fea966d66f9838c140176a7dbd0ed8223e908f44
[linux-2.6.git] / 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-bio-list.h"
11 #include "dm-bio-record.h"
12 #include "dm-uevent.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 <scsi/scsi_dh.h>
23 #include <asm/atomic.h>
24
25 #define DM_MSG_PREFIX "multipath"
26 #define MESG_STR(x) x, sizeof(x)
27
28 /* Path properties */
29 struct pgpath {
30         struct list_head list;
31
32         struct priority_group *pg;      /* Owning PG */
33         unsigned fail_count;            /* Cumulative failure count */
34
35         struct dm_path path;
36 };
37
38 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
39
40 /*
41  * Paths are grouped into Priority Groups and numbered from 1 upwards.
42  * Each has a path selector which controls which path gets used.
43  */
44 struct priority_group {
45         struct list_head list;
46
47         struct multipath *m;            /* Owning multipath instance */
48         struct path_selector ps;
49
50         unsigned pg_num;                /* Reference number */
51         unsigned bypassed;              /* Temporarily bypass this PG? */
52
53         unsigned nr_pgpaths;            /* Number of paths in PG */
54         struct list_head pgpaths;
55 };
56
57 /* Multipath context */
58 struct multipath {
59         struct list_head list;
60         struct dm_target *ti;
61
62         spinlock_t lock;
63
64         const char *hw_handler_name;
65         struct work_struct activate_path;
66         unsigned nr_priority_groups;
67         struct list_head priority_groups;
68         unsigned pg_init_required;      /* pg_init needs calling? */
69         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
70
71         unsigned nr_valid_paths;        /* Total number of usable paths */
72         struct pgpath *current_pgpath;
73         struct priority_group *current_pg;
74         struct priority_group *next_pg; /* Switch to this PG if set */
75         unsigned repeat_count;          /* I/Os left before calling PS again */
76
77         unsigned queue_io;              /* Must we queue all I/O? */
78         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
79         unsigned saved_queue_if_no_path;/* Saved state during suspension */
80         unsigned pg_init_retries;       /* Number of times to retry pg_init */
81         unsigned pg_init_count;         /* Number of times pg_init called */
82
83         struct work_struct process_queued_ios;
84         struct bio_list queued_ios;
85         unsigned queue_size;
86
87         struct work_struct trigger_event;
88
89         /*
90          * We must use a mempool of dm_mpath_io structs so that we
91          * can resubmit bios on error.
92          */
93         mempool_t *mpio_pool;
94 };
95
96 /*
97  * Context information attached to each bio we process.
98  */
99 struct dm_mpath_io {
100         struct pgpath *pgpath;
101         struct dm_bio_details details;
102 };
103
104 typedef int (*action_fn) (struct pgpath *pgpath);
105
106 #define MIN_IOS 256     /* Mempool size */
107
108 static struct kmem_cache *_mpio_cache;
109
110 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
111 static void process_queued_ios(struct work_struct *work);
112 static void trigger_event(struct work_struct *work);
113 static void activate_path(struct work_struct *work);
114
115
116 /*-----------------------------------------------
117  * Allocation routines
118  *-----------------------------------------------*/
119
120 static struct pgpath *alloc_pgpath(void)
121 {
122         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
123
124         if (pgpath)
125                 pgpath->path.is_active = 1;
126
127         return pgpath;
128 }
129
130 static void free_pgpath(struct pgpath *pgpath)
131 {
132         kfree(pgpath);
133 }
134
135 static struct priority_group *alloc_priority_group(void)
136 {
137         struct priority_group *pg;
138
139         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
140
141         if (pg)
142                 INIT_LIST_HEAD(&pg->pgpaths);
143
144         return pg;
145 }
146
147 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
148 {
149         struct pgpath *pgpath, *tmp;
150
151         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
152                 list_del(&pgpath->list);
153                 dm_put_device(ti, pgpath->path.dev);
154                 free_pgpath(pgpath);
155         }
156 }
157
158 static void free_priority_group(struct priority_group *pg,
159                                 struct dm_target *ti)
160 {
161         struct path_selector *ps = &pg->ps;
162
163         if (ps->type) {
164                 ps->type->destroy(ps);
165                 dm_put_path_selector(ps->type);
166         }
167
168         free_pgpaths(&pg->pgpaths, ti);
169         kfree(pg);
170 }
171
172 static struct multipath *alloc_multipath(struct dm_target *ti)
173 {
174         struct multipath *m;
175
176         m = kzalloc(sizeof(*m), GFP_KERNEL);
177         if (m) {
178                 INIT_LIST_HEAD(&m->priority_groups);
179                 spin_lock_init(&m->lock);
180                 m->queue_io = 1;
181                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
182                 INIT_WORK(&m->trigger_event, trigger_event);
183                 INIT_WORK(&m->activate_path, activate_path);
184                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
185                 if (!m->mpio_pool) {
186                         kfree(m);
187                         return NULL;
188                 }
189                 m->ti = ti;
190                 ti->private = m;
191         }
192
193         return m;
194 }
195
196 static void free_multipath(struct multipath *m)
197 {
198         struct priority_group *pg, *tmp;
199
200         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
201                 list_del(&pg->list);
202                 free_priority_group(pg, m->ti);
203         }
204
205         kfree(m->hw_handler_name);
206         mempool_destroy(m->mpio_pool);
207         kfree(m);
208 }
209
210
211 /*-----------------------------------------------
212  * Path selection
213  *-----------------------------------------------*/
214
215 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
216 {
217         m->current_pg = pgpath->pg;
218
219         /* Must we initialise the PG first, and queue I/O till it's ready? */
220         if (m->hw_handler_name) {
221                 m->pg_init_required = 1;
222                 m->queue_io = 1;
223         } else {
224                 m->pg_init_required = 0;
225                 m->queue_io = 0;
226         }
227
228         m->pg_init_count = 0;
229 }
230
231 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
232 {
233         struct dm_path *path;
234
235         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
236         if (!path)
237                 return -ENXIO;
238
239         m->current_pgpath = path_to_pgpath(path);
240
241         if (m->current_pg != pg)
242                 __switch_pg(m, m->current_pgpath);
243
244         return 0;
245 }
246
247 static void __choose_pgpath(struct multipath *m)
248 {
249         struct priority_group *pg;
250         unsigned bypassed = 1;
251
252         if (!m->nr_valid_paths)
253                 goto failed;
254
255         /* Were we instructed to switch PG? */
256         if (m->next_pg) {
257                 pg = m->next_pg;
258                 m->next_pg = NULL;
259                 if (!__choose_path_in_pg(m, pg))
260                         return;
261         }
262
263         /* Don't change PG until it has no remaining paths */
264         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
265                 return;
266
267         /*
268          * Loop through priority groups until we find a valid path.
269          * First time we skip PGs marked 'bypassed'.
270          * Second time we only try the ones we skipped.
271          */
272         do {
273                 list_for_each_entry(pg, &m->priority_groups, list) {
274                         if (pg->bypassed == bypassed)
275                                 continue;
276                         if (!__choose_path_in_pg(m, pg))
277                                 return;
278                 }
279         } while (bypassed--);
280
281 failed:
282         m->current_pgpath = NULL;
283         m->current_pg = NULL;
284 }
285
286 /*
287  * Check whether bios must be queued in the device-mapper core rather
288  * than here in the target.
289  *
290  * m->lock must be held on entry.
291  *
292  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
293  * same value then we are not between multipath_presuspend()
294  * and multipath_resume() calls and we have no need to check
295  * for the DMF_NOFLUSH_SUSPENDING flag.
296  */
297 static int __must_push_back(struct multipath *m)
298 {
299         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
300                 dm_noflush_suspending(m->ti));
301 }
302
303 static int map_io(struct multipath *m, struct bio *bio,
304                   struct dm_mpath_io *mpio, unsigned was_queued)
305 {
306         int r = DM_MAPIO_REMAPPED;
307         unsigned long flags;
308         struct pgpath *pgpath;
309
310         spin_lock_irqsave(&m->lock, flags);
311
312         /* Do we need to select a new pgpath? */
313         if (!m->current_pgpath ||
314             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
315                 __choose_pgpath(m);
316
317         pgpath = m->current_pgpath;
318
319         if (was_queued)
320                 m->queue_size--;
321
322         if ((pgpath && m->queue_io) ||
323             (!pgpath && m->queue_if_no_path)) {
324                 /* Queue for the daemon to resubmit */
325                 bio_list_add(&m->queued_ios, bio);
326                 m->queue_size++;
327                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
328                     !m->queue_io)
329                         queue_work(kmultipathd, &m->process_queued_ios);
330                 pgpath = NULL;
331                 r = DM_MAPIO_SUBMITTED;
332         } else if (pgpath)
333                 bio->bi_bdev = pgpath->path.dev->bdev;
334         else if (__must_push_back(m))
335                 r = DM_MAPIO_REQUEUE;
336         else
337                 r = -EIO;       /* Failed */
338
339         mpio->pgpath = pgpath;
340
341         spin_unlock_irqrestore(&m->lock, flags);
342
343         return r;
344 }
345
346 /*
347  * If we run out of usable paths, should we queue I/O or error it?
348  */
349 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
350                             unsigned save_old_value)
351 {
352         unsigned long flags;
353
354         spin_lock_irqsave(&m->lock, flags);
355
356         if (save_old_value)
357                 m->saved_queue_if_no_path = m->queue_if_no_path;
358         else
359                 m->saved_queue_if_no_path = queue_if_no_path;
360         m->queue_if_no_path = queue_if_no_path;
361         if (!m->queue_if_no_path && m->queue_size)
362                 queue_work(kmultipathd, &m->process_queued_ios);
363
364         spin_unlock_irqrestore(&m->lock, flags);
365
366         return 0;
367 }
368
369 /*-----------------------------------------------------------------
370  * The multipath daemon is responsible for resubmitting queued ios.
371  *---------------------------------------------------------------*/
372
373 static void dispatch_queued_ios(struct multipath *m)
374 {
375         int r;
376         unsigned long flags;
377         struct bio *bio = NULL, *next;
378         struct dm_mpath_io *mpio;
379         union map_info *info;
380
381         spin_lock_irqsave(&m->lock, flags);
382         bio = bio_list_get(&m->queued_ios);
383         spin_unlock_irqrestore(&m->lock, flags);
384
385         while (bio) {
386                 next = bio->bi_next;
387                 bio->bi_next = NULL;
388
389                 info = dm_get_mapinfo(bio);
390                 mpio = info->ptr;
391
392                 r = map_io(m, bio, mpio, 1);
393                 if (r < 0)
394                         bio_endio(bio, r);
395                 else if (r == DM_MAPIO_REMAPPED)
396                         generic_make_request(bio);
397                 else if (r == DM_MAPIO_REQUEUE)
398                         bio_endio(bio, -EIO);
399
400                 bio = next;
401         }
402 }
403
404 static void process_queued_ios(struct work_struct *work)
405 {
406         struct multipath *m =
407                 container_of(work, struct multipath, process_queued_ios);
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_count++;
428                 m->pg_init_required = 0;
429                 m->pg_init_in_progress = 1;
430                 init_required = 1;
431         }
432
433 out:
434         spin_unlock_irqrestore(&m->lock, flags);
435
436         if (init_required)
437                 queue_work(kmpath_handlerd, &m->activate_path);
438
439         if (!must_queue)
440                 dispatch_queued_ios(m);
441 }
442
443 /*
444  * An event is triggered whenever a path is taken out of use.
445  * Includes path failure and PG bypass.
446  */
447 static void trigger_event(struct work_struct *work)
448 {
449         struct multipath *m =
450                 container_of(work, struct multipath, trigger_event);
451
452         dm_table_event(m->ti->table);
453 }
454
455 /*-----------------------------------------------------------------
456  * Constructor/argument parsing:
457  * <#multipath feature args> [<arg>]*
458  * <#hw_handler args> [hw_handler [<arg>]*]
459  * <#priority groups>
460  * <initial priority group>
461  *     [<selector> <#selector args> [<arg>]*
462  *      <#paths> <#per-path selector args>
463  *         [<path> [<arg>]* ]+ ]+
464  *---------------------------------------------------------------*/
465 struct param {
466         unsigned min;
467         unsigned max;
468         char *error;
469 };
470
471 static int read_param(struct param *param, char *str, unsigned *v, char **error)
472 {
473         if (!str ||
474             (sscanf(str, "%u", v) != 1) ||
475             (*v < param->min) ||
476             (*v > param->max)) {
477                 *error = param->error;
478                 return -EINVAL;
479         }
480
481         return 0;
482 }
483
484 struct arg_set {
485         unsigned argc;
486         char **argv;
487 };
488
489 static char *shift(struct arg_set *as)
490 {
491         char *r;
492
493         if (as->argc) {
494                 as->argc--;
495                 r = *as->argv;
496                 as->argv++;
497                 return r;
498         }
499
500         return NULL;
501 }
502
503 static void consume(struct arg_set *as, unsigned n)
504 {
505         BUG_ON (as->argc < n);
506         as->argc -= n;
507         as->argv += n;
508 }
509
510 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
511                                struct dm_target *ti)
512 {
513         int r;
514         struct path_selector_type *pst;
515         unsigned ps_argc;
516
517         static struct param _params[] = {
518                 {0, 1024, "invalid number of path selector args"},
519         };
520
521         pst = dm_get_path_selector(shift(as));
522         if (!pst) {
523                 ti->error = "unknown path selector type";
524                 return -EINVAL;
525         }
526
527         r = read_param(_params, shift(as), &ps_argc, &ti->error);
528         if (r) {
529                 dm_put_path_selector(pst);
530                 return -EINVAL;
531         }
532
533         r = pst->create(&pg->ps, ps_argc, as->argv);
534         if (r) {
535                 dm_put_path_selector(pst);
536                 ti->error = "path selector constructor failed";
537                 return r;
538         }
539
540         pg->ps.type = pst;
541         consume(as, ps_argc);
542
543         return 0;
544 }
545
546 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
547                                struct dm_target *ti)
548 {
549         int r;
550         struct pgpath *p;
551
552         /* we need at least a path arg */
553         if (as->argc < 1) {
554                 ti->error = "no device given";
555                 return NULL;
556         }
557
558         p = alloc_pgpath();
559         if (!p)
560                 return NULL;
561
562         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
563                           dm_table_get_mode(ti->table), &p->path.dev);
564         if (r) {
565                 ti->error = "error getting device";
566                 goto bad;
567         }
568
569         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
570         if (r) {
571                 dm_put_device(ti, p->path.dev);
572                 goto bad;
573         }
574
575         return p;
576
577  bad:
578         free_pgpath(p);
579         return NULL;
580 }
581
582 static struct priority_group *parse_priority_group(struct arg_set *as,
583                                                    struct multipath *m)
584 {
585         static struct param _params[] = {
586                 {1, 1024, "invalid number of paths"},
587                 {0, 1024, "invalid number of selector args"}
588         };
589
590         int r;
591         unsigned i, nr_selector_args, nr_params;
592         struct priority_group *pg;
593         struct dm_target *ti = m->ti;
594
595         if (as->argc < 2) {
596                 as->argc = 0;
597                 ti->error = "not enough priority group aruments";
598                 return NULL;
599         }
600
601         pg = alloc_priority_group();
602         if (!pg) {
603                 ti->error = "couldn't allocate priority group";
604                 return NULL;
605         }
606         pg->m = m;
607
608         r = parse_path_selector(as, pg, ti);
609         if (r)
610                 goto bad;
611
612         /*
613          * read the paths
614          */
615         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
616         if (r)
617                 goto bad;
618
619         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
620         if (r)
621                 goto bad;
622
623         nr_params = 1 + nr_selector_args;
624         for (i = 0; i < pg->nr_pgpaths; i++) {
625                 struct pgpath *pgpath;
626                 struct arg_set path_args;
627
628                 if (as->argc < nr_params) {
629                         ti->error = "not enough path parameters";
630                         goto bad;
631                 }
632
633                 path_args.argc = nr_params;
634                 path_args.argv = as->argv;
635
636                 pgpath = parse_path(&path_args, &pg->ps, ti);
637                 if (!pgpath)
638                         goto bad;
639
640                 pgpath->pg = pg;
641                 list_add_tail(&pgpath->list, &pg->pgpaths);
642                 consume(as, nr_params);
643         }
644
645         return pg;
646
647  bad:
648         free_priority_group(pg, ti);
649         return NULL;
650 }
651
652 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
653 {
654         unsigned hw_argc;
655         struct dm_target *ti = m->ti;
656
657         static struct param _params[] = {
658                 {0, 1024, "invalid number of hardware handler args"},
659         };
660
661         if (read_param(_params, shift(as), &hw_argc, &ti->error))
662                 return -EINVAL;
663
664         if (!hw_argc)
665                 return 0;
666
667         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
668         request_module("scsi_dh_%s", m->hw_handler_name);
669         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
670                 ti->error = "unknown hardware handler type";
671                 kfree(m->hw_handler_name);
672                 m->hw_handler_name = NULL;
673                 return -EINVAL;
674         }
675         consume(as, hw_argc - 1);
676
677         return 0;
678 }
679
680 static int parse_features(struct arg_set *as, struct multipath *m)
681 {
682         int r;
683         unsigned argc;
684         struct dm_target *ti = m->ti;
685         const char *param_name;
686
687         static struct param _params[] = {
688                 {0, 3, "invalid number of feature args"},
689                 {1, 50, "pg_init_retries must be between 1 and 50"},
690         };
691
692         r = read_param(_params, shift(as), &argc, &ti->error);
693         if (r)
694                 return -EINVAL;
695
696         if (!argc)
697                 return 0;
698
699         do {
700                 param_name = shift(as);
701                 argc--;
702
703                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
704                         r = queue_if_no_path(m, 1, 0);
705                         continue;
706                 }
707
708                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
709                     (argc >= 1)) {
710                         r = read_param(_params + 1, shift(as),
711                                        &m->pg_init_retries, &ti->error);
712                         argc--;
713                         continue;
714                 }
715
716                 ti->error = "Unrecognised multipath feature request";
717                 r = -EINVAL;
718         } while (argc && !r);
719
720         return r;
721 }
722
723 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
724                          char **argv)
725 {
726         /* target parameters */
727         static struct param _params[] = {
728                 {1, 1024, "invalid number of priority groups"},
729                 {1, 1024, "invalid initial priority group number"},
730         };
731
732         int r;
733         struct multipath *m;
734         struct arg_set as;
735         unsigned pg_count = 0;
736         unsigned next_pg_num;
737
738         as.argc = argc;
739         as.argv = argv;
740
741         m = alloc_multipath(ti);
742         if (!m) {
743                 ti->error = "can't allocate multipath";
744                 return -EINVAL;
745         }
746
747         r = parse_features(&as, m);
748         if (r)
749                 goto bad;
750
751         r = parse_hw_handler(&as, m);
752         if (r)
753                 goto bad;
754
755         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
756         if (r)
757                 goto bad;
758
759         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
760         if (r)
761                 goto bad;
762
763         /* parse the priority groups */
764         while (as.argc) {
765                 struct priority_group *pg;
766
767                 pg = parse_priority_group(&as, m);
768                 if (!pg) {
769                         r = -EINVAL;
770                         goto bad;
771                 }
772
773                 m->nr_valid_paths += pg->nr_pgpaths;
774                 list_add_tail(&pg->list, &m->priority_groups);
775                 pg_count++;
776                 pg->pg_num = pg_count;
777                 if (!--next_pg_num)
778                         m->next_pg = pg;
779         }
780
781         if (pg_count != m->nr_priority_groups) {
782                 ti->error = "priority group count mismatch";
783                 r = -EINVAL;
784                 goto bad;
785         }
786
787         return 0;
788
789  bad:
790         free_multipath(m);
791         return r;
792 }
793
794 static void multipath_dtr(struct dm_target *ti)
795 {
796         struct multipath *m = (struct multipath *) ti->private;
797
798         flush_workqueue(kmpath_handlerd);
799         flush_workqueue(kmultipathd);
800         free_multipath(m);
801 }
802
803 /*
804  * Map bios, recording original fields for later in case we have to resubmit
805  */
806 static int multipath_map(struct dm_target *ti, struct bio *bio,
807                          union map_info *map_context)
808 {
809         int r;
810         struct dm_mpath_io *mpio;
811         struct multipath *m = (struct multipath *) ti->private;
812
813         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
814         dm_bio_record(&mpio->details, bio);
815
816         map_context->ptr = mpio;
817         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
818         r = map_io(m, bio, mpio, 0);
819         if (r < 0 || r == DM_MAPIO_REQUEUE)
820                 mempool_free(mpio, m->mpio_pool);
821
822         return r;
823 }
824
825 /*
826  * Take a path out of use.
827  */
828 static int fail_path(struct pgpath *pgpath)
829 {
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         DMWARN("Failing path %s.", pgpath->path.dev->name);
839
840         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
841         pgpath->path.is_active = 0;
842         pgpath->fail_count++;
843
844         m->nr_valid_paths--;
845
846         if (pgpath == m->current_pgpath)
847                 m->current_pgpath = NULL;
848
849         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
850                       pgpath->path.dev->name, m->nr_valid_paths);
851
852         queue_work(kmultipathd, &m->trigger_event);
853
854 out:
855         spin_unlock_irqrestore(&m->lock, flags);
856
857         return 0;
858 }
859
860 /*
861  * Reinstate a previously-failed path
862  */
863 static int reinstate_path(struct pgpath *pgpath)
864 {
865         int r = 0;
866         unsigned long flags;
867         struct multipath *m = pgpath->pg->m;
868
869         spin_lock_irqsave(&m->lock, flags);
870
871         if (pgpath->path.is_active)
872                 goto out;
873
874         if (!pgpath->pg->ps.type->reinstate_path) {
875                 DMWARN("Reinstate path not supported by path selector %s",
876                        pgpath->pg->ps.type->name);
877                 r = -EINVAL;
878                 goto out;
879         }
880
881         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
882         if (r)
883                 goto out;
884
885         pgpath->path.is_active = 1;
886
887         m->current_pgpath = NULL;
888         if (!m->nr_valid_paths++ && m->queue_size)
889                 queue_work(kmultipathd, &m->process_queued_ios);
890
891         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
892                       pgpath->path.dev->name, m->nr_valid_paths);
893
894         queue_work(kmultipathd, &m->trigger_event);
895
896 out:
897         spin_unlock_irqrestore(&m->lock, flags);
898
899         return r;
900 }
901
902 /*
903  * Fail or reinstate all paths that match the provided struct dm_dev.
904  */
905 static int action_dev(struct multipath *m, struct dm_dev *dev,
906                       action_fn action)
907 {
908         int r = 0;
909         struct pgpath *pgpath;
910         struct priority_group *pg;
911
912         list_for_each_entry(pg, &m->priority_groups, list) {
913                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
914                         if (pgpath->path.dev == dev)
915                                 r = action(pgpath);
916                 }
917         }
918
919         return r;
920 }
921
922 /*
923  * Temporarily try to avoid having to use the specified PG
924  */
925 static void bypass_pg(struct multipath *m, struct priority_group *pg,
926                       int bypassed)
927 {
928         unsigned long flags;
929
930         spin_lock_irqsave(&m->lock, flags);
931
932         pg->bypassed = bypassed;
933         m->current_pgpath = NULL;
934         m->current_pg = NULL;
935
936         spin_unlock_irqrestore(&m->lock, flags);
937
938         queue_work(kmultipathd, &m->trigger_event);
939 }
940
941 /*
942  * Switch to using the specified PG from the next I/O that gets mapped
943  */
944 static int switch_pg_num(struct multipath *m, const char *pgstr)
945 {
946         struct priority_group *pg;
947         unsigned pgnum;
948         unsigned long flags;
949
950         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
951             (pgnum > m->nr_priority_groups)) {
952                 DMWARN("invalid PG number supplied to switch_pg_num");
953                 return -EINVAL;
954         }
955
956         spin_lock_irqsave(&m->lock, flags);
957         list_for_each_entry(pg, &m->priority_groups, list) {
958                 pg->bypassed = 0;
959                 if (--pgnum)
960                         continue;
961
962                 m->current_pgpath = NULL;
963                 m->current_pg = NULL;
964                 m->next_pg = pg;
965         }
966         spin_unlock_irqrestore(&m->lock, flags);
967
968         queue_work(kmultipathd, &m->trigger_event);
969         return 0;
970 }
971
972 /*
973  * Set/clear bypassed status of a PG.
974  * PGs are numbered upwards from 1 in the order they were declared.
975  */
976 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
977 {
978         struct priority_group *pg;
979         unsigned pgnum;
980
981         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
982             (pgnum > m->nr_priority_groups)) {
983                 DMWARN("invalid PG number supplied to bypass_pg");
984                 return -EINVAL;
985         }
986
987         list_for_each_entry(pg, &m->priority_groups, list) {
988                 if (!--pgnum)
989                         break;
990         }
991
992         bypass_pg(m, pg, bypassed);
993         return 0;
994 }
995
996 /*
997  * Should we retry pg_init immediately?
998  */
999 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1000 {
1001         unsigned long flags;
1002         int limit_reached = 0;
1003
1004         spin_lock_irqsave(&m->lock, flags);
1005
1006         if (m->pg_init_count <= m->pg_init_retries)
1007                 m->pg_init_required = 1;
1008         else
1009                 limit_reached = 1;
1010
1011         spin_unlock_irqrestore(&m->lock, flags);
1012
1013         return limit_reached;
1014 }
1015
1016 static void pg_init_done(struct dm_path *path, int errors)
1017 {
1018         struct pgpath *pgpath = path_to_pgpath(path);
1019         struct priority_group *pg = pgpath->pg;
1020         struct multipath *m = pg->m;
1021         unsigned long flags;
1022
1023         /* device or driver problems */
1024         switch (errors) {
1025         case SCSI_DH_OK:
1026                 break;
1027         case SCSI_DH_NOSYS:
1028                 if (!m->hw_handler_name) {
1029                         errors = 0;
1030                         break;
1031                 }
1032                 DMERR("Cannot failover device because scsi_dh_%s was not "
1033                       "loaded.", m->hw_handler_name);
1034                 /*
1035                  * Fail path for now, so we do not ping pong
1036                  */
1037                 fail_path(pgpath);
1038                 break;
1039         case SCSI_DH_DEV_TEMP_BUSY:
1040                 /*
1041                  * Probably doing something like FW upgrade on the
1042                  * controller so try the other pg.
1043                  */
1044                 bypass_pg(m, pg, 1);
1045                 break;
1046         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1047         case SCSI_DH_RETRY:
1048         case SCSI_DH_IMM_RETRY:
1049         case SCSI_DH_RES_TEMP_UNAVAIL:
1050                 if (pg_init_limit_reached(m, pgpath))
1051                         fail_path(pgpath);
1052                 errors = 0;
1053                 break;
1054         default:
1055                 /*
1056                  * We probably do not want to fail the path for a device
1057                  * error, but this is what the old dm did. In future
1058                  * patches we can do more advanced handling.
1059                  */
1060                 fail_path(pgpath);
1061         }
1062
1063         spin_lock_irqsave(&m->lock, flags);
1064         if (errors) {
1065                 DMERR("Could not failover device. Error %d.", errors);
1066                 m->current_pgpath = NULL;
1067                 m->current_pg = NULL;
1068         } else if (!m->pg_init_required) {
1069                 m->queue_io = 0;
1070                 pg->bypassed = 0;
1071         }
1072
1073         m->pg_init_in_progress = 0;
1074         queue_work(kmultipathd, &m->process_queued_ios);
1075         spin_unlock_irqrestore(&m->lock, flags);
1076 }
1077
1078 static void activate_path(struct work_struct *work)
1079 {
1080         int ret;
1081         struct multipath *m =
1082                 container_of(work, struct multipath, activate_path);
1083         struct dm_path *path = &m->current_pgpath->path;
1084
1085         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1086         pg_init_done(path, ret);
1087 }
1088
1089 /*
1090  * end_io handling
1091  */
1092 static int do_end_io(struct multipath *m, struct bio *bio,
1093                      int error, struct dm_mpath_io *mpio)
1094 {
1095         unsigned long flags;
1096
1097         if (!error)
1098                 return 0;       /* I/O complete */
1099
1100         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1101                 return error;
1102
1103         if (error == -EOPNOTSUPP)
1104                 return error;
1105
1106         spin_lock_irqsave(&m->lock, flags);
1107         if (!m->nr_valid_paths) {
1108                 if (__must_push_back(m)) {
1109                         spin_unlock_irqrestore(&m->lock, flags);
1110                         return DM_ENDIO_REQUEUE;
1111                 } else if (!m->queue_if_no_path) {
1112                         spin_unlock_irqrestore(&m->lock, flags);
1113                         return -EIO;
1114                 } else {
1115                         spin_unlock_irqrestore(&m->lock, flags);
1116                         goto requeue;
1117                 }
1118         }
1119         spin_unlock_irqrestore(&m->lock, flags);
1120
1121         if (mpio->pgpath)
1122                 fail_path(mpio->pgpath);
1123
1124       requeue:
1125         dm_bio_restore(&mpio->details, bio);
1126
1127         /* queue for the daemon to resubmit or fail */
1128         spin_lock_irqsave(&m->lock, flags);
1129         bio_list_add(&m->queued_ios, bio);
1130         m->queue_size++;
1131         if (!m->queue_io)
1132                 queue_work(kmultipathd, &m->process_queued_ios);
1133         spin_unlock_irqrestore(&m->lock, flags);
1134
1135         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1136 }
1137
1138 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1139                             int error, union map_info *map_context)
1140 {
1141         struct multipath *m = ti->private;
1142         struct dm_mpath_io *mpio = map_context->ptr;
1143         struct pgpath *pgpath = mpio->pgpath;
1144         struct path_selector *ps;
1145         int r;
1146
1147         r  = do_end_io(m, bio, error, mpio);
1148         if (pgpath) {
1149                 ps = &pgpath->pg->ps;
1150                 if (ps->type->end_io)
1151                         ps->type->end_io(ps, &pgpath->path);
1152         }
1153         if (r != DM_ENDIO_INCOMPLETE)
1154                 mempool_free(mpio, m->mpio_pool);
1155
1156         return r;
1157 }
1158
1159 /*
1160  * Suspend can't complete until all the I/O is processed so if
1161  * the last path fails we must error any remaining I/O.
1162  * Note that if the freeze_bdev fails while suspending, the
1163  * queue_if_no_path state is lost - userspace should reset it.
1164  */
1165 static void multipath_presuspend(struct dm_target *ti)
1166 {
1167         struct multipath *m = (struct multipath *) ti->private;
1168
1169         queue_if_no_path(m, 0, 1);
1170 }
1171
1172 /*
1173  * Restore the queue_if_no_path setting.
1174  */
1175 static void multipath_resume(struct dm_target *ti)
1176 {
1177         struct multipath *m = (struct multipath *) ti->private;
1178         unsigned long flags;
1179
1180         spin_lock_irqsave(&m->lock, flags);
1181         m->queue_if_no_path = m->saved_queue_if_no_path;
1182         spin_unlock_irqrestore(&m->lock, flags);
1183 }
1184
1185 /*
1186  * Info output has the following format:
1187  * num_multipath_feature_args [multipath_feature_args]*
1188  * num_handler_status_args [handler_status_args]*
1189  * num_groups init_group_number
1190  *            [A|D|E num_ps_status_args [ps_status_args]*
1191  *             num_paths num_selector_args
1192  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1193  *
1194  * Table output has the following format (identical to the constructor string):
1195  * num_feature_args [features_args]*
1196  * num_handler_args hw_handler [hw_handler_args]*
1197  * num_groups init_group_number
1198  *     [priority selector-name num_ps_args [ps_args]*
1199  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1200  */
1201 static int multipath_status(struct dm_target *ti, status_type_t type,
1202                             char *result, unsigned int maxlen)
1203 {
1204         int sz = 0;
1205         unsigned long flags;
1206         struct multipath *m = (struct multipath *) ti->private;
1207         struct priority_group *pg;
1208         struct pgpath *p;
1209         unsigned pg_num;
1210         char state;
1211
1212         spin_lock_irqsave(&m->lock, flags);
1213
1214         /* Features */
1215         if (type == STATUSTYPE_INFO)
1216                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1217         else {
1218                 DMEMIT("%u ", m->queue_if_no_path +
1219                               (m->pg_init_retries > 0) * 2);
1220                 if (m->queue_if_no_path)
1221                         DMEMIT("queue_if_no_path ");
1222                 if (m->pg_init_retries)
1223                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1224         }
1225
1226         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1227                 DMEMIT("0 ");
1228         else
1229                 DMEMIT("1 %s ", m->hw_handler_name);
1230
1231         DMEMIT("%u ", m->nr_priority_groups);
1232
1233         if (m->next_pg)
1234                 pg_num = m->next_pg->pg_num;
1235         else if (m->current_pg)
1236                 pg_num = m->current_pg->pg_num;
1237         else
1238                         pg_num = 1;
1239
1240         DMEMIT("%u ", pg_num);
1241
1242         switch (type) {
1243         case STATUSTYPE_INFO:
1244                 list_for_each_entry(pg, &m->priority_groups, list) {
1245                         if (pg->bypassed)
1246                                 state = 'D';    /* Disabled */
1247                         else if (pg == m->current_pg)
1248                                 state = 'A';    /* Currently Active */
1249                         else
1250                                 state = 'E';    /* Enabled */
1251
1252                         DMEMIT("%c ", state);
1253
1254                         if (pg->ps.type->status)
1255                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1256                                                           result + sz,
1257                                                           maxlen - sz);
1258                         else
1259                                 DMEMIT("0 ");
1260
1261                         DMEMIT("%u %u ", pg->nr_pgpaths,
1262                                pg->ps.type->info_args);
1263
1264                         list_for_each_entry(p, &pg->pgpaths, list) {
1265                                 DMEMIT("%s %s %u ", p->path.dev->name,
1266                                        p->path.is_active ? "A" : "F",
1267                                        p->fail_count);
1268                                 if (pg->ps.type->status)
1269                                         sz += pg->ps.type->status(&pg->ps,
1270                                               &p->path, type, result + sz,
1271                                               maxlen - sz);
1272                         }
1273                 }
1274                 break;
1275
1276         case STATUSTYPE_TABLE:
1277                 list_for_each_entry(pg, &m->priority_groups, list) {
1278                         DMEMIT("%s ", pg->ps.type->name);
1279
1280                         if (pg->ps.type->status)
1281                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1282                                                           result + sz,
1283                                                           maxlen - sz);
1284                         else
1285                                 DMEMIT("0 ");
1286
1287                         DMEMIT("%u %u ", pg->nr_pgpaths,
1288                                pg->ps.type->table_args);
1289
1290                         list_for_each_entry(p, &pg->pgpaths, list) {
1291                                 DMEMIT("%s ", p->path.dev->name);
1292                                 if (pg->ps.type->status)
1293                                         sz += pg->ps.type->status(&pg->ps,
1294                                               &p->path, type, result + sz,
1295                                               maxlen - sz);
1296                         }
1297                 }
1298                 break;
1299         }
1300
1301         spin_unlock_irqrestore(&m->lock, flags);
1302
1303         return 0;
1304 }
1305
1306 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1307 {
1308         int r;
1309         struct dm_dev *dev;
1310         struct multipath *m = (struct multipath *) ti->private;
1311         action_fn action;
1312
1313         if (argc == 1) {
1314                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1315                         return queue_if_no_path(m, 1, 0);
1316                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1317                         return queue_if_no_path(m, 0, 0);
1318         }
1319
1320         if (argc != 2)
1321                 goto error;
1322
1323         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1324                 return bypass_pg_num(m, argv[1], 1);
1325         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1326                 return bypass_pg_num(m, argv[1], 0);
1327         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1328                 return switch_pg_num(m, argv[1]);
1329         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1330                 action = reinstate_path;
1331         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1332                 action = fail_path;
1333         else
1334                 goto error;
1335
1336         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1337                           dm_table_get_mode(ti->table), &dev);
1338         if (r) {
1339                 DMWARN("message: error getting device %s",
1340                        argv[1]);
1341                 return -EINVAL;
1342         }
1343
1344         r = action_dev(m, dev, action);
1345
1346         dm_put_device(ti, dev);
1347
1348         return r;
1349
1350 error:
1351         DMWARN("Unrecognised multipath message received.");
1352         return -EINVAL;
1353 }
1354
1355 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1356                            struct file *filp, unsigned int cmd,
1357                            unsigned long arg)
1358 {
1359         struct multipath *m = (struct multipath *) ti->private;
1360         struct block_device *bdev = NULL;
1361         unsigned long flags;
1362         struct file fake_file = {};
1363         struct dentry fake_dentry = {};
1364         int r = 0;
1365
1366         fake_file.f_path.dentry = &fake_dentry;
1367
1368         spin_lock_irqsave(&m->lock, flags);
1369
1370         if (!m->current_pgpath)
1371                 __choose_pgpath(m);
1372
1373         if (m->current_pgpath) {
1374                 bdev = m->current_pgpath->path.dev->bdev;
1375                 fake_dentry.d_inode = bdev->bd_inode;
1376                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1377         }
1378
1379         if (m->queue_io)
1380                 r = -EAGAIN;
1381         else if (!bdev)
1382                 r = -EIO;
1383
1384         spin_unlock_irqrestore(&m->lock, flags);
1385
1386         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1387                                          bdev->bd_disk, cmd, arg);
1388 }
1389
1390 /*-----------------------------------------------------------------
1391  * Module setup
1392  *---------------------------------------------------------------*/
1393 static struct target_type multipath_target = {
1394         .name = "multipath",
1395         .version = {1, 0, 5},
1396         .module = THIS_MODULE,
1397         .ctr = multipath_ctr,
1398         .dtr = multipath_dtr,
1399         .map = multipath_map,
1400         .end_io = multipath_end_io,
1401         .presuspend = multipath_presuspend,
1402         .resume = multipath_resume,
1403         .status = multipath_status,
1404         .message = multipath_message,
1405         .ioctl  = multipath_ioctl,
1406 };
1407
1408 static int __init dm_multipath_init(void)
1409 {
1410         int r;
1411
1412         /* allocate a slab for the dm_ios */
1413         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1414         if (!_mpio_cache)
1415                 return -ENOMEM;
1416
1417         r = dm_register_target(&multipath_target);
1418         if (r < 0) {
1419                 DMERR("register failed %d", r);
1420                 kmem_cache_destroy(_mpio_cache);
1421                 return -EINVAL;
1422         }
1423
1424         kmultipathd = create_workqueue("kmpathd");
1425         if (!kmultipathd) {
1426                 DMERR("failed to create workqueue kmpathd");
1427                 dm_unregister_target(&multipath_target);
1428                 kmem_cache_destroy(_mpio_cache);
1429                 return -ENOMEM;
1430         }
1431
1432         /*
1433          * A separate workqueue is used to handle the device handlers
1434          * to avoid overloading existing workqueue. Overloading the
1435          * old workqueue would also create a bottleneck in the
1436          * path of the storage hardware device activation.
1437          */
1438         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1439         if (!kmpath_handlerd) {
1440                 DMERR("failed to create workqueue kmpath_handlerd");
1441                 destroy_workqueue(kmultipathd);
1442                 dm_unregister_target(&multipath_target);
1443                 kmem_cache_destroy(_mpio_cache);
1444                 return -ENOMEM;
1445         }
1446
1447         DMINFO("version %u.%u.%u loaded",
1448                multipath_target.version[0], multipath_target.version[1],
1449                multipath_target.version[2]);
1450
1451         return r;
1452 }
1453
1454 static void __exit dm_multipath_exit(void)
1455 {
1456         int r;
1457
1458         destroy_workqueue(kmpath_handlerd);
1459         destroy_workqueue(kmultipathd);
1460
1461         r = dm_unregister_target(&multipath_target);
1462         if (r < 0)
1463                 DMERR("target unregister failed %d", r);
1464         kmem_cache_destroy(_mpio_cache);
1465 }
1466
1467 module_init(dm_multipath_init);
1468 module_exit(dm_multipath_exit);
1469
1470 MODULE_DESCRIPTION(DM_NAME " multipath target");
1471 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1472 MODULE_LICENSE("GPL");