d778563a4ffd2ca829c42e3e480ed89cb908541e
[linux-3.10.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 <linux/delay.h>
22 #include <scsi/scsi_dh.h>
23 #include <linux/atomic.h>
24
25 #define DM_MSG_PREFIX "multipath"
26 #define DM_PG_INIT_DELAY_MSECS 2000
27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
28
29 /* Path properties */
30 struct pgpath {
31         struct list_head list;
32
33         struct priority_group *pg;      /* Owning PG */
34         unsigned is_active;             /* Path status */
35         unsigned fail_count;            /* Cumulative failure count */
36
37         struct dm_path path;
38         struct delayed_work activate_path;
39 };
40
41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
42
43 /*
44  * Paths are grouped into Priority Groups and numbered from 1 upwards.
45  * Each has a path selector which controls which path gets used.
46  */
47 struct priority_group {
48         struct list_head list;
49
50         struct multipath *m;            /* Owning multipath instance */
51         struct path_selector ps;
52
53         unsigned pg_num;                /* Reference number */
54         unsigned bypassed;              /* Temporarily bypass this PG? */
55
56         unsigned nr_pgpaths;            /* Number of paths in PG */
57         struct list_head pgpaths;
58 };
59
60 /* Multipath context */
61 struct multipath {
62         struct list_head list;
63         struct dm_target *ti;
64
65         const char *hw_handler_name;
66         char *hw_handler_params;
67
68         spinlock_t lock;
69
70         unsigned nr_priority_groups;
71         struct list_head priority_groups;
72
73         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
74
75         unsigned pg_init_required;      /* pg_init needs calling? */
76         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
77         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
78
79         unsigned nr_valid_paths;        /* Total number of usable paths */
80         struct pgpath *current_pgpath;
81         struct priority_group *current_pg;
82         struct priority_group *next_pg; /* Switch to this PG if set */
83         unsigned repeat_count;          /* I/Os left before calling PS again */
84
85         unsigned queue_io:1;            /* Must we queue all I/O? */
86         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
87         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
88         unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
89
90         unsigned pg_init_retries;       /* Number of times to retry pg_init */
91         unsigned pg_init_count;         /* Number of times pg_init called */
92         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
93
94         unsigned queue_size;
95         struct work_struct process_queued_ios;
96         struct list_head queued_ios;
97
98         struct work_struct trigger_event;
99
100         /*
101          * We must use a mempool of dm_mpath_io structs so that we
102          * can resubmit bios on error.
103          */
104         mempool_t *mpio_pool;
105
106         struct mutex work_mutex;
107 };
108
109 /*
110  * Context information attached to each bio we process.
111  */
112 struct dm_mpath_io {
113         struct pgpath *pgpath;
114         size_t nr_bytes;
115 };
116
117 typedef int (*action_fn) (struct pgpath *pgpath);
118
119 #define MIN_IOS 256     /* Mempool size */
120
121 static struct kmem_cache *_mpio_cache;
122
123 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
124 static void process_queued_ios(struct work_struct *work);
125 static void trigger_event(struct work_struct *work);
126 static void activate_path(struct work_struct *work);
127
128
129 /*-----------------------------------------------
130  * Allocation routines
131  *-----------------------------------------------*/
132
133 static struct pgpath *alloc_pgpath(void)
134 {
135         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
136
137         if (pgpath) {
138                 pgpath->is_active = 1;
139                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
140         }
141
142         return pgpath;
143 }
144
145 static void free_pgpath(struct pgpath *pgpath)
146 {
147         kfree(pgpath);
148 }
149
150 static struct priority_group *alloc_priority_group(void)
151 {
152         struct priority_group *pg;
153
154         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
155
156         if (pg)
157                 INIT_LIST_HEAD(&pg->pgpaths);
158
159         return pg;
160 }
161
162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
163 {
164         struct pgpath *pgpath, *tmp;
165         struct multipath *m = ti->private;
166
167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168                 list_del(&pgpath->list);
169                 if (m->hw_handler_name)
170                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
171                 dm_put_device(ti, pgpath->path.dev);
172                 free_pgpath(pgpath);
173         }
174 }
175
176 static void free_priority_group(struct priority_group *pg,
177                                 struct dm_target *ti)
178 {
179         struct path_selector *ps = &pg->ps;
180
181         if (ps->type) {
182                 ps->type->destroy(ps);
183                 dm_put_path_selector(ps->type);
184         }
185
186         free_pgpaths(&pg->pgpaths, ti);
187         kfree(pg);
188 }
189
190 static struct multipath *alloc_multipath(struct dm_target *ti)
191 {
192         struct multipath *m;
193
194         m = kzalloc(sizeof(*m), GFP_KERNEL);
195         if (m) {
196                 INIT_LIST_HEAD(&m->priority_groups);
197                 INIT_LIST_HEAD(&m->queued_ios);
198                 spin_lock_init(&m->lock);
199                 m->queue_io = 1;
200                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
201                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
202                 INIT_WORK(&m->trigger_event, trigger_event);
203                 init_waitqueue_head(&m->pg_init_wait);
204                 mutex_init(&m->work_mutex);
205                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
206                 if (!m->mpio_pool) {
207                         kfree(m);
208                         return NULL;
209                 }
210                 m->ti = ti;
211                 ti->private = m;
212         }
213
214         return m;
215 }
216
217 static void free_multipath(struct multipath *m)
218 {
219         struct priority_group *pg, *tmp;
220
221         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
222                 list_del(&pg->list);
223                 free_priority_group(pg, m->ti);
224         }
225
226         kfree(m->hw_handler_name);
227         kfree(m->hw_handler_params);
228         mempool_destroy(m->mpio_pool);
229         kfree(m);
230 }
231
232 static int set_mapinfo(struct multipath *m, union map_info *info)
233 {
234         struct dm_mpath_io *mpio;
235
236         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
237         if (!mpio)
238                 return -ENOMEM;
239
240         memset(mpio, 0, sizeof(*mpio));
241         info->ptr = mpio;
242
243         return 0;
244 }
245
246 static void clear_mapinfo(struct multipath *m, union map_info *info)
247 {
248         struct dm_mpath_io *mpio = info->ptr;
249
250         info->ptr = NULL;
251         mempool_free(mpio, m->mpio_pool);
252 }
253
254 /*-----------------------------------------------
255  * Path selection
256  *-----------------------------------------------*/
257
258 static void __pg_init_all_paths(struct multipath *m)
259 {
260         struct pgpath *pgpath;
261         unsigned long pg_init_delay = 0;
262
263         m->pg_init_count++;
264         m->pg_init_required = 0;
265         if (m->pg_init_delay_retry)
266                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
267                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
268         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
269                 /* Skip failed paths */
270                 if (!pgpath->is_active)
271                         continue;
272                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
273                                        pg_init_delay))
274                         m->pg_init_in_progress++;
275         }
276 }
277
278 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
279 {
280         m->current_pg = pgpath->pg;
281
282         /* Must we initialise the PG first, and queue I/O till it's ready? */
283         if (m->hw_handler_name) {
284                 m->pg_init_required = 1;
285                 m->queue_io = 1;
286         } else {
287                 m->pg_init_required = 0;
288                 m->queue_io = 0;
289         }
290
291         m->pg_init_count = 0;
292 }
293
294 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
295                                size_t nr_bytes)
296 {
297         struct dm_path *path;
298
299         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
300         if (!path)
301                 return -ENXIO;
302
303         m->current_pgpath = path_to_pgpath(path);
304
305         if (m->current_pg != pg)
306                 __switch_pg(m, m->current_pgpath);
307
308         return 0;
309 }
310
311 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
312 {
313         struct priority_group *pg;
314         unsigned bypassed = 1;
315
316         if (!m->nr_valid_paths)
317                 goto failed;
318
319         /* Were we instructed to switch PG? */
320         if (m->next_pg) {
321                 pg = m->next_pg;
322                 m->next_pg = NULL;
323                 if (!__choose_path_in_pg(m, pg, nr_bytes))
324                         return;
325         }
326
327         /* Don't change PG until it has no remaining paths */
328         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
329                 return;
330
331         /*
332          * Loop through priority groups until we find a valid path.
333          * First time we skip PGs marked 'bypassed'.
334          * Second time we only try the ones we skipped, but set
335          * pg_init_delay_retry so we do not hammer controllers.
336          */
337         do {
338                 list_for_each_entry(pg, &m->priority_groups, list) {
339                         if (pg->bypassed == bypassed)
340                                 continue;
341                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
342                                 if (!bypassed)
343                                         m->pg_init_delay_retry = 1;
344                                 return;
345                         }
346                 }
347         } while (bypassed--);
348
349 failed:
350         m->current_pgpath = NULL;
351         m->current_pg = NULL;
352 }
353
354 /*
355  * Check whether bios must be queued in the device-mapper core rather
356  * than here in the target.
357  *
358  * m->lock must be held on entry.
359  *
360  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
361  * same value then we are not between multipath_presuspend()
362  * and multipath_resume() calls and we have no need to check
363  * for the DMF_NOFLUSH_SUSPENDING flag.
364  */
365 static int __must_push_back(struct multipath *m)
366 {
367         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
368                 dm_noflush_suspending(m->ti));
369 }
370
371 static int map_io(struct multipath *m, struct request *clone,
372                   union map_info *map_context, unsigned was_queued)
373 {
374         int r = DM_MAPIO_REMAPPED;
375         size_t nr_bytes = blk_rq_bytes(clone);
376         unsigned long flags;
377         struct pgpath *pgpath;
378         struct block_device *bdev;
379         struct dm_mpath_io *mpio = map_context->ptr;
380
381         spin_lock_irqsave(&m->lock, flags);
382
383         /* Do we need to select a new pgpath? */
384         if (!m->current_pgpath ||
385             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
386                 __choose_pgpath(m, nr_bytes);
387
388         pgpath = m->current_pgpath;
389
390         if (was_queued)
391                 m->queue_size--;
392
393         if ((pgpath && m->queue_io) ||
394             (!pgpath && m->queue_if_no_path)) {
395                 /* Queue for the daemon to resubmit */
396                 list_add_tail(&clone->queuelist, &m->queued_ios);
397                 m->queue_size++;
398                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
399                     !m->queue_io)
400                         queue_work(kmultipathd, &m->process_queued_ios);
401                 pgpath = NULL;
402                 r = DM_MAPIO_SUBMITTED;
403         } else if (pgpath) {
404                 bdev = pgpath->path.dev->bdev;
405                 clone->q = bdev_get_queue(bdev);
406                 clone->rq_disk = bdev->bd_disk;
407         } else if (__must_push_back(m))
408                 r = DM_MAPIO_REQUEUE;
409         else
410                 r = -EIO;       /* Failed */
411
412         mpio->pgpath = pgpath;
413         mpio->nr_bytes = nr_bytes;
414
415         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
416                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
417                                               nr_bytes);
418
419         spin_unlock_irqrestore(&m->lock, flags);
420
421         return r;
422 }
423
424 /*
425  * If we run out of usable paths, should we queue I/O or error it?
426  */
427 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
428                             unsigned save_old_value)
429 {
430         unsigned long flags;
431
432         spin_lock_irqsave(&m->lock, flags);
433
434         if (save_old_value)
435                 m->saved_queue_if_no_path = m->queue_if_no_path;
436         else
437                 m->saved_queue_if_no_path = queue_if_no_path;
438         m->queue_if_no_path = queue_if_no_path;
439         if (!m->queue_if_no_path && m->queue_size)
440                 queue_work(kmultipathd, &m->process_queued_ios);
441
442         spin_unlock_irqrestore(&m->lock, flags);
443
444         return 0;
445 }
446
447 /*-----------------------------------------------------------------
448  * The multipath daemon is responsible for resubmitting queued ios.
449  *---------------------------------------------------------------*/
450
451 static void dispatch_queued_ios(struct multipath *m)
452 {
453         int r;
454         unsigned long flags;
455         union map_info *info;
456         struct request *clone, *n;
457         LIST_HEAD(cl);
458
459         spin_lock_irqsave(&m->lock, flags);
460         list_splice_init(&m->queued_ios, &cl);
461         spin_unlock_irqrestore(&m->lock, flags);
462
463         list_for_each_entry_safe(clone, n, &cl, queuelist) {
464                 list_del_init(&clone->queuelist);
465
466                 info = dm_get_rq_mapinfo(clone);
467
468                 r = map_io(m, clone, info, 1);
469                 if (r < 0) {
470                         clear_mapinfo(m, info);
471                         dm_kill_unmapped_request(clone, r);
472                 } else if (r == DM_MAPIO_REMAPPED)
473                         dm_dispatch_request(clone);
474                 else if (r == DM_MAPIO_REQUEUE) {
475                         clear_mapinfo(m, info);
476                         dm_requeue_unmapped_request(clone);
477                 }
478         }
479 }
480
481 static void process_queued_ios(struct work_struct *work)
482 {
483         struct multipath *m =
484                 container_of(work, struct multipath, process_queued_ios);
485         struct pgpath *pgpath = NULL;
486         unsigned must_queue = 1;
487         unsigned long flags;
488
489         spin_lock_irqsave(&m->lock, flags);
490
491         if (!m->current_pgpath)
492                 __choose_pgpath(m, 0);
493
494         pgpath = m->current_pgpath;
495
496         if ((pgpath && !m->queue_io) ||
497             (!pgpath && !m->queue_if_no_path))
498                 must_queue = 0;
499
500         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
501                 __pg_init_all_paths(m);
502
503         spin_unlock_irqrestore(&m->lock, flags);
504         if (!must_queue)
505                 dispatch_queued_ios(m);
506 }
507
508 /*
509  * An event is triggered whenever a path is taken out of use.
510  * Includes path failure and PG bypass.
511  */
512 static void trigger_event(struct work_struct *work)
513 {
514         struct multipath *m =
515                 container_of(work, struct multipath, trigger_event);
516
517         dm_table_event(m->ti->table);
518 }
519
520 /*-----------------------------------------------------------------
521  * Constructor/argument parsing:
522  * <#multipath feature args> [<arg>]*
523  * <#hw_handler args> [hw_handler [<arg>]*]
524  * <#priority groups>
525  * <initial priority group>
526  *     [<selector> <#selector args> [<arg>]*
527  *      <#paths> <#per-path selector args>
528  *         [<path> [<arg>]* ]+ ]+
529  *---------------------------------------------------------------*/
530 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
531                                struct dm_target *ti)
532 {
533         int r;
534         struct path_selector_type *pst;
535         unsigned ps_argc;
536
537         static struct dm_arg _args[] = {
538                 {0, 1024, "invalid number of path selector args"},
539         };
540
541         pst = dm_get_path_selector(dm_shift_arg(as));
542         if (!pst) {
543                 ti->error = "unknown path selector type";
544                 return -EINVAL;
545         }
546
547         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
548         if (r) {
549                 dm_put_path_selector(pst);
550                 return -EINVAL;
551         }
552
553         r = pst->create(&pg->ps, ps_argc, as->argv);
554         if (r) {
555                 dm_put_path_selector(pst);
556                 ti->error = "path selector constructor failed";
557                 return r;
558         }
559
560         pg->ps.type = pst;
561         dm_consume_args(as, ps_argc);
562
563         return 0;
564 }
565
566 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
567                                struct dm_target *ti)
568 {
569         int r;
570         struct pgpath *p;
571         struct multipath *m = ti->private;
572         struct request_queue *q = NULL;
573         const char *attached_handler_name;
574
575         /* we need at least a path arg */
576         if (as->argc < 1) {
577                 ti->error = "no device given";
578                 return ERR_PTR(-EINVAL);
579         }
580
581         p = alloc_pgpath();
582         if (!p)
583                 return ERR_PTR(-ENOMEM);
584
585         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
586                           &p->path.dev);
587         if (r) {
588                 ti->error = "error getting device";
589                 goto bad;
590         }
591
592         if (m->retain_attached_hw_handler || m->hw_handler_name)
593                 q = bdev_get_queue(p->path.dev->bdev);
594
595         if (m->retain_attached_hw_handler) {
596                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
597                 if (attached_handler_name) {
598                         /*
599                          * Reset hw_handler_name to match the attached handler
600                          * and clear any hw_handler_params associated with the
601                          * ignored handler.
602                          *
603                          * NB. This modifies the table line to show the actual
604                          * handler instead of the original table passed in.
605                          */
606                         kfree(m->hw_handler_name);
607                         m->hw_handler_name = attached_handler_name;
608
609                         kfree(m->hw_handler_params);
610                         m->hw_handler_params = NULL;
611                 }
612         }
613
614         if (m->hw_handler_name) {
615                 /*
616                  * Increments scsi_dh reference, even when using an
617                  * already-attached handler.
618                  */
619                 r = scsi_dh_attach(q, m->hw_handler_name);
620                 if (r == -EBUSY) {
621                         /*
622                          * Already attached to different hw_handler:
623                          * try to reattach with correct one.
624                          */
625                         scsi_dh_detach(q);
626                         r = scsi_dh_attach(q, m->hw_handler_name);
627                 }
628
629                 if (r < 0) {
630                         ti->error = "error attaching hardware handler";
631                         dm_put_device(ti, p->path.dev);
632                         goto bad;
633                 }
634
635                 if (m->hw_handler_params) {
636                         r = scsi_dh_set_params(q, m->hw_handler_params);
637                         if (r < 0) {
638                                 ti->error = "unable to set hardware "
639                                                         "handler parameters";
640                                 scsi_dh_detach(q);
641                                 dm_put_device(ti, p->path.dev);
642                                 goto bad;
643                         }
644                 }
645         }
646
647         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
648         if (r) {
649                 dm_put_device(ti, p->path.dev);
650                 goto bad;
651         }
652
653         return p;
654
655  bad:
656         free_pgpath(p);
657         return ERR_PTR(r);
658 }
659
660 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
661                                                    struct multipath *m)
662 {
663         static struct dm_arg _args[] = {
664                 {1, 1024, "invalid number of paths"},
665                 {0, 1024, "invalid number of selector args"}
666         };
667
668         int r;
669         unsigned i, nr_selector_args, nr_args;
670         struct priority_group *pg;
671         struct dm_target *ti = m->ti;
672
673         if (as->argc < 2) {
674                 as->argc = 0;
675                 ti->error = "not enough priority group arguments";
676                 return ERR_PTR(-EINVAL);
677         }
678
679         pg = alloc_priority_group();
680         if (!pg) {
681                 ti->error = "couldn't allocate priority group";
682                 return ERR_PTR(-ENOMEM);
683         }
684         pg->m = m;
685
686         r = parse_path_selector(as, pg, ti);
687         if (r)
688                 goto bad;
689
690         /*
691          * read the paths
692          */
693         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
694         if (r)
695                 goto bad;
696
697         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
698         if (r)
699                 goto bad;
700
701         nr_args = 1 + nr_selector_args;
702         for (i = 0; i < pg->nr_pgpaths; i++) {
703                 struct pgpath *pgpath;
704                 struct dm_arg_set path_args;
705
706                 if (as->argc < nr_args) {
707                         ti->error = "not enough path parameters";
708                         r = -EINVAL;
709                         goto bad;
710                 }
711
712                 path_args.argc = nr_args;
713                 path_args.argv = as->argv;
714
715                 pgpath = parse_path(&path_args, &pg->ps, ti);
716                 if (IS_ERR(pgpath)) {
717                         r = PTR_ERR(pgpath);
718                         goto bad;
719                 }
720
721                 pgpath->pg = pg;
722                 list_add_tail(&pgpath->list, &pg->pgpaths);
723                 dm_consume_args(as, nr_args);
724         }
725
726         return pg;
727
728  bad:
729         free_priority_group(pg, ti);
730         return ERR_PTR(r);
731 }
732
733 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
734 {
735         unsigned hw_argc;
736         int ret;
737         struct dm_target *ti = m->ti;
738
739         static struct dm_arg _args[] = {
740                 {0, 1024, "invalid number of hardware handler args"},
741         };
742
743         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
744                 return -EINVAL;
745
746         if (!hw_argc)
747                 return 0;
748
749         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
750         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
751                                      "scsi_dh_%s", m->hw_handler_name)) {
752                 ti->error = "unknown hardware handler type";
753                 ret = -EINVAL;
754                 goto fail;
755         }
756
757         if (hw_argc > 1) {
758                 char *p;
759                 int i, j, len = 4;
760
761                 for (i = 0; i <= hw_argc - 2; i++)
762                         len += strlen(as->argv[i]) + 1;
763                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
764                 if (!p) {
765                         ti->error = "memory allocation failed";
766                         ret = -ENOMEM;
767                         goto fail;
768                 }
769                 j = sprintf(p, "%d", hw_argc - 1);
770                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
771                         j = sprintf(p, "%s", as->argv[i]);
772         }
773         dm_consume_args(as, hw_argc - 1);
774
775         return 0;
776 fail:
777         kfree(m->hw_handler_name);
778         m->hw_handler_name = NULL;
779         return ret;
780 }
781
782 static int parse_features(struct dm_arg_set *as, struct multipath *m)
783 {
784         int r;
785         unsigned argc;
786         struct dm_target *ti = m->ti;
787         const char *arg_name;
788
789         static struct dm_arg _args[] = {
790                 {0, 6, "invalid number of feature args"},
791                 {1, 50, "pg_init_retries must be between 1 and 50"},
792                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
793         };
794
795         r = dm_read_arg_group(_args, as, &argc, &ti->error);
796         if (r)
797                 return -EINVAL;
798
799         if (!argc)
800                 return 0;
801
802         do {
803                 arg_name = dm_shift_arg(as);
804                 argc--;
805
806                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
807                         r = queue_if_no_path(m, 1, 0);
808                         continue;
809                 }
810
811                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
812                         m->retain_attached_hw_handler = 1;
813                         continue;
814                 }
815
816                 if (!strcasecmp(arg_name, "pg_init_retries") &&
817                     (argc >= 1)) {
818                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
819                         argc--;
820                         continue;
821                 }
822
823                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
824                     (argc >= 1)) {
825                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
826                         argc--;
827                         continue;
828                 }
829
830                 ti->error = "Unrecognised multipath feature request";
831                 r = -EINVAL;
832         } while (argc && !r);
833
834         return r;
835 }
836
837 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
838                          char **argv)
839 {
840         /* target arguments */
841         static struct dm_arg _args[] = {
842                 {0, 1024, "invalid number of priority groups"},
843                 {0, 1024, "invalid initial priority group number"},
844         };
845
846         int r;
847         struct multipath *m;
848         struct dm_arg_set as;
849         unsigned pg_count = 0;
850         unsigned next_pg_num;
851
852         as.argc = argc;
853         as.argv = argv;
854
855         m = alloc_multipath(ti);
856         if (!m) {
857                 ti->error = "can't allocate multipath";
858                 return -EINVAL;
859         }
860
861         r = parse_features(&as, m);
862         if (r)
863                 goto bad;
864
865         r = parse_hw_handler(&as, m);
866         if (r)
867                 goto bad;
868
869         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
870         if (r)
871                 goto bad;
872
873         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
874         if (r)
875                 goto bad;
876
877         if ((!m->nr_priority_groups && next_pg_num) ||
878             (m->nr_priority_groups && !next_pg_num)) {
879                 ti->error = "invalid initial priority group";
880                 r = -EINVAL;
881                 goto bad;
882         }
883
884         /* parse the priority groups */
885         while (as.argc) {
886                 struct priority_group *pg;
887
888                 pg = parse_priority_group(&as, m);
889                 if (IS_ERR(pg)) {
890                         r = PTR_ERR(pg);
891                         goto bad;
892                 }
893
894                 m->nr_valid_paths += pg->nr_pgpaths;
895                 list_add_tail(&pg->list, &m->priority_groups);
896                 pg_count++;
897                 pg->pg_num = pg_count;
898                 if (!--next_pg_num)
899                         m->next_pg = pg;
900         }
901
902         if (pg_count != m->nr_priority_groups) {
903                 ti->error = "priority group count mismatch";
904                 r = -EINVAL;
905                 goto bad;
906         }
907
908         ti->num_flush_requests = 1;
909         ti->num_discard_requests = 1;
910
911         return 0;
912
913  bad:
914         free_multipath(m);
915         return r;
916 }
917
918 static void multipath_wait_for_pg_init_completion(struct multipath *m)
919 {
920         DECLARE_WAITQUEUE(wait, current);
921         unsigned long flags;
922
923         add_wait_queue(&m->pg_init_wait, &wait);
924
925         while (1) {
926                 set_current_state(TASK_UNINTERRUPTIBLE);
927
928                 spin_lock_irqsave(&m->lock, flags);
929                 if (!m->pg_init_in_progress) {
930                         spin_unlock_irqrestore(&m->lock, flags);
931                         break;
932                 }
933                 spin_unlock_irqrestore(&m->lock, flags);
934
935                 io_schedule();
936         }
937         set_current_state(TASK_RUNNING);
938
939         remove_wait_queue(&m->pg_init_wait, &wait);
940 }
941
942 static void flush_multipath_work(struct multipath *m)
943 {
944         flush_workqueue(kmpath_handlerd);
945         multipath_wait_for_pg_init_completion(m);
946         flush_workqueue(kmultipathd);
947         flush_work(&m->trigger_event);
948 }
949
950 static void multipath_dtr(struct dm_target *ti)
951 {
952         struct multipath *m = ti->private;
953
954         flush_multipath_work(m);
955         free_multipath(m);
956 }
957
958 /*
959  * Map cloned requests
960  */
961 static int multipath_map(struct dm_target *ti, struct request *clone,
962                          union map_info *map_context)
963 {
964         int r;
965         struct multipath *m = (struct multipath *) ti->private;
966
967         if (set_mapinfo(m, map_context) < 0)
968                 /* ENOMEM, requeue */
969                 return DM_MAPIO_REQUEUE;
970
971         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
972         r = map_io(m, clone, map_context, 0);
973         if (r < 0 || r == DM_MAPIO_REQUEUE)
974                 clear_mapinfo(m, map_context);
975
976         return r;
977 }
978
979 /*
980  * Take a path out of use.
981  */
982 static int fail_path(struct pgpath *pgpath)
983 {
984         unsigned long flags;
985         struct multipath *m = pgpath->pg->m;
986
987         spin_lock_irqsave(&m->lock, flags);
988
989         if (!pgpath->is_active)
990                 goto out;
991
992         DMWARN("Failing path %s.", pgpath->path.dev->name);
993
994         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
995         pgpath->is_active = 0;
996         pgpath->fail_count++;
997
998         m->nr_valid_paths--;
999
1000         if (pgpath == m->current_pgpath)
1001                 m->current_pgpath = NULL;
1002
1003         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1004                       pgpath->path.dev->name, m->nr_valid_paths);
1005
1006         schedule_work(&m->trigger_event);
1007
1008 out:
1009         spin_unlock_irqrestore(&m->lock, flags);
1010
1011         return 0;
1012 }
1013
1014 /*
1015  * Reinstate a previously-failed path
1016  */
1017 static int reinstate_path(struct pgpath *pgpath)
1018 {
1019         int r = 0;
1020         unsigned long flags;
1021         struct multipath *m = pgpath->pg->m;
1022
1023         spin_lock_irqsave(&m->lock, flags);
1024
1025         if (pgpath->is_active)
1026                 goto out;
1027
1028         if (!pgpath->pg->ps.type->reinstate_path) {
1029                 DMWARN("Reinstate path not supported by path selector %s",
1030                        pgpath->pg->ps.type->name);
1031                 r = -EINVAL;
1032                 goto out;
1033         }
1034
1035         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1036         if (r)
1037                 goto out;
1038
1039         pgpath->is_active = 1;
1040
1041         if (!m->nr_valid_paths++ && m->queue_size) {
1042                 m->current_pgpath = NULL;
1043                 queue_work(kmultipathd, &m->process_queued_ios);
1044         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1045                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1046                         m->pg_init_in_progress++;
1047         }
1048
1049         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1050                       pgpath->path.dev->name, m->nr_valid_paths);
1051
1052         schedule_work(&m->trigger_event);
1053
1054 out:
1055         spin_unlock_irqrestore(&m->lock, flags);
1056
1057         return r;
1058 }
1059
1060 /*
1061  * Fail or reinstate all paths that match the provided struct dm_dev.
1062  */
1063 static int action_dev(struct multipath *m, struct dm_dev *dev,
1064                       action_fn action)
1065 {
1066         int r = -EINVAL;
1067         struct pgpath *pgpath;
1068         struct priority_group *pg;
1069
1070         list_for_each_entry(pg, &m->priority_groups, list) {
1071                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1072                         if (pgpath->path.dev == dev)
1073                                 r = action(pgpath);
1074                 }
1075         }
1076
1077         return r;
1078 }
1079
1080 /*
1081  * Temporarily try to avoid having to use the specified PG
1082  */
1083 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1084                       int bypassed)
1085 {
1086         unsigned long flags;
1087
1088         spin_lock_irqsave(&m->lock, flags);
1089
1090         pg->bypassed = bypassed;
1091         m->current_pgpath = NULL;
1092         m->current_pg = NULL;
1093
1094         spin_unlock_irqrestore(&m->lock, flags);
1095
1096         schedule_work(&m->trigger_event);
1097 }
1098
1099 /*
1100  * Switch to using the specified PG from the next I/O that gets mapped
1101  */
1102 static int switch_pg_num(struct multipath *m, const char *pgstr)
1103 {
1104         struct priority_group *pg;
1105         unsigned pgnum;
1106         unsigned long flags;
1107         char dummy;
1108
1109         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1110             (pgnum > m->nr_priority_groups)) {
1111                 DMWARN("invalid PG number supplied to switch_pg_num");
1112                 return -EINVAL;
1113         }
1114
1115         spin_lock_irqsave(&m->lock, flags);
1116         list_for_each_entry(pg, &m->priority_groups, list) {
1117                 pg->bypassed = 0;
1118                 if (--pgnum)
1119                         continue;
1120
1121                 m->current_pgpath = NULL;
1122                 m->current_pg = NULL;
1123                 m->next_pg = pg;
1124         }
1125         spin_unlock_irqrestore(&m->lock, flags);
1126
1127         schedule_work(&m->trigger_event);
1128         return 0;
1129 }
1130
1131 /*
1132  * Set/clear bypassed status of a PG.
1133  * PGs are numbered upwards from 1 in the order they were declared.
1134  */
1135 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1136 {
1137         struct priority_group *pg;
1138         unsigned pgnum;
1139         char dummy;
1140
1141         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1142             (pgnum > m->nr_priority_groups)) {
1143                 DMWARN("invalid PG number supplied to bypass_pg");
1144                 return -EINVAL;
1145         }
1146
1147         list_for_each_entry(pg, &m->priority_groups, list) {
1148                 if (!--pgnum)
1149                         break;
1150         }
1151
1152         bypass_pg(m, pg, bypassed);
1153         return 0;
1154 }
1155
1156 /*
1157  * Should we retry pg_init immediately?
1158  */
1159 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1160 {
1161         unsigned long flags;
1162         int limit_reached = 0;
1163
1164         spin_lock_irqsave(&m->lock, flags);
1165
1166         if (m->pg_init_count <= m->pg_init_retries)
1167                 m->pg_init_required = 1;
1168         else
1169                 limit_reached = 1;
1170
1171         spin_unlock_irqrestore(&m->lock, flags);
1172
1173         return limit_reached;
1174 }
1175
1176 static void pg_init_done(void *data, int errors)
1177 {
1178         struct pgpath *pgpath = data;
1179         struct priority_group *pg = pgpath->pg;
1180         struct multipath *m = pg->m;
1181         unsigned long flags;
1182         unsigned delay_retry = 0;
1183
1184         /* device or driver problems */
1185         switch (errors) {
1186         case SCSI_DH_OK:
1187                 break;
1188         case SCSI_DH_NOSYS:
1189                 if (!m->hw_handler_name) {
1190                         errors = 0;
1191                         break;
1192                 }
1193                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1194                       "Error %d.", m->hw_handler_name, errors);
1195                 /*
1196                  * Fail path for now, so we do not ping pong
1197                  */
1198                 fail_path(pgpath);
1199                 break;
1200         case SCSI_DH_DEV_TEMP_BUSY:
1201                 /*
1202                  * Probably doing something like FW upgrade on the
1203                  * controller so try the other pg.
1204                  */
1205                 bypass_pg(m, pg, 1);
1206                 break;
1207         case SCSI_DH_RETRY:
1208                 /* Wait before retrying. */
1209                 delay_retry = 1;
1210         case SCSI_DH_IMM_RETRY:
1211         case SCSI_DH_RES_TEMP_UNAVAIL:
1212                 if (pg_init_limit_reached(m, pgpath))
1213                         fail_path(pgpath);
1214                 errors = 0;
1215                 break;
1216         default:
1217                 /*
1218                  * We probably do not want to fail the path for a device
1219                  * error, but this is what the old dm did. In future
1220                  * patches we can do more advanced handling.
1221                  */
1222                 fail_path(pgpath);
1223         }
1224
1225         spin_lock_irqsave(&m->lock, flags);
1226         if (errors) {
1227                 if (pgpath == m->current_pgpath) {
1228                         DMERR("Could not failover device. Error %d.", errors);
1229                         m->current_pgpath = NULL;
1230                         m->current_pg = NULL;
1231                 }
1232         } else if (!m->pg_init_required)
1233                 pg->bypassed = 0;
1234
1235         if (--m->pg_init_in_progress)
1236                 /* Activations of other paths are still on going */
1237                 goto out;
1238
1239         if (!m->pg_init_required)
1240                 m->queue_io = 0;
1241
1242         m->pg_init_delay_retry = delay_retry;
1243         queue_work(kmultipathd, &m->process_queued_ios);
1244
1245         /*
1246          * Wake up any thread waiting to suspend.
1247          */
1248         wake_up(&m->pg_init_wait);
1249
1250 out:
1251         spin_unlock_irqrestore(&m->lock, flags);
1252 }
1253
1254 static void activate_path(struct work_struct *work)
1255 {
1256         struct pgpath *pgpath =
1257                 container_of(work, struct pgpath, activate_path.work);
1258
1259         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1260                                 pg_init_done, pgpath);
1261 }
1262
1263 /*
1264  * end_io handling
1265  */
1266 static int do_end_io(struct multipath *m, struct request *clone,
1267                      int error, struct dm_mpath_io *mpio)
1268 {
1269         /*
1270          * We don't queue any clone request inside the multipath target
1271          * during end I/O handling, since those clone requests don't have
1272          * bio clones.  If we queue them inside the multipath target,
1273          * we need to make bio clones, that requires memory allocation.
1274          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1275          *  don't have bio clones.)
1276          * Instead of queueing the clone request here, we queue the original
1277          * request into dm core, which will remake a clone request and
1278          * clone bios for it and resubmit it later.
1279          */
1280         int r = DM_ENDIO_REQUEUE;
1281         unsigned long flags;
1282
1283         if (!error && !clone->errors)
1284                 return 0;       /* I/O complete */
1285
1286         if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1287                 return error;
1288
1289         if (mpio->pgpath)
1290                 fail_path(mpio->pgpath);
1291
1292         spin_lock_irqsave(&m->lock, flags);
1293         if (!m->nr_valid_paths) {
1294                 if (!m->queue_if_no_path) {
1295                         if (!__must_push_back(m))
1296                                 r = -EIO;
1297                 } else {
1298                         if (error == -EBADE)
1299                                 r = error;
1300                 }
1301         }
1302         spin_unlock_irqrestore(&m->lock, flags);
1303
1304         return r;
1305 }
1306
1307 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1308                             int error, union map_info *map_context)
1309 {
1310         struct multipath *m = ti->private;
1311         struct dm_mpath_io *mpio = map_context->ptr;
1312         struct pgpath *pgpath = mpio->pgpath;
1313         struct path_selector *ps;
1314         int r;
1315
1316         BUG_ON(!mpio);
1317
1318         r  = do_end_io(m, clone, error, mpio);
1319         if (pgpath) {
1320                 ps = &pgpath->pg->ps;
1321                 if (ps->type->end_io)
1322                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1323         }
1324         clear_mapinfo(m, map_context);
1325
1326         return r;
1327 }
1328
1329 /*
1330  * Suspend can't complete until all the I/O is processed so if
1331  * the last path fails we must error any remaining I/O.
1332  * Note that if the freeze_bdev fails while suspending, the
1333  * queue_if_no_path state is lost - userspace should reset it.
1334  */
1335 static void multipath_presuspend(struct dm_target *ti)
1336 {
1337         struct multipath *m = (struct multipath *) ti->private;
1338
1339         queue_if_no_path(m, 0, 1);
1340 }
1341
1342 static void multipath_postsuspend(struct dm_target *ti)
1343 {
1344         struct multipath *m = ti->private;
1345
1346         mutex_lock(&m->work_mutex);
1347         flush_multipath_work(m);
1348         mutex_unlock(&m->work_mutex);
1349 }
1350
1351 /*
1352  * Restore the queue_if_no_path setting.
1353  */
1354 static void multipath_resume(struct dm_target *ti)
1355 {
1356         struct multipath *m = (struct multipath *) ti->private;
1357         unsigned long flags;
1358
1359         spin_lock_irqsave(&m->lock, flags);
1360         m->queue_if_no_path = m->saved_queue_if_no_path;
1361         spin_unlock_irqrestore(&m->lock, flags);
1362 }
1363
1364 /*
1365  * Info output has the following format:
1366  * num_multipath_feature_args [multipath_feature_args]*
1367  * num_handler_status_args [handler_status_args]*
1368  * num_groups init_group_number
1369  *            [A|D|E num_ps_status_args [ps_status_args]*
1370  *             num_paths num_selector_args
1371  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1372  *
1373  * Table output has the following format (identical to the constructor string):
1374  * num_feature_args [features_args]*
1375  * num_handler_args hw_handler [hw_handler_args]*
1376  * num_groups init_group_number
1377  *     [priority selector-name num_ps_args [ps_args]*
1378  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1379  */
1380 static int multipath_status(struct dm_target *ti, status_type_t type,
1381                             unsigned status_flags, char *result, unsigned maxlen)
1382 {
1383         int sz = 0;
1384         unsigned long flags;
1385         struct multipath *m = (struct multipath *) ti->private;
1386         struct priority_group *pg;
1387         struct pgpath *p;
1388         unsigned pg_num;
1389         char state;
1390
1391         spin_lock_irqsave(&m->lock, flags);
1392
1393         /* Features */
1394         if (type == STATUSTYPE_INFO)
1395                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1396         else {
1397                 DMEMIT("%u ", m->queue_if_no_path +
1398                               (m->pg_init_retries > 0) * 2 +
1399                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1400                               m->retain_attached_hw_handler);
1401                 if (m->queue_if_no_path)
1402                         DMEMIT("queue_if_no_path ");
1403                 if (m->pg_init_retries)
1404                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1405                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1406                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1407                 if (m->retain_attached_hw_handler)
1408                         DMEMIT("retain_attached_hw_handler ");
1409         }
1410
1411         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1412                 DMEMIT("0 ");
1413         else
1414                 DMEMIT("1 %s ", m->hw_handler_name);
1415
1416         DMEMIT("%u ", m->nr_priority_groups);
1417
1418         if (m->next_pg)
1419                 pg_num = m->next_pg->pg_num;
1420         else if (m->current_pg)
1421                 pg_num = m->current_pg->pg_num;
1422         else
1423                 pg_num = (m->nr_priority_groups ? 1 : 0);
1424
1425         DMEMIT("%u ", pg_num);
1426
1427         switch (type) {
1428         case STATUSTYPE_INFO:
1429                 list_for_each_entry(pg, &m->priority_groups, list) {
1430                         if (pg->bypassed)
1431                                 state = 'D';    /* Disabled */
1432                         else if (pg == m->current_pg)
1433                                 state = 'A';    /* Currently Active */
1434                         else
1435                                 state = 'E';    /* Enabled */
1436
1437                         DMEMIT("%c ", state);
1438
1439                         if (pg->ps.type->status)
1440                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1441                                                           result + sz,
1442                                                           maxlen - sz);
1443                         else
1444                                 DMEMIT("0 ");
1445
1446                         DMEMIT("%u %u ", pg->nr_pgpaths,
1447                                pg->ps.type->info_args);
1448
1449                         list_for_each_entry(p, &pg->pgpaths, list) {
1450                                 DMEMIT("%s %s %u ", p->path.dev->name,
1451                                        p->is_active ? "A" : "F",
1452                                        p->fail_count);
1453                                 if (pg->ps.type->status)
1454                                         sz += pg->ps.type->status(&pg->ps,
1455                                               &p->path, type, result + sz,
1456                                               maxlen - sz);
1457                         }
1458                 }
1459                 break;
1460
1461         case STATUSTYPE_TABLE:
1462                 list_for_each_entry(pg, &m->priority_groups, list) {
1463                         DMEMIT("%s ", pg->ps.type->name);
1464
1465                         if (pg->ps.type->status)
1466                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1467                                                           result + sz,
1468                                                           maxlen - sz);
1469                         else
1470                                 DMEMIT("0 ");
1471
1472                         DMEMIT("%u %u ", pg->nr_pgpaths,
1473                                pg->ps.type->table_args);
1474
1475                         list_for_each_entry(p, &pg->pgpaths, list) {
1476                                 DMEMIT("%s ", p->path.dev->name);
1477                                 if (pg->ps.type->status)
1478                                         sz += pg->ps.type->status(&pg->ps,
1479                                               &p->path, type, result + sz,
1480                                               maxlen - sz);
1481                         }
1482                 }
1483                 break;
1484         }
1485
1486         spin_unlock_irqrestore(&m->lock, flags);
1487
1488         return 0;
1489 }
1490
1491 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1492 {
1493         int r = -EINVAL;
1494         struct dm_dev *dev;
1495         struct multipath *m = (struct multipath *) ti->private;
1496         action_fn action;
1497
1498         mutex_lock(&m->work_mutex);
1499
1500         if (dm_suspended(ti)) {
1501                 r = -EBUSY;
1502                 goto out;
1503         }
1504
1505         if (argc == 1) {
1506                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1507                         r = queue_if_no_path(m, 1, 0);
1508                         goto out;
1509                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1510                         r = queue_if_no_path(m, 0, 0);
1511                         goto out;
1512                 }
1513         }
1514
1515         if (argc != 2) {
1516                 DMWARN("Unrecognised multipath message received.");
1517                 goto out;
1518         }
1519
1520         if (!strcasecmp(argv[0], "disable_group")) {
1521                 r = bypass_pg_num(m, argv[1], 1);
1522                 goto out;
1523         } else if (!strcasecmp(argv[0], "enable_group")) {
1524                 r = bypass_pg_num(m, argv[1], 0);
1525                 goto out;
1526         } else if (!strcasecmp(argv[0], "switch_group")) {
1527                 r = switch_pg_num(m, argv[1]);
1528                 goto out;
1529         } else if (!strcasecmp(argv[0], "reinstate_path"))
1530                 action = reinstate_path;
1531         else if (!strcasecmp(argv[0], "fail_path"))
1532                 action = fail_path;
1533         else {
1534                 DMWARN("Unrecognised multipath message received.");
1535                 goto out;
1536         }
1537
1538         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1539         if (r) {
1540                 DMWARN("message: error getting device %s",
1541                        argv[1]);
1542                 goto out;
1543         }
1544
1545         r = action_dev(m, dev, action);
1546
1547         dm_put_device(ti, dev);
1548
1549 out:
1550         mutex_unlock(&m->work_mutex);
1551         return r;
1552 }
1553
1554 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1555                            unsigned long arg)
1556 {
1557         struct multipath *m = ti->private;
1558         struct pgpath *pgpath;
1559         struct block_device *bdev;
1560         fmode_t mode;
1561         unsigned long flags;
1562         int r;
1563
1564 again:
1565         bdev = NULL;
1566         mode = 0;
1567         r = 0;
1568
1569         spin_lock_irqsave(&m->lock, flags);
1570
1571         if (!m->current_pgpath)
1572                 __choose_pgpath(m, 0);
1573
1574         pgpath = m->current_pgpath;
1575
1576         if (pgpath) {
1577                 bdev = pgpath->path.dev->bdev;
1578                 mode = pgpath->path.dev->mode;
1579         }
1580
1581         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1582                 r = -EAGAIN;
1583         else if (!bdev)
1584                 r = -EIO;
1585
1586         spin_unlock_irqrestore(&m->lock, flags);
1587
1588         /*
1589          * Only pass ioctls through if the device sizes match exactly.
1590          */
1591         if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1592                 r = scsi_verify_blk_ioctl(NULL, cmd);
1593
1594         if (r == -EAGAIN && !fatal_signal_pending(current)) {
1595                 queue_work(kmultipathd, &m->process_queued_ios);
1596                 msleep(10);
1597                 goto again;
1598         }
1599
1600         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1601 }
1602
1603 static int multipath_iterate_devices(struct dm_target *ti,
1604                                      iterate_devices_callout_fn fn, void *data)
1605 {
1606         struct multipath *m = ti->private;
1607         struct priority_group *pg;
1608         struct pgpath *p;
1609         int ret = 0;
1610
1611         list_for_each_entry(pg, &m->priority_groups, list) {
1612                 list_for_each_entry(p, &pg->pgpaths, list) {
1613                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1614                         if (ret)
1615                                 goto out;
1616                 }
1617         }
1618
1619 out:
1620         return ret;
1621 }
1622
1623 static int __pgpath_busy(struct pgpath *pgpath)
1624 {
1625         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1626
1627         return dm_underlying_device_busy(q);
1628 }
1629
1630 /*
1631  * We return "busy", only when we can map I/Os but underlying devices
1632  * are busy (so even if we map I/Os now, the I/Os will wait on
1633  * the underlying queue).
1634  * In other words, if we want to kill I/Os or queue them inside us
1635  * due to map unavailability, we don't return "busy".  Otherwise,
1636  * dm core won't give us the I/Os and we can't do what we want.
1637  */
1638 static int multipath_busy(struct dm_target *ti)
1639 {
1640         int busy = 0, has_active = 0;
1641         struct multipath *m = ti->private;
1642         struct priority_group *pg;
1643         struct pgpath *pgpath;
1644         unsigned long flags;
1645
1646         spin_lock_irqsave(&m->lock, flags);
1647
1648         /* Guess which priority_group will be used at next mapping time */
1649         if (unlikely(!m->current_pgpath && m->next_pg))
1650                 pg = m->next_pg;
1651         else if (likely(m->current_pg))
1652                 pg = m->current_pg;
1653         else
1654                 /*
1655                  * We don't know which pg will be used at next mapping time.
1656                  * We don't call __choose_pgpath() here to avoid to trigger
1657                  * pg_init just by busy checking.
1658                  * So we don't know whether underlying devices we will be using
1659                  * at next mapping time are busy or not. Just try mapping.
1660                  */
1661                 goto out;
1662
1663         /*
1664          * If there is one non-busy active path at least, the path selector
1665          * will be able to select it. So we consider such a pg as not busy.
1666          */
1667         busy = 1;
1668         list_for_each_entry(pgpath, &pg->pgpaths, list)
1669                 if (pgpath->is_active) {
1670                         has_active = 1;
1671
1672                         if (!__pgpath_busy(pgpath)) {
1673                                 busy = 0;
1674                                 break;
1675                         }
1676                 }
1677
1678         if (!has_active)
1679                 /*
1680                  * No active path in this pg, so this pg won't be used and
1681                  * the current_pg will be changed at next mapping time.
1682                  * We need to try mapping to determine it.
1683                  */
1684                 busy = 0;
1685
1686 out:
1687         spin_unlock_irqrestore(&m->lock, flags);
1688
1689         return busy;
1690 }
1691
1692 /*-----------------------------------------------------------------
1693  * Module setup
1694  *---------------------------------------------------------------*/
1695 static struct target_type multipath_target = {
1696         .name = "multipath",
1697         .version = {1, 5, 0},
1698         .module = THIS_MODULE,
1699         .ctr = multipath_ctr,
1700         .dtr = multipath_dtr,
1701         .map_rq = multipath_map,
1702         .rq_end_io = multipath_end_io,
1703         .presuspend = multipath_presuspend,
1704         .postsuspend = multipath_postsuspend,
1705         .resume = multipath_resume,
1706         .status = multipath_status,
1707         .message = multipath_message,
1708         .ioctl  = multipath_ioctl,
1709         .iterate_devices = multipath_iterate_devices,
1710         .busy = multipath_busy,
1711 };
1712
1713 static int __init dm_multipath_init(void)
1714 {
1715         int r;
1716
1717         /* allocate a slab for the dm_ios */
1718         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1719         if (!_mpio_cache)
1720                 return -ENOMEM;
1721
1722         r = dm_register_target(&multipath_target);
1723         if (r < 0) {
1724                 DMERR("register failed %d", r);
1725                 kmem_cache_destroy(_mpio_cache);
1726                 return -EINVAL;
1727         }
1728
1729         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1730         if (!kmultipathd) {
1731                 DMERR("failed to create workqueue kmpathd");
1732                 dm_unregister_target(&multipath_target);
1733                 kmem_cache_destroy(_mpio_cache);
1734                 return -ENOMEM;
1735         }
1736
1737         /*
1738          * A separate workqueue is used to handle the device handlers
1739          * to avoid overloading existing workqueue. Overloading the
1740          * old workqueue would also create a bottleneck in the
1741          * path of the storage hardware device activation.
1742          */
1743         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1744                                                   WQ_MEM_RECLAIM);
1745         if (!kmpath_handlerd) {
1746                 DMERR("failed to create workqueue kmpath_handlerd");
1747                 destroy_workqueue(kmultipathd);
1748                 dm_unregister_target(&multipath_target);
1749                 kmem_cache_destroy(_mpio_cache);
1750                 return -ENOMEM;
1751         }
1752
1753         DMINFO("version %u.%u.%u loaded",
1754                multipath_target.version[0], multipath_target.version[1],
1755                multipath_target.version[2]);
1756
1757         return r;
1758 }
1759
1760 static void __exit dm_multipath_exit(void)
1761 {
1762         destroy_workqueue(kmpath_handlerd);
1763         destroy_workqueue(kmultipathd);
1764
1765         dm_unregister_target(&multipath_target);
1766         kmem_cache_destroy(_mpio_cache);
1767 }
1768
1769 module_init(dm_multipath_init);
1770 module_exit(dm_multipath_exit);
1771
1772 MODULE_DESCRIPTION(DM_NAME " multipath target");
1773 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1774 MODULE_LICENSE("GPL");