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