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