Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jbarnes...
[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         m->pgpath_to_activate = m->current_pgpath;
445
446         if ((pgpath && !m->queue_io) ||
447             (!pgpath && !m->queue_if_no_path))
448                 must_queue = 0;
449
450         if (m->pg_init_required && !m->pg_init_in_progress) {
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         consume(as, hw_argc - 1);
712
713         return 0;
714 }
715
716 static int parse_features(struct arg_set *as, struct multipath *m)
717 {
718         int r;
719         unsigned argc;
720         struct dm_target *ti = m->ti;
721         const char *param_name;
722
723         static struct param _params[] = {
724                 {0, 3, "invalid number of feature args"},
725                 {1, 50, "pg_init_retries must be between 1 and 50"},
726         };
727
728         r = read_param(_params, shift(as), &argc, &ti->error);
729         if (r)
730                 return -EINVAL;
731
732         if (!argc)
733                 return 0;
734
735         do {
736                 param_name = shift(as);
737                 argc--;
738
739                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
740                         r = queue_if_no_path(m, 1, 0);
741                         continue;
742                 }
743
744                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
745                     (argc >= 1)) {
746                         r = read_param(_params + 1, shift(as),
747                                        &m->pg_init_retries, &ti->error);
748                         argc--;
749                         continue;
750                 }
751
752                 ti->error = "Unrecognised multipath feature request";
753                 r = -EINVAL;
754         } while (argc && !r);
755
756         return r;
757 }
758
759 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
760                          char **argv)
761 {
762         /* target parameters */
763         static struct param _params[] = {
764                 {1, 1024, "invalid number of priority groups"},
765                 {1, 1024, "invalid initial priority group number"},
766         };
767
768         int r;
769         struct multipath *m;
770         struct arg_set as;
771         unsigned pg_count = 0;
772         unsigned next_pg_num;
773
774         as.argc = argc;
775         as.argv = argv;
776
777         m = alloc_multipath(ti);
778         if (!m) {
779                 ti->error = "can't allocate multipath";
780                 return -EINVAL;
781         }
782
783         r = parse_features(&as, m);
784         if (r)
785                 goto bad;
786
787         r = parse_hw_handler(&as, m);
788         if (r)
789                 goto bad;
790
791         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
792         if (r)
793                 goto bad;
794
795         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
796         if (r)
797                 goto bad;
798
799         /* parse the priority groups */
800         while (as.argc) {
801                 struct priority_group *pg;
802
803                 pg = parse_priority_group(&as, m);
804                 if (IS_ERR(pg)) {
805                         r = PTR_ERR(pg);
806                         goto bad;
807                 }
808
809                 m->nr_valid_paths += pg->nr_pgpaths;
810                 list_add_tail(&pg->list, &m->priority_groups);
811                 pg_count++;
812                 pg->pg_num = pg_count;
813                 if (!--next_pg_num)
814                         m->next_pg = pg;
815         }
816
817         if (pg_count != m->nr_priority_groups) {
818                 ti->error = "priority group count mismatch";
819                 r = -EINVAL;
820                 goto bad;
821         }
822
823         return 0;
824
825  bad:
826         free_multipath(m);
827         return r;
828 }
829
830 static void multipath_dtr(struct dm_target *ti)
831 {
832         struct multipath *m = (struct multipath *) ti->private;
833
834         flush_workqueue(kmpath_handlerd);
835         flush_workqueue(kmultipathd);
836         free_multipath(m);
837 }
838
839 /*
840  * Map bios, recording original fields for later in case we have to resubmit
841  */
842 static int multipath_map(struct dm_target *ti, struct bio *bio,
843                          union map_info *map_context)
844 {
845         int r;
846         struct dm_mpath_io *mpio;
847         struct multipath *m = (struct multipath *) ti->private;
848
849         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
850         dm_bio_record(&mpio->details, bio);
851
852         map_context->ptr = mpio;
853         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
854         r = map_io(m, bio, mpio, 0);
855         if (r < 0 || r == DM_MAPIO_REQUEUE)
856                 mempool_free(mpio, m->mpio_pool);
857
858         return r;
859 }
860
861 /*
862  * Take a path out of use.
863  */
864 static int fail_path(struct pgpath *pgpath)
865 {
866         unsigned long flags;
867         struct multipath *m = pgpath->pg->m;
868
869         spin_lock_irqsave(&m->lock, flags);
870
871         if (!pgpath->is_active)
872                 goto out;
873
874         DMWARN("Failing path %s.", pgpath->path.dev->name);
875
876         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
877         pgpath->is_active = 0;
878         pgpath->fail_count++;
879
880         m->nr_valid_paths--;
881
882         if (pgpath == m->current_pgpath)
883                 m->current_pgpath = NULL;
884
885         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
886                       pgpath->path.dev->name, m->nr_valid_paths);
887
888         queue_work(kmultipathd, &m->trigger_event);
889         queue_work(kmultipathd, &pgpath->deactivate_path);
890
891 out:
892         spin_unlock_irqrestore(&m->lock, flags);
893
894         return 0;
895 }
896
897 /*
898  * Reinstate a previously-failed path
899  */
900 static int reinstate_path(struct pgpath *pgpath)
901 {
902         int r = 0;
903         unsigned long flags;
904         struct multipath *m = pgpath->pg->m;
905
906         spin_lock_irqsave(&m->lock, flags);
907
908         if (pgpath->is_active)
909                 goto out;
910
911         if (!pgpath->pg->ps.type->reinstate_path) {
912                 DMWARN("Reinstate path not supported by path selector %s",
913                        pgpath->pg->ps.type->name);
914                 r = -EINVAL;
915                 goto out;
916         }
917
918         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
919         if (r)
920                 goto out;
921
922         pgpath->is_active = 1;
923
924         m->current_pgpath = NULL;
925         if (!m->nr_valid_paths++ && m->queue_size)
926                 queue_work(kmultipathd, &m->process_queued_ios);
927
928         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
929                       pgpath->path.dev->name, m->nr_valid_paths);
930
931         queue_work(kmultipathd, &m->trigger_event);
932
933 out:
934         spin_unlock_irqrestore(&m->lock, flags);
935
936         return r;
937 }
938
939 /*
940  * Fail or reinstate all paths that match the provided struct dm_dev.
941  */
942 static int action_dev(struct multipath *m, struct dm_dev *dev,
943                       action_fn action)
944 {
945         int r = 0;
946         struct pgpath *pgpath;
947         struct priority_group *pg;
948
949         list_for_each_entry(pg, &m->priority_groups, list) {
950                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
951                         if (pgpath->path.dev == dev)
952                                 r = action(pgpath);
953                 }
954         }
955
956         return r;
957 }
958
959 /*
960  * Temporarily try to avoid having to use the specified PG
961  */
962 static void bypass_pg(struct multipath *m, struct priority_group *pg,
963                       int bypassed)
964 {
965         unsigned long flags;
966
967         spin_lock_irqsave(&m->lock, flags);
968
969         pg->bypassed = bypassed;
970         m->current_pgpath = NULL;
971         m->current_pg = NULL;
972
973         spin_unlock_irqrestore(&m->lock, flags);
974
975         queue_work(kmultipathd, &m->trigger_event);
976 }
977
978 /*
979  * Switch to using the specified PG from the next I/O that gets mapped
980  */
981 static int switch_pg_num(struct multipath *m, const char *pgstr)
982 {
983         struct priority_group *pg;
984         unsigned pgnum;
985         unsigned long flags;
986
987         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
988             (pgnum > m->nr_priority_groups)) {
989                 DMWARN("invalid PG number supplied to switch_pg_num");
990                 return -EINVAL;
991         }
992
993         spin_lock_irqsave(&m->lock, flags);
994         list_for_each_entry(pg, &m->priority_groups, list) {
995                 pg->bypassed = 0;
996                 if (--pgnum)
997                         continue;
998
999                 m->current_pgpath = NULL;
1000                 m->current_pg = NULL;
1001                 m->next_pg = pg;
1002         }
1003         spin_unlock_irqrestore(&m->lock, flags);
1004
1005         queue_work(kmultipathd, &m->trigger_event);
1006         return 0;
1007 }
1008
1009 /*
1010  * Set/clear bypassed status of a PG.
1011  * PGs are numbered upwards from 1 in the order they were declared.
1012  */
1013 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1014 {
1015         struct priority_group *pg;
1016         unsigned pgnum;
1017
1018         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1019             (pgnum > m->nr_priority_groups)) {
1020                 DMWARN("invalid PG number supplied to bypass_pg");
1021                 return -EINVAL;
1022         }
1023
1024         list_for_each_entry(pg, &m->priority_groups, list) {
1025                 if (!--pgnum)
1026                         break;
1027         }
1028
1029         bypass_pg(m, pg, bypassed);
1030         return 0;
1031 }
1032
1033 /*
1034  * Should we retry pg_init immediately?
1035  */
1036 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1037 {
1038         unsigned long flags;
1039         int limit_reached = 0;
1040
1041         spin_lock_irqsave(&m->lock, flags);
1042
1043         if (m->pg_init_count <= m->pg_init_retries)
1044                 m->pg_init_required = 1;
1045         else
1046                 limit_reached = 1;
1047
1048         spin_unlock_irqrestore(&m->lock, flags);
1049
1050         return limit_reached;
1051 }
1052
1053 static void pg_init_done(struct dm_path *path, int errors)
1054 {
1055         struct pgpath *pgpath = path_to_pgpath(path);
1056         struct priority_group *pg = pgpath->pg;
1057         struct multipath *m = pg->m;
1058         unsigned long flags;
1059
1060         /* device or driver problems */
1061         switch (errors) {
1062         case SCSI_DH_OK:
1063                 break;
1064         case SCSI_DH_NOSYS:
1065                 if (!m->hw_handler_name) {
1066                         errors = 0;
1067                         break;
1068                 }
1069                 DMERR("Cannot failover device because scsi_dh_%s was not "
1070                       "loaded.", m->hw_handler_name);
1071                 /*
1072                  * Fail path for now, so we do not ping pong
1073                  */
1074                 fail_path(pgpath);
1075                 break;
1076         case SCSI_DH_DEV_TEMP_BUSY:
1077                 /*
1078                  * Probably doing something like FW upgrade on the
1079                  * controller so try the other pg.
1080                  */
1081                 bypass_pg(m, pg, 1);
1082                 break;
1083         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1084         case SCSI_DH_RETRY:
1085         case SCSI_DH_IMM_RETRY:
1086         case SCSI_DH_RES_TEMP_UNAVAIL:
1087                 if (pg_init_limit_reached(m, pgpath))
1088                         fail_path(pgpath);
1089                 errors = 0;
1090                 break;
1091         default:
1092                 /*
1093                  * We probably do not want to fail the path for a device
1094                  * error, but this is what the old dm did. In future
1095                  * patches we can do more advanced handling.
1096                  */
1097                 fail_path(pgpath);
1098         }
1099
1100         spin_lock_irqsave(&m->lock, flags);
1101         if (errors) {
1102                 DMERR("Could not failover device. Error %d.", errors);
1103                 m->current_pgpath = NULL;
1104                 m->current_pg = NULL;
1105         } else if (!m->pg_init_required) {
1106                 m->queue_io = 0;
1107                 pg->bypassed = 0;
1108         }
1109
1110         m->pg_init_in_progress = 0;
1111         queue_work(kmultipathd, &m->process_queued_ios);
1112         spin_unlock_irqrestore(&m->lock, flags);
1113 }
1114
1115 static void activate_path(struct work_struct *work)
1116 {
1117         int ret;
1118         struct multipath *m =
1119                 container_of(work, struct multipath, activate_path);
1120         struct dm_path *path;
1121         unsigned long flags;
1122
1123         spin_lock_irqsave(&m->lock, flags);
1124         path = &m->pgpath_to_activate->path;
1125         m->pgpath_to_activate = NULL;
1126         spin_unlock_irqrestore(&m->lock, flags);
1127         if (!path)
1128                 return;
1129         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1130         pg_init_done(path, ret);
1131 }
1132
1133 /*
1134  * end_io handling
1135  */
1136 static int do_end_io(struct multipath *m, struct bio *bio,
1137                      int error, struct dm_mpath_io *mpio)
1138 {
1139         unsigned long flags;
1140
1141         if (!error)
1142                 return 0;       /* I/O complete */
1143
1144         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1145                 return error;
1146
1147         if (error == -EOPNOTSUPP)
1148                 return error;
1149
1150         spin_lock_irqsave(&m->lock, flags);
1151         if (!m->nr_valid_paths) {
1152                 if (__must_push_back(m)) {
1153                         spin_unlock_irqrestore(&m->lock, flags);
1154                         return DM_ENDIO_REQUEUE;
1155                 } else if (!m->queue_if_no_path) {
1156                         spin_unlock_irqrestore(&m->lock, flags);
1157                         return -EIO;
1158                 } else {
1159                         spin_unlock_irqrestore(&m->lock, flags);
1160                         goto requeue;
1161                 }
1162         }
1163         spin_unlock_irqrestore(&m->lock, flags);
1164
1165         if (mpio->pgpath)
1166                 fail_path(mpio->pgpath);
1167
1168       requeue:
1169         dm_bio_restore(&mpio->details, bio);
1170
1171         /* queue for the daemon to resubmit or fail */
1172         spin_lock_irqsave(&m->lock, flags);
1173         bio_list_add(&m->queued_ios, bio);
1174         m->queue_size++;
1175         if (!m->queue_io)
1176                 queue_work(kmultipathd, &m->process_queued_ios);
1177         spin_unlock_irqrestore(&m->lock, flags);
1178
1179         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1180 }
1181
1182 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1183                             int error, union map_info *map_context)
1184 {
1185         struct multipath *m = ti->private;
1186         struct dm_mpath_io *mpio = map_context->ptr;
1187         struct pgpath *pgpath = mpio->pgpath;
1188         struct path_selector *ps;
1189         int r;
1190
1191         r  = do_end_io(m, bio, error, mpio);
1192         if (pgpath) {
1193                 ps = &pgpath->pg->ps;
1194                 if (ps->type->end_io)
1195                         ps->type->end_io(ps, &pgpath->path);
1196         }
1197         if (r != DM_ENDIO_INCOMPLETE)
1198                 mempool_free(mpio, m->mpio_pool);
1199
1200         return r;
1201 }
1202
1203 /*
1204  * Suspend can't complete until all the I/O is processed so if
1205  * the last path fails we must error any remaining I/O.
1206  * Note that if the freeze_bdev fails while suspending, the
1207  * queue_if_no_path state is lost - userspace should reset it.
1208  */
1209 static void multipath_presuspend(struct dm_target *ti)
1210 {
1211         struct multipath *m = (struct multipath *) ti->private;
1212
1213         queue_if_no_path(m, 0, 1);
1214 }
1215
1216 /*
1217  * Restore the queue_if_no_path setting.
1218  */
1219 static void multipath_resume(struct dm_target *ti)
1220 {
1221         struct multipath *m = (struct multipath *) ti->private;
1222         unsigned long flags;
1223
1224         spin_lock_irqsave(&m->lock, flags);
1225         m->queue_if_no_path = m->saved_queue_if_no_path;
1226         spin_unlock_irqrestore(&m->lock, flags);
1227 }
1228
1229 /*
1230  * Info output has the following format:
1231  * num_multipath_feature_args [multipath_feature_args]*
1232  * num_handler_status_args [handler_status_args]*
1233  * num_groups init_group_number
1234  *            [A|D|E num_ps_status_args [ps_status_args]*
1235  *             num_paths num_selector_args
1236  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1237  *
1238  * Table output has the following format (identical to the constructor string):
1239  * num_feature_args [features_args]*
1240  * num_handler_args hw_handler [hw_handler_args]*
1241  * num_groups init_group_number
1242  *     [priority selector-name num_ps_args [ps_args]*
1243  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1244  */
1245 static int multipath_status(struct dm_target *ti, status_type_t type,
1246                             char *result, unsigned int maxlen)
1247 {
1248         int sz = 0;
1249         unsigned long flags;
1250         struct multipath *m = (struct multipath *) ti->private;
1251         struct priority_group *pg;
1252         struct pgpath *p;
1253         unsigned pg_num;
1254         char state;
1255
1256         spin_lock_irqsave(&m->lock, flags);
1257
1258         /* Features */
1259         if (type == STATUSTYPE_INFO)
1260                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1261         else {
1262                 DMEMIT("%u ", m->queue_if_no_path +
1263                               (m->pg_init_retries > 0) * 2);
1264                 if (m->queue_if_no_path)
1265                         DMEMIT("queue_if_no_path ");
1266                 if (m->pg_init_retries)
1267                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1268         }
1269
1270         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1271                 DMEMIT("0 ");
1272         else
1273                 DMEMIT("1 %s ", m->hw_handler_name);
1274
1275         DMEMIT("%u ", m->nr_priority_groups);
1276
1277         if (m->next_pg)
1278                 pg_num = m->next_pg->pg_num;
1279         else if (m->current_pg)
1280                 pg_num = m->current_pg->pg_num;
1281         else
1282                         pg_num = 1;
1283
1284         DMEMIT("%u ", pg_num);
1285
1286         switch (type) {
1287         case STATUSTYPE_INFO:
1288                 list_for_each_entry(pg, &m->priority_groups, list) {
1289                         if (pg->bypassed)
1290                                 state = 'D';    /* Disabled */
1291                         else if (pg == m->current_pg)
1292                                 state = 'A';    /* Currently Active */
1293                         else
1294                                 state = 'E';    /* Enabled */
1295
1296                         DMEMIT("%c ", state);
1297
1298                         if (pg->ps.type->status)
1299                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1300                                                           result + sz,
1301                                                           maxlen - sz);
1302                         else
1303                                 DMEMIT("0 ");
1304
1305                         DMEMIT("%u %u ", pg->nr_pgpaths,
1306                                pg->ps.type->info_args);
1307
1308                         list_for_each_entry(p, &pg->pgpaths, list) {
1309                                 DMEMIT("%s %s %u ", p->path.dev->name,
1310                                        p->is_active ? "A" : "F",
1311                                        p->fail_count);
1312                                 if (pg->ps.type->status)
1313                                         sz += pg->ps.type->status(&pg->ps,
1314                                               &p->path, type, result + sz,
1315                                               maxlen - sz);
1316                         }
1317                 }
1318                 break;
1319
1320         case STATUSTYPE_TABLE:
1321                 list_for_each_entry(pg, &m->priority_groups, list) {
1322                         DMEMIT("%s ", pg->ps.type->name);
1323
1324                         if (pg->ps.type->status)
1325                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1326                                                           result + sz,
1327                                                           maxlen - sz);
1328                         else
1329                                 DMEMIT("0 ");
1330
1331                         DMEMIT("%u %u ", pg->nr_pgpaths,
1332                                pg->ps.type->table_args);
1333
1334                         list_for_each_entry(p, &pg->pgpaths, list) {
1335                                 DMEMIT("%s ", p->path.dev->name);
1336                                 if (pg->ps.type->status)
1337                                         sz += pg->ps.type->status(&pg->ps,
1338                                               &p->path, type, result + sz,
1339                                               maxlen - sz);
1340                         }
1341                 }
1342                 break;
1343         }
1344
1345         spin_unlock_irqrestore(&m->lock, flags);
1346
1347         return 0;
1348 }
1349
1350 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1351 {
1352         int r;
1353         struct dm_dev *dev;
1354         struct multipath *m = (struct multipath *) ti->private;
1355         action_fn action;
1356
1357         if (argc == 1) {
1358                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1359                         return queue_if_no_path(m, 1, 0);
1360                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1361                         return queue_if_no_path(m, 0, 0);
1362         }
1363
1364         if (argc != 2)
1365                 goto error;
1366
1367         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1368                 return bypass_pg_num(m, argv[1], 1);
1369         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1370                 return bypass_pg_num(m, argv[1], 0);
1371         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1372                 return switch_pg_num(m, argv[1]);
1373         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1374                 action = reinstate_path;
1375         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1376                 action = fail_path;
1377         else
1378                 goto error;
1379
1380         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1381                           dm_table_get_mode(ti->table), &dev);
1382         if (r) {
1383                 DMWARN("message: error getting device %s",
1384                        argv[1]);
1385                 return -EINVAL;
1386         }
1387
1388         r = action_dev(m, dev, action);
1389
1390         dm_put_device(ti, dev);
1391
1392         return r;
1393
1394 error:
1395         DMWARN("Unrecognised multipath message received.");
1396         return -EINVAL;
1397 }
1398
1399 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1400                            unsigned long arg)
1401 {
1402         struct multipath *m = (struct multipath *) ti->private;
1403         struct block_device *bdev = NULL;
1404         fmode_t mode = 0;
1405         unsigned long flags;
1406         int r = 0;
1407
1408         spin_lock_irqsave(&m->lock, flags);
1409
1410         if (!m->current_pgpath)
1411                 __choose_pgpath(m);
1412
1413         if (m->current_pgpath) {
1414                 bdev = m->current_pgpath->path.dev->bdev;
1415                 mode = m->current_pgpath->path.dev->mode;
1416         }
1417
1418         if (m->queue_io)
1419                 r = -EAGAIN;
1420         else if (!bdev)
1421                 r = -EIO;
1422
1423         spin_unlock_irqrestore(&m->lock, flags);
1424
1425         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1426 }
1427
1428 /*-----------------------------------------------------------------
1429  * Module setup
1430  *---------------------------------------------------------------*/
1431 static struct target_type multipath_target = {
1432         .name = "multipath",
1433         .version = {1, 0, 5},
1434         .module = THIS_MODULE,
1435         .ctr = multipath_ctr,
1436         .dtr = multipath_dtr,
1437         .map = multipath_map,
1438         .end_io = multipath_end_io,
1439         .presuspend = multipath_presuspend,
1440         .resume = multipath_resume,
1441         .status = multipath_status,
1442         .message = multipath_message,
1443         .ioctl  = multipath_ioctl,
1444 };
1445
1446 static int __init dm_multipath_init(void)
1447 {
1448         int r;
1449
1450         /* allocate a slab for the dm_ios */
1451         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1452         if (!_mpio_cache)
1453                 return -ENOMEM;
1454
1455         r = dm_register_target(&multipath_target);
1456         if (r < 0) {
1457                 DMERR("register failed %d", r);
1458                 kmem_cache_destroy(_mpio_cache);
1459                 return -EINVAL;
1460         }
1461
1462         kmultipathd = create_workqueue("kmpathd");
1463         if (!kmultipathd) {
1464                 DMERR("failed to create workqueue kmpathd");
1465                 dm_unregister_target(&multipath_target);
1466                 kmem_cache_destroy(_mpio_cache);
1467                 return -ENOMEM;
1468         }
1469
1470         /*
1471          * A separate workqueue is used to handle the device handlers
1472          * to avoid overloading existing workqueue. Overloading the
1473          * old workqueue would also create a bottleneck in the
1474          * path of the storage hardware device activation.
1475          */
1476         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1477         if (!kmpath_handlerd) {
1478                 DMERR("failed to create workqueue kmpath_handlerd");
1479                 destroy_workqueue(kmultipathd);
1480                 dm_unregister_target(&multipath_target);
1481                 kmem_cache_destroy(_mpio_cache);
1482                 return -ENOMEM;
1483         }
1484
1485         DMINFO("version %u.%u.%u loaded",
1486                multipath_target.version[0], multipath_target.version[1],
1487                multipath_target.version[2]);
1488
1489         return r;
1490 }
1491
1492 static void __exit dm_multipath_exit(void)
1493 {
1494         int r;
1495
1496         destroy_workqueue(kmpath_handlerd);
1497         destroy_workqueue(kmultipathd);
1498
1499         r = dm_unregister_target(&multipath_target);
1500         if (r < 0)
1501                 DMERR("target unregister failed %d", r);
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");