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