]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - drivers/md/dm-mpath.c
Merge branch 'drm-fixes' of git://people.freedesktop.org/~airlied/linux
[mirror_ubuntu-artful-kernel.git] / drivers / md / dm-mpath.c
CommitLineData
1da177e4
LT
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
586e80e6
MP
8#include <linux/device-mapper.h>
9
f4790826 10#include "dm.h"
1da177e4 11#include "dm-path-selector.h"
b15546f9 12#include "dm-uevent.h"
1da177e4
LT
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>
35991652 22#include <linux/delay.h>
cfae5c9b 23#include <scsi/scsi_dh.h>
60063497 24#include <linux/atomic.h>
1da177e4 25
72d94861 26#define DM_MSG_PREFIX "multipath"
4e2d19e4
CS
27#define DM_PG_INIT_DELAY_MSECS 2000
28#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
1da177e4
LT
29
30/* Path properties */
31struct pgpath {
32 struct list_head list;
33
34 struct priority_group *pg; /* Owning PG */
6680073d 35 unsigned is_active; /* Path status */
1da177e4
LT
36 unsigned fail_count; /* Cumulative failure count */
37
c922d5f7 38 struct dm_path path;
4e2d19e4 39 struct delayed_work activate_path;
1da177e4
LT
40};
41
42#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
43
44/*
45 * Paths are grouped into Priority Groups and numbered from 1 upwards.
46 * Each has a path selector which controls which path gets used.
47 */
48struct priority_group {
49 struct list_head list;
50
51 struct multipath *m; /* Owning multipath instance */
52 struct path_selector ps;
53
54 unsigned pg_num; /* Reference number */
55 unsigned bypassed; /* Temporarily bypass this PG? */
56
57 unsigned nr_pgpaths; /* Number of paths in PG */
58 struct list_head pgpaths;
59};
60
61/* Multipath context */
62struct multipath {
63 struct list_head list;
64 struct dm_target *ti;
65
cfae5c9b 66 const char *hw_handler_name;
2bfd2e13 67 char *hw_handler_params;
4e2d19e4 68
1fbdd2b3
MS
69 spinlock_t lock;
70
1da177e4
LT
71 unsigned nr_priority_groups;
72 struct list_head priority_groups;
4e2d19e4
CS
73
74 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
75
1da177e4 76 unsigned pg_init_required; /* pg_init needs calling? */
c3cd4f6b 77 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
4e2d19e4 78 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
1da177e4
LT
79
80 unsigned nr_valid_paths; /* Total number of usable paths */
81 struct pgpath *current_pgpath;
82 struct priority_group *current_pg;
83 struct priority_group *next_pg; /* Switch to this PG if set */
84 unsigned repeat_count; /* I/Os left before calling PS again */
85
1fbdd2b3
MS
86 unsigned queue_io:1; /* Must we queue all I/O? */
87 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
88 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
a58a935d 89 unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
954a73d5 90 unsigned pg_init_disabled:1; /* pg_init is not currently allowed */
1fbdd2b3 91
c9e45581
DW
92 unsigned pg_init_retries; /* Number of times to retry pg_init */
93 unsigned pg_init_count; /* Number of times pg_init called */
4e2d19e4 94 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
1da177e4 95
1da177e4
LT
96 struct work_struct trigger_event;
97
98 /*
028867ac 99 * We must use a mempool of dm_mpath_io structs so that we
1da177e4
LT
100 * can resubmit bios on error.
101 */
102 mempool_t *mpio_pool;
6380f26f
MA
103
104 struct mutex work_mutex;
1da177e4
LT
105};
106
107/*
108 * Context information attached to each bio we process.
109 */
028867ac 110struct dm_mpath_io {
1da177e4 111 struct pgpath *pgpath;
02ab823f 112 size_t nr_bytes;
1da177e4
LT
113};
114
115typedef int (*action_fn) (struct pgpath *pgpath);
116
e18b890b 117static struct kmem_cache *_mpio_cache;
1da177e4 118
bab7cfc7 119static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
c4028958 120static void trigger_event(struct work_struct *work);
bab7cfc7 121static void activate_path(struct work_struct *work);
e8099177 122static int __pgpath_busy(struct pgpath *pgpath);
1da177e4
LT
123
124
125/*-----------------------------------------------
126 * Allocation routines
127 *-----------------------------------------------*/
128
129static struct pgpath *alloc_pgpath(void)
130{
e69fae56 131 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
1da177e4 132
224cb3e9 133 if (pgpath) {
6680073d 134 pgpath->is_active = 1;
4e2d19e4 135 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
224cb3e9 136 }
1da177e4
LT
137
138 return pgpath;
139}
140
028867ac 141static void free_pgpath(struct pgpath *pgpath)
1da177e4
LT
142{
143 kfree(pgpath);
144}
145
146static struct priority_group *alloc_priority_group(void)
147{
148 struct priority_group *pg;
149
e69fae56 150 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
1da177e4 151
e69fae56
MM
152 if (pg)
153 INIT_LIST_HEAD(&pg->pgpaths);
1da177e4
LT
154
155 return pg;
156}
157
158static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
159{
160 struct pgpath *pgpath, *tmp;
ae11b1b3 161 struct multipath *m = ti->private;
1da177e4
LT
162
163 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
164 list_del(&pgpath->list);
ae11b1b3
HR
165 if (m->hw_handler_name)
166 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
1da177e4
LT
167 dm_put_device(ti, pgpath->path.dev);
168 free_pgpath(pgpath);
169 }
170}
171
172static void free_priority_group(struct priority_group *pg,
173 struct dm_target *ti)
174{
175 struct path_selector *ps = &pg->ps;
176
177 if (ps->type) {
178 ps->type->destroy(ps);
179 dm_put_path_selector(ps->type);
180 }
181
182 free_pgpaths(&pg->pgpaths, ti);
183 kfree(pg);
184}
185
28f16c20 186static struct multipath *alloc_multipath(struct dm_target *ti)
1da177e4
LT
187{
188 struct multipath *m;
f4790826 189 unsigned min_ios = dm_get_reserved_rq_based_ios();
1da177e4 190
e69fae56 191 m = kzalloc(sizeof(*m), GFP_KERNEL);
1da177e4 192 if (m) {
1da177e4
LT
193 INIT_LIST_HEAD(&m->priority_groups);
194 spin_lock_init(&m->lock);
195 m->queue_io = 1;
4e2d19e4 196 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
c4028958 197 INIT_WORK(&m->trigger_event, trigger_event);
2bded7bd 198 init_waitqueue_head(&m->pg_init_wait);
6380f26f 199 mutex_init(&m->work_mutex);
f4790826 200 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
1da177e4
LT
201 if (!m->mpio_pool) {
202 kfree(m);
203 return NULL;
204 }
28f16c20
MM
205 m->ti = ti;
206 ti->private = m;
1da177e4
LT
207 }
208
209 return m;
210}
211
212static void free_multipath(struct multipath *m)
213{
214 struct priority_group *pg, *tmp;
1da177e4
LT
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
cfae5c9b 221 kfree(m->hw_handler_name);
2bfd2e13 222 kfree(m->hw_handler_params);
1da177e4
LT
223 mempool_destroy(m->mpio_pool);
224 kfree(m);
225}
226
466891f9
JN
227static int set_mapinfo(struct multipath *m, union map_info *info)
228{
229 struct dm_mpath_io *mpio;
230
231 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
232 if (!mpio)
233 return -ENOMEM;
234
235 memset(mpio, 0, sizeof(*mpio));
236 info->ptr = mpio;
237
238 return 0;
239}
240
241static void clear_mapinfo(struct multipath *m, union map_info *info)
242{
243 struct dm_mpath_io *mpio = info->ptr;
244
245 info->ptr = NULL;
246 mempool_free(mpio, m->mpio_pool);
247}
1da177e4
LT
248
249/*-----------------------------------------------
250 * Path selection
251 *-----------------------------------------------*/
252
3e9f1be1 253static int __pg_init_all_paths(struct multipath *m)
fb612642
KU
254{
255 struct pgpath *pgpath;
4e2d19e4 256 unsigned long pg_init_delay = 0;
fb612642 257
17f4ff45 258 if (m->pg_init_in_progress || m->pg_init_disabled)
3e9f1be1 259 return 0;
17f4ff45 260
fb612642
KU
261 m->pg_init_count++;
262 m->pg_init_required = 0;
3e9f1be1
HR
263
264 /* Check here to reset pg_init_required */
265 if (!m->current_pg)
266 return 0;
267
4e2d19e4
CS
268 if (m->pg_init_delay_retry)
269 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
270 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
fb612642
KU
271 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
272 /* Skip failed paths */
273 if (!pgpath->is_active)
274 continue;
4e2d19e4
CS
275 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
276 pg_init_delay))
fb612642
KU
277 m->pg_init_in_progress++;
278 }
3e9f1be1 279 return m->pg_init_in_progress;
fb612642
KU
280}
281
1da177e4
LT
282static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
283{
1da177e4
LT
284 m->current_pg = pgpath->pg;
285
286 /* Must we initialise the PG first, and queue I/O till it's ready? */
cfae5c9b 287 if (m->hw_handler_name) {
1da177e4
LT
288 m->pg_init_required = 1;
289 m->queue_io = 1;
290 } else {
291 m->pg_init_required = 0;
292 m->queue_io = 0;
293 }
c9e45581
DW
294
295 m->pg_init_count = 0;
1da177e4
LT
296}
297
02ab823f
KU
298static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
299 size_t nr_bytes)
1da177e4 300{
c922d5f7 301 struct dm_path *path;
1da177e4 302
02ab823f 303 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
1da177e4
LT
304 if (!path)
305 return -ENXIO;
306
307 m->current_pgpath = path_to_pgpath(path);
308
309 if (m->current_pg != pg)
310 __switch_pg(m, m->current_pgpath);
311
312 return 0;
313}
314
02ab823f 315static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
1da177e4
LT
316{
317 struct priority_group *pg;
318 unsigned bypassed = 1;
319
320 if (!m->nr_valid_paths)
321 goto failed;
322
323 /* Were we instructed to switch PG? */
324 if (m->next_pg) {
325 pg = m->next_pg;
326 m->next_pg = NULL;
02ab823f 327 if (!__choose_path_in_pg(m, pg, nr_bytes))
1da177e4
LT
328 return;
329 }
330
331 /* Don't change PG until it has no remaining paths */
02ab823f 332 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
1da177e4
LT
333 return;
334
335 /*
336 * Loop through priority groups until we find a valid path.
337 * First time we skip PGs marked 'bypassed'.
f220fd4e
MC
338 * Second time we only try the ones we skipped, but set
339 * pg_init_delay_retry so we do not hammer controllers.
1da177e4
LT
340 */
341 do {
342 list_for_each_entry(pg, &m->priority_groups, list) {
343 if (pg->bypassed == bypassed)
344 continue;
f220fd4e
MC
345 if (!__choose_path_in_pg(m, pg, nr_bytes)) {
346 if (!bypassed)
347 m->pg_init_delay_retry = 1;
1da177e4 348 return;
f220fd4e 349 }
1da177e4
LT
350 }
351 } while (bypassed--);
352
353failed:
354 m->current_pgpath = NULL;
355 m->current_pg = NULL;
356}
357
45e15720
KU
358/*
359 * Check whether bios must be queued in the device-mapper core rather
360 * than here in the target.
361 *
362 * m->lock must be held on entry.
363 *
364 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
365 * same value then we are not between multipath_presuspend()
366 * and multipath_resume() calls and we have no need to check
367 * for the DMF_NOFLUSH_SUSPENDING flag.
368 */
369static int __must_push_back(struct multipath *m)
370{
e8099177
HR
371 return (m->queue_if_no_path ||
372 (m->queue_if_no_path != m->saved_queue_if_no_path &&
373 dm_noflush_suspending(m->ti)));
45e15720
KU
374}
375
36fcffcc
HR
376/*
377 * Map cloned requests
378 */
379static int multipath_map(struct dm_target *ti, struct request *clone,
380 union map_info *map_context)
1da177e4 381{
36fcffcc 382 struct multipath *m = (struct multipath *) ti->private;
e3bde04f 383 int r = DM_MAPIO_REQUEUE;
f40c67f0 384 size_t nr_bytes = blk_rq_bytes(clone);
1da177e4
LT
385 unsigned long flags;
386 struct pgpath *pgpath;
f40c67f0 387 struct block_device *bdev;
e3bde04f 388 struct dm_mpath_io *mpio;
1da177e4
LT
389
390 spin_lock_irqsave(&m->lock, flags);
391
392 /* Do we need to select a new pgpath? */
393 if (!m->current_pgpath ||
394 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
02ab823f 395 __choose_pgpath(m, nr_bytes);
1da177e4
LT
396
397 pgpath = m->current_pgpath;
398
9bf59a61
MS
399 if (!pgpath) {
400 if (!__must_push_back(m))
401 r = -EIO; /* Failed */
402 goto out_unlock;
6afbc01d 403 } else if (m->queue_io || m->pg_init_required) {
e3bde04f 404 __pg_init_all_paths(m);
9bf59a61
MS
405 goto out_unlock;
406 }
6afbc01d 407
9bf59a61
MS
408 if (set_mapinfo(m, map_context) < 0)
409 /* ENOMEM, requeue */
410 goto out_unlock;
411
412 bdev = pgpath->path.dev->bdev;
413 clone->q = bdev_get_queue(bdev);
414 clone->rq_disk = bdev->bd_disk;
415 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
416 mpio = map_context->ptr;
417 mpio->pgpath = pgpath;
418 mpio->nr_bytes = nr_bytes;
419 if (pgpath->pg->ps.type->start_io)
420 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
421 &pgpath->path,
422 nr_bytes);
423 r = DM_MAPIO_REMAPPED;
1da177e4 424
e3bde04f 425out_unlock:
1da177e4
LT
426 spin_unlock_irqrestore(&m->lock, flags);
427
428 return r;
429}
430
431/*
432 * If we run out of usable paths, should we queue I/O or error it?
433 */
485ef69e
AK
434static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
435 unsigned save_old_value)
1da177e4
LT
436{
437 unsigned long flags;
438
439 spin_lock_irqsave(&m->lock, flags);
440
485ef69e
AK
441 if (save_old_value)
442 m->saved_queue_if_no_path = m->queue_if_no_path;
443 else
444 m->saved_queue_if_no_path = queue_if_no_path;
1da177e4 445 m->queue_if_no_path = queue_if_no_path;
1da177e4
LT
446 spin_unlock_irqrestore(&m->lock, flags);
447
63d832c3
HR
448 if (!queue_if_no_path)
449 dm_table_run_md_queue_async(m->ti->table);
450
1da177e4
LT
451 return 0;
452}
453
1da177e4
LT
454/*
455 * An event is triggered whenever a path is taken out of use.
456 * Includes path failure and PG bypass.
457 */
c4028958 458static void trigger_event(struct work_struct *work)
1da177e4 459{
c4028958
DH
460 struct multipath *m =
461 container_of(work, struct multipath, trigger_event);
1da177e4
LT
462
463 dm_table_event(m->ti->table);
464}
465
466/*-----------------------------------------------------------------
467 * Constructor/argument parsing:
468 * <#multipath feature args> [<arg>]*
469 * <#hw_handler args> [hw_handler [<arg>]*]
470 * <#priority groups>
471 * <initial priority group>
472 * [<selector> <#selector args> [<arg>]*
473 * <#paths> <#per-path selector args>
474 * [<path> [<arg>]* ]+ ]+
475 *---------------------------------------------------------------*/
498f0103 476static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
1da177e4
LT
477 struct dm_target *ti)
478{
479 int r;
480 struct path_selector_type *pst;
481 unsigned ps_argc;
482
498f0103 483 static struct dm_arg _args[] = {
72d94861 484 {0, 1024, "invalid number of path selector args"},
1da177e4
LT
485 };
486
498f0103 487 pst = dm_get_path_selector(dm_shift_arg(as));
1da177e4 488 if (!pst) {
72d94861 489 ti->error = "unknown path selector type";
1da177e4
LT
490 return -EINVAL;
491 }
492
498f0103 493 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
371b2e34
MP
494 if (r) {
495 dm_put_path_selector(pst);
1da177e4 496 return -EINVAL;
371b2e34 497 }
1da177e4
LT
498
499 r = pst->create(&pg->ps, ps_argc, as->argv);
500 if (r) {
501 dm_put_path_selector(pst);
72d94861 502 ti->error = "path selector constructor failed";
1da177e4
LT
503 return r;
504 }
505
506 pg->ps.type = pst;
498f0103 507 dm_consume_args(as, ps_argc);
1da177e4
LT
508
509 return 0;
510}
511
498f0103 512static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
1da177e4
LT
513 struct dm_target *ti)
514{
515 int r;
516 struct pgpath *p;
ae11b1b3 517 struct multipath *m = ti->private;
a58a935d
MS
518 struct request_queue *q = NULL;
519 const char *attached_handler_name;
1da177e4
LT
520
521 /* we need at least a path arg */
522 if (as->argc < 1) {
72d94861 523 ti->error = "no device given";
01460f35 524 return ERR_PTR(-EINVAL);
1da177e4
LT
525 }
526
527 p = alloc_pgpath();
528 if (!p)
01460f35 529 return ERR_PTR(-ENOMEM);
1da177e4 530
498f0103 531 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
8215d6ec 532 &p->path.dev);
1da177e4 533 if (r) {
72d94861 534 ti->error = "error getting device";
1da177e4
LT
535 goto bad;
536 }
537
a58a935d
MS
538 if (m->retain_attached_hw_handler || m->hw_handler_name)
539 q = bdev_get_queue(p->path.dev->bdev);
540
541 if (m->retain_attached_hw_handler) {
542 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
543 if (attached_handler_name) {
544 /*
545 * Reset hw_handler_name to match the attached handler
546 * and clear any hw_handler_params associated with the
547 * ignored handler.
548 *
549 * NB. This modifies the table line to show the actual
550 * handler instead of the original table passed in.
551 */
552 kfree(m->hw_handler_name);
553 m->hw_handler_name = attached_handler_name;
554
555 kfree(m->hw_handler_params);
556 m->hw_handler_params = NULL;
557 }
558 }
a0cf7ea9 559
a58a935d
MS
560 if (m->hw_handler_name) {
561 /*
562 * Increments scsi_dh reference, even when using an
563 * already-attached handler.
564 */
a0cf7ea9
HR
565 r = scsi_dh_attach(q, m->hw_handler_name);
566 if (r == -EBUSY) {
567 /*
a58a935d 568 * Already attached to different hw_handler:
a0cf7ea9
HR
569 * try to reattach with correct one.
570 */
571 scsi_dh_detach(q);
572 r = scsi_dh_attach(q, m->hw_handler_name);
573 }
574
ae11b1b3 575 if (r < 0) {
a0cf7ea9 576 ti->error = "error attaching hardware handler";
ae11b1b3
HR
577 dm_put_device(ti, p->path.dev);
578 goto bad;
579 }
2bfd2e13
CS
580
581 if (m->hw_handler_params) {
582 r = scsi_dh_set_params(q, m->hw_handler_params);
583 if (r < 0) {
584 ti->error = "unable to set hardware "
585 "handler parameters";
586 scsi_dh_detach(q);
587 dm_put_device(ti, p->path.dev);
588 goto bad;
589 }
590 }
ae11b1b3
HR
591 }
592
1da177e4
LT
593 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
594 if (r) {
595 dm_put_device(ti, p->path.dev);
596 goto bad;
597 }
598
599 return p;
600
601 bad:
602 free_pgpath(p);
01460f35 603 return ERR_PTR(r);
1da177e4
LT
604}
605
498f0103 606static struct priority_group *parse_priority_group(struct dm_arg_set *as,
28f16c20 607 struct multipath *m)
1da177e4 608{
498f0103 609 static struct dm_arg _args[] = {
72d94861
AK
610 {1, 1024, "invalid number of paths"},
611 {0, 1024, "invalid number of selector args"}
1da177e4
LT
612 };
613
614 int r;
498f0103 615 unsigned i, nr_selector_args, nr_args;
1da177e4 616 struct priority_group *pg;
28f16c20 617 struct dm_target *ti = m->ti;
1da177e4
LT
618
619 if (as->argc < 2) {
620 as->argc = 0;
01460f35
BM
621 ti->error = "not enough priority group arguments";
622 return ERR_PTR(-EINVAL);
1da177e4
LT
623 }
624
625 pg = alloc_priority_group();
626 if (!pg) {
72d94861 627 ti->error = "couldn't allocate priority group";
01460f35 628 return ERR_PTR(-ENOMEM);
1da177e4
LT
629 }
630 pg->m = m;
631
632 r = parse_path_selector(as, pg, ti);
633 if (r)
634 goto bad;
635
636 /*
637 * read the paths
638 */
498f0103 639 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1da177e4
LT
640 if (r)
641 goto bad;
642
498f0103 643 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1da177e4
LT
644 if (r)
645 goto bad;
646
498f0103 647 nr_args = 1 + nr_selector_args;
1da177e4
LT
648 for (i = 0; i < pg->nr_pgpaths; i++) {
649 struct pgpath *pgpath;
498f0103 650 struct dm_arg_set path_args;
1da177e4 651
498f0103 652 if (as->argc < nr_args) {
148acff6 653 ti->error = "not enough path parameters";
6bbf79a1 654 r = -EINVAL;
1da177e4 655 goto bad;
148acff6 656 }
1da177e4 657
498f0103 658 path_args.argc = nr_args;
1da177e4
LT
659 path_args.argv = as->argv;
660
661 pgpath = parse_path(&path_args, &pg->ps, ti);
01460f35
BM
662 if (IS_ERR(pgpath)) {
663 r = PTR_ERR(pgpath);
1da177e4 664 goto bad;
01460f35 665 }
1da177e4
LT
666
667 pgpath->pg = pg;
668 list_add_tail(&pgpath->list, &pg->pgpaths);
498f0103 669 dm_consume_args(as, nr_args);
1da177e4
LT
670 }
671
672 return pg;
673
674 bad:
675 free_priority_group(pg, ti);
01460f35 676 return ERR_PTR(r);
1da177e4
LT
677}
678
498f0103 679static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1da177e4 680{
1da177e4 681 unsigned hw_argc;
2bfd2e13 682 int ret;
28f16c20 683 struct dm_target *ti = m->ti;
1da177e4 684
498f0103 685 static struct dm_arg _args[] = {
72d94861 686 {0, 1024, "invalid number of hardware handler args"},
1da177e4
LT
687 };
688
498f0103 689 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1da177e4
LT
690 return -EINVAL;
691
692 if (!hw_argc)
693 return 0;
694
498f0103 695 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
510193a2
MS
696 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
697 "scsi_dh_%s", m->hw_handler_name)) {
72d94861 698 ti->error = "unknown hardware handler type";
2bfd2e13
CS
699 ret = -EINVAL;
700 goto fail;
1da177e4 701 }
14e98c5c 702
2bfd2e13
CS
703 if (hw_argc > 1) {
704 char *p;
705 int i, j, len = 4;
706
707 for (i = 0; i <= hw_argc - 2; i++)
708 len += strlen(as->argv[i]) + 1;
709 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
710 if (!p) {
711 ti->error = "memory allocation failed";
712 ret = -ENOMEM;
713 goto fail;
714 }
715 j = sprintf(p, "%d", hw_argc - 1);
716 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
717 j = sprintf(p, "%s", as->argv[i]);
718 }
498f0103 719 dm_consume_args(as, hw_argc - 1);
1da177e4
LT
720
721 return 0;
2bfd2e13
CS
722fail:
723 kfree(m->hw_handler_name);
724 m->hw_handler_name = NULL;
725 return ret;
1da177e4
LT
726}
727
498f0103 728static int parse_features(struct dm_arg_set *as, struct multipath *m)
1da177e4
LT
729{
730 int r;
731 unsigned argc;
28f16c20 732 struct dm_target *ti = m->ti;
498f0103 733 const char *arg_name;
1da177e4 734
498f0103 735 static struct dm_arg _args[] = {
a58a935d 736 {0, 6, "invalid number of feature args"},
c9e45581 737 {1, 50, "pg_init_retries must be between 1 and 50"},
4e2d19e4 738 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1da177e4
LT
739 };
740
498f0103 741 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1da177e4
LT
742 if (r)
743 return -EINVAL;
744
745 if (!argc)
746 return 0;
747
c9e45581 748 do {
498f0103 749 arg_name = dm_shift_arg(as);
c9e45581
DW
750 argc--;
751
498f0103 752 if (!strcasecmp(arg_name, "queue_if_no_path")) {
c9e45581
DW
753 r = queue_if_no_path(m, 1, 0);
754 continue;
755 }
756
a58a935d
MS
757 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
758 m->retain_attached_hw_handler = 1;
759 continue;
760 }
761
498f0103 762 if (!strcasecmp(arg_name, "pg_init_retries") &&
c9e45581 763 (argc >= 1)) {
498f0103 764 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
c9e45581
DW
765 argc--;
766 continue;
767 }
768
498f0103 769 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
4e2d19e4 770 (argc >= 1)) {
498f0103 771 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
4e2d19e4
CS
772 argc--;
773 continue;
774 }
775
1da177e4 776 ti->error = "Unrecognised multipath feature request";
c9e45581
DW
777 r = -EINVAL;
778 } while (argc && !r);
779
780 return r;
1da177e4
LT
781}
782
783static int multipath_ctr(struct dm_target *ti, unsigned int argc,
784 char **argv)
785{
498f0103
MS
786 /* target arguments */
787 static struct dm_arg _args[] = {
a490a07a
MS
788 {0, 1024, "invalid number of priority groups"},
789 {0, 1024, "invalid initial priority group number"},
1da177e4
LT
790 };
791
792 int r;
793 struct multipath *m;
498f0103 794 struct dm_arg_set as;
1da177e4
LT
795 unsigned pg_count = 0;
796 unsigned next_pg_num;
797
798 as.argc = argc;
799 as.argv = argv;
800
28f16c20 801 m = alloc_multipath(ti);
1da177e4 802 if (!m) {
72d94861 803 ti->error = "can't allocate multipath";
1da177e4
LT
804 return -EINVAL;
805 }
806
28f16c20 807 r = parse_features(&as, m);
1da177e4
LT
808 if (r)
809 goto bad;
810
28f16c20 811 r = parse_hw_handler(&as, m);
1da177e4
LT
812 if (r)
813 goto bad;
814
498f0103 815 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1da177e4
LT
816 if (r)
817 goto bad;
818
498f0103 819 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1da177e4
LT
820 if (r)
821 goto bad;
822
a490a07a
MS
823 if ((!m->nr_priority_groups && next_pg_num) ||
824 (m->nr_priority_groups && !next_pg_num)) {
825 ti->error = "invalid initial priority group";
826 r = -EINVAL;
827 goto bad;
828 }
829
1da177e4
LT
830 /* parse the priority groups */
831 while (as.argc) {
832 struct priority_group *pg;
833
28f16c20 834 pg = parse_priority_group(&as, m);
01460f35
BM
835 if (IS_ERR(pg)) {
836 r = PTR_ERR(pg);
1da177e4
LT
837 goto bad;
838 }
839
840 m->nr_valid_paths += pg->nr_pgpaths;
841 list_add_tail(&pg->list, &m->priority_groups);
842 pg_count++;
843 pg->pg_num = pg_count;
844 if (!--next_pg_num)
845 m->next_pg = pg;
846 }
847
848 if (pg_count != m->nr_priority_groups) {
72d94861 849 ti->error = "priority group count mismatch";
1da177e4
LT
850 r = -EINVAL;
851 goto bad;
852 }
853
55a62eef
AK
854 ti->num_flush_bios = 1;
855 ti->num_discard_bios = 1;
042bcef8 856 ti->num_write_same_bios = 1;
8627921f 857
1da177e4
LT
858 return 0;
859
860 bad:
861 free_multipath(m);
862 return r;
863}
864
2bded7bd
KU
865static void multipath_wait_for_pg_init_completion(struct multipath *m)
866{
867 DECLARE_WAITQUEUE(wait, current);
868 unsigned long flags;
869
870 add_wait_queue(&m->pg_init_wait, &wait);
871
872 while (1) {
873 set_current_state(TASK_UNINTERRUPTIBLE);
874
875 spin_lock_irqsave(&m->lock, flags);
876 if (!m->pg_init_in_progress) {
877 spin_unlock_irqrestore(&m->lock, flags);
878 break;
879 }
880 spin_unlock_irqrestore(&m->lock, flags);
881
882 io_schedule();
883 }
884 set_current_state(TASK_RUNNING);
885
886 remove_wait_queue(&m->pg_init_wait, &wait);
887}
888
889static void flush_multipath_work(struct multipath *m)
1da177e4 890{
954a73d5
SKM
891 unsigned long flags;
892
893 spin_lock_irqsave(&m->lock, flags);
894 m->pg_init_disabled = 1;
895 spin_unlock_irqrestore(&m->lock, flags);
896
bab7cfc7 897 flush_workqueue(kmpath_handlerd);
2bded7bd 898 multipath_wait_for_pg_init_completion(m);
a044d016 899 flush_workqueue(kmultipathd);
43829731 900 flush_work(&m->trigger_event);
954a73d5
SKM
901
902 spin_lock_irqsave(&m->lock, flags);
903 m->pg_init_disabled = 0;
904 spin_unlock_irqrestore(&m->lock, flags);
6df400ab
KU
905}
906
907static void multipath_dtr(struct dm_target *ti)
908{
909 struct multipath *m = ti->private;
910
2bded7bd 911 flush_multipath_work(m);
1da177e4
LT
912 free_multipath(m);
913}
914
1da177e4
LT
915/*
916 * Take a path out of use.
917 */
918static int fail_path(struct pgpath *pgpath)
919{
920 unsigned long flags;
921 struct multipath *m = pgpath->pg->m;
922
923 spin_lock_irqsave(&m->lock, flags);
924
6680073d 925 if (!pgpath->is_active)
1da177e4
LT
926 goto out;
927
72d94861 928 DMWARN("Failing path %s.", pgpath->path.dev->name);
1da177e4
LT
929
930 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
6680073d 931 pgpath->is_active = 0;
1da177e4
LT
932 pgpath->fail_count++;
933
934 m->nr_valid_paths--;
935
936 if (pgpath == m->current_pgpath)
937 m->current_pgpath = NULL;
938
b15546f9
MA
939 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
940 pgpath->path.dev->name, m->nr_valid_paths);
941
fe9cf30e 942 schedule_work(&m->trigger_event);
1da177e4
LT
943
944out:
945 spin_unlock_irqrestore(&m->lock, flags);
946
947 return 0;
948}
949
950/*
951 * Reinstate a previously-failed path
952 */
953static int reinstate_path(struct pgpath *pgpath)
954{
63d832c3 955 int r = 0, run_queue = 0;
1da177e4
LT
956 unsigned long flags;
957 struct multipath *m = pgpath->pg->m;
958
959 spin_lock_irqsave(&m->lock, flags);
960
6680073d 961 if (pgpath->is_active)
1da177e4
LT
962 goto out;
963
def052d2 964 if (!pgpath->pg->ps.type->reinstate_path) {
1da177e4
LT
965 DMWARN("Reinstate path not supported by path selector %s",
966 pgpath->pg->ps.type->name);
967 r = -EINVAL;
968 goto out;
969 }
970
971 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
972 if (r)
973 goto out;
974
6680073d 975 pgpath->is_active = 1;
1da177e4 976
e8099177 977 if (!m->nr_valid_paths++) {
e54f77dd 978 m->current_pgpath = NULL;
63d832c3 979 run_queue = 1;
e54f77dd 980 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
4e2d19e4 981 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
e54f77dd
CS
982 m->pg_init_in_progress++;
983 }
1da177e4 984
b15546f9
MA
985 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
986 pgpath->path.dev->name, m->nr_valid_paths);
987
fe9cf30e 988 schedule_work(&m->trigger_event);
1da177e4
LT
989
990out:
991 spin_unlock_irqrestore(&m->lock, flags);
63d832c3
HR
992 if (run_queue)
993 dm_table_run_md_queue_async(m->ti->table);
1da177e4
LT
994
995 return r;
996}
997
998/*
999 * Fail or reinstate all paths that match the provided struct dm_dev.
1000 */
1001static int action_dev(struct multipath *m, struct dm_dev *dev,
1002 action_fn action)
1003{
19040c0b 1004 int r = -EINVAL;
1da177e4
LT
1005 struct pgpath *pgpath;
1006 struct priority_group *pg;
1007
1008 list_for_each_entry(pg, &m->priority_groups, list) {
1009 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1010 if (pgpath->path.dev == dev)
1011 r = action(pgpath);
1012 }
1013 }
1014
1015 return r;
1016}
1017
1018/*
1019 * Temporarily try to avoid having to use the specified PG
1020 */
1021static void bypass_pg(struct multipath *m, struct priority_group *pg,
1022 int bypassed)
1023{
1024 unsigned long flags;
1025
1026 spin_lock_irqsave(&m->lock, flags);
1027
1028 pg->bypassed = bypassed;
1029 m->current_pgpath = NULL;
1030 m->current_pg = NULL;
1031
1032 spin_unlock_irqrestore(&m->lock, flags);
1033
fe9cf30e 1034 schedule_work(&m->trigger_event);
1da177e4
LT
1035}
1036
1037/*
1038 * Switch to using the specified PG from the next I/O that gets mapped
1039 */
1040static int switch_pg_num(struct multipath *m, const char *pgstr)
1041{
1042 struct priority_group *pg;
1043 unsigned pgnum;
1044 unsigned long flags;
31998ef1 1045 char dummy;
1da177e4 1046
31998ef1 1047 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1da177e4
LT
1048 (pgnum > m->nr_priority_groups)) {
1049 DMWARN("invalid PG number supplied to switch_pg_num");
1050 return -EINVAL;
1051 }
1052
1053 spin_lock_irqsave(&m->lock, flags);
1054 list_for_each_entry(pg, &m->priority_groups, list) {
1055 pg->bypassed = 0;
1056 if (--pgnum)
1057 continue;
1058
1059 m->current_pgpath = NULL;
1060 m->current_pg = NULL;
1061 m->next_pg = pg;
1062 }
1063 spin_unlock_irqrestore(&m->lock, flags);
1064
fe9cf30e 1065 schedule_work(&m->trigger_event);
1da177e4
LT
1066 return 0;
1067}
1068
1069/*
1070 * Set/clear bypassed status of a PG.
1071 * PGs are numbered upwards from 1 in the order they were declared.
1072 */
1073static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1074{
1075 struct priority_group *pg;
1076 unsigned pgnum;
31998ef1 1077 char dummy;
1da177e4 1078
31998ef1 1079 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1da177e4
LT
1080 (pgnum > m->nr_priority_groups)) {
1081 DMWARN("invalid PG number supplied to bypass_pg");
1082 return -EINVAL;
1083 }
1084
1085 list_for_each_entry(pg, &m->priority_groups, list) {
1086 if (!--pgnum)
1087 break;
1088 }
1089
1090 bypass_pg(m, pg, bypassed);
1091 return 0;
1092}
1093
c9e45581
DW
1094/*
1095 * Should we retry pg_init immediately?
1096 */
1097static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1098{
1099 unsigned long flags;
1100 int limit_reached = 0;
1101
1102 spin_lock_irqsave(&m->lock, flags);
1103
954a73d5 1104 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
c9e45581
DW
1105 m->pg_init_required = 1;
1106 else
1107 limit_reached = 1;
1108
1109 spin_unlock_irqrestore(&m->lock, flags);
1110
1111 return limit_reached;
1112}
1113
3ae31f6a 1114static void pg_init_done(void *data, int errors)
cfae5c9b 1115{
83c0d5d5 1116 struct pgpath *pgpath = data;
cfae5c9b
CS
1117 struct priority_group *pg = pgpath->pg;
1118 struct multipath *m = pg->m;
1119 unsigned long flags;
4e2d19e4 1120 unsigned delay_retry = 0;
cfae5c9b
CS
1121
1122 /* device or driver problems */
1123 switch (errors) {
1124 case SCSI_DH_OK:
1125 break;
1126 case SCSI_DH_NOSYS:
1127 if (!m->hw_handler_name) {
1128 errors = 0;
1129 break;
1130 }
f7b934c8
MB
1131 DMERR("Could not failover the device: Handler scsi_dh_%s "
1132 "Error %d.", m->hw_handler_name, errors);
cfae5c9b
CS
1133 /*
1134 * Fail path for now, so we do not ping pong
1135 */
1136 fail_path(pgpath);
1137 break;
1138 case SCSI_DH_DEV_TEMP_BUSY:
1139 /*
1140 * Probably doing something like FW upgrade on the
1141 * controller so try the other pg.
1142 */
1143 bypass_pg(m, pg, 1);
1144 break;
cfae5c9b 1145 case SCSI_DH_RETRY:
4e2d19e4
CS
1146 /* Wait before retrying. */
1147 delay_retry = 1;
cfae5c9b
CS
1148 case SCSI_DH_IMM_RETRY:
1149 case SCSI_DH_RES_TEMP_UNAVAIL:
1150 if (pg_init_limit_reached(m, pgpath))
1151 fail_path(pgpath);
1152 errors = 0;
1153 break;
1154 default:
1155 /*
1156 * We probably do not want to fail the path for a device
1157 * error, but this is what the old dm did. In future
1158 * patches we can do more advanced handling.
1159 */
1160 fail_path(pgpath);
1161 }
1162
1163 spin_lock_irqsave(&m->lock, flags);
1164 if (errors) {
e54f77dd
CS
1165 if (pgpath == m->current_pgpath) {
1166 DMERR("Could not failover device. Error %d.", errors);
1167 m->current_pgpath = NULL;
1168 m->current_pg = NULL;
1169 }
d0259bf0 1170 } else if (!m->pg_init_required)
cfae5c9b 1171 pg->bypassed = 0;
cfae5c9b 1172
d0259bf0
KU
1173 if (--m->pg_init_in_progress)
1174 /* Activations of other paths are still on going */
1175 goto out;
1176
3e9f1be1
HR
1177 if (m->pg_init_required) {
1178 m->pg_init_delay_retry = delay_retry;
1179 if (__pg_init_all_paths(m))
1180 goto out;
1181 }
1182 m->queue_io = 0;
d0259bf0 1183
2bded7bd
KU
1184 /*
1185 * Wake up any thread waiting to suspend.
1186 */
1187 wake_up(&m->pg_init_wait);
1188
d0259bf0 1189out:
cfae5c9b
CS
1190 spin_unlock_irqrestore(&m->lock, flags);
1191}
1192
bab7cfc7
CS
1193static void activate_path(struct work_struct *work)
1194{
e54f77dd 1195 struct pgpath *pgpath =
4e2d19e4 1196 container_of(work, struct pgpath, activate_path.work);
bab7cfc7 1197
3a017509
HR
1198 if (pgpath->is_active)
1199 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1200 pg_init_done, pgpath);
1201 else
1202 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
bab7cfc7
CS
1203}
1204
7e782af5
HR
1205static int noretry_error(int error)
1206{
1207 switch (error) {
1208 case -EOPNOTSUPP:
1209 case -EREMOTEIO:
1210 case -EILSEQ:
1211 case -ENODATA:
cc9d3c38 1212 case -ENOSPC:
7e782af5
HR
1213 return 1;
1214 }
1215
1216 /* Anything else could be a path failure, so should be retried */
1217 return 0;
1218}
1219
1da177e4
LT
1220/*
1221 * end_io handling
1222 */
f40c67f0 1223static int do_end_io(struct multipath *m, struct request *clone,
028867ac 1224 int error, struct dm_mpath_io *mpio)
1da177e4 1225{
f40c67f0
KU
1226 /*
1227 * We don't queue any clone request inside the multipath target
1228 * during end I/O handling, since those clone requests don't have
1229 * bio clones. If we queue them inside the multipath target,
1230 * we need to make bio clones, that requires memory allocation.
1231 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1232 * don't have bio clones.)
1233 * Instead of queueing the clone request here, we queue the original
1234 * request into dm core, which will remake a clone request and
1235 * clone bios for it and resubmit it later.
1236 */
1237 int r = DM_ENDIO_REQUEUE;
640eb3b0 1238 unsigned long flags;
1da177e4 1239
f40c67f0 1240 if (!error && !clone->errors)
1da177e4
LT
1241 return 0; /* I/O complete */
1242
7eee4ae2 1243 if (noretry_error(error))
959eb4e5
MS
1244 return error;
1245
cfae5c9b
CS
1246 if (mpio->pgpath)
1247 fail_path(mpio->pgpath);
1da177e4 1248
640eb3b0 1249 spin_lock_irqsave(&m->lock, flags);
751b2a7d
HR
1250 if (!m->nr_valid_paths) {
1251 if (!m->queue_if_no_path) {
1252 if (!__must_push_back(m))
1253 r = -EIO;
1254 } else {
1255 if (error == -EBADE)
1256 r = error;
1257 }
1258 }
640eb3b0 1259 spin_unlock_irqrestore(&m->lock, flags);
1da177e4 1260
f40c67f0 1261 return r;
1da177e4
LT
1262}
1263
f40c67f0 1264static int multipath_end_io(struct dm_target *ti, struct request *clone,
1da177e4
LT
1265 int error, union map_info *map_context)
1266{
028867ac
AK
1267 struct multipath *m = ti->private;
1268 struct dm_mpath_io *mpio = map_context->ptr;
a71a261f 1269 struct pgpath *pgpath;
1da177e4
LT
1270 struct path_selector *ps;
1271 int r;
1272
466891f9
JN
1273 BUG_ON(!mpio);
1274
f40c67f0 1275 r = do_end_io(m, clone, error, mpio);
a71a261f 1276 pgpath = mpio->pgpath;
1da177e4
LT
1277 if (pgpath) {
1278 ps = &pgpath->pg->ps;
1279 if (ps->type->end_io)
02ab823f 1280 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1da177e4 1281 }
466891f9 1282 clear_mapinfo(m, map_context);
1da177e4
LT
1283
1284 return r;
1285}
1286
1287/*
1288 * Suspend can't complete until all the I/O is processed so if
436d4108
AK
1289 * the last path fails we must error any remaining I/O.
1290 * Note that if the freeze_bdev fails while suspending, the
1291 * queue_if_no_path state is lost - userspace should reset it.
1da177e4
LT
1292 */
1293static void multipath_presuspend(struct dm_target *ti)
1294{
1295 struct multipath *m = (struct multipath *) ti->private;
1da177e4 1296
485ef69e 1297 queue_if_no_path(m, 0, 1);
1da177e4
LT
1298}
1299
6df400ab
KU
1300static void multipath_postsuspend(struct dm_target *ti)
1301{
6380f26f
MA
1302 struct multipath *m = ti->private;
1303
1304 mutex_lock(&m->work_mutex);
2bded7bd 1305 flush_multipath_work(m);
6380f26f 1306 mutex_unlock(&m->work_mutex);
6df400ab
KU
1307}
1308
436d4108
AK
1309/*
1310 * Restore the queue_if_no_path setting.
1311 */
1da177e4
LT
1312static void multipath_resume(struct dm_target *ti)
1313{
1314 struct multipath *m = (struct multipath *) ti->private;
1315 unsigned long flags;
1316
1317 spin_lock_irqsave(&m->lock, flags);
436d4108 1318 m->queue_if_no_path = m->saved_queue_if_no_path;
1da177e4
LT
1319 spin_unlock_irqrestore(&m->lock, flags);
1320}
1321
1322/*
1323 * Info output has the following format:
1324 * num_multipath_feature_args [multipath_feature_args]*
1325 * num_handler_status_args [handler_status_args]*
1326 * num_groups init_group_number
1327 * [A|D|E num_ps_status_args [ps_status_args]*
1328 * num_paths num_selector_args
1329 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1330 *
1331 * Table output has the following format (identical to the constructor string):
1332 * num_feature_args [features_args]*
1333 * num_handler_args hw_handler [hw_handler_args]*
1334 * num_groups init_group_number
1335 * [priority selector-name num_ps_args [ps_args]*
1336 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1337 */
fd7c092e
MP
1338static void multipath_status(struct dm_target *ti, status_type_t type,
1339 unsigned status_flags, char *result, unsigned maxlen)
1da177e4
LT
1340{
1341 int sz = 0;
1342 unsigned long flags;
1343 struct multipath *m = (struct multipath *) ti->private;
1da177e4
LT
1344 struct priority_group *pg;
1345 struct pgpath *p;
1346 unsigned pg_num;
1347 char state;
1348
1349 spin_lock_irqsave(&m->lock, flags);
1350
1351 /* Features */
1352 if (type == STATUSTYPE_INFO)
e8099177 1353 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
c9e45581
DW
1354 else {
1355 DMEMIT("%u ", m->queue_if_no_path +
4e2d19e4 1356 (m->pg_init_retries > 0) * 2 +
a58a935d
MS
1357 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1358 m->retain_attached_hw_handler);
c9e45581
DW
1359 if (m->queue_if_no_path)
1360 DMEMIT("queue_if_no_path ");
1361 if (m->pg_init_retries)
1362 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
4e2d19e4
CS
1363 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1364 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
a58a935d
MS
1365 if (m->retain_attached_hw_handler)
1366 DMEMIT("retain_attached_hw_handler ");
c9e45581 1367 }
1da177e4 1368
cfae5c9b 1369 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1da177e4
LT
1370 DMEMIT("0 ");
1371 else
cfae5c9b 1372 DMEMIT("1 %s ", m->hw_handler_name);
1da177e4
LT
1373
1374 DMEMIT("%u ", m->nr_priority_groups);
1375
1376 if (m->next_pg)
1377 pg_num = m->next_pg->pg_num;
1378 else if (m->current_pg)
1379 pg_num = m->current_pg->pg_num;
1380 else
a490a07a 1381 pg_num = (m->nr_priority_groups ? 1 : 0);
1da177e4
LT
1382
1383 DMEMIT("%u ", pg_num);
1384
1385 switch (type) {
1386 case STATUSTYPE_INFO:
1387 list_for_each_entry(pg, &m->priority_groups, list) {
1388 if (pg->bypassed)
1389 state = 'D'; /* Disabled */
1390 else if (pg == m->current_pg)
1391 state = 'A'; /* Currently Active */
1392 else
1393 state = 'E'; /* Enabled */
1394
1395 DMEMIT("%c ", state);
1396
1397 if (pg->ps.type->status)
1398 sz += pg->ps.type->status(&pg->ps, NULL, type,
1399 result + sz,
1400 maxlen - sz);
1401 else
1402 DMEMIT("0 ");
1403
1404 DMEMIT("%u %u ", pg->nr_pgpaths,
1405 pg->ps.type->info_args);
1406
1407 list_for_each_entry(p, &pg->pgpaths, list) {
1408 DMEMIT("%s %s %u ", p->path.dev->name,
6680073d 1409 p->is_active ? "A" : "F",
1da177e4
LT
1410 p->fail_count);
1411 if (pg->ps.type->status)
1412 sz += pg->ps.type->status(&pg->ps,
1413 &p->path, type, result + sz,
1414 maxlen - sz);
1415 }
1416 }
1417 break;
1418
1419 case STATUSTYPE_TABLE:
1420 list_for_each_entry(pg, &m->priority_groups, list) {
1421 DMEMIT("%s ", pg->ps.type->name);
1422
1423 if (pg->ps.type->status)
1424 sz += pg->ps.type->status(&pg->ps, NULL, type,
1425 result + sz,
1426 maxlen - sz);
1427 else
1428 DMEMIT("0 ");
1429
1430 DMEMIT("%u %u ", pg->nr_pgpaths,
1431 pg->ps.type->table_args);
1432
1433 list_for_each_entry(p, &pg->pgpaths, list) {
1434 DMEMIT("%s ", p->path.dev->name);
1435 if (pg->ps.type->status)
1436 sz += pg->ps.type->status(&pg->ps,
1437 &p->path, type, result + sz,
1438 maxlen - sz);
1439 }
1440 }
1441 break;
1442 }
1443
1444 spin_unlock_irqrestore(&m->lock, flags);
1da177e4
LT
1445}
1446
1447static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1448{
6380f26f 1449 int r = -EINVAL;
1da177e4
LT
1450 struct dm_dev *dev;
1451 struct multipath *m = (struct multipath *) ti->private;
1452 action_fn action;
1453
6380f26f
MA
1454 mutex_lock(&m->work_mutex);
1455
c2f3d24b
KU
1456 if (dm_suspended(ti)) {
1457 r = -EBUSY;
1458 goto out;
1459 }
1460
1da177e4 1461 if (argc == 1) {
498f0103 1462 if (!strcasecmp(argv[0], "queue_if_no_path")) {
6380f26f
MA
1463 r = queue_if_no_path(m, 1, 0);
1464 goto out;
498f0103 1465 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
6380f26f
MA
1466 r = queue_if_no_path(m, 0, 0);
1467 goto out;
1468 }
1da177e4
LT
1469 }
1470
6380f26f 1471 if (argc != 2) {
a356e426 1472 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
6380f26f
MA
1473 goto out;
1474 }
1da177e4 1475
498f0103 1476 if (!strcasecmp(argv[0], "disable_group")) {
6380f26f
MA
1477 r = bypass_pg_num(m, argv[1], 1);
1478 goto out;
498f0103 1479 } else if (!strcasecmp(argv[0], "enable_group")) {
6380f26f
MA
1480 r = bypass_pg_num(m, argv[1], 0);
1481 goto out;
498f0103 1482 } else if (!strcasecmp(argv[0], "switch_group")) {
6380f26f
MA
1483 r = switch_pg_num(m, argv[1]);
1484 goto out;
498f0103 1485 } else if (!strcasecmp(argv[0], "reinstate_path"))
1da177e4 1486 action = reinstate_path;
498f0103 1487 else if (!strcasecmp(argv[0], "fail_path"))
1da177e4 1488 action = fail_path;
6380f26f 1489 else {
a356e426 1490 DMWARN("Unrecognised multipath message received: %s", argv[0]);
6380f26f
MA
1491 goto out;
1492 }
1da177e4 1493
8215d6ec 1494 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1da177e4 1495 if (r) {
72d94861 1496 DMWARN("message: error getting device %s",
1da177e4 1497 argv[1]);
6380f26f 1498 goto out;
1da177e4
LT
1499 }
1500
1501 r = action_dev(m, dev, action);
1502
1503 dm_put_device(ti, dev);
1504
6380f26f
MA
1505out:
1506 mutex_unlock(&m->work_mutex);
1da177e4 1507 return r;
1da177e4
LT
1508}
1509
647b3d00 1510static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
9af4aa30
MB
1511 unsigned long arg)
1512{
35991652 1513 struct multipath *m = ti->private;
7ba10aa6 1514 struct pgpath *pgpath;
35991652
MP
1515 struct block_device *bdev;
1516 fmode_t mode;
9af4aa30 1517 unsigned long flags;
35991652
MP
1518 int r;
1519
35991652
MP
1520 bdev = NULL;
1521 mode = 0;
1522 r = 0;
9af4aa30
MB
1523
1524 spin_lock_irqsave(&m->lock, flags);
1525
1526 if (!m->current_pgpath)
02ab823f 1527 __choose_pgpath(m, 0);
9af4aa30 1528
7ba10aa6
MS
1529 pgpath = m->current_pgpath;
1530
1531 if (pgpath) {
1532 bdev = pgpath->path.dev->bdev;
1533 mode = pgpath->path.dev->mode;
e90dae1f 1534 }
9af4aa30 1535
7ba10aa6 1536 if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
6c182cd8 1537 r = -ENOTCONN;
9af4aa30
MB
1538 else if (!bdev)
1539 r = -EIO;
1540
1541 spin_unlock_irqrestore(&m->lock, flags);
1542
ec8013be
PB
1543 /*
1544 * Only pass ioctls through if the device sizes match exactly.
1545 */
a1989b33
HR
1546 if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1547 int err = scsi_verify_blk_ioctl(NULL, cmd);
1548 if (err)
1549 r = err;
1550 }
ec8013be 1551
3e9f1be1
HR
1552 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1553 spin_lock_irqsave(&m->lock, flags);
1554 if (!m->current_pg) {
1555 /* Path status changed, redo selection */
1556 __choose_pgpath(m, 0);
1557 }
1558 if (m->pg_init_required)
1559 __pg_init_all_paths(m);
4cdd2ad7 1560 spin_unlock_irqrestore(&m->lock, flags);
63d832c3 1561 dm_table_run_md_queue_async(m->ti->table);
3e9f1be1 1562 }
35991652 1563
633a08b8 1564 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
9af4aa30
MB
1565}
1566
af4874e0
MS
1567static int multipath_iterate_devices(struct dm_target *ti,
1568 iterate_devices_callout_fn fn, void *data)
1569{
1570 struct multipath *m = ti->private;
1571 struct priority_group *pg;
1572 struct pgpath *p;
1573 int ret = 0;
1574
1575 list_for_each_entry(pg, &m->priority_groups, list) {
1576 list_for_each_entry(p, &pg->pgpaths, list) {
5dea271b 1577 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
af4874e0
MS
1578 if (ret)
1579 goto out;
1580 }
1581 }
1582
1583out:
1584 return ret;
1585}
1586
f40c67f0
KU
1587static int __pgpath_busy(struct pgpath *pgpath)
1588{
1589 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1590
1591 return dm_underlying_device_busy(q);
1592}
1593
1594/*
1595 * We return "busy", only when we can map I/Os but underlying devices
1596 * are busy (so even if we map I/Os now, the I/Os will wait on
1597 * the underlying queue).
1598 * In other words, if we want to kill I/Os or queue them inside us
1599 * due to map unavailability, we don't return "busy". Otherwise,
1600 * dm core won't give us the I/Os and we can't do what we want.
1601 */
1602static int multipath_busy(struct dm_target *ti)
1603{
1604 int busy = 0, has_active = 0;
1605 struct multipath *m = ti->private;
1606 struct priority_group *pg;
1607 struct pgpath *pgpath;
1608 unsigned long flags;
1609
1610 spin_lock_irqsave(&m->lock, flags);
1611
7a7a3b45
JN
1612 /* pg_init in progress or no paths available */
1613 if (m->pg_init_in_progress ||
1614 (!m->nr_valid_paths && m->queue_if_no_path)) {
b63349a7
HR
1615 busy = 1;
1616 goto out;
1617 }
f40c67f0
KU
1618 /* Guess which priority_group will be used at next mapping time */
1619 if (unlikely(!m->current_pgpath && m->next_pg))
1620 pg = m->next_pg;
1621 else if (likely(m->current_pg))
1622 pg = m->current_pg;
1623 else
1624 /*
1625 * We don't know which pg will be used at next mapping time.
1626 * We don't call __choose_pgpath() here to avoid to trigger
1627 * pg_init just by busy checking.
1628 * So we don't know whether underlying devices we will be using
1629 * at next mapping time are busy or not. Just try mapping.
1630 */
1631 goto out;
1632
1633 /*
1634 * If there is one non-busy active path at least, the path selector
1635 * will be able to select it. So we consider such a pg as not busy.
1636 */
1637 busy = 1;
1638 list_for_each_entry(pgpath, &pg->pgpaths, list)
1639 if (pgpath->is_active) {
1640 has_active = 1;
1641
1642 if (!__pgpath_busy(pgpath)) {
1643 busy = 0;
1644 break;
1645 }
1646 }
1647
1648 if (!has_active)
1649 /*
1650 * No active path in this pg, so this pg won't be used and
1651 * the current_pg will be changed at next mapping time.
1652 * We need to try mapping to determine it.
1653 */
1654 busy = 0;
1655
1656out:
1657 spin_unlock_irqrestore(&m->lock, flags);
1658
1659 return busy;
1660}
1661
1da177e4
LT
1662/*-----------------------------------------------------------------
1663 * Module setup
1664 *---------------------------------------------------------------*/
1665static struct target_type multipath_target = {
1666 .name = "multipath",
e8099177 1667 .version = {1, 7, 0},
1da177e4
LT
1668 .module = THIS_MODULE,
1669 .ctr = multipath_ctr,
1670 .dtr = multipath_dtr,
f40c67f0
KU
1671 .map_rq = multipath_map,
1672 .rq_end_io = multipath_end_io,
1da177e4 1673 .presuspend = multipath_presuspend,
6df400ab 1674 .postsuspend = multipath_postsuspend,
1da177e4
LT
1675 .resume = multipath_resume,
1676 .status = multipath_status,
1677 .message = multipath_message,
9af4aa30 1678 .ioctl = multipath_ioctl,
af4874e0 1679 .iterate_devices = multipath_iterate_devices,
f40c67f0 1680 .busy = multipath_busy,
1da177e4
LT
1681};
1682
1683static int __init dm_multipath_init(void)
1684{
1685 int r;
1686
1687 /* allocate a slab for the dm_ios */
028867ac 1688 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1da177e4
LT
1689 if (!_mpio_cache)
1690 return -ENOMEM;
1691
1692 r = dm_register_target(&multipath_target);
1693 if (r < 0) {
0cd33124 1694 DMERR("register failed %d", r);
1da177e4
LT
1695 kmem_cache_destroy(_mpio_cache);
1696 return -EINVAL;
1697 }
1698
4d4d66ab 1699 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
c557308e 1700 if (!kmultipathd) {
0cd33124 1701 DMERR("failed to create workqueue kmpathd");
c557308e
AK
1702 dm_unregister_target(&multipath_target);
1703 kmem_cache_destroy(_mpio_cache);
1704 return -ENOMEM;
1705 }
1706
bab7cfc7
CS
1707 /*
1708 * A separate workqueue is used to handle the device handlers
1709 * to avoid overloading existing workqueue. Overloading the
1710 * old workqueue would also create a bottleneck in the
1711 * path of the storage hardware device activation.
1712 */
4d4d66ab
TH
1713 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1714 WQ_MEM_RECLAIM);
bab7cfc7
CS
1715 if (!kmpath_handlerd) {
1716 DMERR("failed to create workqueue kmpath_handlerd");
1717 destroy_workqueue(kmultipathd);
1718 dm_unregister_target(&multipath_target);
1719 kmem_cache_destroy(_mpio_cache);
1720 return -ENOMEM;
1721 }
1722
72d94861 1723 DMINFO("version %u.%u.%u loaded",
1da177e4
LT
1724 multipath_target.version[0], multipath_target.version[1],
1725 multipath_target.version[2]);
1726
1727 return r;
1728}
1729
1730static void __exit dm_multipath_exit(void)
1731{
bab7cfc7 1732 destroy_workqueue(kmpath_handlerd);
c557308e
AK
1733 destroy_workqueue(kmultipathd);
1734
10d3bd09 1735 dm_unregister_target(&multipath_target);
1da177e4
LT
1736 kmem_cache_destroy(_mpio_cache);
1737}
1738
1da177e4
LT
1739module_init(dm_multipath_init);
1740module_exit(dm_multipath_exit);
1741
1742MODULE_DESCRIPTION(DM_NAME " multipath target");
1743MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1744MODULE_LICENSE("GPL");