dm log: use standard kernel module refcount
[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 <linux/device-mapper.h>
9
10 #include "dm-path-selector.h"
11 #include "dm-bio-list.h"
12 #include "dm-bio-record.h"
13 #include "dm-uevent.h"
14
15 #include <linux/ctype.h>
16 #include <linux/init.h>
17 #include <linux/mempool.h>
18 #include <linux/module.h>
19 #include <linux/pagemap.h>
20 #include <linux/slab.h>
21 #include <linux/time.h>
22 #include <linux/workqueue.h>
23 #include <scsi/scsi_dh.h>
24 #include <asm/atomic.h>
25
26 #define DM_MSG_PREFIX "multipath"
27 #define MESG_STR(x) x, sizeof(x)
28
29 /* Path properties */
30 struct pgpath {
31         struct list_head list;
32
33         struct priority_group *pg;      /* Owning PG */
34         unsigned is_active;             /* Path status */
35         unsigned fail_count;            /* Cumulative failure count */
36
37         struct dm_path path;
38         struct work_struct deactivate_path;
39 };
40
41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
42
43 /*
44  * Paths are grouped into Priority Groups and numbered from 1 upwards.
45  * Each has a path selector which controls which path gets used.
46  */
47 struct priority_group {
48         struct list_head list;
49
50         struct multipath *m;            /* Owning multipath instance */
51         struct path_selector ps;
52
53         unsigned pg_num;                /* Reference number */
54         unsigned bypassed;              /* Temporarily bypass this PG? */
55
56         unsigned nr_pgpaths;            /* Number of paths in PG */
57         struct list_head pgpaths;
58 };
59
60 /* Multipath context */
61 struct multipath {
62         struct list_head list;
63         struct dm_target *ti;
64
65         spinlock_t lock;
66
67         const char *hw_handler_name;
68         struct work_struct activate_path;
69         struct pgpath *pgpath_to_activate;
70         unsigned nr_priority_groups;
71         struct list_head priority_groups;
72         unsigned pg_init_required;      /* pg_init needs calling? */
73         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
74
75         unsigned nr_valid_paths;        /* Total number of usable paths */
76         struct pgpath *current_pgpath;
77         struct priority_group *current_pg;
78         struct priority_group *next_pg; /* Switch to this PG if set */
79         unsigned repeat_count;          /* I/Os left before calling PS again */
80
81         unsigned queue_io;              /* Must we queue all I/O? */
82         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
83         unsigned saved_queue_if_no_path;/* Saved state during suspension */
84         unsigned pg_init_retries;       /* Number of times to retry pg_init */
85         unsigned pg_init_count;         /* Number of times pg_init called */
86
87         struct work_struct process_queued_ios;
88         struct bio_list queued_ios;
89         unsigned queue_size;
90
91         struct work_struct trigger_event;
92
93         /*
94          * We must use a mempool of dm_mpath_io structs so that we
95          * can resubmit bios on error.
96          */
97         mempool_t *mpio_pool;
98 };
99
100 /*
101  * Context information attached to each bio we process.
102  */
103 struct dm_mpath_io {
104         struct pgpath *pgpath;
105         struct dm_bio_details details;
106 };
107
108 typedef int (*action_fn) (struct pgpath *pgpath);
109
110 #define MIN_IOS 256     /* Mempool size */
111
112 static struct kmem_cache *_mpio_cache;
113
114 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
115 static void process_queued_ios(struct work_struct *work);
116 static void trigger_event(struct work_struct *work);
117 static void activate_path(struct work_struct *work);
118 static void deactivate_path(struct work_struct *work);
119
120
121 /*-----------------------------------------------
122  * Allocation routines
123  *-----------------------------------------------*/
124
125 static struct pgpath *alloc_pgpath(void)
126 {
127         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
128
129         if (pgpath) {
130                 pgpath->is_active = 1;
131                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
132         }
133
134         return pgpath;
135 }
136
137 static void free_pgpath(struct pgpath *pgpath)
138 {
139         kfree(pgpath);
140 }
141
142 static void deactivate_path(struct work_struct *work)
143 {
144         struct pgpath *pgpath =
145                 container_of(work, struct pgpath, deactivate_path);
146
147         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
148 }
149
150 static struct priority_group *alloc_priority_group(void)
151 {
152         struct priority_group *pg;
153
154         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
155
156         if (pg)
157                 INIT_LIST_HEAD(&pg->pgpaths);
158
159         return pg;
160 }
161
162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
163 {
164         unsigned long flags;
165         struct pgpath *pgpath, *tmp;
166         struct multipath *m = ti->private;
167
168         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
169                 list_del(&pgpath->list);
170                 if (m->hw_handler_name)
171                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
172                 dm_put_device(ti, pgpath->path.dev);
173                 spin_lock_irqsave(&m->lock, flags);
174                 if (m->pgpath_to_activate == pgpath)
175                         m->pgpath_to_activate = NULL;
176                 spin_unlock_irqrestore(&m->lock, flags);
177                 free_pgpath(pgpath);
178         }
179 }
180
181 static void free_priority_group(struct priority_group *pg,
182                                 struct dm_target *ti)
183 {
184         struct path_selector *ps = &pg->ps;
185
186         if (ps->type) {
187                 ps->type->destroy(ps);
188                 dm_put_path_selector(ps->type);
189         }
190
191         free_pgpaths(&pg->pgpaths, ti);
192         kfree(pg);
193 }
194
195 static struct multipath *alloc_multipath(struct dm_target *ti)
196 {
197         struct multipath *m;
198
199         m = kzalloc(sizeof(*m), GFP_KERNEL);
200         if (m) {
201                 INIT_LIST_HEAD(&m->priority_groups);
202                 spin_lock_init(&m->lock);
203                 m->queue_io = 1;
204                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
205                 INIT_WORK(&m->trigger_event, trigger_event);
206                 INIT_WORK(&m->activate_path, activate_path);
207                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
208                 if (!m->mpio_pool) {
209                         kfree(m);
210                         return NULL;
211                 }
212                 m->ti = ti;
213                 ti->private = m;
214         }
215
216         return m;
217 }
218
219 static void free_multipath(struct multipath *m)
220 {
221         struct priority_group *pg, *tmp;
222
223         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
224                 list_del(&pg->list);
225                 free_priority_group(pg, m->ti);
226         }
227
228         kfree(m->hw_handler_name);
229         mempool_destroy(m->mpio_pool);
230         kfree(m);
231 }
232
233
234 /*-----------------------------------------------
235  * Path selection
236  *-----------------------------------------------*/
237
238 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
239 {
240         m->current_pg = pgpath->pg;
241
242         /* Must we initialise the PG first, and queue I/O till it's ready? */
243         if (m->hw_handler_name) {
244                 m->pg_init_required = 1;
245                 m->queue_io = 1;
246         } else {
247                 m->pg_init_required = 0;
248                 m->queue_io = 0;
249         }
250
251         m->pg_init_count = 0;
252 }
253
254 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
255 {
256         struct dm_path *path;
257
258         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
259         if (!path)
260                 return -ENXIO;
261
262         m->current_pgpath = path_to_pgpath(path);
263
264         if (m->current_pg != pg)
265                 __switch_pg(m, m->current_pgpath);
266
267         return 0;
268 }
269
270 static void __choose_pgpath(struct multipath *m)
271 {
272         struct priority_group *pg;
273         unsigned bypassed = 1;
274
275         if (!m->nr_valid_paths)
276                 goto failed;
277
278         /* Were we instructed to switch PG? */
279         if (m->next_pg) {
280                 pg = m->next_pg;
281                 m->next_pg = NULL;
282                 if (!__choose_path_in_pg(m, pg))
283                         return;
284         }
285
286         /* Don't change PG until it has no remaining paths */
287         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
288                 return;
289
290         /*
291          * Loop through priority groups until we find a valid path.
292          * First time we skip PGs marked 'bypassed'.
293          * Second time we only try the ones we skipped.
294          */
295         do {
296                 list_for_each_entry(pg, &m->priority_groups, list) {
297                         if (pg->bypassed == bypassed)
298                                 continue;
299                         if (!__choose_path_in_pg(m, pg))
300                                 return;
301                 }
302         } while (bypassed--);
303
304 failed:
305         m->current_pgpath = NULL;
306         m->current_pg = NULL;
307 }
308
309 /*
310  * Check whether bios must be queued in the device-mapper core rather
311  * than here in the target.
312  *
313  * m->lock must be held on entry.
314  *
315  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
316  * same value then we are not between multipath_presuspend()
317  * and multipath_resume() calls and we have no need to check
318  * for the DMF_NOFLUSH_SUSPENDING flag.
319  */
320 static int __must_push_back(struct multipath *m)
321 {
322         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
323                 dm_noflush_suspending(m->ti));
324 }
325
326 static int map_io(struct multipath *m, struct bio *bio,
327                   struct dm_mpath_io *mpio, unsigned was_queued)
328 {
329         int r = DM_MAPIO_REMAPPED;
330         unsigned long flags;
331         struct pgpath *pgpath;
332
333         spin_lock_irqsave(&m->lock, flags);
334
335         /* Do we need to select a new pgpath? */
336         if (!m->current_pgpath ||
337             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
338                 __choose_pgpath(m);
339
340         pgpath = m->current_pgpath;
341
342         if (was_queued)
343                 m->queue_size--;
344
345         if ((pgpath && m->queue_io) ||
346             (!pgpath && m->queue_if_no_path)) {
347                 /* Queue for the daemon to resubmit */
348                 bio_list_add(&m->queued_ios, bio);
349                 m->queue_size++;
350                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
351                     !m->queue_io)
352                         queue_work(kmultipathd, &m->process_queued_ios);
353                 pgpath = NULL;
354                 r = DM_MAPIO_SUBMITTED;
355         } else if (pgpath)
356                 bio->bi_bdev = pgpath->path.dev->bdev;
357         else if (__must_push_back(m))
358                 r = DM_MAPIO_REQUEUE;
359         else
360                 r = -EIO;       /* Failed */
361
362         mpio->pgpath = pgpath;
363
364         spin_unlock_irqrestore(&m->lock, flags);
365
366         return r;
367 }
368
369 /*
370  * If we run out of usable paths, should we queue I/O or error it?
371  */
372 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
373                             unsigned save_old_value)
374 {
375         unsigned long flags;
376
377         spin_lock_irqsave(&m->lock, flags);
378
379         if (save_old_value)
380                 m->saved_queue_if_no_path = m->queue_if_no_path;
381         else
382                 m->saved_queue_if_no_path = queue_if_no_path;
383         m->queue_if_no_path = queue_if_no_path;
384         if (!m->queue_if_no_path && m->queue_size)
385                 queue_work(kmultipathd, &m->process_queued_ios);
386
387         spin_unlock_irqrestore(&m->lock, flags);
388
389         return 0;
390 }
391
392 /*-----------------------------------------------------------------
393  * The multipath daemon is responsible for resubmitting queued ios.
394  *---------------------------------------------------------------*/
395
396 static void dispatch_queued_ios(struct multipath *m)
397 {
398         int r;
399         unsigned long flags;
400         struct bio *bio = NULL, *next;
401         struct dm_mpath_io *mpio;
402         union map_info *info;
403
404         spin_lock_irqsave(&m->lock, flags);
405         bio = bio_list_get(&m->queued_ios);
406         spin_unlock_irqrestore(&m->lock, flags);
407
408         while (bio) {
409                 next = bio->bi_next;
410                 bio->bi_next = NULL;
411
412                 info = dm_get_mapinfo(bio);
413                 mpio = info->ptr;
414
415                 r = map_io(m, bio, mpio, 1);
416                 if (r < 0)
417                         bio_endio(bio, r);
418                 else if (r == DM_MAPIO_REMAPPED)
419                         generic_make_request(bio);
420                 else if (r == DM_MAPIO_REQUEUE)
421                         bio_endio(bio, -EIO);
422
423                 bio = next;
424         }
425 }
426
427 static void process_queued_ios(struct work_struct *work)
428 {
429         struct multipath *m =
430                 container_of(work, struct multipath, process_queued_ios);
431         struct pgpath *pgpath = NULL;
432         unsigned init_required = 0, must_queue = 1;
433         unsigned long flags;
434
435         spin_lock_irqsave(&m->lock, flags);
436
437         if (!m->queue_size)
438                 goto out;
439
440         if (!m->current_pgpath)
441                 __choose_pgpath(m);
442
443         pgpath = m->current_pgpath;
444
445         if ((pgpath && !m->queue_io) ||
446             (!pgpath && !m->queue_if_no_path))
447                 must_queue = 0;
448
449         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
450                 m->pgpath_to_activate = pgpath;
451                 m->pg_init_count++;
452                 m->pg_init_required = 0;
453                 m->pg_init_in_progress = 1;
454                 init_required = 1;
455         }
456
457 out:
458         spin_unlock_irqrestore(&m->lock, flags);
459
460         if (init_required)
461                 queue_work(kmpath_handlerd, &m->activate_path);
462
463         if (!must_queue)
464                 dispatch_queued_ios(m);
465 }
466
467 /*
468  * An event is triggered whenever a path is taken out of use.
469  * Includes path failure and PG bypass.
470  */
471 static void trigger_event(struct work_struct *work)
472 {
473         struct multipath *m =
474                 container_of(work, struct multipath, trigger_event);
475
476         dm_table_event(m->ti->table);
477 }
478
479 /*-----------------------------------------------------------------
480  * Constructor/argument parsing:
481  * <#multipath feature args> [<arg>]*
482  * <#hw_handler args> [hw_handler [<arg>]*]
483  * <#priority groups>
484  * <initial priority group>
485  *     [<selector> <#selector args> [<arg>]*
486  *      <#paths> <#per-path selector args>
487  *         [<path> [<arg>]* ]+ ]+
488  *---------------------------------------------------------------*/
489 struct param {
490         unsigned min;
491         unsigned max;
492         char *error;
493 };
494
495 static int read_param(struct param *param, char *str, unsigned *v, char **error)
496 {
497         if (!str ||
498             (sscanf(str, "%u", v) != 1) ||
499             (*v < param->min) ||
500             (*v > param->max)) {
501                 *error = param->error;
502                 return -EINVAL;
503         }
504
505         return 0;
506 }
507
508 struct arg_set {
509         unsigned argc;
510         char **argv;
511 };
512
513 static char *shift(struct arg_set *as)
514 {
515         char *r;
516
517         if (as->argc) {
518                 as->argc--;
519                 r = *as->argv;
520                 as->argv++;
521                 return r;
522         }
523
524         return NULL;
525 }
526
527 static void consume(struct arg_set *as, unsigned n)
528 {
529         BUG_ON (as->argc < n);
530         as->argc -= n;
531         as->argv += n;
532 }
533
534 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
535                                struct dm_target *ti)
536 {
537         int r;
538         struct path_selector_type *pst;
539         unsigned ps_argc;
540
541         static struct param _params[] = {
542                 {0, 1024, "invalid number of path selector args"},
543         };
544
545         pst = dm_get_path_selector(shift(as));
546         if (!pst) {
547                 ti->error = "unknown path selector type";
548                 return -EINVAL;
549         }
550
551         r = read_param(_params, shift(as), &ps_argc, &ti->error);
552         if (r) {
553                 dm_put_path_selector(pst);
554                 return -EINVAL;
555         }
556
557         r = pst->create(&pg->ps, ps_argc, as->argv);
558         if (r) {
559                 dm_put_path_selector(pst);
560                 ti->error = "path selector constructor failed";
561                 return r;
562         }
563
564         pg->ps.type = pst;
565         consume(as, ps_argc);
566
567         return 0;
568 }
569
570 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
571                                struct dm_target *ti)
572 {
573         int r;
574         struct pgpath *p;
575         struct multipath *m = ti->private;
576
577         /* we need at least a path arg */
578         if (as->argc < 1) {
579                 ti->error = "no device given";
580                 return ERR_PTR(-EINVAL);
581         }
582
583         p = alloc_pgpath();
584         if (!p)
585                 return ERR_PTR(-ENOMEM);
586
587         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
588                           dm_table_get_mode(ti->table), &p->path.dev);
589         if (r) {
590                 ti->error = "error getting device";
591                 goto bad;
592         }
593
594         if (m->hw_handler_name) {
595                 r = scsi_dh_attach(bdev_get_queue(p->path.dev->bdev),
596                                    m->hw_handler_name);
597                 if (r < 0) {
598                         dm_put_device(ti, p->path.dev);
599                         goto bad;
600                 }
601         }
602
603         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
604         if (r) {
605                 dm_put_device(ti, p->path.dev);
606                 goto bad;
607         }
608
609         return p;
610
611  bad:
612         free_pgpath(p);
613         return ERR_PTR(r);
614 }
615
616 static struct priority_group *parse_priority_group(struct arg_set *as,
617                                                    struct multipath *m)
618 {
619         static struct param _params[] = {
620                 {1, 1024, "invalid number of paths"},
621                 {0, 1024, "invalid number of selector args"}
622         };
623
624         int r;
625         unsigned i, nr_selector_args, nr_params;
626         struct priority_group *pg;
627         struct dm_target *ti = m->ti;
628
629         if (as->argc < 2) {
630                 as->argc = 0;
631                 ti->error = "not enough priority group arguments";
632                 return ERR_PTR(-EINVAL);
633         }
634
635         pg = alloc_priority_group();
636         if (!pg) {
637                 ti->error = "couldn't allocate priority group";
638                 return ERR_PTR(-ENOMEM);
639         }
640         pg->m = m;
641
642         r = parse_path_selector(as, pg, ti);
643         if (r)
644                 goto bad;
645
646         /*
647          * read the paths
648          */
649         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
650         if (r)
651                 goto bad;
652
653         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
654         if (r)
655                 goto bad;
656
657         nr_params = 1 + nr_selector_args;
658         for (i = 0; i < pg->nr_pgpaths; i++) {
659                 struct pgpath *pgpath;
660                 struct arg_set path_args;
661
662                 if (as->argc < nr_params) {
663                         ti->error = "not enough path parameters";
664                         goto bad;
665                 }
666
667                 path_args.argc = nr_params;
668                 path_args.argv = as->argv;
669
670                 pgpath = parse_path(&path_args, &pg->ps, ti);
671                 if (IS_ERR(pgpath)) {
672                         r = PTR_ERR(pgpath);
673                         goto bad;
674                 }
675
676                 pgpath->pg = pg;
677                 list_add_tail(&pgpath->list, &pg->pgpaths);
678                 consume(as, nr_params);
679         }
680
681         return pg;
682
683  bad:
684         free_priority_group(pg, ti);
685         return ERR_PTR(r);
686 }
687
688 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
689 {
690         unsigned hw_argc;
691         struct dm_target *ti = m->ti;
692
693         static struct param _params[] = {
694                 {0, 1024, "invalid number of hardware handler args"},
695         };
696
697         if (read_param(_params, shift(as), &hw_argc, &ti->error))
698                 return -EINVAL;
699
700         if (!hw_argc)
701                 return 0;
702
703         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
704         request_module("scsi_dh_%s", m->hw_handler_name);
705         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
706                 ti->error = "unknown hardware handler type";
707                 kfree(m->hw_handler_name);
708                 m->hw_handler_name = NULL;
709                 return -EINVAL;
710         }
711
712         if (hw_argc > 1)
713                 DMWARN("Ignoring user-specified arguments for "
714                        "hardware handler \"%s\"", m->hw_handler_name);
715         consume(as, hw_argc - 1);
716
717         return 0;
718 }
719
720 static int parse_features(struct arg_set *as, struct multipath *m)
721 {
722         int r;
723         unsigned argc;
724         struct dm_target *ti = m->ti;
725         const char *param_name;
726
727         static struct param _params[] = {
728                 {0, 3, "invalid number of feature args"},
729                 {1, 50, "pg_init_retries must be between 1 and 50"},
730         };
731
732         r = read_param(_params, shift(as), &argc, &ti->error);
733         if (r)
734                 return -EINVAL;
735
736         if (!argc)
737                 return 0;
738
739         do {
740                 param_name = shift(as);
741                 argc--;
742
743                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
744                         r = queue_if_no_path(m, 1, 0);
745                         continue;
746                 }
747
748                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
749                     (argc >= 1)) {
750                         r = read_param(_params + 1, shift(as),
751                                        &m->pg_init_retries, &ti->error);
752                         argc--;
753                         continue;
754                 }
755
756                 ti->error = "Unrecognised multipath feature request";
757                 r = -EINVAL;
758         } while (argc && !r);
759
760         return r;
761 }
762
763 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
764                          char **argv)
765 {
766         /* target parameters */
767         static struct param _params[] = {
768                 {1, 1024, "invalid number of priority groups"},
769                 {1, 1024, "invalid initial priority group number"},
770         };
771
772         int r;
773         struct multipath *m;
774         struct arg_set as;
775         unsigned pg_count = 0;
776         unsigned next_pg_num;
777
778         as.argc = argc;
779         as.argv = argv;
780
781         m = alloc_multipath(ti);
782         if (!m) {
783                 ti->error = "can't allocate multipath";
784                 return -EINVAL;
785         }
786
787         r = parse_features(&as, m);
788         if (r)
789                 goto bad;
790
791         r = parse_hw_handler(&as, m);
792         if (r)
793                 goto bad;
794
795         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
796         if (r)
797                 goto bad;
798
799         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
800         if (r)
801                 goto bad;
802
803         /* parse the priority groups */
804         while (as.argc) {
805                 struct priority_group *pg;
806
807                 pg = parse_priority_group(&as, m);
808                 if (IS_ERR(pg)) {
809                         r = PTR_ERR(pg);
810                         goto bad;
811                 }
812
813                 m->nr_valid_paths += pg->nr_pgpaths;
814                 list_add_tail(&pg->list, &m->priority_groups);
815                 pg_count++;
816                 pg->pg_num = pg_count;
817                 if (!--next_pg_num)
818                         m->next_pg = pg;
819         }
820
821         if (pg_count != m->nr_priority_groups) {
822                 ti->error = "priority group count mismatch";
823                 r = -EINVAL;
824                 goto bad;
825         }
826
827         return 0;
828
829  bad:
830         free_multipath(m);
831         return r;
832 }
833
834 static void multipath_dtr(struct dm_target *ti)
835 {
836         struct multipath *m = (struct multipath *) ti->private;
837
838         flush_workqueue(kmpath_handlerd);
839         flush_workqueue(kmultipathd);
840         free_multipath(m);
841 }
842
843 /*
844  * Map bios, recording original fields for later in case we have to resubmit
845  */
846 static int multipath_map(struct dm_target *ti, struct bio *bio,
847                          union map_info *map_context)
848 {
849         int r;
850         struct dm_mpath_io *mpio;
851         struct multipath *m = (struct multipath *) ti->private;
852
853         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
854         dm_bio_record(&mpio->details, bio);
855
856         map_context->ptr = mpio;
857         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
858         r = map_io(m, bio, mpio, 0);
859         if (r < 0 || r == DM_MAPIO_REQUEUE)
860                 mempool_free(mpio, m->mpio_pool);
861
862         return r;
863 }
864
865 /*
866  * Take a path out of use.
867  */
868 static int fail_path(struct pgpath *pgpath)
869 {
870         unsigned long flags;
871         struct multipath *m = pgpath->pg->m;
872
873         spin_lock_irqsave(&m->lock, flags);
874
875         if (!pgpath->is_active)
876                 goto out;
877
878         DMWARN("Failing path %s.", pgpath->path.dev->name);
879
880         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
881         pgpath->is_active = 0;
882         pgpath->fail_count++;
883
884         m->nr_valid_paths--;
885
886         if (pgpath == m->current_pgpath)
887                 m->current_pgpath = NULL;
888
889         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
890                       pgpath->path.dev->name, m->nr_valid_paths);
891
892         schedule_work(&m->trigger_event);
893         queue_work(kmultipathd, &pgpath->deactivate_path);
894
895 out:
896         spin_unlock_irqrestore(&m->lock, flags);
897
898         return 0;
899 }
900
901 /*
902  * Reinstate a previously-failed path
903  */
904 static int reinstate_path(struct pgpath *pgpath)
905 {
906         int r = 0;
907         unsigned long flags;
908         struct multipath *m = pgpath->pg->m;
909
910         spin_lock_irqsave(&m->lock, flags);
911
912         if (pgpath->is_active)
913                 goto out;
914
915         if (!pgpath->pg->ps.type->reinstate_path) {
916                 DMWARN("Reinstate path not supported by path selector %s",
917                        pgpath->pg->ps.type->name);
918                 r = -EINVAL;
919                 goto out;
920         }
921
922         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
923         if (r)
924                 goto out;
925
926         pgpath->is_active = 1;
927
928         m->current_pgpath = NULL;
929         if (!m->nr_valid_paths++ && m->queue_size)
930                 queue_work(kmultipathd, &m->process_queued_ios);
931
932         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
933                       pgpath->path.dev->name, m->nr_valid_paths);
934
935         schedule_work(&m->trigger_event);
936
937 out:
938         spin_unlock_irqrestore(&m->lock, flags);
939
940         return r;
941 }
942
943 /*
944  * Fail or reinstate all paths that match the provided struct dm_dev.
945  */
946 static int action_dev(struct multipath *m, struct dm_dev *dev,
947                       action_fn action)
948 {
949         int r = 0;
950         struct pgpath *pgpath;
951         struct priority_group *pg;
952
953         list_for_each_entry(pg, &m->priority_groups, list) {
954                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
955                         if (pgpath->path.dev == dev)
956                                 r = action(pgpath);
957                 }
958         }
959
960         return r;
961 }
962
963 /*
964  * Temporarily try to avoid having to use the specified PG
965  */
966 static void bypass_pg(struct multipath *m, struct priority_group *pg,
967                       int bypassed)
968 {
969         unsigned long flags;
970
971         spin_lock_irqsave(&m->lock, flags);
972
973         pg->bypassed = bypassed;
974         m->current_pgpath = NULL;
975         m->current_pg = NULL;
976
977         spin_unlock_irqrestore(&m->lock, flags);
978
979         schedule_work(&m->trigger_event);
980 }
981
982 /*
983  * Switch to using the specified PG from the next I/O that gets mapped
984  */
985 static int switch_pg_num(struct multipath *m, const char *pgstr)
986 {
987         struct priority_group *pg;
988         unsigned pgnum;
989         unsigned long flags;
990
991         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
992             (pgnum > m->nr_priority_groups)) {
993                 DMWARN("invalid PG number supplied to switch_pg_num");
994                 return -EINVAL;
995         }
996
997         spin_lock_irqsave(&m->lock, flags);
998         list_for_each_entry(pg, &m->priority_groups, list) {
999                 pg->bypassed = 0;
1000                 if (--pgnum)
1001                         continue;
1002
1003                 m->current_pgpath = NULL;
1004                 m->current_pg = NULL;
1005                 m->next_pg = pg;
1006         }
1007         spin_unlock_irqrestore(&m->lock, flags);
1008
1009         schedule_work(&m->trigger_event);
1010         return 0;
1011 }
1012
1013 /*
1014  * Set/clear bypassed status of a PG.
1015  * PGs are numbered upwards from 1 in the order they were declared.
1016  */
1017 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1018 {
1019         struct priority_group *pg;
1020         unsigned pgnum;
1021
1022         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1023             (pgnum > m->nr_priority_groups)) {
1024                 DMWARN("invalid PG number supplied to bypass_pg");
1025                 return -EINVAL;
1026         }
1027
1028         list_for_each_entry(pg, &m->priority_groups, list) {
1029                 if (!--pgnum)
1030                         break;
1031         }
1032
1033         bypass_pg(m, pg, bypassed);
1034         return 0;
1035 }
1036
1037 /*
1038  * Should we retry pg_init immediately?
1039  */
1040 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1041 {
1042         unsigned long flags;
1043         int limit_reached = 0;
1044
1045         spin_lock_irqsave(&m->lock, flags);
1046
1047         if (m->pg_init_count <= m->pg_init_retries)
1048                 m->pg_init_required = 1;
1049         else
1050                 limit_reached = 1;
1051
1052         spin_unlock_irqrestore(&m->lock, flags);
1053
1054         return limit_reached;
1055 }
1056
1057 static void pg_init_done(struct dm_path *path, int errors)
1058 {
1059         struct pgpath *pgpath = path_to_pgpath(path);
1060         struct priority_group *pg = pgpath->pg;
1061         struct multipath *m = pg->m;
1062         unsigned long flags;
1063
1064         /* device or driver problems */
1065         switch (errors) {
1066         case SCSI_DH_OK:
1067                 break;
1068         case SCSI_DH_NOSYS:
1069                 if (!m->hw_handler_name) {
1070                         errors = 0;
1071                         break;
1072                 }
1073                 DMERR("Cannot failover device because scsi_dh_%s was not "
1074                       "loaded.", m->hw_handler_name);
1075                 /*
1076                  * Fail path for now, so we do not ping pong
1077                  */
1078                 fail_path(pgpath);
1079                 break;
1080         case SCSI_DH_DEV_TEMP_BUSY:
1081                 /*
1082                  * Probably doing something like FW upgrade on the
1083                  * controller so try the other pg.
1084                  */
1085                 bypass_pg(m, pg, 1);
1086                 break;
1087         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1088         case SCSI_DH_RETRY:
1089         case SCSI_DH_IMM_RETRY:
1090         case SCSI_DH_RES_TEMP_UNAVAIL:
1091                 if (pg_init_limit_reached(m, pgpath))
1092                         fail_path(pgpath);
1093                 errors = 0;
1094                 break;
1095         default:
1096                 /*
1097                  * We probably do not want to fail the path for a device
1098                  * error, but this is what the old dm did. In future
1099                  * patches we can do more advanced handling.
1100                  */
1101                 fail_path(pgpath);
1102         }
1103
1104         spin_lock_irqsave(&m->lock, flags);
1105         if (errors) {
1106                 DMERR("Could not failover device. Error %d.", errors);
1107                 m->current_pgpath = NULL;
1108                 m->current_pg = NULL;
1109         } else if (!m->pg_init_required) {
1110                 m->queue_io = 0;
1111                 pg->bypassed = 0;
1112         }
1113
1114         m->pg_init_in_progress = 0;
1115         queue_work(kmultipathd, &m->process_queued_ios);
1116         spin_unlock_irqrestore(&m->lock, flags);
1117 }
1118
1119 static void activate_path(struct work_struct *work)
1120 {
1121         int ret;
1122         struct multipath *m =
1123                 container_of(work, struct multipath, activate_path);
1124         struct dm_path *path;
1125         unsigned long flags;
1126
1127         spin_lock_irqsave(&m->lock, flags);
1128         path = &m->pgpath_to_activate->path;
1129         m->pgpath_to_activate = NULL;
1130         spin_unlock_irqrestore(&m->lock, flags);
1131         if (!path)
1132                 return;
1133         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1134         pg_init_done(path, ret);
1135 }
1136
1137 /*
1138  * end_io handling
1139  */
1140 static int do_end_io(struct multipath *m, struct bio *bio,
1141                      int error, struct dm_mpath_io *mpio)
1142 {
1143         unsigned long flags;
1144
1145         if (!error)
1146                 return 0;       /* I/O complete */
1147
1148         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1149                 return error;
1150
1151         if (error == -EOPNOTSUPP)
1152                 return error;
1153
1154         spin_lock_irqsave(&m->lock, flags);
1155         if (!m->nr_valid_paths) {
1156                 if (__must_push_back(m)) {
1157                         spin_unlock_irqrestore(&m->lock, flags);
1158                         return DM_ENDIO_REQUEUE;
1159                 } else if (!m->queue_if_no_path) {
1160                         spin_unlock_irqrestore(&m->lock, flags);
1161                         return -EIO;
1162                 } else {
1163                         spin_unlock_irqrestore(&m->lock, flags);
1164                         goto requeue;
1165                 }
1166         }
1167         spin_unlock_irqrestore(&m->lock, flags);
1168
1169         if (mpio->pgpath)
1170                 fail_path(mpio->pgpath);
1171
1172       requeue:
1173         dm_bio_restore(&mpio->details, bio);
1174
1175         /* queue for the daemon to resubmit or fail */
1176         spin_lock_irqsave(&m->lock, flags);
1177         bio_list_add(&m->queued_ios, bio);
1178         m->queue_size++;
1179         if (!m->queue_io)
1180                 queue_work(kmultipathd, &m->process_queued_ios);
1181         spin_unlock_irqrestore(&m->lock, flags);
1182
1183         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1184 }
1185
1186 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1187                             int error, union map_info *map_context)
1188 {
1189         struct multipath *m = ti->private;
1190         struct dm_mpath_io *mpio = map_context->ptr;
1191         struct pgpath *pgpath = mpio->pgpath;
1192         struct path_selector *ps;
1193         int r;
1194
1195         r  = do_end_io(m, bio, error, mpio);
1196         if (pgpath) {
1197                 ps = &pgpath->pg->ps;
1198                 if (ps->type->end_io)
1199                         ps->type->end_io(ps, &pgpath->path);
1200         }
1201         if (r != DM_ENDIO_INCOMPLETE)
1202                 mempool_free(mpio, m->mpio_pool);
1203
1204         return r;
1205 }
1206
1207 /*
1208  * Suspend can't complete until all the I/O is processed so if
1209  * the last path fails we must error any remaining I/O.
1210  * Note that if the freeze_bdev fails while suspending, the
1211  * queue_if_no_path state is lost - userspace should reset it.
1212  */
1213 static void multipath_presuspend(struct dm_target *ti)
1214 {
1215         struct multipath *m = (struct multipath *) ti->private;
1216
1217         queue_if_no_path(m, 0, 1);
1218 }
1219
1220 /*
1221  * Restore the queue_if_no_path setting.
1222  */
1223 static void multipath_resume(struct dm_target *ti)
1224 {
1225         struct multipath *m = (struct multipath *) ti->private;
1226         unsigned long flags;
1227
1228         spin_lock_irqsave(&m->lock, flags);
1229         m->queue_if_no_path = m->saved_queue_if_no_path;
1230         spin_unlock_irqrestore(&m->lock, flags);
1231 }
1232
1233 /*
1234  * Info output has the following format:
1235  * num_multipath_feature_args [multipath_feature_args]*
1236  * num_handler_status_args [handler_status_args]*
1237  * num_groups init_group_number
1238  *            [A|D|E num_ps_status_args [ps_status_args]*
1239  *             num_paths num_selector_args
1240  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1241  *
1242  * Table output has the following format (identical to the constructor string):
1243  * num_feature_args [features_args]*
1244  * num_handler_args hw_handler [hw_handler_args]*
1245  * num_groups init_group_number
1246  *     [priority selector-name num_ps_args [ps_args]*
1247  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1248  */
1249 static int multipath_status(struct dm_target *ti, status_type_t type,
1250                             char *result, unsigned int maxlen)
1251 {
1252         int sz = 0;
1253         unsigned long flags;
1254         struct multipath *m = (struct multipath *) ti->private;
1255         struct priority_group *pg;
1256         struct pgpath *p;
1257         unsigned pg_num;
1258         char state;
1259
1260         spin_lock_irqsave(&m->lock, flags);
1261
1262         /* Features */
1263         if (type == STATUSTYPE_INFO)
1264                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1265         else {
1266                 DMEMIT("%u ", m->queue_if_no_path +
1267                               (m->pg_init_retries > 0) * 2);
1268                 if (m->queue_if_no_path)
1269                         DMEMIT("queue_if_no_path ");
1270                 if (m->pg_init_retries)
1271                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1272         }
1273
1274         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1275                 DMEMIT("0 ");
1276         else
1277                 DMEMIT("1 %s ", m->hw_handler_name);
1278
1279         DMEMIT("%u ", m->nr_priority_groups);
1280
1281         if (m->next_pg)
1282                 pg_num = m->next_pg->pg_num;
1283         else if (m->current_pg)
1284                 pg_num = m->current_pg->pg_num;
1285         else
1286                         pg_num = 1;
1287
1288         DMEMIT("%u ", pg_num);
1289
1290         switch (type) {
1291         case STATUSTYPE_INFO:
1292                 list_for_each_entry(pg, &m->priority_groups, list) {
1293                         if (pg->bypassed)
1294                                 state = 'D';    /* Disabled */
1295                         else if (pg == m->current_pg)
1296                                 state = 'A';    /* Currently Active */
1297                         else
1298                                 state = 'E';    /* Enabled */
1299
1300                         DMEMIT("%c ", state);
1301
1302                         if (pg->ps.type->status)
1303                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1304                                                           result + sz,
1305                                                           maxlen - sz);
1306                         else
1307                                 DMEMIT("0 ");
1308
1309                         DMEMIT("%u %u ", pg->nr_pgpaths,
1310                                pg->ps.type->info_args);
1311
1312                         list_for_each_entry(p, &pg->pgpaths, list) {
1313                                 DMEMIT("%s %s %u ", p->path.dev->name,
1314                                        p->is_active ? "A" : "F",
1315                                        p->fail_count);
1316                                 if (pg->ps.type->status)
1317                                         sz += pg->ps.type->status(&pg->ps,
1318                                               &p->path, type, result + sz,
1319                                               maxlen - sz);
1320                         }
1321                 }
1322                 break;
1323
1324         case STATUSTYPE_TABLE:
1325                 list_for_each_entry(pg, &m->priority_groups, list) {
1326                         DMEMIT("%s ", pg->ps.type->name);
1327
1328                         if (pg->ps.type->status)
1329                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1330                                                           result + sz,
1331                                                           maxlen - sz);
1332                         else
1333                                 DMEMIT("0 ");
1334
1335                         DMEMIT("%u %u ", pg->nr_pgpaths,
1336                                pg->ps.type->table_args);
1337
1338                         list_for_each_entry(p, &pg->pgpaths, list) {
1339                                 DMEMIT("%s ", p->path.dev->name);
1340                                 if (pg->ps.type->status)
1341                                         sz += pg->ps.type->status(&pg->ps,
1342                                               &p->path, type, result + sz,
1343                                               maxlen - sz);
1344                         }
1345                 }
1346                 break;
1347         }
1348
1349         spin_unlock_irqrestore(&m->lock, flags);
1350
1351         return 0;
1352 }
1353
1354 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1355 {
1356         int r;
1357         struct dm_dev *dev;
1358         struct multipath *m = (struct multipath *) ti->private;
1359         action_fn action;
1360
1361         if (argc == 1) {
1362                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1363                         return queue_if_no_path(m, 1, 0);
1364                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1365                         return queue_if_no_path(m, 0, 0);
1366         }
1367
1368         if (argc != 2)
1369                 goto error;
1370
1371         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1372                 return bypass_pg_num(m, argv[1], 1);
1373         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1374                 return bypass_pg_num(m, argv[1], 0);
1375         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1376                 return switch_pg_num(m, argv[1]);
1377         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1378                 action = reinstate_path;
1379         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1380                 action = fail_path;
1381         else
1382                 goto error;
1383
1384         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1385                           dm_table_get_mode(ti->table), &dev);
1386         if (r) {
1387                 DMWARN("message: error getting device %s",
1388                        argv[1]);
1389                 return -EINVAL;
1390         }
1391
1392         r = action_dev(m, dev, action);
1393
1394         dm_put_device(ti, dev);
1395
1396         return r;
1397
1398 error:
1399         DMWARN("Unrecognised multipath message received.");
1400         return -EINVAL;
1401 }
1402
1403 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1404                            unsigned long arg)
1405 {
1406         struct multipath *m = (struct multipath *) ti->private;
1407         struct block_device *bdev = NULL;
1408         fmode_t mode = 0;
1409         unsigned long flags;
1410         int r = 0;
1411
1412         spin_lock_irqsave(&m->lock, flags);
1413
1414         if (!m->current_pgpath)
1415                 __choose_pgpath(m);
1416
1417         if (m->current_pgpath) {
1418                 bdev = m->current_pgpath->path.dev->bdev;
1419                 mode = m->current_pgpath->path.dev->mode;
1420         }
1421
1422         if (m->queue_io)
1423                 r = -EAGAIN;
1424         else if (!bdev)
1425                 r = -EIO;
1426
1427         spin_unlock_irqrestore(&m->lock, flags);
1428
1429         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1430 }
1431
1432 /*-----------------------------------------------------------------
1433  * Module setup
1434  *---------------------------------------------------------------*/
1435 static struct target_type multipath_target = {
1436         .name = "multipath",
1437         .version = {1, 0, 5},
1438         .module = THIS_MODULE,
1439         .ctr = multipath_ctr,
1440         .dtr = multipath_dtr,
1441         .map = multipath_map,
1442         .end_io = multipath_end_io,
1443         .presuspend = multipath_presuspend,
1444         .resume = multipath_resume,
1445         .status = multipath_status,
1446         .message = multipath_message,
1447         .ioctl  = multipath_ioctl,
1448 };
1449
1450 static int __init dm_multipath_init(void)
1451 {
1452         int r;
1453
1454         /* allocate a slab for the dm_ios */
1455         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1456         if (!_mpio_cache)
1457                 return -ENOMEM;
1458
1459         r = dm_register_target(&multipath_target);
1460         if (r < 0) {
1461                 DMERR("register failed %d", r);
1462                 kmem_cache_destroy(_mpio_cache);
1463                 return -EINVAL;
1464         }
1465
1466         kmultipathd = create_workqueue("kmpathd");
1467         if (!kmultipathd) {
1468                 DMERR("failed to create workqueue kmpathd");
1469                 dm_unregister_target(&multipath_target);
1470                 kmem_cache_destroy(_mpio_cache);
1471                 return -ENOMEM;
1472         }
1473
1474         /*
1475          * A separate workqueue is used to handle the device handlers
1476          * to avoid overloading existing workqueue. Overloading the
1477          * old workqueue would also create a bottleneck in the
1478          * path of the storage hardware device activation.
1479          */
1480         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1481         if (!kmpath_handlerd) {
1482                 DMERR("failed to create workqueue kmpath_handlerd");
1483                 destroy_workqueue(kmultipathd);
1484                 dm_unregister_target(&multipath_target);
1485                 kmem_cache_destroy(_mpio_cache);
1486                 return -ENOMEM;
1487         }
1488
1489         DMINFO("version %u.%u.%u loaded",
1490                multipath_target.version[0], multipath_target.version[1],
1491                multipath_target.version[2]);
1492
1493         return r;
1494 }
1495
1496 static void __exit dm_multipath_exit(void)
1497 {
1498         destroy_workqueue(kmpath_handlerd);
1499         destroy_workqueue(kmultipathd);
1500
1501         dm_unregister_target(&multipath_target);
1502         kmem_cache_destroy(_mpio_cache);
1503 }
1504
1505 module_init(dm_multipath_init);
1506 module_exit(dm_multipath_exit);
1507
1508 MODULE_DESCRIPTION(DM_NAME " multipath target");
1509 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1510 MODULE_LICENSE("GPL");