Merge branch 'bkl-removal' of git://git.lwn.net/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-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 is_active;             /* Path status */
34         unsigned fail_count;            /* Cumulative failure count */
35
36         struct dm_path path;
37         struct work_struct deactivate_path;
38 };
39
40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
41
42 /*
43  * Paths are grouped into Priority Groups and numbered from 1 upwards.
44  * Each has a path selector which controls which path gets used.
45  */
46 struct priority_group {
47         struct list_head list;
48
49         struct multipath *m;            /* Owning multipath instance */
50         struct path_selector ps;
51
52         unsigned pg_num;                /* Reference number */
53         unsigned bypassed;              /* Temporarily bypass this PG? */
54
55         unsigned nr_pgpaths;            /* Number of paths in PG */
56         struct list_head pgpaths;
57 };
58
59 /* Multipath context */
60 struct multipath {
61         struct list_head list;
62         struct dm_target *ti;
63
64         spinlock_t lock;
65
66         const char *hw_handler_name;
67         struct work_struct activate_path;
68         struct pgpath *pgpath_to_activate;
69         unsigned nr_priority_groups;
70         struct list_head priority_groups;
71         unsigned pg_init_required;      /* pg_init needs calling? */
72         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
73
74         unsigned nr_valid_paths;        /* Total number of usable paths */
75         struct pgpath *current_pgpath;
76         struct priority_group *current_pg;
77         struct priority_group *next_pg; /* Switch to this PG if set */
78         unsigned repeat_count;          /* I/Os left before calling PS again */
79
80         unsigned queue_io;              /* Must we queue all I/O? */
81         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
82         unsigned saved_queue_if_no_path;/* Saved state during suspension */
83         unsigned pg_init_retries;       /* Number of times to retry pg_init */
84         unsigned pg_init_count;         /* Number of times pg_init called */
85
86         struct work_struct process_queued_ios;
87         struct bio_list queued_ios;
88         unsigned queue_size;
89
90         struct work_struct trigger_event;
91
92         /*
93          * We must use a mempool of dm_mpath_io structs so that we
94          * can resubmit bios on error.
95          */
96         mempool_t *mpio_pool;
97 };
98
99 /*
100  * Context information attached to each bio we process.
101  */
102 struct dm_mpath_io {
103         struct pgpath *pgpath;
104         struct dm_bio_details details;
105 };
106
107 typedef int (*action_fn) (struct pgpath *pgpath);
108
109 #define MIN_IOS 256     /* Mempool size */
110
111 static struct kmem_cache *_mpio_cache;
112
113 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
114 static void process_queued_ios(struct work_struct *work);
115 static void trigger_event(struct work_struct *work);
116 static void activate_path(struct work_struct *work);
117 static void deactivate_path(struct work_struct *work);
118
119
120 /*-----------------------------------------------
121  * Allocation routines
122  *-----------------------------------------------*/
123
124 static struct pgpath *alloc_pgpath(void)
125 {
126         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
127
128         if (pgpath) {
129                 pgpath->is_active = 1;
130                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
131         }
132
133         return pgpath;
134 }
135
136 static void free_pgpath(struct pgpath *pgpath)
137 {
138         kfree(pgpath);
139 }
140
141 static void deactivate_path(struct work_struct *work)
142 {
143         struct pgpath *pgpath =
144                 container_of(work, struct pgpath, deactivate_path);
145
146         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
147 }
148
149 static struct priority_group *alloc_priority_group(void)
150 {
151         struct priority_group *pg;
152
153         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
154
155         if (pg)
156                 INIT_LIST_HEAD(&pg->pgpaths);
157
158         return pg;
159 }
160
161 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
162 {
163         unsigned long flags;
164         struct pgpath *pgpath, *tmp;
165         struct multipath *m = ti->private;
166
167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168                 list_del(&pgpath->list);
169                 if (m->hw_handler_name)
170                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
171                 dm_put_device(ti, pgpath->path.dev);
172                 spin_lock_irqsave(&m->lock, flags);
173                 if (m->pgpath_to_activate == pgpath)
174                         m->pgpath_to_activate = NULL;
175                 spin_unlock_irqrestore(&m->lock, flags);
176                 free_pgpath(pgpath);
177         }
178 }
179
180 static void free_priority_group(struct priority_group *pg,
181                                 struct dm_target *ti)
182 {
183         struct path_selector *ps = &pg->ps;
184
185         if (ps->type) {
186                 ps->type->destroy(ps);
187                 dm_put_path_selector(ps->type);
188         }
189
190         free_pgpaths(&pg->pgpaths, ti);
191         kfree(pg);
192 }
193
194 static struct multipath *alloc_multipath(struct dm_target *ti)
195 {
196         struct multipath *m;
197
198         m = kzalloc(sizeof(*m), GFP_KERNEL);
199         if (m) {
200                 INIT_LIST_HEAD(&m->priority_groups);
201                 spin_lock_init(&m->lock);
202                 m->queue_io = 1;
203                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
204                 INIT_WORK(&m->trigger_event, trigger_event);
205                 INIT_WORK(&m->activate_path, activate_path);
206                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
207                 if (!m->mpio_pool) {
208                         kfree(m);
209                         return NULL;
210                 }
211                 m->ti = ti;
212                 ti->private = m;
213         }
214
215         return m;
216 }
217
218 static void free_multipath(struct multipath *m)
219 {
220         struct priority_group *pg, *tmp;
221
222         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
223                 list_del(&pg->list);
224                 free_priority_group(pg, m->ti);
225         }
226
227         kfree(m->hw_handler_name);
228         mempool_destroy(m->mpio_pool);
229         kfree(m);
230 }
231
232
233 /*-----------------------------------------------
234  * Path selection
235  *-----------------------------------------------*/
236
237 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
238 {
239         m->current_pg = pgpath->pg;
240
241         /* Must we initialise the PG first, and queue I/O till it's ready? */
242         if (m->hw_handler_name) {
243                 m->pg_init_required = 1;
244                 m->queue_io = 1;
245         } else {
246                 m->pg_init_required = 0;
247                 m->queue_io = 0;
248         }
249
250         m->pg_init_count = 0;
251 }
252
253 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
254 {
255         struct dm_path *path;
256
257         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
258         if (!path)
259                 return -ENXIO;
260
261         m->current_pgpath = path_to_pgpath(path);
262
263         if (m->current_pg != pg)
264                 __switch_pg(m, m->current_pgpath);
265
266         return 0;
267 }
268
269 static void __choose_pgpath(struct multipath *m)
270 {
271         struct priority_group *pg;
272         unsigned bypassed = 1;
273
274         if (!m->nr_valid_paths)
275                 goto failed;
276
277         /* Were we instructed to switch PG? */
278         if (m->next_pg) {
279                 pg = m->next_pg;
280                 m->next_pg = NULL;
281                 if (!__choose_path_in_pg(m, pg))
282                         return;
283         }
284
285         /* Don't change PG until it has no remaining paths */
286         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
287                 return;
288
289         /*
290          * Loop through priority groups until we find a valid path.
291          * First time we skip PGs marked 'bypassed'.
292          * Second time we only try the ones we skipped.
293          */
294         do {
295                 list_for_each_entry(pg, &m->priority_groups, list) {
296                         if (pg->bypassed == bypassed)
297                                 continue;
298                         if (!__choose_path_in_pg(m, pg))
299                                 return;
300                 }
301         } while (bypassed--);
302
303 failed:
304         m->current_pgpath = NULL;
305         m->current_pg = NULL;
306 }
307
308 /*
309  * Check whether bios must be queued in the device-mapper core rather
310  * than here in the target.
311  *
312  * m->lock must be held on entry.
313  *
314  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
315  * same value then we are not between multipath_presuspend()
316  * and multipath_resume() calls and we have no need to check
317  * for the DMF_NOFLUSH_SUSPENDING flag.
318  */
319 static int __must_push_back(struct multipath *m)
320 {
321         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
322                 dm_noflush_suspending(m->ti));
323 }
324
325 static int map_io(struct multipath *m, struct bio *bio,
326                   struct dm_mpath_io *mpio, unsigned was_queued)
327 {
328         int r = DM_MAPIO_REMAPPED;
329         unsigned long flags;
330         struct pgpath *pgpath;
331
332         spin_lock_irqsave(&m->lock, flags);
333
334         /* Do we need to select a new pgpath? */
335         if (!m->current_pgpath ||
336             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
337                 __choose_pgpath(m);
338
339         pgpath = m->current_pgpath;
340
341         if (was_queued)
342                 m->queue_size--;
343
344         if ((pgpath && m->queue_io) ||
345             (!pgpath && m->queue_if_no_path)) {
346                 /* Queue for the daemon to resubmit */
347                 bio_list_add(&m->queued_ios, bio);
348                 m->queue_size++;
349                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
350                     !m->queue_io)
351                         queue_work(kmultipathd, &m->process_queued_ios);
352                 pgpath = NULL;
353                 r = DM_MAPIO_SUBMITTED;
354         } else if (pgpath)
355                 bio->bi_bdev = pgpath->path.dev->bdev;
356         else if (__must_push_back(m))
357                 r = DM_MAPIO_REQUEUE;
358         else
359                 r = -EIO;       /* Failed */
360
361         mpio->pgpath = pgpath;
362
363         spin_unlock_irqrestore(&m->lock, flags);
364
365         return r;
366 }
367
368 /*
369  * If we run out of usable paths, should we queue I/O or error it?
370  */
371 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
372                             unsigned save_old_value)
373 {
374         unsigned long flags;
375
376         spin_lock_irqsave(&m->lock, flags);
377
378         if (save_old_value)
379                 m->saved_queue_if_no_path = m->queue_if_no_path;
380         else
381                 m->saved_queue_if_no_path = queue_if_no_path;
382         m->queue_if_no_path = queue_if_no_path;
383         if (!m->queue_if_no_path && m->queue_size)
384                 queue_work(kmultipathd, &m->process_queued_ios);
385
386         spin_unlock_irqrestore(&m->lock, flags);
387
388         return 0;
389 }
390
391 /*-----------------------------------------------------------------
392  * The multipath daemon is responsible for resubmitting queued ios.
393  *---------------------------------------------------------------*/
394
395 static void dispatch_queued_ios(struct multipath *m)
396 {
397         int r;
398         unsigned long flags;
399         struct bio *bio = NULL, *next;
400         struct dm_mpath_io *mpio;
401         union map_info *info;
402
403         spin_lock_irqsave(&m->lock, flags);
404         bio = bio_list_get(&m->queued_ios);
405         spin_unlock_irqrestore(&m->lock, flags);
406
407         while (bio) {
408                 next = bio->bi_next;
409                 bio->bi_next = NULL;
410
411                 info = dm_get_mapinfo(bio);
412                 mpio = info->ptr;
413
414                 r = map_io(m, bio, mpio, 1);
415                 if (r < 0)
416                         bio_endio(bio, r);
417                 else if (r == DM_MAPIO_REMAPPED)
418                         generic_make_request(bio);
419                 else if (r == DM_MAPIO_REQUEUE)
420                         bio_endio(bio, -EIO);
421
422                 bio = next;
423         }
424 }
425
426 static void process_queued_ios(struct work_struct *work)
427 {
428         struct multipath *m =
429                 container_of(work, struct multipath, process_queued_ios);
430         struct pgpath *pgpath = NULL;
431         unsigned init_required = 0, must_queue = 1;
432         unsigned long flags;
433
434         spin_lock_irqsave(&m->lock, flags);
435
436         if (!m->queue_size)
437                 goto out;
438
439         if (!m->current_pgpath)
440                 __choose_pgpath(m);
441
442         pgpath = m->current_pgpath;
443         m->pgpath_to_activate = m->current_pgpath;
444
445         if ((pgpath && !m->queue_io) ||
446             (!pgpath && !m->queue_if_no_path))
447                 must_queue = 0;
448
449         if (m->pg_init_required && !m->pg_init_in_progress) {
450                 m->pg_init_count++;
451                 m->pg_init_required = 0;
452                 m->pg_init_in_progress = 1;
453                 init_required = 1;
454         }
455
456 out:
457         spin_unlock_irqrestore(&m->lock, flags);
458
459         if (init_required)
460                 queue_work(kmpath_handlerd, &m->activate_path);
461
462         if (!must_queue)
463                 dispatch_queued_ios(m);
464 }
465
466 /*
467  * An event is triggered whenever a path is taken out of use.
468  * Includes path failure and PG bypass.
469  */
470 static void trigger_event(struct work_struct *work)
471 {
472         struct multipath *m =
473                 container_of(work, struct multipath, trigger_event);
474
475         dm_table_event(m->ti->table);
476 }
477
478 /*-----------------------------------------------------------------
479  * Constructor/argument parsing:
480  * <#multipath feature args> [<arg>]*
481  * <#hw_handler args> [hw_handler [<arg>]*]
482  * <#priority groups>
483  * <initial priority group>
484  *     [<selector> <#selector args> [<arg>]*
485  *      <#paths> <#per-path selector args>
486  *         [<path> [<arg>]* ]+ ]+
487  *---------------------------------------------------------------*/
488 struct param {
489         unsigned min;
490         unsigned max;
491         char *error;
492 };
493
494 static int read_param(struct param *param, char *str, unsigned *v, char **error)
495 {
496         if (!str ||
497             (sscanf(str, "%u", v) != 1) ||
498             (*v < param->min) ||
499             (*v > param->max)) {
500                 *error = param->error;
501                 return -EINVAL;
502         }
503
504         return 0;
505 }
506
507 struct arg_set {
508         unsigned argc;
509         char **argv;
510 };
511
512 static char *shift(struct arg_set *as)
513 {
514         char *r;
515
516         if (as->argc) {
517                 as->argc--;
518                 r = *as->argv;
519                 as->argv++;
520                 return r;
521         }
522
523         return NULL;
524 }
525
526 static void consume(struct arg_set *as, unsigned n)
527 {
528         BUG_ON (as->argc < n);
529         as->argc -= n;
530         as->argv += n;
531 }
532
533 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
534                                struct dm_target *ti)
535 {
536         int r;
537         struct path_selector_type *pst;
538         unsigned ps_argc;
539
540         static struct param _params[] = {
541                 {0, 1024, "invalid number of path selector args"},
542         };
543
544         pst = dm_get_path_selector(shift(as));
545         if (!pst) {
546                 ti->error = "unknown path selector type";
547                 return -EINVAL;
548         }
549
550         r = read_param(_params, shift(as), &ps_argc, &ti->error);
551         if (r) {
552                 dm_put_path_selector(pst);
553                 return -EINVAL;
554         }
555
556         r = pst->create(&pg->ps, ps_argc, as->argv);
557         if (r) {
558                 dm_put_path_selector(pst);
559                 ti->error = "path selector constructor failed";
560                 return r;
561         }
562
563         pg->ps.type = pst;
564         consume(as, ps_argc);
565
566         return 0;
567 }
568
569 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
570                                struct dm_target *ti)
571 {
572         int r;
573         struct pgpath *p;
574         struct multipath *m = ti->private;
575
576         /* we need at least a path arg */
577         if (as->argc < 1) {
578                 ti->error = "no device given";
579                 return ERR_PTR(-EINVAL);
580         }
581
582         p = alloc_pgpath();
583         if (!p)
584                 return ERR_PTR(-ENOMEM);
585
586         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
587                           dm_table_get_mode(ti->table), &p->path.dev);
588         if (r) {
589                 ti->error = "error getting device";
590                 goto bad;
591         }
592
593         if (m->hw_handler_name) {
594                 r = scsi_dh_attach(bdev_get_queue(p->path.dev->bdev),
595                                    m->hw_handler_name);
596                 if (r < 0) {
597                         dm_put_device(ti, p->path.dev);
598                         goto bad;
599                 }
600         }
601
602         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
603         if (r) {
604                 dm_put_device(ti, p->path.dev);
605                 goto bad;
606         }
607
608         return p;
609
610  bad:
611         free_pgpath(p);
612         return ERR_PTR(r);
613 }
614
615 static struct priority_group *parse_priority_group(struct arg_set *as,
616                                                    struct multipath *m)
617 {
618         static struct param _params[] = {
619                 {1, 1024, "invalid number of paths"},
620                 {0, 1024, "invalid number of selector args"}
621         };
622
623         int r;
624         unsigned i, nr_selector_args, nr_params;
625         struct priority_group *pg;
626         struct dm_target *ti = m->ti;
627
628         if (as->argc < 2) {
629                 as->argc = 0;
630                 ti->error = "not enough priority group arguments";
631                 return ERR_PTR(-EINVAL);
632         }
633
634         pg = alloc_priority_group();
635         if (!pg) {
636                 ti->error = "couldn't allocate priority group";
637                 return ERR_PTR(-ENOMEM);
638         }
639         pg->m = m;
640
641         r = parse_path_selector(as, pg, ti);
642         if (r)
643                 goto bad;
644
645         /*
646          * read the paths
647          */
648         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
649         if (r)
650                 goto bad;
651
652         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
653         if (r)
654                 goto bad;
655
656         nr_params = 1 + nr_selector_args;
657         for (i = 0; i < pg->nr_pgpaths; i++) {
658                 struct pgpath *pgpath;
659                 struct arg_set path_args;
660
661                 if (as->argc < nr_params) {
662                         ti->error = "not enough path parameters";
663                         goto bad;
664                 }
665
666                 path_args.argc = nr_params;
667                 path_args.argv = as->argv;
668
669                 pgpath = parse_path(&path_args, &pg->ps, ti);
670                 if (IS_ERR(pgpath)) {
671                         r = PTR_ERR(pgpath);
672                         goto bad;
673                 }
674
675                 pgpath->pg = pg;
676                 list_add_tail(&pgpath->list, &pg->pgpaths);
677                 consume(as, nr_params);
678         }
679
680         return pg;
681
682  bad:
683         free_priority_group(pg, ti);
684         return ERR_PTR(r);
685 }
686
687 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
688 {
689         unsigned hw_argc;
690         struct dm_target *ti = m->ti;
691
692         static struct param _params[] = {
693                 {0, 1024, "invalid number of hardware handler args"},
694         };
695
696         if (read_param(_params, shift(as), &hw_argc, &ti->error))
697                 return -EINVAL;
698
699         if (!hw_argc)
700                 return 0;
701
702         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
703         request_module("scsi_dh_%s", m->hw_handler_name);
704         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
705                 ti->error = "unknown hardware handler type";
706                 kfree(m->hw_handler_name);
707                 m->hw_handler_name = NULL;
708                 return -EINVAL;
709         }
710         consume(as, hw_argc - 1);
711
712         return 0;
713 }
714
715 static int parse_features(struct arg_set *as, struct multipath *m)
716 {
717         int r;
718         unsigned argc;
719         struct dm_target *ti = m->ti;
720         const char *param_name;
721
722         static struct param _params[] = {
723                 {0, 3, "invalid number of feature args"},
724                 {1, 50, "pg_init_retries must be between 1 and 50"},
725         };
726
727         r = read_param(_params, shift(as), &argc, &ti->error);
728         if (r)
729                 return -EINVAL;
730
731         if (!argc)
732                 return 0;
733
734         do {
735                 param_name = shift(as);
736                 argc--;
737
738                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
739                         r = queue_if_no_path(m, 1, 0);
740                         continue;
741                 }
742
743                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
744                     (argc >= 1)) {
745                         r = read_param(_params + 1, shift(as),
746                                        &m->pg_init_retries, &ti->error);
747                         argc--;
748                         continue;
749                 }
750
751                 ti->error = "Unrecognised multipath feature request";
752                 r = -EINVAL;
753         } while (argc && !r);
754
755         return r;
756 }
757
758 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
759                          char **argv)
760 {
761         /* target parameters */
762         static struct param _params[] = {
763                 {1, 1024, "invalid number of priority groups"},
764                 {1, 1024, "invalid initial priority group number"},
765         };
766
767         int r;
768         struct multipath *m;
769         struct arg_set as;
770         unsigned pg_count = 0;
771         unsigned next_pg_num;
772
773         as.argc = argc;
774         as.argv = argv;
775
776         m = alloc_multipath(ti);
777         if (!m) {
778                 ti->error = "can't allocate multipath";
779                 return -EINVAL;
780         }
781
782         r = parse_features(&as, m);
783         if (r)
784                 goto bad;
785
786         r = parse_hw_handler(&as, m);
787         if (r)
788                 goto bad;
789
790         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
791         if (r)
792                 goto bad;
793
794         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
795         if (r)
796                 goto bad;
797
798         /* parse the priority groups */
799         while (as.argc) {
800                 struct priority_group *pg;
801
802                 pg = parse_priority_group(&as, m);
803                 if (IS_ERR(pg)) {
804                         r = PTR_ERR(pg);
805                         goto bad;
806                 }
807
808                 m->nr_valid_paths += pg->nr_pgpaths;
809                 list_add_tail(&pg->list, &m->priority_groups);
810                 pg_count++;
811                 pg->pg_num = pg_count;
812                 if (!--next_pg_num)
813                         m->next_pg = pg;
814         }
815
816         if (pg_count != m->nr_priority_groups) {
817                 ti->error = "priority group count mismatch";
818                 r = -EINVAL;
819                 goto bad;
820         }
821
822         return 0;
823
824  bad:
825         free_multipath(m);
826         return r;
827 }
828
829 static void multipath_dtr(struct dm_target *ti)
830 {
831         struct multipath *m = (struct multipath *) ti->private;
832
833         flush_workqueue(kmpath_handlerd);
834         flush_workqueue(kmultipathd);
835         free_multipath(m);
836 }
837
838 /*
839  * Map bios, recording original fields for later in case we have to resubmit
840  */
841 static int multipath_map(struct dm_target *ti, struct bio *bio,
842                          union map_info *map_context)
843 {
844         int r;
845         struct dm_mpath_io *mpio;
846         struct multipath *m = (struct multipath *) ti->private;
847
848         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
849         dm_bio_record(&mpio->details, bio);
850
851         map_context->ptr = mpio;
852         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
853         r = map_io(m, bio, mpio, 0);
854         if (r < 0 || r == DM_MAPIO_REQUEUE)
855                 mempool_free(mpio, m->mpio_pool);
856
857         return r;
858 }
859
860 /*
861  * Take a path out of use.
862  */
863 static int fail_path(struct pgpath *pgpath)
864 {
865         unsigned long flags;
866         struct multipath *m = pgpath->pg->m;
867
868         spin_lock_irqsave(&m->lock, flags);
869
870         if (!pgpath->is_active)
871                 goto out;
872
873         DMWARN("Failing path %s.", pgpath->path.dev->name);
874
875         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
876         pgpath->is_active = 0;
877         pgpath->fail_count++;
878
879         m->nr_valid_paths--;
880
881         if (pgpath == m->current_pgpath)
882                 m->current_pgpath = NULL;
883
884         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
885                       pgpath->path.dev->name, m->nr_valid_paths);
886
887         queue_work(kmultipathd, &m->trigger_event);
888         queue_work(kmultipathd, &pgpath->deactivate_path);
889
890 out:
891         spin_unlock_irqrestore(&m->lock, flags);
892
893         return 0;
894 }
895
896 /*
897  * Reinstate a previously-failed path
898  */
899 static int reinstate_path(struct pgpath *pgpath)
900 {
901         int r = 0;
902         unsigned long flags;
903         struct multipath *m = pgpath->pg->m;
904
905         spin_lock_irqsave(&m->lock, flags);
906
907         if (pgpath->is_active)
908                 goto out;
909
910         if (!pgpath->pg->ps.type->reinstate_path) {
911                 DMWARN("Reinstate path not supported by path selector %s",
912                        pgpath->pg->ps.type->name);
913                 r = -EINVAL;
914                 goto out;
915         }
916
917         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
918         if (r)
919                 goto out;
920
921         pgpath->is_active = 1;
922
923         m->current_pgpath = NULL;
924         if (!m->nr_valid_paths++ && m->queue_size)
925                 queue_work(kmultipathd, &m->process_queued_ios);
926
927         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
928                       pgpath->path.dev->name, m->nr_valid_paths);
929
930         queue_work(kmultipathd, &m->trigger_event);
931
932 out:
933         spin_unlock_irqrestore(&m->lock, flags);
934
935         return r;
936 }
937
938 /*
939  * Fail or reinstate all paths that match the provided struct dm_dev.
940  */
941 static int action_dev(struct multipath *m, struct dm_dev *dev,
942                       action_fn action)
943 {
944         int r = 0;
945         struct pgpath *pgpath;
946         struct priority_group *pg;
947
948         list_for_each_entry(pg, &m->priority_groups, list) {
949                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
950                         if (pgpath->path.dev == dev)
951                                 r = action(pgpath);
952                 }
953         }
954
955         return r;
956 }
957
958 /*
959  * Temporarily try to avoid having to use the specified PG
960  */
961 static void bypass_pg(struct multipath *m, struct priority_group *pg,
962                       int bypassed)
963 {
964         unsigned long flags;
965
966         spin_lock_irqsave(&m->lock, flags);
967
968         pg->bypassed = bypassed;
969         m->current_pgpath = NULL;
970         m->current_pg = NULL;
971
972         spin_unlock_irqrestore(&m->lock, flags);
973
974         queue_work(kmultipathd, &m->trigger_event);
975 }
976
977 /*
978  * Switch to using the specified PG from the next I/O that gets mapped
979  */
980 static int switch_pg_num(struct multipath *m, const char *pgstr)
981 {
982         struct priority_group *pg;
983         unsigned pgnum;
984         unsigned long flags;
985
986         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
987             (pgnum > m->nr_priority_groups)) {
988                 DMWARN("invalid PG number supplied to switch_pg_num");
989                 return -EINVAL;
990         }
991
992         spin_lock_irqsave(&m->lock, flags);
993         list_for_each_entry(pg, &m->priority_groups, list) {
994                 pg->bypassed = 0;
995                 if (--pgnum)
996                         continue;
997
998                 m->current_pgpath = NULL;
999                 m->current_pg = NULL;
1000                 m->next_pg = pg;
1001         }
1002         spin_unlock_irqrestore(&m->lock, flags);
1003
1004         queue_work(kmultipathd, &m->trigger_event);
1005         return 0;
1006 }
1007
1008 /*
1009  * Set/clear bypassed status of a PG.
1010  * PGs are numbered upwards from 1 in the order they were declared.
1011  */
1012 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1013 {
1014         struct priority_group *pg;
1015         unsigned pgnum;
1016
1017         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1018             (pgnum > m->nr_priority_groups)) {
1019                 DMWARN("invalid PG number supplied to bypass_pg");
1020                 return -EINVAL;
1021         }
1022
1023         list_for_each_entry(pg, &m->priority_groups, list) {
1024                 if (!--pgnum)
1025                         break;
1026         }
1027
1028         bypass_pg(m, pg, bypassed);
1029         return 0;
1030 }
1031
1032 /*
1033  * Should we retry pg_init immediately?
1034  */
1035 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1036 {
1037         unsigned long flags;
1038         int limit_reached = 0;
1039
1040         spin_lock_irqsave(&m->lock, flags);
1041
1042         if (m->pg_init_count <= m->pg_init_retries)
1043                 m->pg_init_required = 1;
1044         else
1045                 limit_reached = 1;
1046
1047         spin_unlock_irqrestore(&m->lock, flags);
1048
1049         return limit_reached;
1050 }
1051
1052 static void pg_init_done(struct dm_path *path, int errors)
1053 {
1054         struct pgpath *pgpath = path_to_pgpath(path);
1055         struct priority_group *pg = pgpath->pg;
1056         struct multipath *m = pg->m;
1057         unsigned long flags;
1058
1059         /* device or driver problems */
1060         switch (errors) {
1061         case SCSI_DH_OK:
1062                 break;
1063         case SCSI_DH_NOSYS:
1064                 if (!m->hw_handler_name) {
1065                         errors = 0;
1066                         break;
1067                 }
1068                 DMERR("Cannot failover device because scsi_dh_%s was not "
1069                       "loaded.", m->hw_handler_name);
1070                 /*
1071                  * Fail path for now, so we do not ping pong
1072                  */
1073                 fail_path(pgpath);
1074                 break;
1075         case SCSI_DH_DEV_TEMP_BUSY:
1076                 /*
1077                  * Probably doing something like FW upgrade on the
1078                  * controller so try the other pg.
1079                  */
1080                 bypass_pg(m, pg, 1);
1081                 break;
1082         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1083         case SCSI_DH_RETRY:
1084         case SCSI_DH_IMM_RETRY:
1085         case SCSI_DH_RES_TEMP_UNAVAIL:
1086                 if (pg_init_limit_reached(m, pgpath))
1087                         fail_path(pgpath);
1088                 errors = 0;
1089                 break;
1090         default:
1091                 /*
1092                  * We probably do not want to fail the path for a device
1093                  * error, but this is what the old dm did. In future
1094                  * patches we can do more advanced handling.
1095                  */
1096                 fail_path(pgpath);
1097         }
1098
1099         spin_lock_irqsave(&m->lock, flags);
1100         if (errors) {
1101                 DMERR("Could not failover device. Error %d.", errors);
1102                 m->current_pgpath = NULL;
1103                 m->current_pg = NULL;
1104         } else if (!m->pg_init_required) {
1105                 m->queue_io = 0;
1106                 pg->bypassed = 0;
1107         }
1108
1109         m->pg_init_in_progress = 0;
1110         queue_work(kmultipathd, &m->process_queued_ios);
1111         spin_unlock_irqrestore(&m->lock, flags);
1112 }
1113
1114 static void activate_path(struct work_struct *work)
1115 {
1116         int ret;
1117         struct multipath *m =
1118                 container_of(work, struct multipath, activate_path);
1119         struct dm_path *path;
1120         unsigned long flags;
1121
1122         spin_lock_irqsave(&m->lock, flags);
1123         path = &m->pgpath_to_activate->path;
1124         m->pgpath_to_activate = NULL;
1125         spin_unlock_irqrestore(&m->lock, flags);
1126         if (!path)
1127                 return;
1128         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1129         pg_init_done(path, ret);
1130 }
1131
1132 /*
1133  * end_io handling
1134  */
1135 static int do_end_io(struct multipath *m, struct bio *bio,
1136                      int error, struct dm_mpath_io *mpio)
1137 {
1138         unsigned long flags;
1139
1140         if (!error)
1141                 return 0;       /* I/O complete */
1142
1143         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1144                 return error;
1145
1146         if (error == -EOPNOTSUPP)
1147                 return error;
1148
1149         spin_lock_irqsave(&m->lock, flags);
1150         if (!m->nr_valid_paths) {
1151                 if (__must_push_back(m)) {
1152                         spin_unlock_irqrestore(&m->lock, flags);
1153                         return DM_ENDIO_REQUEUE;
1154                 } else if (!m->queue_if_no_path) {
1155                         spin_unlock_irqrestore(&m->lock, flags);
1156                         return -EIO;
1157                 } else {
1158                         spin_unlock_irqrestore(&m->lock, flags);
1159                         goto requeue;
1160                 }
1161         }
1162         spin_unlock_irqrestore(&m->lock, flags);
1163
1164         if (mpio->pgpath)
1165                 fail_path(mpio->pgpath);
1166
1167       requeue:
1168         dm_bio_restore(&mpio->details, bio);
1169
1170         /* queue for the daemon to resubmit or fail */
1171         spin_lock_irqsave(&m->lock, flags);
1172         bio_list_add(&m->queued_ios, bio);
1173         m->queue_size++;
1174         if (!m->queue_io)
1175                 queue_work(kmultipathd, &m->process_queued_ios);
1176         spin_unlock_irqrestore(&m->lock, flags);
1177
1178         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1179 }
1180
1181 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1182                             int error, union map_info *map_context)
1183 {
1184         struct multipath *m = ti->private;
1185         struct dm_mpath_io *mpio = map_context->ptr;
1186         struct pgpath *pgpath = mpio->pgpath;
1187         struct path_selector *ps;
1188         int r;
1189
1190         r  = do_end_io(m, bio, error, mpio);
1191         if (pgpath) {
1192                 ps = &pgpath->pg->ps;
1193                 if (ps->type->end_io)
1194                         ps->type->end_io(ps, &pgpath->path);
1195         }
1196         if (r != DM_ENDIO_INCOMPLETE)
1197                 mempool_free(mpio, m->mpio_pool);
1198
1199         return r;
1200 }
1201
1202 /*
1203  * Suspend can't complete until all the I/O is processed so if
1204  * the last path fails we must error any remaining I/O.
1205  * Note that if the freeze_bdev fails while suspending, the
1206  * queue_if_no_path state is lost - userspace should reset it.
1207  */
1208 static void multipath_presuspend(struct dm_target *ti)
1209 {
1210         struct multipath *m = (struct multipath *) ti->private;
1211
1212         queue_if_no_path(m, 0, 1);
1213 }
1214
1215 /*
1216  * Restore the queue_if_no_path setting.
1217  */
1218 static void multipath_resume(struct dm_target *ti)
1219 {
1220         struct multipath *m = (struct multipath *) ti->private;
1221         unsigned long flags;
1222
1223         spin_lock_irqsave(&m->lock, flags);
1224         m->queue_if_no_path = m->saved_queue_if_no_path;
1225         spin_unlock_irqrestore(&m->lock, flags);
1226 }
1227
1228 /*
1229  * Info output has the following format:
1230  * num_multipath_feature_args [multipath_feature_args]*
1231  * num_handler_status_args [handler_status_args]*
1232  * num_groups init_group_number
1233  *            [A|D|E num_ps_status_args [ps_status_args]*
1234  *             num_paths num_selector_args
1235  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1236  *
1237  * Table output has the following format (identical to the constructor string):
1238  * num_feature_args [features_args]*
1239  * num_handler_args hw_handler [hw_handler_args]*
1240  * num_groups init_group_number
1241  *     [priority selector-name num_ps_args [ps_args]*
1242  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1243  */
1244 static int multipath_status(struct dm_target *ti, status_type_t type,
1245                             char *result, unsigned int maxlen)
1246 {
1247         int sz = 0;
1248         unsigned long flags;
1249         struct multipath *m = (struct multipath *) ti->private;
1250         struct priority_group *pg;
1251         struct pgpath *p;
1252         unsigned pg_num;
1253         char state;
1254
1255         spin_lock_irqsave(&m->lock, flags);
1256
1257         /* Features */
1258         if (type == STATUSTYPE_INFO)
1259                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1260         else {
1261                 DMEMIT("%u ", m->queue_if_no_path +
1262                               (m->pg_init_retries > 0) * 2);
1263                 if (m->queue_if_no_path)
1264                         DMEMIT("queue_if_no_path ");
1265                 if (m->pg_init_retries)
1266                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1267         }
1268
1269         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1270                 DMEMIT("0 ");
1271         else
1272                 DMEMIT("1 %s ", m->hw_handler_name);
1273
1274         DMEMIT("%u ", m->nr_priority_groups);
1275
1276         if (m->next_pg)
1277                 pg_num = m->next_pg->pg_num;
1278         else if (m->current_pg)
1279                 pg_num = m->current_pg->pg_num;
1280         else
1281                         pg_num = 1;
1282
1283         DMEMIT("%u ", pg_num);
1284
1285         switch (type) {
1286         case STATUSTYPE_INFO:
1287                 list_for_each_entry(pg, &m->priority_groups, list) {
1288                         if (pg->bypassed)
1289                                 state = 'D';    /* Disabled */
1290                         else if (pg == m->current_pg)
1291                                 state = 'A';    /* Currently Active */
1292                         else
1293                                 state = 'E';    /* Enabled */
1294
1295                         DMEMIT("%c ", state);
1296
1297                         if (pg->ps.type->status)
1298                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1299                                                           result + sz,
1300                                                           maxlen - sz);
1301                         else
1302                                 DMEMIT("0 ");
1303
1304                         DMEMIT("%u %u ", pg->nr_pgpaths,
1305                                pg->ps.type->info_args);
1306
1307                         list_for_each_entry(p, &pg->pgpaths, list) {
1308                                 DMEMIT("%s %s %u ", p->path.dev->name,
1309                                        p->is_active ? "A" : "F",
1310                                        p->fail_count);
1311                                 if (pg->ps.type->status)
1312                                         sz += pg->ps.type->status(&pg->ps,
1313                                               &p->path, type, result + sz,
1314                                               maxlen - sz);
1315                         }
1316                 }
1317                 break;
1318
1319         case STATUSTYPE_TABLE:
1320                 list_for_each_entry(pg, &m->priority_groups, list) {
1321                         DMEMIT("%s ", pg->ps.type->name);
1322
1323                         if (pg->ps.type->status)
1324                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1325                                                           result + sz,
1326                                                           maxlen - sz);
1327                         else
1328                                 DMEMIT("0 ");
1329
1330                         DMEMIT("%u %u ", pg->nr_pgpaths,
1331                                pg->ps.type->table_args);
1332
1333                         list_for_each_entry(p, &pg->pgpaths, list) {
1334                                 DMEMIT("%s ", p->path.dev->name);
1335                                 if (pg->ps.type->status)
1336                                         sz += pg->ps.type->status(&pg->ps,
1337                                               &p->path, type, result + sz,
1338                                               maxlen - sz);
1339                         }
1340                 }
1341                 break;
1342         }
1343
1344         spin_unlock_irqrestore(&m->lock, flags);
1345
1346         return 0;
1347 }
1348
1349 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1350 {
1351         int r;
1352         struct dm_dev *dev;
1353         struct multipath *m = (struct multipath *) ti->private;
1354         action_fn action;
1355
1356         if (argc == 1) {
1357                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1358                         return queue_if_no_path(m, 1, 0);
1359                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1360                         return queue_if_no_path(m, 0, 0);
1361         }
1362
1363         if (argc != 2)
1364                 goto error;
1365
1366         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1367                 return bypass_pg_num(m, argv[1], 1);
1368         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1369                 return bypass_pg_num(m, argv[1], 0);
1370         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1371                 return switch_pg_num(m, argv[1]);
1372         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1373                 action = reinstate_path;
1374         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1375                 action = fail_path;
1376         else
1377                 goto error;
1378
1379         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1380                           dm_table_get_mode(ti->table), &dev);
1381         if (r) {
1382                 DMWARN("message: error getting device %s",
1383                        argv[1]);
1384                 return -EINVAL;
1385         }
1386
1387         r = action_dev(m, dev, action);
1388
1389         dm_put_device(ti, dev);
1390
1391         return r;
1392
1393 error:
1394         DMWARN("Unrecognised multipath message received.");
1395         return -EINVAL;
1396 }
1397
1398 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1399                            struct file *filp, unsigned int cmd,
1400                            unsigned long arg)
1401 {
1402         struct multipath *m = (struct multipath *) ti->private;
1403         struct block_device *bdev = NULL;
1404         unsigned long flags;
1405         struct file fake_file = {};
1406         struct dentry fake_dentry = {};
1407         int r = 0;
1408
1409         fake_file.f_path.dentry = &fake_dentry;
1410
1411         spin_lock_irqsave(&m->lock, flags);
1412
1413         if (!m->current_pgpath)
1414                 __choose_pgpath(m);
1415
1416         if (m->current_pgpath) {
1417                 bdev = m->current_pgpath->path.dev->bdev;
1418                 fake_dentry.d_inode = bdev->bd_inode;
1419                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1420         }
1421
1422         if (m->queue_io)
1423                 r = -EAGAIN;
1424         else if (!bdev)
1425                 r = -EIO;
1426
1427         spin_unlock_irqrestore(&m->lock, flags);
1428
1429         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1430                                          bdev->bd_disk, cmd, arg);
1431 }
1432
1433 /*-----------------------------------------------------------------
1434  * Module setup
1435  *---------------------------------------------------------------*/
1436 static struct target_type multipath_target = {
1437         .name = "multipath",
1438         .version = {1, 0, 5},
1439         .module = THIS_MODULE,
1440         .ctr = multipath_ctr,
1441         .dtr = multipath_dtr,
1442         .map = multipath_map,
1443         .end_io = multipath_end_io,
1444         .presuspend = multipath_presuspend,
1445         .resume = multipath_resume,
1446         .status = multipath_status,
1447         .message = multipath_message,
1448         .ioctl  = multipath_ioctl,
1449 };
1450
1451 static int __init dm_multipath_init(void)
1452 {
1453         int r;
1454
1455         /* allocate a slab for the dm_ios */
1456         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1457         if (!_mpio_cache)
1458                 return -ENOMEM;
1459
1460         r = dm_register_target(&multipath_target);
1461         if (r < 0) {
1462                 DMERR("register failed %d", r);
1463                 kmem_cache_destroy(_mpio_cache);
1464                 return -EINVAL;
1465         }
1466
1467         kmultipathd = create_workqueue("kmpathd");
1468         if (!kmultipathd) {
1469                 DMERR("failed to create workqueue kmpathd");
1470                 dm_unregister_target(&multipath_target);
1471                 kmem_cache_destroy(_mpio_cache);
1472                 return -ENOMEM;
1473         }
1474
1475         /*
1476          * A separate workqueue is used to handle the device handlers
1477          * to avoid overloading existing workqueue. Overloading the
1478          * old workqueue would also create a bottleneck in the
1479          * path of the storage hardware device activation.
1480          */
1481         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1482         if (!kmpath_handlerd) {
1483                 DMERR("failed to create workqueue kmpath_handlerd");
1484                 destroy_workqueue(kmultipathd);
1485                 dm_unregister_target(&multipath_target);
1486                 kmem_cache_destroy(_mpio_cache);
1487                 return -ENOMEM;
1488         }
1489
1490         DMINFO("version %u.%u.%u loaded",
1491                multipath_target.version[0], multipath_target.version[1],
1492                multipath_target.version[2]);
1493
1494         return r;
1495 }
1496
1497 static void __exit dm_multipath_exit(void)
1498 {
1499         int r;
1500
1501         destroy_workqueue(kmpath_handlerd);
1502         destroy_workqueue(kmultipathd);
1503
1504         r = dm_unregister_target(&multipath_target);
1505         if (r < 0)
1506                 DMERR("target unregister failed %d", r);
1507         kmem_cache_destroy(_mpio_cache);
1508 }
1509
1510 module_init(dm_multipath_init);
1511 module_exit(dm_multipath_exit);
1512
1513 MODULE_DESCRIPTION(DM_NAME " multipath target");
1514 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1515 MODULE_LICENSE("GPL");