dm mpath: validate table argument count
[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-bio-record.h"
12 #include "dm-uevent.h"
13
14 #include <linux/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <scsi/scsi_dh.h>
23 #include <asm/atomic.h>
24
25 #define DM_MSG_PREFIX "multipath"
26 #define MESG_STR(x) x, sizeof(x)
27
28 /* Path properties */
29 struct pgpath {
30         struct list_head list;
31
32         struct priority_group *pg;      /* Owning PG */
33         unsigned is_active;             /* Path status */
34         unsigned fail_count;            /* Cumulative failure count */
35
36         struct dm_path path;
37         struct work_struct deactivate_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         struct work_struct activate_path;
68         struct pgpath *pgpath_to_activate;
69         unsigned nr_priority_groups;
70         struct list_head priority_groups;
71         unsigned pg_init_required;      /* pg_init needs calling? */
72         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
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 bio_list 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
99 /*
100  * Context information attached to each bio we process.
101  */
102 struct dm_mpath_io {
103         struct pgpath *pgpath;
104         struct dm_bio_details details;
105 };
106
107 typedef int (*action_fn) (struct pgpath *pgpath);
108
109 #define MIN_IOS 256     /* Mempool size */
110
111 static struct kmem_cache *_mpio_cache;
112
113 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
114 static void process_queued_ios(struct work_struct *work);
115 static void trigger_event(struct work_struct *work);
116 static void activate_path(struct work_struct *work);
117 static void deactivate_path(struct work_struct *work);
118
119
120 /*-----------------------------------------------
121  * Allocation routines
122  *-----------------------------------------------*/
123
124 static struct pgpath *alloc_pgpath(void)
125 {
126         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
127
128         if (pgpath) {
129                 pgpath->is_active = 1;
130                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
131         }
132
133         return pgpath;
134 }
135
136 static void free_pgpath(struct pgpath *pgpath)
137 {
138         kfree(pgpath);
139 }
140
141 static void deactivate_path(struct work_struct *work)
142 {
143         struct pgpath *pgpath =
144                 container_of(work, struct pgpath, deactivate_path);
145
146         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
147 }
148
149 static struct priority_group *alloc_priority_group(void)
150 {
151         struct priority_group *pg;
152
153         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
154
155         if (pg)
156                 INIT_LIST_HEAD(&pg->pgpaths);
157
158         return pg;
159 }
160
161 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
162 {
163         unsigned long flags;
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                 spin_lock_irqsave(&m->lock, flags);
173                 if (m->pgpath_to_activate == pgpath)
174                         m->pgpath_to_activate = NULL;
175                 spin_unlock_irqrestore(&m->lock, flags);
176                 free_pgpath(pgpath);
177         }
178 }
179
180 static void free_priority_group(struct priority_group *pg,
181                                 struct dm_target *ti)
182 {
183         struct path_selector *ps = &pg->ps;
184
185         if (ps->type) {
186                 ps->type->destroy(ps);
187                 dm_put_path_selector(ps->type);
188         }
189
190         free_pgpaths(&pg->pgpaths, ti);
191         kfree(pg);
192 }
193
194 static struct multipath *alloc_multipath(struct dm_target *ti)
195 {
196         struct multipath *m;
197
198         m = kzalloc(sizeof(*m), GFP_KERNEL);
199         if (m) {
200                 INIT_LIST_HEAD(&m->priority_groups);
201                 spin_lock_init(&m->lock);
202                 m->queue_io = 1;
203                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
204                 INIT_WORK(&m->trigger_event, trigger_event);
205                 INIT_WORK(&m->activate_path, activate_path);
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         mempool_destroy(m->mpio_pool);
229         kfree(m);
230 }
231
232
233 /*-----------------------------------------------
234  * Path selection
235  *-----------------------------------------------*/
236
237 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
238 {
239         m->current_pg = pgpath->pg;
240
241         /* Must we initialise the PG first, and queue I/O till it's ready? */
242         if (m->hw_handler_name) {
243                 m->pg_init_required = 1;
244                 m->queue_io = 1;
245         } else {
246                 m->pg_init_required = 0;
247                 m->queue_io = 0;
248         }
249
250         m->pg_init_count = 0;
251 }
252
253 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
254 {
255         struct dm_path *path;
256
257         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
258         if (!path)
259                 return -ENXIO;
260
261         m->current_pgpath = path_to_pgpath(path);
262
263         if (m->current_pg != pg)
264                 __switch_pg(m, m->current_pgpath);
265
266         return 0;
267 }
268
269 static void __choose_pgpath(struct multipath *m)
270 {
271         struct priority_group *pg;
272         unsigned bypassed = 1;
273
274         if (!m->nr_valid_paths)
275                 goto failed;
276
277         /* Were we instructed to switch PG? */
278         if (m->next_pg) {
279                 pg = m->next_pg;
280                 m->next_pg = NULL;
281                 if (!__choose_path_in_pg(m, pg))
282                         return;
283         }
284
285         /* Don't change PG until it has no remaining paths */
286         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
287                 return;
288
289         /*
290          * Loop through priority groups until we find a valid path.
291          * First time we skip PGs marked 'bypassed'.
292          * Second time we only try the ones we skipped.
293          */
294         do {
295                 list_for_each_entry(pg, &m->priority_groups, list) {
296                         if (pg->bypassed == bypassed)
297                                 continue;
298                         if (!__choose_path_in_pg(m, pg))
299                                 return;
300                 }
301         } while (bypassed--);
302
303 failed:
304         m->current_pgpath = NULL;
305         m->current_pg = NULL;
306 }
307
308 /*
309  * Check whether bios must be queued in the device-mapper core rather
310  * than here in the target.
311  *
312  * m->lock must be held on entry.
313  *
314  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
315  * same value then we are not between multipath_presuspend()
316  * and multipath_resume() calls and we have no need to check
317  * for the DMF_NOFLUSH_SUSPENDING flag.
318  */
319 static int __must_push_back(struct multipath *m)
320 {
321         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
322                 dm_noflush_suspending(m->ti));
323 }
324
325 static int map_io(struct multipath *m, struct bio *bio,
326                   struct dm_mpath_io *mpio, unsigned was_queued)
327 {
328         int r = DM_MAPIO_REMAPPED;
329         unsigned long flags;
330         struct pgpath *pgpath;
331
332         spin_lock_irqsave(&m->lock, flags);
333
334         /* Do we need to select a new pgpath? */
335         if (!m->current_pgpath ||
336             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
337                 __choose_pgpath(m);
338
339         pgpath = m->current_pgpath;
340
341         if (was_queued)
342                 m->queue_size--;
343
344         if ((pgpath && m->queue_io) ||
345             (!pgpath && m->queue_if_no_path)) {
346                 /* Queue for the daemon to resubmit */
347                 bio_list_add(&m->queued_ios, bio);
348                 m->queue_size++;
349                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
350                     !m->queue_io)
351                         queue_work(kmultipathd, &m->process_queued_ios);
352                 pgpath = NULL;
353                 r = DM_MAPIO_SUBMITTED;
354         } else if (pgpath)
355                 bio->bi_bdev = pgpath->path.dev->bdev;
356         else if (__must_push_back(m))
357                 r = DM_MAPIO_REQUEUE;
358         else
359                 r = -EIO;       /* Failed */
360
361         mpio->pgpath = pgpath;
362
363         spin_unlock_irqrestore(&m->lock, flags);
364
365         return r;
366 }
367
368 /*
369  * If we run out of usable paths, should we queue I/O or error it?
370  */
371 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
372                             unsigned save_old_value)
373 {
374         unsigned long flags;
375
376         spin_lock_irqsave(&m->lock, flags);
377
378         if (save_old_value)
379                 m->saved_queue_if_no_path = m->queue_if_no_path;
380         else
381                 m->saved_queue_if_no_path = queue_if_no_path;
382         m->queue_if_no_path = queue_if_no_path;
383         if (!m->queue_if_no_path && m->queue_size)
384                 queue_work(kmultipathd, &m->process_queued_ios);
385
386         spin_unlock_irqrestore(&m->lock, flags);
387
388         return 0;
389 }
390
391 /*-----------------------------------------------------------------
392  * The multipath daemon is responsible for resubmitting queued ios.
393  *---------------------------------------------------------------*/
394
395 static void dispatch_queued_ios(struct multipath *m)
396 {
397         int r;
398         unsigned long flags;
399         struct bio *bio = NULL, *next;
400         struct dm_mpath_io *mpio;
401         union map_info *info;
402
403         spin_lock_irqsave(&m->lock, flags);
404         bio = bio_list_get(&m->queued_ios);
405         spin_unlock_irqrestore(&m->lock, flags);
406
407         while (bio) {
408                 next = bio->bi_next;
409                 bio->bi_next = NULL;
410
411                 info = dm_get_mapinfo(bio);
412                 mpio = info->ptr;
413
414                 r = map_io(m, bio, mpio, 1);
415                 if (r < 0)
416                         bio_endio(bio, r);
417                 else if (r == DM_MAPIO_REMAPPED)
418                         generic_make_request(bio);
419                 else if (r == DM_MAPIO_REQUEUE)
420                         bio_endio(bio, -EIO);
421
422                 bio = next;
423         }
424 }
425
426 static void process_queued_ios(struct work_struct *work)
427 {
428         struct multipath *m =
429                 container_of(work, struct multipath, process_queued_ios);
430         struct pgpath *pgpath = NULL;
431         unsigned init_required = 0, must_queue = 1;
432         unsigned long flags;
433
434         spin_lock_irqsave(&m->lock, flags);
435
436         if (!m->queue_size)
437                 goto out;
438
439         if (!m->current_pgpath)
440                 __choose_pgpath(m);
441
442         pgpath = m->current_pgpath;
443
444         if ((pgpath && !m->queue_io) ||
445             (!pgpath && !m->queue_if_no_path))
446                 must_queue = 0;
447
448         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
449                 m->pgpath_to_activate = pgpath;
450                 m->pg_init_count++;
451                 m->pg_init_required = 0;
452                 m->pg_init_in_progress = 1;
453                 init_required = 1;
454         }
455
456 out:
457         spin_unlock_irqrestore(&m->lock, flags);
458
459         if (init_required)
460                 queue_work(kmpath_handlerd, &m->activate_path);
461
462         if (!must_queue)
463                 dispatch_queued_ios(m);
464 }
465
466 /*
467  * An event is triggered whenever a path is taken out of use.
468  * Includes path failure and PG bypass.
469  */
470 static void trigger_event(struct work_struct *work)
471 {
472         struct multipath *m =
473                 container_of(work, struct multipath, trigger_event);
474
475         dm_table_event(m->ti->table);
476 }
477
478 /*-----------------------------------------------------------------
479  * Constructor/argument parsing:
480  * <#multipath feature args> [<arg>]*
481  * <#hw_handler args> [hw_handler [<arg>]*]
482  * <#priority groups>
483  * <initial priority group>
484  *     [<selector> <#selector args> [<arg>]*
485  *      <#paths> <#per-path selector args>
486  *         [<path> [<arg>]* ]+ ]+
487  *---------------------------------------------------------------*/
488 struct param {
489         unsigned min;
490         unsigned max;
491         char *error;
492 };
493
494 static int read_param(struct param *param, char *str, unsigned *v, char **error)
495 {
496         if (!str ||
497             (sscanf(str, "%u", v) != 1) ||
498             (*v < param->min) ||
499             (*v > param->max)) {
500                 *error = param->error;
501                 return -EINVAL;
502         }
503
504         return 0;
505 }
506
507 struct arg_set {
508         unsigned argc;
509         char **argv;
510 };
511
512 static char *shift(struct arg_set *as)
513 {
514         char *r;
515
516         if (as->argc) {
517                 as->argc--;
518                 r = *as->argv;
519                 as->argv++;
520                 return r;
521         }
522
523         return NULL;
524 }
525
526 static void consume(struct arg_set *as, unsigned n)
527 {
528         BUG_ON (as->argc < n);
529         as->argc -= n;
530         as->argv += n;
531 }
532
533 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
534                                struct dm_target *ti)
535 {
536         int r;
537         struct path_selector_type *pst;
538         unsigned ps_argc;
539
540         static struct param _params[] = {
541                 {0, 1024, "invalid number of path selector args"},
542         };
543
544         pst = dm_get_path_selector(shift(as));
545         if (!pst) {
546                 ti->error = "unknown path selector type";
547                 return -EINVAL;
548         }
549
550         r = read_param(_params, shift(as), &ps_argc, &ti->error);
551         if (r) {
552                 dm_put_path_selector(pst);
553                 return -EINVAL;
554         }
555
556         if (ps_argc > as->argc) {
557                 dm_put_path_selector(pst);
558                 ti->error = "not enough arguments for path selector";
559                 return -EINVAL;
560         }
561
562         r = pst->create(&pg->ps, ps_argc, as->argv);
563         if (r) {
564                 dm_put_path_selector(pst);
565                 ti->error = "path selector constructor failed";
566                 return r;
567         }
568
569         pg->ps.type = pst;
570         consume(as, ps_argc);
571
572         return 0;
573 }
574
575 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
576                                struct dm_target *ti)
577 {
578         int r;
579         struct pgpath *p;
580         struct multipath *m = ti->private;
581
582         /* we need at least a path arg */
583         if (as->argc < 1) {
584                 ti->error = "no device given";
585                 return ERR_PTR(-EINVAL);
586         }
587
588         p = alloc_pgpath();
589         if (!p)
590                 return ERR_PTR(-ENOMEM);
591
592         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
593                           dm_table_get_mode(ti->table), &p->path.dev);
594         if (r) {
595                 ti->error = "error getting device";
596                 goto bad;
597         }
598
599         if (m->hw_handler_name) {
600                 r = scsi_dh_attach(bdev_get_queue(p->path.dev->bdev),
601                                    m->hw_handler_name);
602                 if (r < 0) {
603                         dm_put_device(ti, p->path.dev);
604                         goto bad;
605                 }
606         }
607
608         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
609         if (r) {
610                 dm_put_device(ti, p->path.dev);
611                 goto bad;
612         }
613
614         return p;
615
616  bad:
617         free_pgpath(p);
618         return ERR_PTR(r);
619 }
620
621 static struct priority_group *parse_priority_group(struct arg_set *as,
622                                                    struct multipath *m)
623 {
624         static struct param _params[] = {
625                 {1, 1024, "invalid number of paths"},
626                 {0, 1024, "invalid number of selector args"}
627         };
628
629         int r;
630         unsigned i, nr_selector_args, nr_params;
631         struct priority_group *pg;
632         struct dm_target *ti = m->ti;
633
634         if (as->argc < 2) {
635                 as->argc = 0;
636                 ti->error = "not enough priority group arguments";
637                 return ERR_PTR(-EINVAL);
638         }
639
640         pg = alloc_priority_group();
641         if (!pg) {
642                 ti->error = "couldn't allocate priority group";
643                 return ERR_PTR(-ENOMEM);
644         }
645         pg->m = m;
646
647         r = parse_path_selector(as, pg, ti);
648         if (r)
649                 goto bad;
650
651         /*
652          * read the paths
653          */
654         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
655         if (r)
656                 goto bad;
657
658         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
659         if (r)
660                 goto bad;
661
662         nr_params = 1 + nr_selector_args;
663         for (i = 0; i < pg->nr_pgpaths; i++) {
664                 struct pgpath *pgpath;
665                 struct arg_set path_args;
666
667                 if (as->argc < nr_params) {
668                         ti->error = "not enough path parameters";
669                         goto bad;
670                 }
671
672                 path_args.argc = nr_params;
673                 path_args.argv = as->argv;
674
675                 pgpath = parse_path(&path_args, &pg->ps, ti);
676                 if (IS_ERR(pgpath)) {
677                         r = PTR_ERR(pgpath);
678                         goto bad;
679                 }
680
681                 pgpath->pg = pg;
682                 list_add_tail(&pgpath->list, &pg->pgpaths);
683                 consume(as, nr_params);
684         }
685
686         return pg;
687
688  bad:
689         free_priority_group(pg, ti);
690         return ERR_PTR(r);
691 }
692
693 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
694 {
695         unsigned hw_argc;
696         struct dm_target *ti = m->ti;
697
698         static struct param _params[] = {
699                 {0, 1024, "invalid number of hardware handler args"},
700         };
701
702         if (read_param(_params, shift(as), &hw_argc, &ti->error))
703                 return -EINVAL;
704
705         if (!hw_argc)
706                 return 0;
707
708         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
709         request_module("scsi_dh_%s", m->hw_handler_name);
710         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
711                 ti->error = "unknown hardware handler type";
712                 kfree(m->hw_handler_name);
713                 m->hw_handler_name = NULL;
714                 return -EINVAL;
715         }
716
717         if (hw_argc > 1)
718                 DMWARN("Ignoring user-specified arguments for "
719                        "hardware handler \"%s\"", m->hw_handler_name);
720         consume(as, hw_argc - 1);
721
722         return 0;
723 }
724
725 static int parse_features(struct arg_set *as, struct multipath *m)
726 {
727         int r;
728         unsigned argc;
729         struct dm_target *ti = m->ti;
730         const char *param_name;
731
732         static struct param _params[] = {
733                 {0, 3, "invalid number of feature args"},
734                 {1, 50, "pg_init_retries must be between 1 and 50"},
735         };
736
737         r = read_param(_params, shift(as), &argc, &ti->error);
738         if (r)
739                 return -EINVAL;
740
741         if (!argc)
742                 return 0;
743
744         do {
745                 param_name = shift(as);
746                 argc--;
747
748                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
749                         r = queue_if_no_path(m, 1, 0);
750                         continue;
751                 }
752
753                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
754                     (argc >= 1)) {
755                         r = read_param(_params + 1, shift(as),
756                                        &m->pg_init_retries, &ti->error);
757                         argc--;
758                         continue;
759                 }
760
761                 ti->error = "Unrecognised multipath feature request";
762                 r = -EINVAL;
763         } while (argc && !r);
764
765         return r;
766 }
767
768 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
769                          char **argv)
770 {
771         /* target parameters */
772         static struct param _params[] = {
773                 {1, 1024, "invalid number of priority groups"},
774                 {1, 1024, "invalid initial priority group number"},
775         };
776
777         int r;
778         struct multipath *m;
779         struct arg_set as;
780         unsigned pg_count = 0;
781         unsigned next_pg_num;
782
783         as.argc = argc;
784         as.argv = argv;
785
786         m = alloc_multipath(ti);
787         if (!m) {
788                 ti->error = "can't allocate multipath";
789                 return -EINVAL;
790         }
791
792         r = parse_features(&as, m);
793         if (r)
794                 goto bad;
795
796         r = parse_hw_handler(&as, m);
797         if (r)
798                 goto bad;
799
800         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
801         if (r)
802                 goto bad;
803
804         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
805         if (r)
806                 goto bad;
807
808         /* parse the priority groups */
809         while (as.argc) {
810                 struct priority_group *pg;
811
812                 pg = parse_priority_group(&as, m);
813                 if (IS_ERR(pg)) {
814                         r = PTR_ERR(pg);
815                         goto bad;
816                 }
817
818                 m->nr_valid_paths += pg->nr_pgpaths;
819                 list_add_tail(&pg->list, &m->priority_groups);
820                 pg_count++;
821                 pg->pg_num = pg_count;
822                 if (!--next_pg_num)
823                         m->next_pg = pg;
824         }
825
826         if (pg_count != m->nr_priority_groups) {
827                 ti->error = "priority group count mismatch";
828                 r = -EINVAL;
829                 goto bad;
830         }
831
832         return 0;
833
834  bad:
835         free_multipath(m);
836         return r;
837 }
838
839 static void multipath_dtr(struct dm_target *ti)
840 {
841         struct multipath *m = (struct multipath *) ti->private;
842
843         flush_workqueue(kmpath_handlerd);
844         flush_workqueue(kmultipathd);
845         free_multipath(m);
846 }
847
848 /*
849  * Map bios, recording original fields for later in case we have to resubmit
850  */
851 static int multipath_map(struct dm_target *ti, struct bio *bio,
852                          union map_info *map_context)
853 {
854         int r;
855         struct dm_mpath_io *mpio;
856         struct multipath *m = (struct multipath *) ti->private;
857
858         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
859         dm_bio_record(&mpio->details, bio);
860
861         map_context->ptr = mpio;
862         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
863         r = map_io(m, bio, mpio, 0);
864         if (r < 0 || r == DM_MAPIO_REQUEUE)
865                 mempool_free(mpio, m->mpio_pool);
866
867         return r;
868 }
869
870 /*
871  * Take a path out of use.
872  */
873 static int fail_path(struct pgpath *pgpath)
874 {
875         unsigned long flags;
876         struct multipath *m = pgpath->pg->m;
877
878         spin_lock_irqsave(&m->lock, flags);
879
880         if (!pgpath->is_active)
881                 goto out;
882
883         DMWARN("Failing path %s.", pgpath->path.dev->name);
884
885         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
886         pgpath->is_active = 0;
887         pgpath->fail_count++;
888
889         m->nr_valid_paths--;
890
891         if (pgpath == m->current_pgpath)
892                 m->current_pgpath = NULL;
893
894         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
895                       pgpath->path.dev->name, m->nr_valid_paths);
896
897         schedule_work(&m->trigger_event);
898         queue_work(kmultipathd, &pgpath->deactivate_path);
899
900 out:
901         spin_unlock_irqrestore(&m->lock, flags);
902
903         return 0;
904 }
905
906 /*
907  * Reinstate a previously-failed path
908  */
909 static int reinstate_path(struct pgpath *pgpath)
910 {
911         int r = 0;
912         unsigned long flags;
913         struct multipath *m = pgpath->pg->m;
914
915         spin_lock_irqsave(&m->lock, flags);
916
917         if (pgpath->is_active)
918                 goto out;
919
920         if (!pgpath->pg->ps.type->reinstate_path) {
921                 DMWARN("Reinstate path not supported by path selector %s",
922                        pgpath->pg->ps.type->name);
923                 r = -EINVAL;
924                 goto out;
925         }
926
927         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
928         if (r)
929                 goto out;
930
931         pgpath->is_active = 1;
932
933         m->current_pgpath = NULL;
934         if (!m->nr_valid_paths++ && m->queue_size)
935                 queue_work(kmultipathd, &m->process_queued_ios);
936
937         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
938                       pgpath->path.dev->name, m->nr_valid_paths);
939
940         schedule_work(&m->trigger_event);
941
942 out:
943         spin_unlock_irqrestore(&m->lock, flags);
944
945         return r;
946 }
947
948 /*
949  * Fail or reinstate all paths that match the provided struct dm_dev.
950  */
951 static int action_dev(struct multipath *m, struct dm_dev *dev,
952                       action_fn action)
953 {
954         int r = 0;
955         struct pgpath *pgpath;
956         struct priority_group *pg;
957
958         list_for_each_entry(pg, &m->priority_groups, list) {
959                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
960                         if (pgpath->path.dev == dev)
961                                 r = action(pgpath);
962                 }
963         }
964
965         return r;
966 }
967
968 /*
969  * Temporarily try to avoid having to use the specified PG
970  */
971 static void bypass_pg(struct multipath *m, struct priority_group *pg,
972                       int bypassed)
973 {
974         unsigned long flags;
975
976         spin_lock_irqsave(&m->lock, flags);
977
978         pg->bypassed = bypassed;
979         m->current_pgpath = NULL;
980         m->current_pg = NULL;
981
982         spin_unlock_irqrestore(&m->lock, flags);
983
984         schedule_work(&m->trigger_event);
985 }
986
987 /*
988  * Switch to using the specified PG from the next I/O that gets mapped
989  */
990 static int switch_pg_num(struct multipath *m, const char *pgstr)
991 {
992         struct priority_group *pg;
993         unsigned pgnum;
994         unsigned long flags;
995
996         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
997             (pgnum > m->nr_priority_groups)) {
998                 DMWARN("invalid PG number supplied to switch_pg_num");
999                 return -EINVAL;
1000         }
1001
1002         spin_lock_irqsave(&m->lock, flags);
1003         list_for_each_entry(pg, &m->priority_groups, list) {
1004                 pg->bypassed = 0;
1005                 if (--pgnum)
1006                         continue;
1007
1008                 m->current_pgpath = NULL;
1009                 m->current_pg = NULL;
1010                 m->next_pg = pg;
1011         }
1012         spin_unlock_irqrestore(&m->lock, flags);
1013
1014         schedule_work(&m->trigger_event);
1015         return 0;
1016 }
1017
1018 /*
1019  * Set/clear bypassed status of a PG.
1020  * PGs are numbered upwards from 1 in the order they were declared.
1021  */
1022 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1023 {
1024         struct priority_group *pg;
1025         unsigned pgnum;
1026
1027         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1028             (pgnum > m->nr_priority_groups)) {
1029                 DMWARN("invalid PG number supplied to bypass_pg");
1030                 return -EINVAL;
1031         }
1032
1033         list_for_each_entry(pg, &m->priority_groups, list) {
1034                 if (!--pgnum)
1035                         break;
1036         }
1037
1038         bypass_pg(m, pg, bypassed);
1039         return 0;
1040 }
1041
1042 /*
1043  * Should we retry pg_init immediately?
1044  */
1045 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1046 {
1047         unsigned long flags;
1048         int limit_reached = 0;
1049
1050         spin_lock_irqsave(&m->lock, flags);
1051
1052         if (m->pg_init_count <= m->pg_init_retries)
1053                 m->pg_init_required = 1;
1054         else
1055                 limit_reached = 1;
1056
1057         spin_unlock_irqrestore(&m->lock, flags);
1058
1059         return limit_reached;
1060 }
1061
1062 static void pg_init_done(struct dm_path *path, int errors)
1063 {
1064         struct pgpath *pgpath = path_to_pgpath(path);
1065         struct priority_group *pg = pgpath->pg;
1066         struct multipath *m = pg->m;
1067         unsigned long flags;
1068
1069         /* device or driver problems */
1070         switch (errors) {
1071         case SCSI_DH_OK:
1072                 break;
1073         case SCSI_DH_NOSYS:
1074                 if (!m->hw_handler_name) {
1075                         errors = 0;
1076                         break;
1077                 }
1078                 DMERR("Cannot failover device because scsi_dh_%s was not "
1079                       "loaded.", m->hw_handler_name);
1080                 /*
1081                  * Fail path for now, so we do not ping pong
1082                  */
1083                 fail_path(pgpath);
1084                 break;
1085         case SCSI_DH_DEV_TEMP_BUSY:
1086                 /*
1087                  * Probably doing something like FW upgrade on the
1088                  * controller so try the other pg.
1089                  */
1090                 bypass_pg(m, pg, 1);
1091                 break;
1092         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1093         case SCSI_DH_RETRY:
1094         case SCSI_DH_IMM_RETRY:
1095         case SCSI_DH_RES_TEMP_UNAVAIL:
1096                 if (pg_init_limit_reached(m, pgpath))
1097                         fail_path(pgpath);
1098                 errors = 0;
1099                 break;
1100         default:
1101                 /*
1102                  * We probably do not want to fail the path for a device
1103                  * error, but this is what the old dm did. In future
1104                  * patches we can do more advanced handling.
1105                  */
1106                 fail_path(pgpath);
1107         }
1108
1109         spin_lock_irqsave(&m->lock, flags);
1110         if (errors) {
1111                 DMERR("Could not failover device. Error %d.", errors);
1112                 m->current_pgpath = NULL;
1113                 m->current_pg = NULL;
1114         } else if (!m->pg_init_required) {
1115                 m->queue_io = 0;
1116                 pg->bypassed = 0;
1117         }
1118
1119         m->pg_init_in_progress = 0;
1120         queue_work(kmultipathd, &m->process_queued_ios);
1121         spin_unlock_irqrestore(&m->lock, flags);
1122 }
1123
1124 static void activate_path(struct work_struct *work)
1125 {
1126         int ret;
1127         struct multipath *m =
1128                 container_of(work, struct multipath, activate_path);
1129         struct dm_path *path;
1130         unsigned long flags;
1131
1132         spin_lock_irqsave(&m->lock, flags);
1133         path = &m->pgpath_to_activate->path;
1134         m->pgpath_to_activate = NULL;
1135         spin_unlock_irqrestore(&m->lock, flags);
1136         if (!path)
1137                 return;
1138         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1139         pg_init_done(path, ret);
1140 }
1141
1142 /*
1143  * end_io handling
1144  */
1145 static int do_end_io(struct multipath *m, struct bio *bio,
1146                      int error, struct dm_mpath_io *mpio)
1147 {
1148         unsigned long flags;
1149
1150         if (!error)
1151                 return 0;       /* I/O complete */
1152
1153         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1154                 return error;
1155
1156         if (error == -EOPNOTSUPP)
1157                 return error;
1158
1159         spin_lock_irqsave(&m->lock, flags);
1160         if (!m->nr_valid_paths) {
1161                 if (__must_push_back(m)) {
1162                         spin_unlock_irqrestore(&m->lock, flags);
1163                         return DM_ENDIO_REQUEUE;
1164                 } else if (!m->queue_if_no_path) {
1165                         spin_unlock_irqrestore(&m->lock, flags);
1166                         return -EIO;
1167                 } else {
1168                         spin_unlock_irqrestore(&m->lock, flags);
1169                         goto requeue;
1170                 }
1171         }
1172         spin_unlock_irqrestore(&m->lock, flags);
1173
1174         if (mpio->pgpath)
1175                 fail_path(mpio->pgpath);
1176
1177       requeue:
1178         dm_bio_restore(&mpio->details, bio);
1179
1180         /* queue for the daemon to resubmit or fail */
1181         spin_lock_irqsave(&m->lock, flags);
1182         bio_list_add(&m->queued_ios, bio);
1183         m->queue_size++;
1184         if (!m->queue_io)
1185                 queue_work(kmultipathd, &m->process_queued_ios);
1186         spin_unlock_irqrestore(&m->lock, flags);
1187
1188         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1189 }
1190
1191 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1192                             int error, union map_info *map_context)
1193 {
1194         struct multipath *m = ti->private;
1195         struct dm_mpath_io *mpio = map_context->ptr;
1196         struct pgpath *pgpath = mpio->pgpath;
1197         struct path_selector *ps;
1198         int r;
1199
1200         r  = do_end_io(m, bio, error, mpio);
1201         if (pgpath) {
1202                 ps = &pgpath->pg->ps;
1203                 if (ps->type->end_io)
1204                         ps->type->end_io(ps, &pgpath->path);
1205         }
1206         if (r != DM_ENDIO_INCOMPLETE)
1207                 mempool_free(mpio, m->mpio_pool);
1208
1209         return r;
1210 }
1211
1212 /*
1213  * Suspend can't complete until all the I/O is processed so if
1214  * the last path fails we must error any remaining I/O.
1215  * Note that if the freeze_bdev fails while suspending, the
1216  * queue_if_no_path state is lost - userspace should reset it.
1217  */
1218 static void multipath_presuspend(struct dm_target *ti)
1219 {
1220         struct multipath *m = (struct multipath *) ti->private;
1221
1222         queue_if_no_path(m, 0, 1);
1223 }
1224
1225 /*
1226  * Restore the queue_if_no_path setting.
1227  */
1228 static void multipath_resume(struct dm_target *ti)
1229 {
1230         struct multipath *m = (struct multipath *) ti->private;
1231         unsigned long flags;
1232
1233         spin_lock_irqsave(&m->lock, flags);
1234         m->queue_if_no_path = m->saved_queue_if_no_path;
1235         spin_unlock_irqrestore(&m->lock, flags);
1236 }
1237
1238 /*
1239  * Info output has the following format:
1240  * num_multipath_feature_args [multipath_feature_args]*
1241  * num_handler_status_args [handler_status_args]*
1242  * num_groups init_group_number
1243  *            [A|D|E num_ps_status_args [ps_status_args]*
1244  *             num_paths num_selector_args
1245  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1246  *
1247  * Table output has the following format (identical to the constructor string):
1248  * num_feature_args [features_args]*
1249  * num_handler_args hw_handler [hw_handler_args]*
1250  * num_groups init_group_number
1251  *     [priority selector-name num_ps_args [ps_args]*
1252  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1253  */
1254 static int multipath_status(struct dm_target *ti, status_type_t type,
1255                             char *result, unsigned int maxlen)
1256 {
1257         int sz = 0;
1258         unsigned long flags;
1259         struct multipath *m = (struct multipath *) ti->private;
1260         struct priority_group *pg;
1261         struct pgpath *p;
1262         unsigned pg_num;
1263         char state;
1264
1265         spin_lock_irqsave(&m->lock, flags);
1266
1267         /* Features */
1268         if (type == STATUSTYPE_INFO)
1269                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1270         else {
1271                 DMEMIT("%u ", m->queue_if_no_path +
1272                               (m->pg_init_retries > 0) * 2);
1273                 if (m->queue_if_no_path)
1274                         DMEMIT("queue_if_no_path ");
1275                 if (m->pg_init_retries)
1276                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1277         }
1278
1279         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1280                 DMEMIT("0 ");
1281         else
1282                 DMEMIT("1 %s ", m->hw_handler_name);
1283
1284         DMEMIT("%u ", m->nr_priority_groups);
1285
1286         if (m->next_pg)
1287                 pg_num = m->next_pg->pg_num;
1288         else if (m->current_pg)
1289                 pg_num = m->current_pg->pg_num;
1290         else
1291                         pg_num = 1;
1292
1293         DMEMIT("%u ", pg_num);
1294
1295         switch (type) {
1296         case STATUSTYPE_INFO:
1297                 list_for_each_entry(pg, &m->priority_groups, list) {
1298                         if (pg->bypassed)
1299                                 state = 'D';    /* Disabled */
1300                         else if (pg == m->current_pg)
1301                                 state = 'A';    /* Currently Active */
1302                         else
1303                                 state = 'E';    /* Enabled */
1304
1305                         DMEMIT("%c ", state);
1306
1307                         if (pg->ps.type->status)
1308                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1309                                                           result + sz,
1310                                                           maxlen - sz);
1311                         else
1312                                 DMEMIT("0 ");
1313
1314                         DMEMIT("%u %u ", pg->nr_pgpaths,
1315                                pg->ps.type->info_args);
1316
1317                         list_for_each_entry(p, &pg->pgpaths, list) {
1318                                 DMEMIT("%s %s %u ", p->path.dev->name,
1319                                        p->is_active ? "A" : "F",
1320                                        p->fail_count);
1321                                 if (pg->ps.type->status)
1322                                         sz += pg->ps.type->status(&pg->ps,
1323                                               &p->path, type, result + sz,
1324                                               maxlen - sz);
1325                         }
1326                 }
1327                 break;
1328
1329         case STATUSTYPE_TABLE:
1330                 list_for_each_entry(pg, &m->priority_groups, list) {
1331                         DMEMIT("%s ", pg->ps.type->name);
1332
1333                         if (pg->ps.type->status)
1334                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1335                                                           result + sz,
1336                                                           maxlen - sz);
1337                         else
1338                                 DMEMIT("0 ");
1339
1340                         DMEMIT("%u %u ", pg->nr_pgpaths,
1341                                pg->ps.type->table_args);
1342
1343                         list_for_each_entry(p, &pg->pgpaths, list) {
1344                                 DMEMIT("%s ", p->path.dev->name);
1345                                 if (pg->ps.type->status)
1346                                         sz += pg->ps.type->status(&pg->ps,
1347                                               &p->path, type, result + sz,
1348                                               maxlen - sz);
1349                         }
1350                 }
1351                 break;
1352         }
1353
1354         spin_unlock_irqrestore(&m->lock, flags);
1355
1356         return 0;
1357 }
1358
1359 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1360 {
1361         int r;
1362         struct dm_dev *dev;
1363         struct multipath *m = (struct multipath *) ti->private;
1364         action_fn action;
1365
1366         if (argc == 1) {
1367                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1368                         return queue_if_no_path(m, 1, 0);
1369                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1370                         return queue_if_no_path(m, 0, 0);
1371         }
1372
1373         if (argc != 2)
1374                 goto error;
1375
1376         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1377                 return bypass_pg_num(m, argv[1], 1);
1378         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1379                 return bypass_pg_num(m, argv[1], 0);
1380         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1381                 return switch_pg_num(m, argv[1]);
1382         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1383                 action = reinstate_path;
1384         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1385                 action = fail_path;
1386         else
1387                 goto error;
1388
1389         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1390                           dm_table_get_mode(ti->table), &dev);
1391         if (r) {
1392                 DMWARN("message: error getting device %s",
1393                        argv[1]);
1394                 return -EINVAL;
1395         }
1396
1397         r = action_dev(m, dev, action);
1398
1399         dm_put_device(ti, dev);
1400
1401         return r;
1402
1403 error:
1404         DMWARN("Unrecognised multipath message received.");
1405         return -EINVAL;
1406 }
1407
1408 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1409                            unsigned long arg)
1410 {
1411         struct multipath *m = (struct multipath *) ti->private;
1412         struct block_device *bdev = NULL;
1413         fmode_t mode = 0;
1414         unsigned long flags;
1415         int r = 0;
1416
1417         spin_lock_irqsave(&m->lock, flags);
1418
1419         if (!m->current_pgpath)
1420                 __choose_pgpath(m);
1421
1422         if (m->current_pgpath) {
1423                 bdev = m->current_pgpath->path.dev->bdev;
1424                 mode = m->current_pgpath->path.dev->mode;
1425         }
1426
1427         if (m->queue_io)
1428                 r = -EAGAIN;
1429         else if (!bdev)
1430                 r = -EIO;
1431
1432         spin_unlock_irqrestore(&m->lock, flags);
1433
1434         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1435 }
1436
1437 /*-----------------------------------------------------------------
1438  * Module setup
1439  *---------------------------------------------------------------*/
1440 static struct target_type multipath_target = {
1441         .name = "multipath",
1442         .version = {1, 0, 5},
1443         .module = THIS_MODULE,
1444         .ctr = multipath_ctr,
1445         .dtr = multipath_dtr,
1446         .map = multipath_map,
1447         .end_io = multipath_end_io,
1448         .presuspend = multipath_presuspend,
1449         .resume = multipath_resume,
1450         .status = multipath_status,
1451         .message = multipath_message,
1452         .ioctl  = multipath_ioctl,
1453 };
1454
1455 static int __init dm_multipath_init(void)
1456 {
1457         int r;
1458
1459         /* allocate a slab for the dm_ios */
1460         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1461         if (!_mpio_cache)
1462                 return -ENOMEM;
1463
1464         r = dm_register_target(&multipath_target);
1465         if (r < 0) {
1466                 DMERR("register failed %d", r);
1467                 kmem_cache_destroy(_mpio_cache);
1468                 return -EINVAL;
1469         }
1470
1471         kmultipathd = create_workqueue("kmpathd");
1472         if (!kmultipathd) {
1473                 DMERR("failed to create workqueue kmpathd");
1474                 dm_unregister_target(&multipath_target);
1475                 kmem_cache_destroy(_mpio_cache);
1476                 return -ENOMEM;
1477         }
1478
1479         /*
1480          * A separate workqueue is used to handle the device handlers
1481          * to avoid overloading existing workqueue. Overloading the
1482          * old workqueue would also create a bottleneck in the
1483          * path of the storage hardware device activation.
1484          */
1485         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1486         if (!kmpath_handlerd) {
1487                 DMERR("failed to create workqueue kmpath_handlerd");
1488                 destroy_workqueue(kmultipathd);
1489                 dm_unregister_target(&multipath_target);
1490                 kmem_cache_destroy(_mpio_cache);
1491                 return -ENOMEM;
1492         }
1493
1494         DMINFO("version %u.%u.%u loaded",
1495                multipath_target.version[0], multipath_target.version[1],
1496                multipath_target.version[2]);
1497
1498         return r;
1499 }
1500
1501 static void __exit dm_multipath_exit(void)
1502 {
1503         destroy_workqueue(kmpath_handlerd);
1504         destroy_workqueue(kmultipathd);
1505
1506         dm_unregister_target(&multipath_target);
1507         kmem_cache_destroy(_mpio_cache);
1508 }
1509
1510 module_init(dm_multipath_init);
1511 module_exit(dm_multipath_exit);
1512
1513 MODULE_DESCRIPTION(DM_NAME " multipath target");
1514 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1515 MODULE_LICENSE("GPL");