dm mpath: fail message ioctl if specified path is not valid
[linux-2.6.git] / drivers / md / dm-mpath.c
1 /*
2  * Copyright (C) 2003 Sistina Software Limited.
3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
4  *
5  * This file is released under the GPL.
6  */
7
8 #include <linux/device-mapper.h>
9
10 #include "dm-path-selector.h"
11 #include "dm-uevent.h"
12
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/workqueue.h>
21 #include <scsi/scsi_dh.h>
22 #include <asm/atomic.h>
23
24 #define DM_MSG_PREFIX "multipath"
25 #define MESG_STR(x) x, sizeof(x)
26 #define DM_PG_INIT_DELAY_MSECS 2000
27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
28
29 /* Path properties */
30 struct pgpath {
31         struct list_head list;
32
33         struct priority_group *pg;      /* Owning PG */
34         unsigned is_active;             /* Path status */
35         unsigned fail_count;            /* Cumulative failure count */
36
37         struct dm_path path;
38         struct delayed_work activate_path;
39 };
40
41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
42
43 /*
44  * Paths are grouped into Priority Groups and numbered from 1 upwards.
45  * Each has a path selector which controls which path gets used.
46  */
47 struct priority_group {
48         struct list_head list;
49
50         struct multipath *m;            /* Owning multipath instance */
51         struct path_selector ps;
52
53         unsigned pg_num;                /* Reference number */
54         unsigned bypassed;              /* Temporarily bypass this PG? */
55
56         unsigned nr_pgpaths;            /* Number of paths in PG */
57         struct list_head pgpaths;
58 };
59
60 /* Multipath context */
61 struct multipath {
62         struct list_head list;
63         struct dm_target *ti;
64
65         spinlock_t lock;
66
67         const char *hw_handler_name;
68         char *hw_handler_params;
69
70         unsigned nr_priority_groups;
71         struct list_head priority_groups;
72
73         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
74
75         unsigned pg_init_required;      /* pg_init needs calling? */
76         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
77         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
78
79         unsigned nr_valid_paths;        /* Total number of usable paths */
80         struct pgpath *current_pgpath;
81         struct priority_group *current_pg;
82         struct priority_group *next_pg; /* Switch to this PG if set */
83         unsigned repeat_count;          /* I/Os left before calling PS again */
84
85         unsigned queue_io;              /* Must we queue all I/O? */
86         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
87         unsigned saved_queue_if_no_path;/* Saved state during suspension */
88         unsigned pg_init_retries;       /* Number of times to retry pg_init */
89         unsigned pg_init_count;         /* Number of times pg_init called */
90         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
91
92         struct work_struct process_queued_ios;
93         struct list_head queued_ios;
94         unsigned queue_size;
95
96         struct work_struct trigger_event;
97
98         /*
99          * We must use a mempool of dm_mpath_io structs so that we
100          * can resubmit bios on error.
101          */
102         mempool_t *mpio_pool;
103
104         struct mutex work_mutex;
105 };
106
107 /*
108  * Context information attached to each bio we process.
109  */
110 struct dm_mpath_io {
111         struct pgpath *pgpath;
112         size_t nr_bytes;
113 };
114
115 typedef int (*action_fn) (struct pgpath *pgpath);
116
117 #define MIN_IOS 256     /* Mempool size */
118
119 static struct kmem_cache *_mpio_cache;
120
121 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
122 static void process_queued_ios(struct work_struct *work);
123 static void trigger_event(struct work_struct *work);
124 static void activate_path(struct work_struct *work);
125
126
127 /*-----------------------------------------------
128  * Allocation routines
129  *-----------------------------------------------*/
130
131 static struct pgpath *alloc_pgpath(void)
132 {
133         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
134
135         if (pgpath) {
136                 pgpath->is_active = 1;
137                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
138         }
139
140         return pgpath;
141 }
142
143 static void free_pgpath(struct pgpath *pgpath)
144 {
145         kfree(pgpath);
146 }
147
148 static struct priority_group *alloc_priority_group(void)
149 {
150         struct priority_group *pg;
151
152         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
153
154         if (pg)
155                 INIT_LIST_HEAD(&pg->pgpaths);
156
157         return pg;
158 }
159
160 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
161 {
162         struct pgpath *pgpath, *tmp;
163         struct multipath *m = ti->private;
164
165         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
166                 list_del(&pgpath->list);
167                 if (m->hw_handler_name)
168                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
169                 dm_put_device(ti, pgpath->path.dev);
170                 free_pgpath(pgpath);
171         }
172 }
173
174 static void free_priority_group(struct priority_group *pg,
175                                 struct dm_target *ti)
176 {
177         struct path_selector *ps = &pg->ps;
178
179         if (ps->type) {
180                 ps->type->destroy(ps);
181                 dm_put_path_selector(ps->type);
182         }
183
184         free_pgpaths(&pg->pgpaths, ti);
185         kfree(pg);
186 }
187
188 static struct multipath *alloc_multipath(struct dm_target *ti)
189 {
190         struct multipath *m;
191
192         m = kzalloc(sizeof(*m), GFP_KERNEL);
193         if (m) {
194                 INIT_LIST_HEAD(&m->priority_groups);
195                 INIT_LIST_HEAD(&m->queued_ios);
196                 spin_lock_init(&m->lock);
197                 m->queue_io = 1;
198                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
199                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
200                 INIT_WORK(&m->trigger_event, trigger_event);
201                 init_waitqueue_head(&m->pg_init_wait);
202                 mutex_init(&m->work_mutex);
203                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
204                 if (!m->mpio_pool) {
205                         kfree(m);
206                         return NULL;
207                 }
208                 m->ti = ti;
209                 ti->private = m;
210         }
211
212         return m;
213 }
214
215 static void free_multipath(struct multipath *m)
216 {
217         struct priority_group *pg, *tmp;
218
219         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
220                 list_del(&pg->list);
221                 free_priority_group(pg, m->ti);
222         }
223
224         kfree(m->hw_handler_name);
225         kfree(m->hw_handler_params);
226         mempool_destroy(m->mpio_pool);
227         kfree(m);
228 }
229
230
231 /*-----------------------------------------------
232  * Path selection
233  *-----------------------------------------------*/
234
235 static void __pg_init_all_paths(struct multipath *m)
236 {
237         struct pgpath *pgpath;
238         unsigned long pg_init_delay = 0;
239
240         m->pg_init_count++;
241         m->pg_init_required = 0;
242         if (m->pg_init_delay_retry)
243                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
244                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
245         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
246                 /* Skip failed paths */
247                 if (!pgpath->is_active)
248                         continue;
249                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
250                                        pg_init_delay))
251                         m->pg_init_in_progress++;
252         }
253 }
254
255 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
256 {
257         m->current_pg = pgpath->pg;
258
259         /* Must we initialise the PG first, and queue I/O till it's ready? */
260         if (m->hw_handler_name) {
261                 m->pg_init_required = 1;
262                 m->queue_io = 1;
263         } else {
264                 m->pg_init_required = 0;
265                 m->queue_io = 0;
266         }
267
268         m->pg_init_count = 0;
269 }
270
271 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
272                                size_t nr_bytes)
273 {
274         struct dm_path *path;
275
276         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
277         if (!path)
278                 return -ENXIO;
279
280         m->current_pgpath = path_to_pgpath(path);
281
282         if (m->current_pg != pg)
283                 __switch_pg(m, m->current_pgpath);
284
285         return 0;
286 }
287
288 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
289 {
290         struct priority_group *pg;
291         unsigned bypassed = 1;
292
293         if (!m->nr_valid_paths)
294                 goto failed;
295
296         /* Were we instructed to switch PG? */
297         if (m->next_pg) {
298                 pg = m->next_pg;
299                 m->next_pg = NULL;
300                 if (!__choose_path_in_pg(m, pg, nr_bytes))
301                         return;
302         }
303
304         /* Don't change PG until it has no remaining paths */
305         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
306                 return;
307
308         /*
309          * Loop through priority groups until we find a valid path.
310          * First time we skip PGs marked 'bypassed'.
311          * Second time we only try the ones we skipped.
312          */
313         do {
314                 list_for_each_entry(pg, &m->priority_groups, list) {
315                         if (pg->bypassed == bypassed)
316                                 continue;
317                         if (!__choose_path_in_pg(m, pg, nr_bytes))
318                                 return;
319                 }
320         } while (bypassed--);
321
322 failed:
323         m->current_pgpath = NULL;
324         m->current_pg = NULL;
325 }
326
327 /*
328  * Check whether bios must be queued in the device-mapper core rather
329  * than here in the target.
330  *
331  * m->lock must be held on entry.
332  *
333  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
334  * same value then we are not between multipath_presuspend()
335  * and multipath_resume() calls and we have no need to check
336  * for the DMF_NOFLUSH_SUSPENDING flag.
337  */
338 static int __must_push_back(struct multipath *m)
339 {
340         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
341                 dm_noflush_suspending(m->ti));
342 }
343
344 static int map_io(struct multipath *m, struct request *clone,
345                   struct dm_mpath_io *mpio, unsigned was_queued)
346 {
347         int r = DM_MAPIO_REMAPPED;
348         size_t nr_bytes = blk_rq_bytes(clone);
349         unsigned long flags;
350         struct pgpath *pgpath;
351         struct block_device *bdev;
352
353         spin_lock_irqsave(&m->lock, flags);
354
355         /* Do we need to select a new pgpath? */
356         if (!m->current_pgpath ||
357             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
358                 __choose_pgpath(m, nr_bytes);
359
360         pgpath = m->current_pgpath;
361
362         if (was_queued)
363                 m->queue_size--;
364
365         if ((pgpath && m->queue_io) ||
366             (!pgpath && m->queue_if_no_path)) {
367                 /* Queue for the daemon to resubmit */
368                 list_add_tail(&clone->queuelist, &m->queued_ios);
369                 m->queue_size++;
370                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
371                     !m->queue_io)
372                         queue_work(kmultipathd, &m->process_queued_ios);
373                 pgpath = NULL;
374                 r = DM_MAPIO_SUBMITTED;
375         } else if (pgpath) {
376                 bdev = pgpath->path.dev->bdev;
377                 clone->q = bdev_get_queue(bdev);
378                 clone->rq_disk = bdev->bd_disk;
379         } else if (__must_push_back(m))
380                 r = DM_MAPIO_REQUEUE;
381         else
382                 r = -EIO;       /* Failed */
383
384         mpio->pgpath = pgpath;
385         mpio->nr_bytes = nr_bytes;
386
387         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
388                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
389                                               nr_bytes);
390
391         spin_unlock_irqrestore(&m->lock, flags);
392
393         return r;
394 }
395
396 /*
397  * If we run out of usable paths, should we queue I/O or error it?
398  */
399 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
400                             unsigned save_old_value)
401 {
402         unsigned long flags;
403
404         spin_lock_irqsave(&m->lock, flags);
405
406         if (save_old_value)
407                 m->saved_queue_if_no_path = m->queue_if_no_path;
408         else
409                 m->saved_queue_if_no_path = queue_if_no_path;
410         m->queue_if_no_path = queue_if_no_path;
411         if (!m->queue_if_no_path && m->queue_size)
412                 queue_work(kmultipathd, &m->process_queued_ios);
413
414         spin_unlock_irqrestore(&m->lock, flags);
415
416         return 0;
417 }
418
419 /*-----------------------------------------------------------------
420  * The multipath daemon is responsible for resubmitting queued ios.
421  *---------------------------------------------------------------*/
422
423 static void dispatch_queued_ios(struct multipath *m)
424 {
425         int r;
426         unsigned long flags;
427         struct dm_mpath_io *mpio;
428         union map_info *info;
429         struct request *clone, *n;
430         LIST_HEAD(cl);
431
432         spin_lock_irqsave(&m->lock, flags);
433         list_splice_init(&m->queued_ios, &cl);
434         spin_unlock_irqrestore(&m->lock, flags);
435
436         list_for_each_entry_safe(clone, n, &cl, queuelist) {
437                 list_del_init(&clone->queuelist);
438
439                 info = dm_get_rq_mapinfo(clone);
440                 mpio = info->ptr;
441
442                 r = map_io(m, clone, mpio, 1);
443                 if (r < 0) {
444                         mempool_free(mpio, m->mpio_pool);
445                         dm_kill_unmapped_request(clone, r);
446                 } else if (r == DM_MAPIO_REMAPPED)
447                         dm_dispatch_request(clone);
448                 else if (r == DM_MAPIO_REQUEUE) {
449                         mempool_free(mpio, m->mpio_pool);
450                         dm_requeue_unmapped_request(clone);
451                 }
452         }
453 }
454
455 static void process_queued_ios(struct work_struct *work)
456 {
457         struct multipath *m =
458                 container_of(work, struct multipath, process_queued_ios);
459         struct pgpath *pgpath = NULL;
460         unsigned must_queue = 1;
461         unsigned long flags;
462
463         spin_lock_irqsave(&m->lock, flags);
464
465         if (!m->queue_size)
466                 goto out;
467
468         if (!m->current_pgpath)
469                 __choose_pgpath(m, 0);
470
471         pgpath = m->current_pgpath;
472
473         if ((pgpath && !m->queue_io) ||
474             (!pgpath && !m->queue_if_no_path))
475                 must_queue = 0;
476
477         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
478                 __pg_init_all_paths(m);
479
480 out:
481         spin_unlock_irqrestore(&m->lock, flags);
482         if (!must_queue)
483                 dispatch_queued_ios(m);
484 }
485
486 /*
487  * An event is triggered whenever a path is taken out of use.
488  * Includes path failure and PG bypass.
489  */
490 static void trigger_event(struct work_struct *work)
491 {
492         struct multipath *m =
493                 container_of(work, struct multipath, trigger_event);
494
495         dm_table_event(m->ti->table);
496 }
497
498 /*-----------------------------------------------------------------
499  * Constructor/argument parsing:
500  * <#multipath feature args> [<arg>]*
501  * <#hw_handler args> [hw_handler [<arg>]*]
502  * <#priority groups>
503  * <initial priority group>
504  *     [<selector> <#selector args> [<arg>]*
505  *      <#paths> <#per-path selector args>
506  *         [<path> [<arg>]* ]+ ]+
507  *---------------------------------------------------------------*/
508 struct param {
509         unsigned min;
510         unsigned max;
511         char *error;
512 };
513
514 static int read_param(struct param *param, char *str, unsigned *v, char **error)
515 {
516         if (!str ||
517             (sscanf(str, "%u", v) != 1) ||
518             (*v < param->min) ||
519             (*v > param->max)) {
520                 *error = param->error;
521                 return -EINVAL;
522         }
523
524         return 0;
525 }
526
527 struct arg_set {
528         unsigned argc;
529         char **argv;
530 };
531
532 static char *shift(struct arg_set *as)
533 {
534         char *r;
535
536         if (as->argc) {
537                 as->argc--;
538                 r = *as->argv;
539                 as->argv++;
540                 return r;
541         }
542
543         return NULL;
544 }
545
546 static void consume(struct arg_set *as, unsigned n)
547 {
548         BUG_ON (as->argc < n);
549         as->argc -= n;
550         as->argv += n;
551 }
552
553 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
554                                struct dm_target *ti)
555 {
556         int r;
557         struct path_selector_type *pst;
558         unsigned ps_argc;
559
560         static struct param _params[] = {
561                 {0, 1024, "invalid number of path selector args"},
562         };
563
564         pst = dm_get_path_selector(shift(as));
565         if (!pst) {
566                 ti->error = "unknown path selector type";
567                 return -EINVAL;
568         }
569
570         r = read_param(_params, shift(as), &ps_argc, &ti->error);
571         if (r) {
572                 dm_put_path_selector(pst);
573                 return -EINVAL;
574         }
575
576         if (ps_argc > as->argc) {
577                 dm_put_path_selector(pst);
578                 ti->error = "not enough arguments for path selector";
579                 return -EINVAL;
580         }
581
582         r = pst->create(&pg->ps, ps_argc, as->argv);
583         if (r) {
584                 dm_put_path_selector(pst);
585                 ti->error = "path selector constructor failed";
586                 return r;
587         }
588
589         pg->ps.type = pst;
590         consume(as, ps_argc);
591
592         return 0;
593 }
594
595 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
596                                struct dm_target *ti)
597 {
598         int r;
599         struct pgpath *p;
600         struct multipath *m = ti->private;
601
602         /* we need at least a path arg */
603         if (as->argc < 1) {
604                 ti->error = "no device given";
605                 return ERR_PTR(-EINVAL);
606         }
607
608         p = alloc_pgpath();
609         if (!p)
610                 return ERR_PTR(-ENOMEM);
611
612         r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
613                           &p->path.dev);
614         if (r) {
615                 ti->error = "error getting device";
616                 goto bad;
617         }
618
619         if (m->hw_handler_name) {
620                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
621
622                 r = scsi_dh_attach(q, m->hw_handler_name);
623                 if (r == -EBUSY) {
624                         /*
625                          * Already attached to different hw_handler,
626                          * try to reattach with correct one.
627                          */
628                         scsi_dh_detach(q);
629                         r = scsi_dh_attach(q, m->hw_handler_name);
630                 }
631
632                 if (r < 0) {
633                         ti->error = "error attaching hardware handler";
634                         dm_put_device(ti, p->path.dev);
635                         goto bad;
636                 }
637
638                 if (m->hw_handler_params) {
639                         r = scsi_dh_set_params(q, m->hw_handler_params);
640                         if (r < 0) {
641                                 ti->error = "unable to set hardware "
642                                                         "handler parameters";
643                                 scsi_dh_detach(q);
644                                 dm_put_device(ti, p->path.dev);
645                                 goto bad;
646                         }
647                 }
648         }
649
650         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
651         if (r) {
652                 dm_put_device(ti, p->path.dev);
653                 goto bad;
654         }
655
656         return p;
657
658  bad:
659         free_pgpath(p);
660         return ERR_PTR(r);
661 }
662
663 static struct priority_group *parse_priority_group(struct arg_set *as,
664                                                    struct multipath *m)
665 {
666         static struct param _params[] = {
667                 {1, 1024, "invalid number of paths"},
668                 {0, 1024, "invalid number of selector args"}
669         };
670
671         int r;
672         unsigned i, nr_selector_args, nr_params;
673         struct priority_group *pg;
674         struct dm_target *ti = m->ti;
675
676         if (as->argc < 2) {
677                 as->argc = 0;
678                 ti->error = "not enough priority group arguments";
679                 return ERR_PTR(-EINVAL);
680         }
681
682         pg = alloc_priority_group();
683         if (!pg) {
684                 ti->error = "couldn't allocate priority group";
685                 return ERR_PTR(-ENOMEM);
686         }
687         pg->m = m;
688
689         r = parse_path_selector(as, pg, ti);
690         if (r)
691                 goto bad;
692
693         /*
694          * read the paths
695          */
696         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
697         if (r)
698                 goto bad;
699
700         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
701         if (r)
702                 goto bad;
703
704         nr_params = 1 + nr_selector_args;
705         for (i = 0; i < pg->nr_pgpaths; i++) {
706                 struct pgpath *pgpath;
707                 struct arg_set path_args;
708
709                 if (as->argc < nr_params) {
710                         ti->error = "not enough path parameters";
711                         r = -EINVAL;
712                         goto bad;
713                 }
714
715                 path_args.argc = nr_params;
716                 path_args.argv = as->argv;
717
718                 pgpath = parse_path(&path_args, &pg->ps, ti);
719                 if (IS_ERR(pgpath)) {
720                         r = PTR_ERR(pgpath);
721                         goto bad;
722                 }
723
724                 pgpath->pg = pg;
725                 list_add_tail(&pgpath->list, &pg->pgpaths);
726                 consume(as, nr_params);
727         }
728
729         return pg;
730
731  bad:
732         free_priority_group(pg, ti);
733         return ERR_PTR(r);
734 }
735
736 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
737 {
738         unsigned hw_argc;
739         int ret;
740         struct dm_target *ti = m->ti;
741
742         static struct param _params[] = {
743                 {0, 1024, "invalid number of hardware handler args"},
744         };
745
746         if (read_param(_params, shift(as), &hw_argc, &ti->error))
747                 return -EINVAL;
748
749         if (!hw_argc)
750                 return 0;
751
752         if (hw_argc > as->argc) {
753                 ti->error = "not enough arguments for hardware handler";
754                 return -EINVAL;
755         }
756
757         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
758         request_module("scsi_dh_%s", m->hw_handler_name);
759         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
760                 ti->error = "unknown hardware handler type";
761                 ret = -EINVAL;
762                 goto fail;
763         }
764
765         if (hw_argc > 1) {
766                 char *p;
767                 int i, j, len = 4;
768
769                 for (i = 0; i <= hw_argc - 2; i++)
770                         len += strlen(as->argv[i]) + 1;
771                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
772                 if (!p) {
773                         ti->error = "memory allocation failed";
774                         ret = -ENOMEM;
775                         goto fail;
776                 }
777                 j = sprintf(p, "%d", hw_argc - 1);
778                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
779                         j = sprintf(p, "%s", as->argv[i]);
780         }
781         consume(as, hw_argc - 1);
782
783         return 0;
784 fail:
785         kfree(m->hw_handler_name);
786         m->hw_handler_name = NULL;
787         return ret;
788 }
789
790 static int parse_features(struct arg_set *as, struct multipath *m)
791 {
792         int r;
793         unsigned argc;
794         struct dm_target *ti = m->ti;
795         const char *param_name;
796
797         static struct param _params[] = {
798                 {0, 5, "invalid number of feature args"},
799                 {1, 50, "pg_init_retries must be between 1 and 50"},
800                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
801         };
802
803         r = read_param(_params, shift(as), &argc, &ti->error);
804         if (r)
805                 return -EINVAL;
806
807         if (!argc)
808                 return 0;
809
810         do {
811                 param_name = shift(as);
812                 argc--;
813
814                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
815                         r = queue_if_no_path(m, 1, 0);
816                         continue;
817                 }
818
819                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
820                     (argc >= 1)) {
821                         r = read_param(_params + 1, shift(as),
822                                        &m->pg_init_retries, &ti->error);
823                         argc--;
824                         continue;
825                 }
826
827                 if (!strnicmp(param_name, MESG_STR("pg_init_delay_msecs")) &&
828                     (argc >= 1)) {
829                         r = read_param(_params + 2, shift(as),
830                                        &m->pg_init_delay_msecs, &ti->error);
831                         argc--;
832                         continue;
833                 }
834
835                 ti->error = "Unrecognised multipath feature request";
836                 r = -EINVAL;
837         } while (argc && !r);
838
839         return r;
840 }
841
842 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
843                          char **argv)
844 {
845         /* target parameters */
846         static struct param _params[] = {
847                 {1, 1024, "invalid number of priority groups"},
848                 {1, 1024, "invalid initial priority group number"},
849         };
850
851         int r;
852         struct multipath *m;
853         struct arg_set as;
854         unsigned pg_count = 0;
855         unsigned next_pg_num;
856
857         as.argc = argc;
858         as.argv = argv;
859
860         m = alloc_multipath(ti);
861         if (!m) {
862                 ti->error = "can't allocate multipath";
863                 return -EINVAL;
864         }
865
866         r = parse_features(&as, m);
867         if (r)
868                 goto bad;
869
870         r = parse_hw_handler(&as, m);
871         if (r)
872                 goto bad;
873
874         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
875         if (r)
876                 goto bad;
877
878         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
879         if (r)
880                 goto bad;
881
882         /* parse the priority groups */
883         while (as.argc) {
884                 struct priority_group *pg;
885
886                 pg = parse_priority_group(&as, m);
887                 if (IS_ERR(pg)) {
888                         r = PTR_ERR(pg);
889                         goto bad;
890                 }
891
892                 m->nr_valid_paths += pg->nr_pgpaths;
893                 list_add_tail(&pg->list, &m->priority_groups);
894                 pg_count++;
895                 pg->pg_num = pg_count;
896                 if (!--next_pg_num)
897                         m->next_pg = pg;
898         }
899
900         if (pg_count != m->nr_priority_groups) {
901                 ti->error = "priority group count mismatch";
902                 r = -EINVAL;
903                 goto bad;
904         }
905
906         ti->num_flush_requests = 1;
907         ti->num_discard_requests = 1;
908
909         return 0;
910
911  bad:
912         free_multipath(m);
913         return r;
914 }
915
916 static void multipath_wait_for_pg_init_completion(struct multipath *m)
917 {
918         DECLARE_WAITQUEUE(wait, current);
919         unsigned long flags;
920
921         add_wait_queue(&m->pg_init_wait, &wait);
922
923         while (1) {
924                 set_current_state(TASK_UNINTERRUPTIBLE);
925
926                 spin_lock_irqsave(&m->lock, flags);
927                 if (!m->pg_init_in_progress) {
928                         spin_unlock_irqrestore(&m->lock, flags);
929                         break;
930                 }
931                 spin_unlock_irqrestore(&m->lock, flags);
932
933                 io_schedule();
934         }
935         set_current_state(TASK_RUNNING);
936
937         remove_wait_queue(&m->pg_init_wait, &wait);
938 }
939
940 static void flush_multipath_work(struct multipath *m)
941 {
942         flush_workqueue(kmpath_handlerd);
943         multipath_wait_for_pg_init_completion(m);
944         flush_workqueue(kmultipathd);
945         flush_work_sync(&m->trigger_event);
946 }
947
948 static void multipath_dtr(struct dm_target *ti)
949 {
950         struct multipath *m = ti->private;
951
952         flush_multipath_work(m);
953         free_multipath(m);
954 }
955
956 /*
957  * Map cloned requests
958  */
959 static int multipath_map(struct dm_target *ti, struct request *clone,
960                          union map_info *map_context)
961 {
962         int r;
963         struct dm_mpath_io *mpio;
964         struct multipath *m = (struct multipath *) ti->private;
965
966         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
967         if (!mpio)
968                 /* ENOMEM, requeue */
969                 return DM_MAPIO_REQUEUE;
970         memset(mpio, 0, sizeof(*mpio));
971
972         map_context->ptr = mpio;
973         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
974         r = map_io(m, clone, mpio, 0);
975         if (r < 0 || r == DM_MAPIO_REQUEUE)
976                 mempool_free(mpio, m->mpio_pool);
977
978         return r;
979 }
980
981 /*
982  * Take a path out of use.
983  */
984 static int fail_path(struct pgpath *pgpath)
985 {
986         unsigned long flags;
987         struct multipath *m = pgpath->pg->m;
988
989         spin_lock_irqsave(&m->lock, flags);
990
991         if (!pgpath->is_active)
992                 goto out;
993
994         DMWARN("Failing path %s.", pgpath->path.dev->name);
995
996         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
997         pgpath->is_active = 0;
998         pgpath->fail_count++;
999
1000         m->nr_valid_paths--;
1001
1002         if (pgpath == m->current_pgpath)
1003                 m->current_pgpath = NULL;
1004
1005         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1006                       pgpath->path.dev->name, m->nr_valid_paths);
1007
1008         schedule_work(&m->trigger_event);
1009
1010 out:
1011         spin_unlock_irqrestore(&m->lock, flags);
1012
1013         return 0;
1014 }
1015
1016 /*
1017  * Reinstate a previously-failed path
1018  */
1019 static int reinstate_path(struct pgpath *pgpath)
1020 {
1021         int r = 0;
1022         unsigned long flags;
1023         struct multipath *m = pgpath->pg->m;
1024
1025         spin_lock_irqsave(&m->lock, flags);
1026
1027         if (pgpath->is_active)
1028                 goto out;
1029
1030         if (!pgpath->pg->ps.type->reinstate_path) {
1031                 DMWARN("Reinstate path not supported by path selector %s",
1032                        pgpath->pg->ps.type->name);
1033                 r = -EINVAL;
1034                 goto out;
1035         }
1036
1037         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1038         if (r)
1039                 goto out;
1040
1041         pgpath->is_active = 1;
1042
1043         if (!m->nr_valid_paths++ && m->queue_size) {
1044                 m->current_pgpath = NULL;
1045                 queue_work(kmultipathd, &m->process_queued_ios);
1046         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1047                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1048                         m->pg_init_in_progress++;
1049         }
1050
1051         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1052                       pgpath->path.dev->name, m->nr_valid_paths);
1053
1054         schedule_work(&m->trigger_event);
1055
1056 out:
1057         spin_unlock_irqrestore(&m->lock, flags);
1058
1059         return r;
1060 }
1061
1062 /*
1063  * Fail or reinstate all paths that match the provided struct dm_dev.
1064  */
1065 static int action_dev(struct multipath *m, struct dm_dev *dev,
1066                       action_fn action)
1067 {
1068         int r = -EINVAL;
1069         struct pgpath *pgpath;
1070         struct priority_group *pg;
1071
1072         list_for_each_entry(pg, &m->priority_groups, list) {
1073                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1074                         if (pgpath->path.dev == dev)
1075                                 r = action(pgpath);
1076                 }
1077         }
1078
1079         return r;
1080 }
1081
1082 /*
1083  * Temporarily try to avoid having to use the specified PG
1084  */
1085 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1086                       int bypassed)
1087 {
1088         unsigned long flags;
1089
1090         spin_lock_irqsave(&m->lock, flags);
1091
1092         pg->bypassed = bypassed;
1093         m->current_pgpath = NULL;
1094         m->current_pg = NULL;
1095
1096         spin_unlock_irqrestore(&m->lock, flags);
1097
1098         schedule_work(&m->trigger_event);
1099 }
1100
1101 /*
1102  * Switch to using the specified PG from the next I/O that gets mapped
1103  */
1104 static int switch_pg_num(struct multipath *m, const char *pgstr)
1105 {
1106         struct priority_group *pg;
1107         unsigned pgnum;
1108         unsigned long flags;
1109
1110         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1111             (pgnum > m->nr_priority_groups)) {
1112                 DMWARN("invalid PG number supplied to switch_pg_num");
1113                 return -EINVAL;
1114         }
1115
1116         spin_lock_irqsave(&m->lock, flags);
1117         list_for_each_entry(pg, &m->priority_groups, list) {
1118                 pg->bypassed = 0;
1119                 if (--pgnum)
1120                         continue;
1121
1122                 m->current_pgpath = NULL;
1123                 m->current_pg = NULL;
1124                 m->next_pg = pg;
1125         }
1126         spin_unlock_irqrestore(&m->lock, flags);
1127
1128         schedule_work(&m->trigger_event);
1129         return 0;
1130 }
1131
1132 /*
1133  * Set/clear bypassed status of a PG.
1134  * PGs are numbered upwards from 1 in the order they were declared.
1135  */
1136 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1137 {
1138         struct priority_group *pg;
1139         unsigned pgnum;
1140
1141         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1142             (pgnum > m->nr_priority_groups)) {
1143                 DMWARN("invalid PG number supplied to bypass_pg");
1144                 return -EINVAL;
1145         }
1146
1147         list_for_each_entry(pg, &m->priority_groups, list) {
1148                 if (!--pgnum)
1149                         break;
1150         }
1151
1152         bypass_pg(m, pg, bypassed);
1153         return 0;
1154 }
1155
1156 /*
1157  * Should we retry pg_init immediately?
1158  */
1159 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1160 {
1161         unsigned long flags;
1162         int limit_reached = 0;
1163
1164         spin_lock_irqsave(&m->lock, flags);
1165
1166         if (m->pg_init_count <= m->pg_init_retries)
1167                 m->pg_init_required = 1;
1168         else
1169                 limit_reached = 1;
1170
1171         spin_unlock_irqrestore(&m->lock, flags);
1172
1173         return limit_reached;
1174 }
1175
1176 static void pg_init_done(void *data, int errors)
1177 {
1178         struct pgpath *pgpath = data;
1179         struct priority_group *pg = pgpath->pg;
1180         struct multipath *m = pg->m;
1181         unsigned long flags;
1182         unsigned delay_retry = 0;
1183
1184         /* device or driver problems */
1185         switch (errors) {
1186         case SCSI_DH_OK:
1187                 break;
1188         case SCSI_DH_NOSYS:
1189                 if (!m->hw_handler_name) {
1190                         errors = 0;
1191                         break;
1192                 }
1193                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1194                       "Error %d.", m->hw_handler_name, errors);
1195                 /*
1196                  * Fail path for now, so we do not ping pong
1197                  */
1198                 fail_path(pgpath);
1199                 break;
1200         case SCSI_DH_DEV_TEMP_BUSY:
1201                 /*
1202                  * Probably doing something like FW upgrade on the
1203                  * controller so try the other pg.
1204                  */
1205                 bypass_pg(m, pg, 1);
1206                 break;
1207         case SCSI_DH_RETRY:
1208                 /* Wait before retrying. */
1209                 delay_retry = 1;
1210         case SCSI_DH_IMM_RETRY:
1211         case SCSI_DH_RES_TEMP_UNAVAIL:
1212                 if (pg_init_limit_reached(m, pgpath))
1213                         fail_path(pgpath);
1214                 errors = 0;
1215                 break;
1216         default:
1217                 /*
1218                  * We probably do not want to fail the path for a device
1219                  * error, but this is what the old dm did. In future
1220                  * patches we can do more advanced handling.
1221                  */
1222                 fail_path(pgpath);
1223         }
1224
1225         spin_lock_irqsave(&m->lock, flags);
1226         if (errors) {
1227                 if (pgpath == m->current_pgpath) {
1228                         DMERR("Could not failover device. Error %d.", errors);
1229                         m->current_pgpath = NULL;
1230                         m->current_pg = NULL;
1231                 }
1232         } else if (!m->pg_init_required)
1233                 pg->bypassed = 0;
1234
1235         if (--m->pg_init_in_progress)
1236                 /* Activations of other paths are still on going */
1237                 goto out;
1238
1239         if (!m->pg_init_required)
1240                 m->queue_io = 0;
1241
1242         m->pg_init_delay_retry = delay_retry;
1243         queue_work(kmultipathd, &m->process_queued_ios);
1244
1245         /*
1246          * Wake up any thread waiting to suspend.
1247          */
1248         wake_up(&m->pg_init_wait);
1249
1250 out:
1251         spin_unlock_irqrestore(&m->lock, flags);
1252 }
1253
1254 static void activate_path(struct work_struct *work)
1255 {
1256         struct pgpath *pgpath =
1257                 container_of(work, struct pgpath, activate_path.work);
1258
1259         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1260                                 pg_init_done, pgpath);
1261 }
1262
1263 /*
1264  * end_io handling
1265  */
1266 static int do_end_io(struct multipath *m, struct request *clone,
1267                      int error, struct dm_mpath_io *mpio)
1268 {
1269         /*
1270          * We don't queue any clone request inside the multipath target
1271          * during end I/O handling, since those clone requests don't have
1272          * bio clones.  If we queue them inside the multipath target,
1273          * we need to make bio clones, that requires memory allocation.
1274          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1275          *  don't have bio clones.)
1276          * Instead of queueing the clone request here, we queue the original
1277          * request into dm core, which will remake a clone request and
1278          * clone bios for it and resubmit it later.
1279          */
1280         int r = DM_ENDIO_REQUEUE;
1281         unsigned long flags;
1282
1283         if (!error && !clone->errors)
1284                 return 0;       /* I/O complete */
1285
1286         if (error == -EOPNOTSUPP || error == -EREMOTEIO)
1287                 return error;
1288
1289         if (mpio->pgpath)
1290                 fail_path(mpio->pgpath);
1291
1292         spin_lock_irqsave(&m->lock, flags);
1293         if (!m->nr_valid_paths) {
1294                 if (!m->queue_if_no_path) {
1295                         if (!__must_push_back(m))
1296                                 r = -EIO;
1297                 } else {
1298                         if (error == -EBADE)
1299                                 r = error;
1300                 }
1301         }
1302         spin_unlock_irqrestore(&m->lock, flags);
1303
1304         return r;
1305 }
1306
1307 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1308                             int error, union map_info *map_context)
1309 {
1310         struct multipath *m = ti->private;
1311         struct dm_mpath_io *mpio = map_context->ptr;
1312         struct pgpath *pgpath = mpio->pgpath;
1313         struct path_selector *ps;
1314         int r;
1315
1316         r  = do_end_io(m, clone, error, mpio);
1317         if (pgpath) {
1318                 ps = &pgpath->pg->ps;
1319                 if (ps->type->end_io)
1320                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1321         }
1322         mempool_free(mpio, m->mpio_pool);
1323
1324         return r;
1325 }
1326
1327 /*
1328  * Suspend can't complete until all the I/O is processed so if
1329  * the last path fails we must error any remaining I/O.
1330  * Note that if the freeze_bdev fails while suspending, the
1331  * queue_if_no_path state is lost - userspace should reset it.
1332  */
1333 static void multipath_presuspend(struct dm_target *ti)
1334 {
1335         struct multipath *m = (struct multipath *) ti->private;
1336
1337         queue_if_no_path(m, 0, 1);
1338 }
1339
1340 static void multipath_postsuspend(struct dm_target *ti)
1341 {
1342         struct multipath *m = ti->private;
1343
1344         mutex_lock(&m->work_mutex);
1345         flush_multipath_work(m);
1346         mutex_unlock(&m->work_mutex);
1347 }
1348
1349 /*
1350  * Restore the queue_if_no_path setting.
1351  */
1352 static void multipath_resume(struct dm_target *ti)
1353 {
1354         struct multipath *m = (struct multipath *) ti->private;
1355         unsigned long flags;
1356
1357         spin_lock_irqsave(&m->lock, flags);
1358         m->queue_if_no_path = m->saved_queue_if_no_path;
1359         spin_unlock_irqrestore(&m->lock, flags);
1360 }
1361
1362 /*
1363  * Info output has the following format:
1364  * num_multipath_feature_args [multipath_feature_args]*
1365  * num_handler_status_args [handler_status_args]*
1366  * num_groups init_group_number
1367  *            [A|D|E num_ps_status_args [ps_status_args]*
1368  *             num_paths num_selector_args
1369  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1370  *
1371  * Table output has the following format (identical to the constructor string):
1372  * num_feature_args [features_args]*
1373  * num_handler_args hw_handler [hw_handler_args]*
1374  * num_groups init_group_number
1375  *     [priority selector-name num_ps_args [ps_args]*
1376  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1377  */
1378 static int multipath_status(struct dm_target *ti, status_type_t type,
1379                             char *result, unsigned int maxlen)
1380 {
1381         int sz = 0;
1382         unsigned long flags;
1383         struct multipath *m = (struct multipath *) ti->private;
1384         struct priority_group *pg;
1385         struct pgpath *p;
1386         unsigned pg_num;
1387         char state;
1388
1389         spin_lock_irqsave(&m->lock, flags);
1390
1391         /* Features */
1392         if (type == STATUSTYPE_INFO)
1393                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1394         else {
1395                 DMEMIT("%u ", m->queue_if_no_path +
1396                               (m->pg_init_retries > 0) * 2 +
1397                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1398                 if (m->queue_if_no_path)
1399                         DMEMIT("queue_if_no_path ");
1400                 if (m->pg_init_retries)
1401                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1402                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1403                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1404         }
1405
1406         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1407                 DMEMIT("0 ");
1408         else
1409                 DMEMIT("1 %s ", m->hw_handler_name);
1410
1411         DMEMIT("%u ", m->nr_priority_groups);
1412
1413         if (m->next_pg)
1414                 pg_num = m->next_pg->pg_num;
1415         else if (m->current_pg)
1416                 pg_num = m->current_pg->pg_num;
1417         else
1418                         pg_num = 1;
1419
1420         DMEMIT("%u ", pg_num);
1421
1422         switch (type) {
1423         case STATUSTYPE_INFO:
1424                 list_for_each_entry(pg, &m->priority_groups, list) {
1425                         if (pg->bypassed)
1426                                 state = 'D';    /* Disabled */
1427                         else if (pg == m->current_pg)
1428                                 state = 'A';    /* Currently Active */
1429                         else
1430                                 state = 'E';    /* Enabled */
1431
1432                         DMEMIT("%c ", state);
1433
1434                         if (pg->ps.type->status)
1435                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1436                                                           result + sz,
1437                                                           maxlen - sz);
1438                         else
1439                                 DMEMIT("0 ");
1440
1441                         DMEMIT("%u %u ", pg->nr_pgpaths,
1442                                pg->ps.type->info_args);
1443
1444                         list_for_each_entry(p, &pg->pgpaths, list) {
1445                                 DMEMIT("%s %s %u ", p->path.dev->name,
1446                                        p->is_active ? "A" : "F",
1447                                        p->fail_count);
1448                                 if (pg->ps.type->status)
1449                                         sz += pg->ps.type->status(&pg->ps,
1450                                               &p->path, type, result + sz,
1451                                               maxlen - sz);
1452                         }
1453                 }
1454                 break;
1455
1456         case STATUSTYPE_TABLE:
1457                 list_for_each_entry(pg, &m->priority_groups, list) {
1458                         DMEMIT("%s ", pg->ps.type->name);
1459
1460                         if (pg->ps.type->status)
1461                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1462                                                           result + sz,
1463                                                           maxlen - sz);
1464                         else
1465                                 DMEMIT("0 ");
1466
1467                         DMEMIT("%u %u ", pg->nr_pgpaths,
1468                                pg->ps.type->table_args);
1469
1470                         list_for_each_entry(p, &pg->pgpaths, list) {
1471                                 DMEMIT("%s ", p->path.dev->name);
1472                                 if (pg->ps.type->status)
1473                                         sz += pg->ps.type->status(&pg->ps,
1474                                               &p->path, type, result + sz,
1475                                               maxlen - sz);
1476                         }
1477                 }
1478                 break;
1479         }
1480
1481         spin_unlock_irqrestore(&m->lock, flags);
1482
1483         return 0;
1484 }
1485
1486 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1487 {
1488         int r = -EINVAL;
1489         struct dm_dev *dev;
1490         struct multipath *m = (struct multipath *) ti->private;
1491         action_fn action;
1492
1493         mutex_lock(&m->work_mutex);
1494
1495         if (dm_suspended(ti)) {
1496                 r = -EBUSY;
1497                 goto out;
1498         }
1499
1500         if (argc == 1) {
1501                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1502                         r = queue_if_no_path(m, 1, 0);
1503                         goto out;
1504                 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1505                         r = queue_if_no_path(m, 0, 0);
1506                         goto out;
1507                 }
1508         }
1509
1510         if (argc != 2) {
1511                 DMWARN("Unrecognised multipath message received.");
1512                 goto out;
1513         }
1514
1515         if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1516                 r = bypass_pg_num(m, argv[1], 1);
1517                 goto out;
1518         } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1519                 r = bypass_pg_num(m, argv[1], 0);
1520                 goto out;
1521         } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1522                 r = switch_pg_num(m, argv[1]);
1523                 goto out;
1524         } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1525                 action = reinstate_path;
1526         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1527                 action = fail_path;
1528         else {
1529                 DMWARN("Unrecognised multipath message received.");
1530                 goto out;
1531         }
1532
1533         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1534         if (r) {
1535                 DMWARN("message: error getting device %s",
1536                        argv[1]);
1537                 goto out;
1538         }
1539
1540         r = action_dev(m, dev, action);
1541
1542         dm_put_device(ti, dev);
1543
1544 out:
1545         mutex_unlock(&m->work_mutex);
1546         return r;
1547 }
1548
1549 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1550                            unsigned long arg)
1551 {
1552         struct multipath *m = (struct multipath *) ti->private;
1553         struct block_device *bdev = NULL;
1554         fmode_t mode = 0;
1555         unsigned long flags;
1556         int r = 0;
1557
1558         spin_lock_irqsave(&m->lock, flags);
1559
1560         if (!m->current_pgpath)
1561                 __choose_pgpath(m, 0);
1562
1563         if (m->current_pgpath) {
1564                 bdev = m->current_pgpath->path.dev->bdev;
1565                 mode = m->current_pgpath->path.dev->mode;
1566         }
1567
1568         if (m->queue_io)
1569                 r = -EAGAIN;
1570         else if (!bdev)
1571                 r = -EIO;
1572
1573         spin_unlock_irqrestore(&m->lock, flags);
1574
1575         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1576 }
1577
1578 static int multipath_iterate_devices(struct dm_target *ti,
1579                                      iterate_devices_callout_fn fn, void *data)
1580 {
1581         struct multipath *m = ti->private;
1582         struct priority_group *pg;
1583         struct pgpath *p;
1584         int ret = 0;
1585
1586         list_for_each_entry(pg, &m->priority_groups, list) {
1587                 list_for_each_entry(p, &pg->pgpaths, list) {
1588                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1589                         if (ret)
1590                                 goto out;
1591                 }
1592         }
1593
1594 out:
1595         return ret;
1596 }
1597
1598 static int __pgpath_busy(struct pgpath *pgpath)
1599 {
1600         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1601
1602         return dm_underlying_device_busy(q);
1603 }
1604
1605 /*
1606  * We return "busy", only when we can map I/Os but underlying devices
1607  * are busy (so even if we map I/Os now, the I/Os will wait on
1608  * the underlying queue).
1609  * In other words, if we want to kill I/Os or queue them inside us
1610  * due to map unavailability, we don't return "busy".  Otherwise,
1611  * dm core won't give us the I/Os and we can't do what we want.
1612  */
1613 static int multipath_busy(struct dm_target *ti)
1614 {
1615         int busy = 0, has_active = 0;
1616         struct multipath *m = ti->private;
1617         struct priority_group *pg;
1618         struct pgpath *pgpath;
1619         unsigned long flags;
1620
1621         spin_lock_irqsave(&m->lock, flags);
1622
1623         /* Guess which priority_group will be used at next mapping time */
1624         if (unlikely(!m->current_pgpath && m->next_pg))
1625                 pg = m->next_pg;
1626         else if (likely(m->current_pg))
1627                 pg = m->current_pg;
1628         else
1629                 /*
1630                  * We don't know which pg will be used at next mapping time.
1631                  * We don't call __choose_pgpath() here to avoid to trigger
1632                  * pg_init just by busy checking.
1633                  * So we don't know whether underlying devices we will be using
1634                  * at next mapping time are busy or not. Just try mapping.
1635                  */
1636                 goto out;
1637
1638         /*
1639          * If there is one non-busy active path at least, the path selector
1640          * will be able to select it. So we consider such a pg as not busy.
1641          */
1642         busy = 1;
1643         list_for_each_entry(pgpath, &pg->pgpaths, list)
1644                 if (pgpath->is_active) {
1645                         has_active = 1;
1646
1647                         if (!__pgpath_busy(pgpath)) {
1648                                 busy = 0;
1649                                 break;
1650                         }
1651                 }
1652
1653         if (!has_active)
1654                 /*
1655                  * No active path in this pg, so this pg won't be used and
1656                  * the current_pg will be changed at next mapping time.
1657                  * We need to try mapping to determine it.
1658                  */
1659                 busy = 0;
1660
1661 out:
1662         spin_unlock_irqrestore(&m->lock, flags);
1663
1664         return busy;
1665 }
1666
1667 /*-----------------------------------------------------------------
1668  * Module setup
1669  *---------------------------------------------------------------*/
1670 static struct target_type multipath_target = {
1671         .name = "multipath",
1672         .version = {1, 3, 0},
1673         .module = THIS_MODULE,
1674         .ctr = multipath_ctr,
1675         .dtr = multipath_dtr,
1676         .map_rq = multipath_map,
1677         .rq_end_io = multipath_end_io,
1678         .presuspend = multipath_presuspend,
1679         .postsuspend = multipath_postsuspend,
1680         .resume = multipath_resume,
1681         .status = multipath_status,
1682         .message = multipath_message,
1683         .ioctl  = multipath_ioctl,
1684         .iterate_devices = multipath_iterate_devices,
1685         .busy = multipath_busy,
1686 };
1687
1688 static int __init dm_multipath_init(void)
1689 {
1690         int r;
1691
1692         /* allocate a slab for the dm_ios */
1693         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1694         if (!_mpio_cache)
1695                 return -ENOMEM;
1696
1697         r = dm_register_target(&multipath_target);
1698         if (r < 0) {
1699                 DMERR("register failed %d", r);
1700                 kmem_cache_destroy(_mpio_cache);
1701                 return -EINVAL;
1702         }
1703
1704         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1705         if (!kmultipathd) {
1706                 DMERR("failed to create workqueue kmpathd");
1707                 dm_unregister_target(&multipath_target);
1708                 kmem_cache_destroy(_mpio_cache);
1709                 return -ENOMEM;
1710         }
1711
1712         /*
1713          * A separate workqueue is used to handle the device handlers
1714          * to avoid overloading existing workqueue. Overloading the
1715          * old workqueue would also create a bottleneck in the
1716          * path of the storage hardware device activation.
1717          */
1718         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1719                                                   WQ_MEM_RECLAIM);
1720         if (!kmpath_handlerd) {
1721                 DMERR("failed to create workqueue kmpath_handlerd");
1722                 destroy_workqueue(kmultipathd);
1723                 dm_unregister_target(&multipath_target);
1724                 kmem_cache_destroy(_mpio_cache);
1725                 return -ENOMEM;
1726         }
1727
1728         DMINFO("version %u.%u.%u loaded",
1729                multipath_target.version[0], multipath_target.version[1],
1730                multipath_target.version[2]);
1731
1732         return r;
1733 }
1734
1735 static void __exit dm_multipath_exit(void)
1736 {
1737         destroy_workqueue(kmpath_handlerd);
1738         destroy_workqueue(kmultipathd);
1739
1740         dm_unregister_target(&multipath_target);
1741         kmem_cache_destroy(_mpio_cache);
1742 }
1743
1744 module_init(dm_multipath_init);
1745 module_exit(dm_multipath_exit);
1746
1747 MODULE_DESCRIPTION(DM_NAME " multipath target");
1748 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1749 MODULE_LICENSE("GPL");