queue_size);
1155         else if (m->queue_if_no_path)
1156                 DMEMIT("1 queue_if_no_path ");
1157         else
1158                 DMEMIT("0 ");
1159
1160         if (hwh->type && hwh->type->status)
1161                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1162         else if (!hwh->type || type == STATUSTYPE_INFO)
1163                 DMEMIT("0 ");
1164         else
1165                 DMEMIT("1 %s ", hwh->type->name);
1166
1167         DMEMIT("%u ", m->nr_priority_groups);
1168
1169         if (m->next_pg)
1170                 pg_num = m->next_pg->pg_num;
1171         else if (m->current_pg)
1172                 pg_num = m->current_pg->pg_num;
1173         else
1174                         pg_num = 1;
1175
1176         DMEMIT("%u ", pg_num);
1177
1178         switch (type) {
1179         case STATUSTYPE_INFO:
1180                 list_for_each_entry(pg, &m->priority_groups, list) {
1181                         if (pg->bypassed)
1182                                 state = 'D';    /* Disabled */
1183                         else if (pg == m->current_pg)
1184                                 state = 'A';    /* Currently Active */
1185                         else
1186                                 state = 'E';    /* Enabled */
1187
1188                         DMEMIT("%c ", state);
1189
1190                         if (pg->ps.type->status)
1191                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1192                                                           result + sz,
1193                                                           maxlen - sz);
1194                         else
1195                                 DMEMIT("0 ");
1196
1197                         DMEMIT("%u %u ", pg->nr_pgpaths,
1198                                pg->ps.type->info_args);
1199
1200                         list_for_each_entry(p, &pg->pgpaths, list) {
1201                                 DMEMIT("%s %s %u ", p->path.dev->name,
1202                                        p->path.is_active ? "A" : "F",
1203                                        p->fail_count);
1204                                 if (pg->ps.type->status)
1205                                         sz += pg->ps.type->status(&pg->ps,
1206                                               &p->path, type, result + sz,
1207                                               maxlen - sz);
1208                         }
1209                 }
1210                 break;
1211
1212         case STATUSTYPE_TABLE:
1213                 list_for_each_entry(pg, &m->priority_groups, list) {
1214                         DMEMIT("%s ", pg->ps.type->name);
1215
1216                         if (pg->ps.type->status)
1217                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1218                                                           result + sz,
1219                                                           maxlen - sz);
1220                         else
1221                                 DMEMIT("0 ");
1222
1223                         DMEMIT("%u %u ", pg->nr_pgpaths,
1224                                pg->ps.type->table_args);
1225
1226                         list_for_each_entry(p, &pg->pgpaths, list) {
1227                                 DMEMIT("%s ", p->path.dev->name);
1228                                 if (pg->ps.type->status)
1229                                         sz += pg->ps.type->status(&pg->ps,
1230                                               &p->path, type, result + sz,
1231                                               maxlen - sz);
1232                         }
1233                 }
1234                 break;
1235         }
1236
1237         spin_unlock_irqrestore(&m->lock, flags);
1238
1239         return 0;
1240 }
1241
1242 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1243 {
1244         int r;
1245         struct dm_dev *dev;
1246         struct multipath *m = (struct multipath *) ti->private;
1247         action_fn action;
1248
1249         if (argc == 1) {
1250                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1251                         return queue_if_no_path(m, 1, 0);
1252                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1253                         return queue_if_no_path(m, 0, 0);
1254         }
1255
1256         if (argc != 2)
1257                 goto error;
1258
1259         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1260                 return bypass_pg_num(m, argv[1], 1);
1261         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1262                 return bypass_pg_num(m, argv[1], 0);
1263         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1264                 return switch_pg_num(m, argv[1]);
1265         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1266                 action = reinstate_path;
1267         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1268                 action = fail_path;
1269         else
1270                 goto error;
1271
1272         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1273                           dm_table_get_mode(ti->table), &dev);
1274         if (r) {
1275                 DMWARN("message: error getting device %s",
1276                        argv[1]);
1277                 return -EINVAL;
1278         }
1279
1280         r = action_dev(m, dev, action);
1281
1282         dm_put_device(ti, dev);
1283
1284         return r;
1285
1286 error:
1287         DMWARN("Unrecognised multipath message received.");
1288         return -EINVAL;
1289 }
1290
1291 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1292                            struct file *filp, unsigned int cmd,
1293                            unsigned long arg)
1294 {
1295         struct multipath *m = (struct multipath *) ti->private;
1296         struct block_device *bdev = NULL;
1297         unsigned long flags;
1298         struct file fake_file = {};
1299         struct dentry fake_dentry = {};
1300         int r = 0;
1301
1302         fake_file.f_path.dentry = &fake_dentry;
1303
1304         spin_lock_irqsave(&m->lock, flags);
1305
1306         if (!m->current_pgpath)
1307                 __choose_pgpath(m);
1308
1309         if (m->current_pgpath) {
1310                 bdev = m->current_pgpath->path.dev->bdev;
1311                 fake_dentry.d_inode = bdev->bd_inode;
1312                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1313         }
1314
1315         if (m->queue_io)
1316                 r = -EAGAIN;
1317         else if (!bdev)
1318                 r = -EIO;
1319
1320         spin_unlock_irqrestore(&m->lock, flags);
1321
1322         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1323                                          bdev->bd_disk, cmd, arg);
1324 }
1325
1326 /*-----------------------------------------------------------------
1327  * Module setup
1328  *---------------------------------------------------------------*/
1329 static struct target_type multipath_target = {
1330         .name = "multipath",
1331         .version = {1, 0, 5},
1332         .module = THIS_MODULE,
1333         .ctr = multipath_ctr,
1334         .dtr = multipath_dtr,
1335         .map = multipath_map,
1336         .end_io = multipath_end_io,
1337         .presuspend = multipath_presuspend,
1338         .resume = multipath_resume,
1339         .status = multipath_status,
1340         .message = multipath_message,
1341         .ioctl  = multipath_ioctl,
1342 };
1343
1344 static int __init dm_multipath_init(void)
1345 {
1346         int r;
1347
1348         /* allocate a slab for the dm_ios */
1349         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1350                                         0, 0, NULL, NULL);
1351         if (!_mpio_cache)
1352                 return -ENOMEM;
1353
1354         r = dm_register_target(&multipath_target);
1355         if (r < 0) {
1356                 DMERR("%s: register failed %d", multipath_target.name, r);
1357                 kmem_cache_destroy(_mpio_cache);
1358                 return -EINVAL;
1359         }
1360
1361         kmultipathd = create_workqueue("kmpathd");
1362         if (!kmultipathd) {
1363                 DMERR("%s: failed to create workqueue kmpathd",
1364                                 multipath_target.name);
1365                 dm_unregister_target(&multipath_target);
1366                 kmem_cache_destroy(_mpio_cache);
1367                 return -ENOMEM;
1368         }
1369
1370         DMINFO("version %u.%u.%u loaded",
1371                multipath_target.version[0], multipath_target.version[1],
1372                multipath_target.version[2]);
1373
1374         return r;
1375 }
1376
1377 static void __exit dm_multipath_exit(void)
1378 {
1379         int r;
1380
1381         destroy_workqueue(kmultipathd);
1382
1383         r = dm_unregister_target(&multipath_target);
1384         if (r < 0)
1385                 DMERR("%s: target unregister failed %d",
1386                       multipath_target.name, r);
1387         kmem_cache_destroy(_mpio_cache);
1388 }
1389
1390 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1391
1392 module_init(dm_multipath_init);
1393 module_exit(dm_multipath_exit);
1394
1395 MODULE_DESCRIPTION(DM_NAME " multipath target");
1396 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1397 MODULE_LICENSE("GPL");