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