Merge /spare/repo/linux-2.6/
[linux-2.6] / drivers / md / dm-mpath.c
1 /*
2  * Copyright (C) 2003 Sistina Software Limited.
3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
4  *
5  * This file is released under the GPL.
6  */
7
8 #include "dm.h"
9 #include "dm-path-selector.h"
10 #include "dm-hw-handler.h"
11 #include "dm-bio-list.h"
12 #include "dm-bio-record.h"
13
14 #include <linux/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <asm/atomic.h>
23
24 #define MESG_STR(x) x, sizeof(x)
25
26 /* Path properties */
27 struct pgpath {
28         struct list_head list;
29
30         struct priority_group *pg;      /* Owning PG */
31         unsigned fail_count;            /* Cumulative failure count */
32
33         struct path path;
34 };
35
36 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
37
38 /*
39  * Paths are grouped into Priority Groups and numbered from 1 upwards.
40  * Each has a path selector which controls which path gets used.
41  */
42 struct priority_group {
43         struct list_head list;
44
45         struct multipath *m;            /* Owning multipath instance */
46         struct path_selector ps;
47
48         unsigned pg_num;                /* Reference number */
49         unsigned bypassed;              /* Temporarily bypass this PG? */
50
51         unsigned nr_pgpaths;            /* Number of paths in PG */
52         struct list_head pgpaths;
53 };
54
55 /* Multipath context */
56 struct multipath {
57         struct list_head list;
58         struct dm_target *ti;
59
60         spinlock_t lock;
61
62         struct hw_handler hw_handler;
63         unsigned nr_priority_groups;
64         struct list_head priority_groups;
65         unsigned pg_init_required;      /* pg_init needs calling? */
66
67         unsigned nr_valid_paths;        /* Total number of usable paths */
68         struct pgpath *current_pgpath;
69         struct priority_group *current_pg;
70         struct priority_group *next_pg; /* Switch to this PG if set */
71         unsigned repeat_count;          /* I/Os left before calling PS again */
72
73         unsigned queue_io;              /* Must we queue all I/O? */
74         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
75         unsigned suspended;             /* Has dm core suspended our I/O? */
76
77         struct work_struct process_queued_ios;
78         struct bio_list queued_ios;
79         unsigned queue_size;
80
81         struct work_struct trigger_event;
82
83         /*
84          * We must use a mempool of mpath_io structs so that we
85          * can resubmit bios on error.
86          */
87         mempool_t *mpio_pool;
88 };
89
90 /*
91  * Context information attached to each bio we process.
92  */
93 struct mpath_io {
94         struct pgpath *pgpath;
95         struct dm_bio_details details;
96 };
97
98 typedef int (*action_fn) (struct pgpath *pgpath);
99
100 #define MIN_IOS 256     /* Mempool size */
101
102 static kmem_cache_t *_mpio_cache;
103
104 struct workqueue_struct *kmultipathd;
105 static void process_queued_ios(void *data);
106 static void trigger_event(void *data);
107
108
109 /*-----------------------------------------------
110  * Allocation routines
111  *-----------------------------------------------*/
112
113 static struct pgpath *alloc_pgpath(void)
114 {
115         struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
116
117         if (pgpath) {
118                 memset(pgpath, 0, sizeof(*pgpath));
119                 pgpath->path.is_active = 1;
120         }
121
122         return pgpath;
123 }
124
125 static inline void free_pgpath(struct pgpath *pgpath)
126 {
127         kfree(pgpath);
128 }
129
130 static struct priority_group *alloc_priority_group(void)
131 {
132         struct priority_group *pg;
133
134         pg = kmalloc(sizeof(*pg), GFP_KERNEL);
135         if (!pg)
136                 return NULL;
137
138         memset(pg, 0, sizeof(*pg));
139         INIT_LIST_HEAD(&pg->pgpaths);
140
141         return pg;
142 }
143
144 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
145 {
146         struct pgpath *pgpath, *tmp;
147
148         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
149                 list_del(&pgpath->list);
150                 dm_put_device(ti, pgpath->path.dev);
151                 free_pgpath(pgpath);
152         }
153 }
154
155 static void free_priority_group(struct priority_group *pg,
156                                 struct dm_target *ti)
157 {
158         struct path_selector *ps = &pg->ps;
159
160         if (ps->type) {
161                 ps->type->destroy(ps);
162                 dm_put_path_selector(ps->type);
163         }
164
165         free_pgpaths(&pg->pgpaths, ti);
166         kfree(pg);
167 }
168
169 static struct multipath *alloc_multipath(void)
170 {
171         struct multipath *m;
172
173         m = kmalloc(sizeof(*m), GFP_KERNEL);
174         if (m) {
175                 memset(m, 0, sizeof(*m));
176                 INIT_LIST_HEAD(&m->priority_groups);
177                 spin_lock_init(&m->lock);
178                 m->queue_io = 1;
179                 INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
180                 INIT_WORK(&m->trigger_event, trigger_event, m);
181                 m->mpio_pool = mempool_create(MIN_IOS, mempool_alloc_slab,
182                                               mempool_free_slab, _mpio_cache);
183                 if (!m->mpio_pool) {
184                         kfree(m);
185                         return NULL;
186                 }
187         }
188
189         return m;
190 }
191
192 static void free_multipath(struct multipath *m)
193 {
194         struct priority_group *pg, *tmp;
195         struct hw_handler *hwh = &m->hw_handler;
196
197         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
198                 list_del(&pg->list);
199                 free_priority_group(pg, m->ti);
200         }
201
202         if (hwh->type) {
203                 hwh->type->destroy(hwh);
204                 dm_put_hw_handler(hwh->type);
205         }
206
207         mempool_destroy(m->mpio_pool);
208         kfree(m);
209 }
210
211
212 /*-----------------------------------------------
213  * Path selection
214  *-----------------------------------------------*/
215
216 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
217 {
218         struct hw_handler *hwh = &m->hw_handler;
219
220         m->current_pg = pgpath->pg;
221
222         /* Must we initialise the PG first, and queue I/O till it's ready? */
223         if (hwh->type && hwh->type->pg_init) {
224                 m->pg_init_required = 1;
225                 m->queue_io = 1;
226         } else {
227                 m->pg_init_required = 0;
228                 m->queue_io = 0;
229         }
230 }
231
232 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
233 {
234         struct path *path;
235
236         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
237         if (!path)
238                 return -ENXIO;
239
240         m->current_pgpath = path_to_pgpath(path);
241
242         if (m->current_pg != pg)
243                 __switch_pg(m, m->current_pgpath);
244
245         return 0;
246 }
247
248 static void __choose_pgpath(struct multipath *m)
249 {
250         struct priority_group *pg;
251         unsigned bypassed = 1;
252
253         if (!m->nr_valid_paths)
254                 goto failed;
255
256         /* Were we instructed to switch PG? */
257         if (m->next_pg) {
258                 pg = m->next_pg;
259                 m->next_pg = NULL;
260                 if (!__choose_path_in_pg(m, pg))
261                         return;
262         }
263
264         /* Don't change PG until it has no remaining paths */
265         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
266                 return;
267
268         /*
269          * Loop through priority groups until we find a valid path.
270          * First time we skip PGs marked 'bypassed'.
271          * Second time we only try the ones we skipped.
272          */
273         do {
274                 list_for_each_entry(pg, &m->priority_groups, list) {
275                         if (pg->bypassed == bypassed)
276                                 continue;
277                         if (!__choose_path_in_pg(m, pg))
278                                 return;
279                 }
280         } while (bypassed--);
281
282 failed:
283         m->current_pgpath = NULL;
284         m->current_pg = NULL;
285 }
286
287 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
288                   unsigned was_queued)
289 {
290         int r = 1;
291         unsigned long flags;
292         struct pgpath *pgpath;
293
294         spin_lock_irqsave(&m->lock, flags);
295
296         /* Do we need to select a new pgpath? */
297         if (!m->current_pgpath ||
298             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
299                 __choose_pgpath(m);
300
301         pgpath = m->current_pgpath;
302
303         if (was_queued)
304                 m->queue_size--;
305
306         if ((pgpath && m->queue_io) ||
307             (!pgpath && m->queue_if_no_path && !m->suspended)) {
308                 /* Queue for the daemon to resubmit */
309                 bio_list_add(&m->queued_ios, bio);
310                 m->queue_size++;
311                 if (m->pg_init_required || !m->queue_io)
312                         queue_work(kmultipathd, &m->process_queued_ios);
313                 pgpath = NULL;
314                 r = 0;
315         } else if (!pgpath)
316                 r = -EIO;               /* Failed */
317         else
318                 bio->bi_bdev = pgpath->path.dev->bdev;
319
320         mpio->pgpath = pgpath;
321
322         spin_unlock_irqrestore(&m->lock, flags);
323
324         return r;
325 }
326
327 /*
328  * If we run out of usable paths, should we queue I/O or error it?
329  */
330 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path)
331 {
332         unsigned long flags;
333
334         spin_lock_irqsave(&m->lock, flags);
335
336         m->queue_if_no_path = queue_if_no_path;
337         if (!m->queue_if_no_path)
338                 queue_work(kmultipathd, &m->process_queued_ios);
339
340         spin_unlock_irqrestore(&m->lock, flags);
341
342         return 0;
343 }
344
345 /*-----------------------------------------------------------------
346  * The multipath daemon is responsible for resubmitting queued ios.
347  *---------------------------------------------------------------*/
348
349 static void dispatch_queued_ios(struct multipath *m)
350 {
351         int r;
352         unsigned long flags;
353         struct bio *bio = NULL, *next;
354         struct mpath_io *mpio;
355         union map_info *info;
356
357         spin_lock_irqsave(&m->lock, flags);
358         bio = bio_list_get(&m->queued_ios);
359         spin_unlock_irqrestore(&m->lock, flags);
360
361         while (bio) {
362                 next = bio->bi_next;
363                 bio->bi_next = NULL;
364
365                 info = dm_get_mapinfo(bio);
366                 mpio = info->ptr;
367
368                 r = map_io(m, bio, mpio, 1);
369                 if (r < 0)
370                         bio_endio(bio, bio->bi_size, r);
371                 else if (r == 1)
372                         generic_make_request(bio);
373
374                 bio = next;
375         }
376 }
377
378 static void process_queued_ios(void *data)
379 {
380         struct multipath *m = (struct multipath *) data;
381         struct hw_handler *hwh = &m->hw_handler;
382         struct pgpath *pgpath;
383         unsigned init_required, must_queue = 0;
384         unsigned long flags;
385
386         spin_lock_irqsave(&m->lock, flags);
387
388         if (!m->current_pgpath)
389                 __choose_pgpath(m);
390
391         pgpath = m->current_pgpath;
392
393         if ((pgpath && m->queue_io) ||
394             (!pgpath && m->queue_if_no_path && !m->suspended))
395                 must_queue = 1;
396
397         init_required = m->pg_init_required;
398         if (init_required)
399                 m->pg_init_required = 0;
400
401         spin_unlock_irqrestore(&m->lock, flags);
402
403         if (init_required)
404                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
405
406         if (!must_queue)
407                 dispatch_queued_ios(m);
408 }
409
410 /*
411  * An event is triggered whenever a path is taken out of use.
412  * Includes path failure and PG bypass.
413  */
414 static void trigger_event(void *data)
415 {
416         struct multipath *m = (struct multipath *) data;
417
418         dm_table_event(m->ti->table);
419 }
420
421 /*-----------------------------------------------------------------
422  * Constructor/argument parsing:
423  * <#multipath feature args> [<arg>]*
424  * <#hw_handler args> [hw_handler [<arg>]*]
425  * <#priority groups>
426  * <initial priority group>
427  *     [<selector> <#selector args> [<arg>]*
428  *      <#paths> <#per-path selector args>
429  *         [<path> [<arg>]* ]+ ]+
430  *---------------------------------------------------------------*/
431 struct param {
432         unsigned min;
433         unsigned max;
434         char *error;
435 };
436
437 #define ESTR(s) ("dm-multipath: " s)
438
439 static int read_param(struct param *param, char *str, unsigned *v, char **error)
440 {
441         if (!str ||
442             (sscanf(str, "%u", v) != 1) ||
443             (*v < param->min) ||
444             (*v > param->max)) {
445                 *error = param->error;
446                 return -EINVAL;
447         }
448
449         return 0;
450 }
451
452 struct arg_set {
453         unsigned argc;
454         char **argv;
455 };
456
457 static char *shift(struct arg_set *as)
458 {
459         char *r;
460
461         if (as->argc) {
462                 as->argc--;
463                 r = *as->argv;
464                 as->argv++;
465                 return r;
466         }
467
468         return NULL;
469 }
470
471 static void consume(struct arg_set *as, unsigned n)
472 {
473         BUG_ON (as->argc < n);
474         as->argc -= n;
475         as->argv += n;
476 }
477
478 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
479                                struct dm_target *ti)
480 {
481         int r;
482         struct path_selector_type *pst;
483         unsigned ps_argc;
484
485         static struct param _params[] = {
486                 {0, 1024, ESTR("invalid number of path selector args")},
487         };
488
489         pst = dm_get_path_selector(shift(as));
490         if (!pst) {
491                 ti->error = ESTR("unknown path selector type");
492                 return -EINVAL;
493         }
494
495         r = read_param(_params, shift(as), &ps_argc, &ti->error);
496         if (r)
497                 return -EINVAL;
498
499         r = pst->create(&pg->ps, ps_argc, as->argv);
500         if (r) {
501                 dm_put_path_selector(pst);
502                 ti->error = ESTR("path selector constructor failed");
503                 return r;
504         }
505
506         pg->ps.type = pst;
507         consume(as, ps_argc);
508
509         return 0;
510 }
511
512 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
513                                struct dm_target *ti)
514 {
515         int r;
516         struct pgpath *p;
517
518         /* we need at least a path arg */
519         if (as->argc < 1) {
520                 ti->error = ESTR("no device given");
521                 return NULL;
522         }
523
524         p = alloc_pgpath();
525         if (!p)
526                 return NULL;
527
528         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
529                           dm_table_get_mode(ti->table), &p->path.dev);
530         if (r) {
531                 ti->error = ESTR("error getting device");
532                 goto bad;
533         }
534
535         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
536         if (r) {
537                 dm_put_device(ti, p->path.dev);
538                 goto bad;
539         }
540
541         return p;
542
543  bad:
544         free_pgpath(p);
545         return NULL;
546 }
547
548 static struct priority_group *parse_priority_group(struct arg_set *as,
549                                                    struct multipath *m,
550                                                    struct dm_target *ti)
551 {
552         static struct param _params[] = {
553                 {1, 1024, ESTR("invalid number of paths")},
554                 {0, 1024, ESTR("invalid number of selector args")}
555         };
556
557         int r;
558         unsigned i, nr_selector_args, nr_params;
559         struct priority_group *pg;
560
561         if (as->argc < 2) {
562                 as->argc = 0;
563                 ti->error = ESTR("not enough priority group aruments");
564                 return NULL;
565         }
566
567         pg = alloc_priority_group();
568         if (!pg) {
569                 ti->error = ESTR("couldn't allocate priority group");
570                 return NULL;
571         }
572         pg->m = m;
573
574         r = parse_path_selector(as, pg, ti);
575         if (r)
576                 goto bad;
577
578         /*
579          * read the paths
580          */
581         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
582         if (r)
583                 goto bad;
584
585         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
586         if (r)
587                 goto bad;
588
589         nr_params = 1 + nr_selector_args;
590         for (i = 0; i < pg->nr_pgpaths; i++) {
591                 struct pgpath *pgpath;
592                 struct arg_set path_args;
593
594                 if (as->argc < nr_params)
595                         goto bad;
596
597                 path_args.argc = nr_params;
598                 path_args.argv = as->argv;
599
600                 pgpath = parse_path(&path_args, &pg->ps, ti);
601                 if (!pgpath)
602                         goto bad;
603
604                 pgpath->pg = pg;
605                 list_add_tail(&pgpath->list, &pg->pgpaths);
606                 consume(as, nr_params);
607         }
608
609         return pg;
610
611  bad:
612         free_priority_group(pg, ti);
613         return NULL;
614 }
615
616 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
617                             struct dm_target *ti)
618 {
619         int r;
620         struct hw_handler_type *hwht;
621         unsigned hw_argc;
622
623         static struct param _params[] = {
624                 {0, 1024, ESTR("invalid number of hardware handler args")},
625         };
626
627         r = read_param(_params, shift(as), &hw_argc, &ti->error);
628         if (r)
629                 return -EINVAL;
630
631         if (!hw_argc)
632                 return 0;
633
634         hwht = dm_get_hw_handler(shift(as));
635         if (!hwht) {
636                 ti->error = ESTR("unknown hardware handler type");
637                 return -EINVAL;
638         }
639
640         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
641         if (r) {
642                 dm_put_hw_handler(hwht);
643                 ti->error = ESTR("hardware handler constructor failed");
644                 return r;
645         }
646
647         m->hw_handler.type = hwht;
648         consume(as, hw_argc - 1);
649
650         return 0;
651 }
652
653 static int parse_features(struct arg_set *as, struct multipath *m,
654                           struct dm_target *ti)
655 {
656         int r;
657         unsigned argc;
658
659         static struct param _params[] = {
660                 {0, 1, ESTR("invalid number of feature args")},
661         };
662
663         r = read_param(_params, shift(as), &argc, &ti->error);
664         if (r)
665                 return -EINVAL;
666
667         if (!argc)
668                 return 0;
669
670         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
671                 return queue_if_no_path(m, 1);
672         else {
673                 ti->error = "Unrecognised multipath feature request";
674                 return -EINVAL;
675         }
676 }
677
678 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
679                          char **argv)
680 {
681         /* target parameters */
682         static struct param _params[] = {
683                 {1, 1024, ESTR("invalid number of priority groups")},
684                 {1, 1024, ESTR("invalid initial priority group number")},
685         };
686
687         int r;
688         struct multipath *m;
689         struct arg_set as;
690         unsigned pg_count = 0;
691         unsigned next_pg_num;
692
693         as.argc = argc;
694         as.argv = argv;
695
696         m = alloc_multipath();
697         if (!m) {
698                 ti->error = ESTR("can't allocate multipath");
699                 return -EINVAL;
700         }
701
702         r = parse_features(&as, m, ti);
703         if (r)
704                 goto bad;
705
706         r = parse_hw_handler(&as, m, ti);
707         if (r)
708                 goto bad;
709
710         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
711         if (r)
712                 goto bad;
713
714         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
715         if (r)
716                 goto bad;
717
718         /* parse the priority groups */
719         while (as.argc) {
720                 struct priority_group *pg;
721
722                 pg = parse_priority_group(&as, m, ti);
723                 if (!pg) {
724                         r = -EINVAL;
725                         goto bad;
726                 }
727
728                 m->nr_valid_paths += pg->nr_pgpaths;
729                 list_add_tail(&pg->list, &m->priority_groups);
730                 pg_count++;
731                 pg->pg_num = pg_count;
732                 if (!--next_pg_num)
733                         m->next_pg = pg;
734         }
735
736         if (pg_count != m->nr_priority_groups) {
737                 ti->error = ESTR("priority group count mismatch");
738                 r = -EINVAL;
739                 goto bad;
740         }
741
742         ti->private = m;
743         m->ti = ti;
744
745         return 0;
746
747  bad:
748         free_multipath(m);
749         return r;
750 }
751
752 static void multipath_dtr(struct dm_target *ti)
753 {
754         struct multipath *m = (struct multipath *) ti->private;
755         free_multipath(m);
756 }
757
758 /*
759  * Map bios, recording original fields for later in case we have to resubmit
760  */
761 static int multipath_map(struct dm_target *ti, struct bio *bio,
762                          union map_info *map_context)
763 {
764         int r;
765         struct mpath_io *mpio;
766         struct multipath *m = (struct multipath *) ti->private;
767
768         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
769         dm_bio_record(&mpio->details, bio);
770
771         map_context->ptr = mpio;
772         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
773         r = map_io(m, bio, mpio, 0);
774         if (r < 0)
775                 mempool_free(mpio, m->mpio_pool);
776
777         return r;
778 }
779
780 /*
781  * Take a path out of use.
782  */
783 static int fail_path(struct pgpath *pgpath)
784 {
785         unsigned long flags;
786         struct multipath *m = pgpath->pg->m;
787
788         spin_lock_irqsave(&m->lock, flags);
789
790         if (!pgpath->path.is_active)
791                 goto out;
792
793         DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
794
795         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
796         pgpath->path.is_active = 0;
797         pgpath->fail_count++;
798
799         m->nr_valid_paths--;
800
801         if (pgpath == m->current_pgpath)
802                 m->current_pgpath = NULL;
803
804         queue_work(kmultipathd, &m->trigger_event);
805
806 out:
807         spin_unlock_irqrestore(&m->lock, flags);
808
809         return 0;
810 }
811
812 /*
813  * Reinstate a previously-failed path
814  */
815 static int reinstate_path(struct pgpath *pgpath)
816 {
817         int r = 0;
818         unsigned long flags;
819         struct multipath *m = pgpath->pg->m;
820
821         spin_lock_irqsave(&m->lock, flags);
822
823         if (pgpath->path.is_active)
824                 goto out;
825
826         if (!pgpath->pg->ps.type) {
827                 DMWARN("Reinstate path not supported by path selector %s",
828                        pgpath->pg->ps.type->name);
829                 r = -EINVAL;
830                 goto out;
831         }
832
833         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
834         if (r)
835                 goto out;
836
837         pgpath->path.is_active = 1;
838
839         m->current_pgpath = NULL;
840         if (!m->nr_valid_paths++)
841                 queue_work(kmultipathd, &m->process_queued_ios);
842
843         queue_work(kmultipathd, &m->trigger_event);
844
845 out:
846         spin_unlock_irqrestore(&m->lock, flags);
847
848         return r;
849 }
850
851 /*
852  * Fail or reinstate all paths that match the provided struct dm_dev.
853  */
854 static int action_dev(struct multipath *m, struct dm_dev *dev,
855                       action_fn action)
856 {
857         int r = 0;
858         struct pgpath *pgpath;
859         struct priority_group *pg;
860
861         list_for_each_entry(pg, &m->priority_groups, list) {
862                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
863                         if (pgpath->path.dev == dev)
864                                 r = action(pgpath);
865                 }
866         }
867
868         return r;
869 }
870
871 /*
872  * Temporarily try to avoid having to use the specified PG
873  */
874 static void bypass_pg(struct multipath *m, struct priority_group *pg,
875                       int bypassed)
876 {
877         unsigned long flags;
878
879         spin_lock_irqsave(&m->lock, flags);
880
881         pg->bypassed = bypassed;
882         m->current_pgpath = NULL;
883         m->current_pg = NULL;
884
885         spin_unlock_irqrestore(&m->lock, flags);
886
887         queue_work(kmultipathd, &m->trigger_event);
888 }
889
890 /*
891  * Switch to using the specified PG from the next I/O that gets mapped
892  */
893 static int switch_pg_num(struct multipath *m, const char *pgstr)
894 {
895         struct priority_group *pg;
896         unsigned pgnum;
897         unsigned long flags;
898
899         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
900             (pgnum > m->nr_priority_groups)) {
901                 DMWARN("invalid PG number supplied to switch_pg_num");
902                 return -EINVAL;
903         }
904
905         spin_lock_irqsave(&m->lock, flags);
906         list_for_each_entry(pg, &m->priority_groups, list) {
907                 pg->bypassed = 0;
908                 if (--pgnum)
909                         continue;
910
911                 m->current_pgpath = NULL;
912                 m->current_pg = NULL;
913                 m->next_pg = pg;
914         }
915         spin_unlock_irqrestore(&m->lock, flags);
916
917         queue_work(kmultipathd, &m->trigger_event);
918         return 0;
919 }
920
921 /*
922  * Set/clear bypassed status of a PG.
923  * PGs are numbered upwards from 1 in the order they were declared.
924  */
925 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
926 {
927         struct priority_group *pg;
928         unsigned pgnum;
929
930         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
931             (pgnum > m->nr_priority_groups)) {
932                 DMWARN("invalid PG number supplied to bypass_pg");
933                 return -EINVAL;
934         }
935
936         list_for_each_entry(pg, &m->priority_groups, list) {
937                 if (!--pgnum)
938                         break;
939         }
940
941         bypass_pg(m, pg, bypassed);
942         return 0;
943 }
944
945 /*
946  * pg_init must call this when it has completed its initialisation
947  */
948 void dm_pg_init_complete(struct path *path, unsigned err_flags)
949 {
950         struct pgpath *pgpath = path_to_pgpath(path);
951         struct priority_group *pg = pgpath->pg;
952         struct multipath *m = pg->m;
953         unsigned long flags;
954
955         /* We insist on failing the path if the PG is already bypassed. */
956         if (err_flags && pg->bypassed)
957                 err_flags |= MP_FAIL_PATH;
958
959         if (err_flags & MP_FAIL_PATH)
960                 fail_path(pgpath);
961
962         if (err_flags & MP_BYPASS_PG)
963                 bypass_pg(m, pg, 1);
964
965         spin_lock_irqsave(&m->lock, flags);
966         if (!err_flags)
967                 m->queue_io = 0;
968         else {
969                 m->current_pgpath = NULL;
970                 m->current_pg = NULL;
971         }
972         queue_work(kmultipathd, &m->process_queued_ios);
973         spin_unlock_irqrestore(&m->lock, flags);
974 }
975
976 /*
977  * end_io handling
978  */
979 static int do_end_io(struct multipath *m, struct bio *bio,
980                      int error, struct mpath_io *mpio)
981 {
982         struct hw_handler *hwh = &m->hw_handler;
983         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
984
985         if (!error)
986                 return 0;       /* I/O complete */
987
988         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
989                 return error;
990
991         spin_lock(&m->lock);
992         if (!m->nr_valid_paths) {
993                 if (!m->queue_if_no_path || m->suspended) {
994                         spin_unlock(&m->lock);
995                         return -EIO;
996                 } else {
997                         spin_unlock(&m->lock);
998                         goto requeue;
999                 }
1000         }
1001         spin_unlock(&m->lock);
1002
1003         if (hwh->type && hwh->type->error)
1004                 err_flags = hwh->type->error(hwh, bio);
1005
1006         if (mpio->pgpath) {
1007                 if (err_flags & MP_FAIL_PATH)
1008                         fail_path(mpio->pgpath);
1009
1010                 if (err_flags & MP_BYPASS_PG)
1011                         bypass_pg(m, mpio->pgpath->pg, 1);
1012         }
1013
1014         if (err_flags & MP_ERROR_IO)
1015                 return -EIO;
1016
1017       requeue:
1018         dm_bio_restore(&mpio->details, bio);
1019
1020         /* queue for the daemon to resubmit or fail */
1021         spin_lock(&m->lock);
1022         bio_list_add(&m->queued_ios, bio);
1023         m->queue_size++;
1024         if (!m->queue_io)
1025                 queue_work(kmultipathd, &m->process_queued_ios);
1026         spin_unlock(&m->lock);
1027
1028         return 1;       /* io not complete */
1029 }
1030
1031 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1032                             int error, union map_info *map_context)
1033 {
1034         struct multipath *m = (struct multipath *) ti->private;
1035         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1036         struct pgpath *pgpath = mpio->pgpath;
1037         struct path_selector *ps;
1038         int r;
1039
1040         r  = do_end_io(m, bio, error, mpio);
1041         if (pgpath) {
1042                 ps = &pgpath->pg->ps;
1043                 if (ps->type->end_io)
1044                         ps->type->end_io(ps, &pgpath->path);
1045         }
1046         if (r <= 0)
1047                 mempool_free(mpio, m->mpio_pool);
1048
1049         return r;
1050 }
1051
1052 /*
1053  * Suspend can't complete until all the I/O is processed so if
1054  * the last path failed we will now error any queued I/O.
1055  */
1056 static void multipath_presuspend(struct dm_target *ti)
1057 {
1058         struct multipath *m = (struct multipath *) ti->private;
1059         unsigned long flags;
1060
1061         spin_lock_irqsave(&m->lock, flags);
1062         m->suspended = 1;
1063         if (m->queue_if_no_path)
1064                 queue_work(kmultipathd, &m->process_queued_ios);
1065         spin_unlock_irqrestore(&m->lock, flags);
1066 }
1067
1068 static void multipath_resume(struct dm_target *ti)
1069 {
1070         struct multipath *m = (struct multipath *) ti->private;
1071         unsigned long flags;
1072
1073         spin_lock_irqsave(&m->lock, flags);
1074         m->suspended = 0;
1075         spin_unlock_irqrestore(&m->lock, flags);
1076 }
1077
1078 /*
1079  * Info output has the following format:
1080  * num_multipath_feature_args [multipath_feature_args]*
1081  * num_handler_status_args [handler_status_args]*
1082  * num_groups init_group_number
1083  *            [A|D|E num_ps_status_args [ps_status_args]*
1084  *             num_paths num_selector_args
1085  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1086  *
1087  * Table output has the following format (identical to the constructor string):
1088  * num_feature_args [features_args]*
1089  * num_handler_args hw_handler [hw_handler_args]*
1090  * num_groups init_group_number
1091  *     [priority selector-name num_ps_args [ps_args]*
1092  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1093  */
1094 static int multipath_status(struct dm_target *ti, status_type_t type,
1095                             char *result, unsigned int maxlen)
1096 {
1097         int sz = 0;
1098         unsigned long flags;
1099         struct multipath *m = (struct multipath *) ti->private;
1100         struct hw_handler *hwh = &m->hw_handler;
1101         struct priority_group *pg;
1102         struct pgpath *p;
1103         unsigned pg_num;
1104         char state;
1105
1106         spin_lock_irqsave(&m->lock, flags);
1107
1108         /* Features */
1109         if (type == STATUSTYPE_INFO)
1110                 DMEMIT("1 %u ", m->queue_size);
1111         else if (m->queue_if_no_path)
1112                 DMEMIT("1 queue_if_no_path ");
1113         else
1114                 DMEMIT("0 ");
1115
1116         if (hwh->type && hwh->type->status)
1117                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1118         else if (!hwh->type || type == STATUSTYPE_INFO)
1119                 DMEMIT("0 ");
1120         else
1121                 DMEMIT("1 %s ", hwh->type->name);
1122
1123         DMEMIT("%u ", m->nr_priority_groups);
1124
1125         if (m->next_pg)
1126                 pg_num = m->next_pg->pg_num;
1127         else if (m->current_pg)
1128                 pg_num = m->current_pg->pg_num;
1129         else
1130                         pg_num = 1;
1131
1132         DMEMIT("%u ", pg_num);
1133
1134         switch (type) {
1135         case STATUSTYPE_INFO:
1136                 list_for_each_entry(pg, &m->priority_groups, list) {
1137                         if (pg->bypassed)
1138                                 state = 'D';    /* Disabled */
1139                         else if (pg == m->current_pg)
1140                                 state = 'A';    /* Currently Active */
1141                         else
1142                                 state = 'E';    /* Enabled */
1143
1144                         DMEMIT("%c ", state);
1145
1146                         if (pg->ps.type->status)
1147                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1148                                                           result + sz,
1149                                                           maxlen - sz);
1150                         else
1151                                 DMEMIT("0 ");
1152
1153                         DMEMIT("%u %u ", pg->nr_pgpaths,
1154                                pg->ps.type->info_args);
1155
1156                         list_for_each_entry(p, &pg->pgpaths, list) {
1157                                 DMEMIT("%s %s %u ", p->path.dev->name,
1158                                        p->path.is_active ? "A" : "F",
1159                                        p->fail_count);
1160                                 if (pg->ps.type->status)
1161                                         sz += pg->ps.type->status(&pg->ps,
1162                                               &p->path, type, result + sz,
1163                                               maxlen - sz);
1164                         }
1165                 }
1166                 break;
1167
1168         case STATUSTYPE_TABLE:
1169                 list_for_each_entry(pg, &m->priority_groups, list) {
1170                         DMEMIT("%s ", pg->ps.type->name);
1171
1172                         if (pg->ps.type->status)
1173                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1174                                                           result + sz,
1175                                                           maxlen - sz);
1176                         else
1177                                 DMEMIT("0 ");
1178
1179                         DMEMIT("%u %u ", pg->nr_pgpaths,
1180                                pg->ps.type->table_args);
1181
1182                         list_for_each_entry(p, &pg->pgpaths, list) {
1183                                 DMEMIT("%s ", p->path.dev->name);
1184                                 if (pg->ps.type->status)
1185                                         sz += pg->ps.type->status(&pg->ps,
1186                                               &p->path, type, result + sz,
1187                                               maxlen - sz);
1188                         }
1189                 }
1190                 break;
1191         }
1192
1193         spin_unlock_irqrestore(&m->lock, flags);
1194
1195         return 0;
1196 }
1197
1198 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1199 {
1200         int r;
1201         struct dm_dev *dev;
1202         struct multipath *m = (struct multipath *) ti->private;
1203         action_fn action;
1204
1205         if (argc == 1) {
1206                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1207                         return queue_if_no_path(m, 1);
1208                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1209                         return queue_if_no_path(m, 0);
1210         }
1211
1212         if (argc != 2)
1213                 goto error;
1214
1215         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1216                 return bypass_pg_num(m, argv[1], 1);
1217         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1218                 return bypass_pg_num(m, argv[1], 0);
1219         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1220                 return switch_pg_num(m, argv[1]);
1221         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1222                 action = reinstate_path;
1223         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1224                 action = fail_path;
1225         else
1226                 goto error;
1227
1228         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1229                           dm_table_get_mode(ti->table), &dev);
1230         if (r) {
1231                 DMWARN("dm-multipath message: error getting device %s",
1232                        argv[1]);
1233                 return -EINVAL;
1234         }
1235
1236         r = action_dev(m, dev, action);
1237
1238         dm_put_device(ti, dev);
1239
1240         return r;
1241
1242 error:
1243         DMWARN("Unrecognised multipath message received.");
1244         return -EINVAL;
1245 }
1246
1247 /*-----------------------------------------------------------------
1248  * Module setup
1249  *---------------------------------------------------------------*/
1250 static struct target_type multipath_target = {
1251         .name = "multipath",
1252         .version = {1, 0, 4},
1253         .module = THIS_MODULE,
1254         .ctr = multipath_ctr,
1255         .dtr = multipath_dtr,
1256         .map = multipath_map,
1257         .end_io = multipath_end_io,
1258         .presuspend = multipath_presuspend,
1259         .resume = multipath_resume,
1260         .status = multipath_status,
1261         .message = multipath_message,
1262 };
1263
1264 static int __init dm_multipath_init(void)
1265 {
1266         int r;
1267
1268         /* allocate a slab for the dm_ios */
1269         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1270                                         0, 0, NULL, NULL);
1271         if (!_mpio_cache)
1272                 return -ENOMEM;
1273
1274         r = dm_register_target(&multipath_target);
1275         if (r < 0) {
1276                 DMERR("%s: register failed %d", multipath_target.name, r);
1277                 kmem_cache_destroy(_mpio_cache);
1278                 return -EINVAL;
1279         }
1280
1281         kmultipathd = create_workqueue("kmpathd");
1282         if (!kmultipathd) {
1283                 DMERR("%s: failed to create workqueue kmpathd",
1284                                 multipath_target.name);
1285                 dm_unregister_target(&multipath_target);
1286                 kmem_cache_destroy(_mpio_cache);
1287                 return -ENOMEM;
1288         }
1289
1290         DMINFO("dm-multipath version %u.%u.%u loaded",
1291                multipath_target.version[0], multipath_target.version[1],
1292                multipath_target.version[2]);
1293
1294         return r;
1295 }
1296
1297 static void __exit dm_multipath_exit(void)
1298 {
1299         int r;
1300
1301         destroy_workqueue(kmultipathd);
1302
1303         r = dm_unregister_target(&multipath_target);
1304         if (r < 0)
1305                 DMERR("%s: target unregister failed %d",
1306                       multipath_target.name, r);
1307         kmem_cache_destroy(_mpio_cache);
1308 }
1309
1310 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1311
1312 module_init(dm_multipath_init);
1313 module_exit(dm_multipath_exit);
1314
1315 MODULE_DESCRIPTION(DM_NAME " multipath target");
1316 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1317 MODULE_LICENSE("GPL");