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