]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - drivers/md/dm-mpath.c
projects / mirror_ubuntu-artful-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
dm mpath: make it easier to detect unintended I/O request flushes
[mirror_ubuntu-artful-kernel.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-rq.h"
11 #include "dm-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
14
15 #include <linux/blkdev.h>
16 #include <linux/ctype.h>
17 #include <linux/init.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/time.h>
23 #include <linux/workqueue.h>
24 #include <linux/delay.h>
25 #include <scsi/scsi_dh.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-mq.h>
28
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32
33 /* Path properties */
34 struct pgpath {
35 struct list_head list;
36
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
39
40 struct dm_path path;
41 struct delayed_work activate_path;
42
43 bool is_active:1; /* Path status */
44 };
45
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
47
48 /*
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
51 */
52 struct priority_group {
53 struct list_head list;
54
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
57
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
61
62 bool bypassed:1; /* Temporarily bypass this PG? */
63 };
64
65 /* Multipath context */
66 struct multipath {
67 struct list_head list;
68 struct dm_target *ti;
69
70 const char *hw_handler_name;
71 char *hw_handler_params;
72
73 spinlock_t lock;
74
75 unsigned nr_priority_groups;
76 struct list_head priority_groups;
77
78 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
79
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
83
84 unsigned long flags; /* Multipath state flags */
85
86 unsigned pg_init_retries; /* Number of times to retry pg_init */
87 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
88
89 atomic_t nr_valid_paths; /* Total number of usable paths */
90 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
91 atomic_t pg_init_count; /* Number of times pg_init called */
92
93 enum dm_queue_mode queue_mode;
94
95 struct mutex work_mutex;
96 struct work_struct trigger_event;
97
98 struct work_struct process_queued_bios;
99 struct bio_list queued_bios;
100 };
101
102 /*
103 * Context information attached to each io we process.
104 */
105 struct dm_mpath_io {
106 struct pgpath *pgpath;
107 size_t nr_bytes;
108 };
109
110 typedef int (*action_fn) (struct pgpath *pgpath);
111
112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113 static void trigger_event(struct work_struct *work);
114 static void activate_or_offline_path(struct pgpath *pgpath);
115 static void activate_path_work(struct work_struct *work);
116 static void process_queued_bios(struct work_struct *work);
117
118 /*-----------------------------------------------
119 * Multipath state flags.
120 *-----------------------------------------------*/
121
122 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
123 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
126 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
127 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
128 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
129
130 /*-----------------------------------------------
131 * Allocation routines
132 *-----------------------------------------------*/
133
134 static struct pgpath *alloc_pgpath(void)
135 {
136 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
137
138 if (pgpath) {
139 pgpath->is_active = true;
140 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
141 }
142
143 return pgpath;
144 }
145
146 static void free_pgpath(struct pgpath *pgpath)
147 {
148 kfree(pgpath);
149 }
150
151 static struct priority_group *alloc_priority_group(void)
152 {
153 struct priority_group *pg;
154
155 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
156
157 if (pg)
158 INIT_LIST_HEAD(&pg->pgpaths);
159
160 return pg;
161 }
162
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
164 {
165 struct pgpath *pgpath, *tmp;
166
167 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168 list_del(&pgpath->list);
169 dm_put_device(ti, pgpath->path.dev);
170 free_pgpath(pgpath);
171 }
172 }
173
174 static void free_priority_group(struct priority_group *pg,
175 struct dm_target *ti)
176 {
177 struct path_selector *ps = &pg->ps;
178
179 if (ps->type) {
180 ps->type->destroy(ps);
181 dm_put_path_selector(ps->type);
182 }
183
184 free_pgpaths(&pg->pgpaths, ti);
185 kfree(pg);
186 }
187
188 static struct multipath *alloc_multipath(struct dm_target *ti)
189 {
190 struct multipath *m;
191
192 m = kzalloc(sizeof(*m), GFP_KERNEL);
193 if (m) {
194 INIT_LIST_HEAD(&m->priority_groups);
195 spin_lock_init(&m->lock);
196 set_bit(MPATHF_QUEUE_IO, &m->flags);
197 atomic_set(&m->nr_valid_paths, 0);
198 atomic_set(&m->pg_init_in_progress, 0);
199 atomic_set(&m->pg_init_count, 0);
200 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
201 INIT_WORK(&m->trigger_event, trigger_event);
202 init_waitqueue_head(&m->pg_init_wait);
203 mutex_init(&m->work_mutex);
204
205 m->queue_mode = DM_TYPE_NONE;
206
207 m->ti = ti;
208 ti->private = m;
209 }
210
211 return m;
212 }
213
214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
215 {
216 if (m->queue_mode == DM_TYPE_NONE) {
217 /*
218 * Default to request-based.
219 */
220 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
221 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
222 else
223 m->queue_mode = DM_TYPE_REQUEST_BASED;
224 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
225 INIT_WORK(&m->process_queued_bios, process_queued_bios);
226 /*
227 * bio-based doesn't support any direct scsi_dh management;
228 * it just discovers if a scsi_dh is attached.
229 */
230 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
231 }
232
233 dm_table_set_type(ti->table, m->queue_mode);
234
235 return 0;
236 }
237
238 static void free_multipath(struct multipath *m)
239 {
240 struct priority_group *pg, *tmp;
241
242 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
243 list_del(&pg->list);
244 free_priority_group(pg, m->ti);
245 }
246
247 kfree(m->hw_handler_name);
248 kfree(m->hw_handler_params);
249 kfree(m);
250 }
251
252 static struct dm_mpath_io *get_mpio(union map_info *info)
253 {
254 return info->ptr;
255 }
256
257 static size_t multipath_per_bio_data_size(void)
258 {
259 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
260 }
261
262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
263 {
264 return dm_per_bio_data(bio, multipath_per_bio_data_size());
265 }
266
267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
268 {
269 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
271 void *bio_details = mpio + 1;
272
273 return bio_details;
274 }
275
276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
277 struct dm_bio_details **bio_details_p)
278 {
279 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
280 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
281
282 memset(mpio, 0, sizeof(*mpio));
283 memset(bio_details, 0, sizeof(*bio_details));
284 dm_bio_record(bio_details, bio);
285
286 if (mpio_p)
287 *mpio_p = mpio;
288 if (bio_details_p)
289 *bio_details_p = bio_details;
290 }
291
292 /*-----------------------------------------------
293 * Path selection
294 *-----------------------------------------------*/
295
296 static int __pg_init_all_paths(struct multipath *m)
297 {
298 struct pgpath *pgpath;
299 unsigned long pg_init_delay = 0;
300
301 lockdep_assert_held(&m->lock);
302
303 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
304 return 0;
305
306 atomic_inc(&m->pg_init_count);
307 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
308
309 /* Check here to reset pg_init_required */
310 if (!m->current_pg)
311 return 0;
312
313 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
314 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
315 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
316 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
317 /* Skip failed paths */
318 if (!pgpath->is_active)
319 continue;
320 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
321 pg_init_delay))
322 atomic_inc(&m->pg_init_in_progress);
323 }
324 return atomic_read(&m->pg_init_in_progress);
325 }
326
327 static int pg_init_all_paths(struct multipath *m)
328 {
329 int ret;
330 unsigned long flags;
331
332 spin_lock_irqsave(&m->lock, flags);
333 ret = __pg_init_all_paths(m);
334 spin_unlock_irqrestore(&m->lock, flags);
335
336 return ret;
337 }
338
339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
340 {
341 m->current_pg = pg;
342
343 /* Must we initialise the PG first, and queue I/O till it's ready? */
344 if (m->hw_handler_name) {
345 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
346 set_bit(MPATHF_QUEUE_IO, &m->flags);
347 } else {
348 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
349 clear_bit(MPATHF_QUEUE_IO, &m->flags);
350 }
351
352 atomic_set(&m->pg_init_count, 0);
353 }
354
355 static struct pgpath *choose_path_in_pg(struct multipath *m,
356 struct priority_group *pg,
357 size_t nr_bytes)
358 {
359 unsigned long flags;
360 struct dm_path *path;
361 struct pgpath *pgpath;
362
363 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
364 if (!path)
365 return ERR_PTR(-ENXIO);
366
367 pgpath = path_to_pgpath(path);
368
369 if (unlikely(lockless_dereference(m->current_pg) != pg)) {
370 /* Only update current_pgpath if pg changed */
371 spin_lock_irqsave(&m->lock, flags);
372 m->current_pgpath = pgpath;
373 __switch_pg(m, pg);
374 spin_unlock_irqrestore(&m->lock, flags);
375 }
376
377 return pgpath;
378 }
379
380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
381 {
382 unsigned long flags;
383 struct priority_group *pg;
384 struct pgpath *pgpath;
385 unsigned bypassed = 1;
386
387 if (!atomic_read(&m->nr_valid_paths)) {
388 clear_bit(MPATHF_QUEUE_IO, &m->flags);
389 goto failed;
390 }
391
392 /* Were we instructed to switch PG? */
393 if (lockless_dereference(m->next_pg)) {
394 spin_lock_irqsave(&m->lock, flags);
395 pg = m->next_pg;
396 if (!pg) {
397 spin_unlock_irqrestore(&m->lock, flags);
398 goto check_current_pg;
399 }
400 m->next_pg = NULL;
401 spin_unlock_irqrestore(&m->lock, flags);
402 pgpath = choose_path_in_pg(m, pg, nr_bytes);
403 if (!IS_ERR_OR_NULL(pgpath))
404 return pgpath;
405 }
406
407 /* Don't change PG until it has no remaining paths */
408 check_current_pg:
409 pg = lockless_dereference(m->current_pg);
410 if (pg) {
411 pgpath = choose_path_in_pg(m, pg, nr_bytes);
412 if (!IS_ERR_OR_NULL(pgpath))
413 return pgpath;
414 }
415
416 /*
417 * Loop through priority groups until we find a valid path.
418 * First time we skip PGs marked 'bypassed'.
419 * Second time we only try the ones we skipped, but set
420 * pg_init_delay_retry so we do not hammer controllers.
421 */
422 do {
423 list_for_each_entry(pg, &m->priority_groups, list) {
424 if (pg->bypassed == !!bypassed)
425 continue;
426 pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 if (!IS_ERR_OR_NULL(pgpath)) {
428 if (!bypassed)
429 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
430 return pgpath;
431 }
432 }
433 } while (bypassed--);
434
435 failed:
436 spin_lock_irqsave(&m->lock, flags);
437 m->current_pgpath = NULL;
438 m->current_pg = NULL;
439 spin_unlock_irqrestore(&m->lock, flags);
440
441 return NULL;
442 }
443
444 /*
445 * dm_report_EIO() is a macro instead of a function to make pr_debug()
446 * report the function name and line number of the function from which
447 * it has been invoked.
448 */
449 #define dm_report_EIO(m) \
450 ({ \
451 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
452 \
453 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
454 dm_device_name(md), \
455 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
456 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
457 dm_noflush_suspending((m)->ti)); \
458 -EIO; \
459 })
460
461 /*
462 * Map cloned requests (request-based multipath)
463 */
464 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
465 union map_info *map_context,
466 struct request **__clone)
467 {
468 struct multipath *m = ti->private;
469 size_t nr_bytes = blk_rq_bytes(rq);
470 struct pgpath *pgpath;
471 struct block_device *bdev;
472 struct dm_mpath_io *mpio = get_mpio(map_context);
473 struct request_queue *q;
474 struct request *clone;
475
476 /* Do we need to select a new pgpath? */
477 pgpath = lockless_dereference(m->current_pgpath);
478 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
479 pgpath = choose_pgpath(m, nr_bytes);
480
481 if (!pgpath) {
482 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
483 return DM_MAPIO_DELAY_REQUEUE;
484 return dm_report_EIO(m); /* Failed */
485 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
486 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
487 if (pg_init_all_paths(m))
488 return DM_MAPIO_DELAY_REQUEUE;
489 return DM_MAPIO_REQUEUE;
490 }
491
492 memset(mpio, 0, sizeof(*mpio));
493 mpio->pgpath = pgpath;
494 mpio->nr_bytes = nr_bytes;
495
496 bdev = pgpath->path.dev->bdev;
497 q = bdev_get_queue(bdev);
498 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
499 if (IS_ERR(clone)) {
500 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
501 bool queue_dying = blk_queue_dying(q);
502 DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
503 PTR_ERR(clone), queue_dying ? " (path offline)" : "");
504 if (queue_dying) {
505 atomic_inc(&m->pg_init_in_progress);
506 activate_or_offline_path(pgpath);
507 return DM_MAPIO_REQUEUE;
508 }
509 return DM_MAPIO_DELAY_REQUEUE;
510 }
511 clone->bio = clone->biotail = NULL;
512 clone->rq_disk = bdev->bd_disk;
513 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
514 *__clone = clone;
515
516 if (pgpath->pg->ps.type->start_io)
517 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
518 &pgpath->path,
519 nr_bytes);
520 return DM_MAPIO_REMAPPED;
521 }
522
523 static void multipath_release_clone(struct request *clone)
524 {
525 blk_put_request(clone);
526 }
527
528 /*
529 * Map cloned bios (bio-based multipath)
530 */
531 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
532 {
533 size_t nr_bytes = bio->bi_iter.bi_size;
534 struct pgpath *pgpath;
535 unsigned long flags;
536 bool queue_io;
537
538 /* Do we need to select a new pgpath? */
539 pgpath = lockless_dereference(m->current_pgpath);
540 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
541 if (!pgpath || !queue_io)
542 pgpath = choose_pgpath(m, nr_bytes);
543
544 if ((pgpath && queue_io) ||
545 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
546 /* Queue for the daemon to resubmit */
547 spin_lock_irqsave(&m->lock, flags);
548 bio_list_add(&m->queued_bios, bio);
549 spin_unlock_irqrestore(&m->lock, flags);
550 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
551 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
552 pg_init_all_paths(m);
553 else if (!queue_io)
554 queue_work(kmultipathd, &m->process_queued_bios);
555 return DM_MAPIO_SUBMITTED;
556 }
557
558 if (!pgpath) {
559 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
560 return DM_MAPIO_REQUEUE;
561 return dm_report_EIO(m);
562 }
563
564 mpio->pgpath = pgpath;
565 mpio->nr_bytes = nr_bytes;
566
567 bio->bi_error = 0;
568 bio->bi_bdev = pgpath->path.dev->bdev;
569 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
570
571 if (pgpath->pg->ps.type->start_io)
572 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
573 &pgpath->path,
574 nr_bytes);
575 return DM_MAPIO_REMAPPED;
576 }
577
578 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
579 {
580 struct multipath *m = ti->private;
581 struct dm_mpath_io *mpio = NULL;
582
583 multipath_init_per_bio_data(bio, &mpio, NULL);
584
585 return __multipath_map_bio(m, bio, mpio);
586 }
587
588 static void process_queued_io_list(struct multipath *m)
589 {
590 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
591 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
592 else if (m->queue_mode == DM_TYPE_BIO_BASED)
593 queue_work(kmultipathd, &m->process_queued_bios);
594 }
595
596 static void process_queued_bios(struct work_struct *work)
597 {
598 int r;
599 unsigned long flags;
600 struct bio *bio;
601 struct bio_list bios;
602 struct blk_plug plug;
603 struct multipath *m =
604 container_of(work, struct multipath, process_queued_bios);
605
606 bio_list_init(&bios);
607
608 spin_lock_irqsave(&m->lock, flags);
609
610 if (bio_list_empty(&m->queued_bios)) {
611 spin_unlock_irqrestore(&m->lock, flags);
612 return;
613 }
614
615 bio_list_merge(&bios, &m->queued_bios);
616 bio_list_init(&m->queued_bios);
617
618 spin_unlock_irqrestore(&m->lock, flags);
619
620 blk_start_plug(&plug);
621 while ((bio = bio_list_pop(&bios))) {
622 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
623 if (r < 0 || r == DM_MAPIO_REQUEUE) {
624 bio->bi_error = r;
625 bio_endio(bio);
626 } else if (r == DM_MAPIO_REMAPPED)
627 generic_make_request(bio);
628 }
629 blk_finish_plug(&plug);
630 }
631
632 static void assign_bit(bool value, long nr, unsigned long *addr)
633 {
634 if (value)
635 set_bit(nr, addr);
636 else
637 clear_bit(nr, addr);
638 }
639
640 /*
641 * If we run out of usable paths, should we queue I/O or error it?
642 */
643 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
644 bool save_old_value)
645 {
646 unsigned long flags;
647
648 spin_lock_irqsave(&m->lock, flags);
649 assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
650 (!save_old_value && queue_if_no_path),
651 MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
652 assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
653 MPATHF_QUEUE_IF_NO_PATH, &m->flags);
654 spin_unlock_irqrestore(&m->lock, flags);
655
656 if (!queue_if_no_path) {
657 dm_table_run_md_queue_async(m->ti->table);
658 process_queued_io_list(m);
659 }
660
661 return 0;
662 }
663
664 /*
665 * An event is triggered whenever a path is taken out of use.
666 * Includes path failure and PG bypass.
667 */
668 static void trigger_event(struct work_struct *work)
669 {
670 struct multipath *m =
671 container_of(work, struct multipath, trigger_event);
672
673 dm_table_event(m->ti->table);
674 }
675
676 /*-----------------------------------------------------------------
677 * Constructor/argument parsing:
678 * <#multipath feature args> [<arg>]*
679 * <#hw_handler args> [hw_handler [<arg>]*]
680 * <#priority groups>
681 * <initial priority group>
682 * [<selector> <#selector args> [<arg>]*
683 * <#paths> <#per-path selector args>
684 * [<path> [<arg>]* ]+ ]+
685 *---------------------------------------------------------------*/
686 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
687 struct dm_target *ti)
688 {
689 int r;
690 struct path_selector_type *pst;
691 unsigned ps_argc;
692
693 static struct dm_arg _args[] = {
694 {0, 1024, "invalid number of path selector args"},
695 };
696
697 pst = dm_get_path_selector(dm_shift_arg(as));
698 if (!pst) {
699 ti->error = "unknown path selector type";
700 return -EINVAL;
701 }
702
703 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
704 if (r) {
705 dm_put_path_selector(pst);
706 return -EINVAL;
707 }
708
709 r = pst->create(&pg->ps, ps_argc, as->argv);
710 if (r) {
711 dm_put_path_selector(pst);
712 ti->error = "path selector constructor failed";
713 return r;
714 }
715
716 pg->ps.type = pst;
717 dm_consume_args(as, ps_argc);
718
719 return 0;
720 }
721
722 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
723 struct dm_target *ti)
724 {
725 int r;
726 struct pgpath *p;
727 struct multipath *m = ti->private;
728 struct request_queue *q = NULL;
729 const char *attached_handler_name;
730
731 /* we need at least a path arg */
732 if (as->argc < 1) {
733 ti->error = "no device given";
734 return ERR_PTR(-EINVAL);
735 }
736
737 p = alloc_pgpath();
738 if (!p)
739 return ERR_PTR(-ENOMEM);
740
741 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
742 &p->path.dev);
743 if (r) {
744 ti->error = "error getting device";
745 goto bad;
746 }
747
748 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
749 q = bdev_get_queue(p->path.dev->bdev);
750
751 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
752 retain:
753 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
754 if (attached_handler_name) {
755 /*
756 * Clear any hw_handler_params associated with a
757 * handler that isn't already attached.
758 */
759 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
760 kfree(m->hw_handler_params);
761 m->hw_handler_params = NULL;
762 }
763
764 /*
765 * Reset hw_handler_name to match the attached handler
766 *
767 * NB. This modifies the table line to show the actual
768 * handler instead of the original table passed in.
769 */
770 kfree(m->hw_handler_name);
771 m->hw_handler_name = attached_handler_name;
772 }
773 }
774
775 if (m->hw_handler_name) {
776 r = scsi_dh_attach(q, m->hw_handler_name);
777 if (r == -EBUSY) {
778 char b[BDEVNAME_SIZE];
779
780 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
781 bdevname(p->path.dev->bdev, b));
782 goto retain;
783 }
784 if (r < 0) {
785 ti->error = "error attaching hardware handler";
786 dm_put_device(ti, p->path.dev);
787 goto bad;
788 }
789
790 if (m->hw_handler_params) {
791 r = scsi_dh_set_params(q, m->hw_handler_params);
792 if (r < 0) {
793 ti->error = "unable to set hardware "
794 "handler parameters";
795 dm_put_device(ti, p->path.dev);
796 goto bad;
797 }
798 }
799 }
800
801 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
802 if (r) {
803 dm_put_device(ti, p->path.dev);
804 goto bad;
805 }
806
807 return p;
808
809 bad:
810 free_pgpath(p);
811 return ERR_PTR(r);
812 }
813
814 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
815 struct multipath *m)
816 {
817 static struct dm_arg _args[] = {
818 {1, 1024, "invalid number of paths"},
819 {0, 1024, "invalid number of selector args"}
820 };
821
822 int r;
823 unsigned i, nr_selector_args, nr_args;
824 struct priority_group *pg;
825 struct dm_target *ti = m->ti;
826
827 if (as->argc < 2) {
828 as->argc = 0;
829 ti->error = "not enough priority group arguments";
830 return ERR_PTR(-EINVAL);
831 }
832
833 pg = alloc_priority_group();
834 if (!pg) {
835 ti->error = "couldn't allocate priority group";
836 return ERR_PTR(-ENOMEM);
837 }
838 pg->m = m;
839
840 r = parse_path_selector(as, pg, ti);
841 if (r)
842 goto bad;
843
844 /*
845 * read the paths
846 */
847 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
848 if (r)
849 goto bad;
850
851 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
852 if (r)
853 goto bad;
854
855 nr_args = 1 + nr_selector_args;
856 for (i = 0; i < pg->nr_pgpaths; i++) {
857 struct pgpath *pgpath;
858 struct dm_arg_set path_args;
859
860 if (as->argc < nr_args) {
861 ti->error = "not enough path parameters";
862 r = -EINVAL;
863 goto bad;
864 }
865
866 path_args.argc = nr_args;
867 path_args.argv = as->argv;
868
869 pgpath = parse_path(&path_args, &pg->ps, ti);
870 if (IS_ERR(pgpath)) {
871 r = PTR_ERR(pgpath);
872 goto bad;
873 }
874
875 pgpath->pg = pg;
876 list_add_tail(&pgpath->list, &pg->pgpaths);
877 dm_consume_args(as, nr_args);
878 }
879
880 return pg;
881
882 bad:
883 free_priority_group(pg, ti);
884 return ERR_PTR(r);
885 }
886
887 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
888 {
889 unsigned hw_argc;
890 int ret;
891 struct dm_target *ti = m->ti;
892
893 static struct dm_arg _args[] = {
894 {0, 1024, "invalid number of hardware handler args"},
895 };
896
897 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
898 return -EINVAL;
899
900 if (!hw_argc)
901 return 0;
902
903 if (m->queue_mode == DM_TYPE_BIO_BASED) {
904 dm_consume_args(as, hw_argc);
905 DMERR("bio-based multipath doesn't allow hardware handler args");
906 return 0;
907 }
908
909 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
910 if (!m->hw_handler_name)
911 return -EINVAL;
912
913 if (hw_argc > 1) {
914 char *p;
915 int i, j, len = 4;
916
917 for (i = 0; i <= hw_argc - 2; i++)
918 len += strlen(as->argv[i]) + 1;
919 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
920 if (!p) {
921 ti->error = "memory allocation failed";
922 ret = -ENOMEM;
923 goto fail;
924 }
925 j = sprintf(p, "%d", hw_argc - 1);
926 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
927 j = sprintf(p, "%s", as->argv[i]);
928 }
929 dm_consume_args(as, hw_argc - 1);
930
931 return 0;
932 fail:
933 kfree(m->hw_handler_name);
934 m->hw_handler_name = NULL;
935 return ret;
936 }
937
938 static int parse_features(struct dm_arg_set *as, struct multipath *m)
939 {
940 int r;
941 unsigned argc;
942 struct dm_target *ti = m->ti;
943 const char *arg_name;
944
945 static struct dm_arg _args[] = {
946 {0, 8, "invalid number of feature args"},
947 {1, 50, "pg_init_retries must be between 1 and 50"},
948 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
949 };
950
951 r = dm_read_arg_group(_args, as, &argc, &ti->error);
952 if (r)
953 return -EINVAL;
954
955 if (!argc)
956 return 0;
957
958 do {
959 arg_name = dm_shift_arg(as);
960 argc--;
961
962 if (!strcasecmp(arg_name, "queue_if_no_path")) {
963 r = queue_if_no_path(m, true, false);
964 continue;
965 }
966
967 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
968 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
969 continue;
970 }
971
972 if (!strcasecmp(arg_name, "pg_init_retries") &&
973 (argc >= 1)) {
974 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
975 argc--;
976 continue;
977 }
978
979 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
980 (argc >= 1)) {
981 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
982 argc--;
983 continue;
984 }
985
986 if (!strcasecmp(arg_name, "queue_mode") &&
987 (argc >= 1)) {
988 const char *queue_mode_name = dm_shift_arg(as);
989
990 if (!strcasecmp(queue_mode_name, "bio"))
991 m->queue_mode = DM_TYPE_BIO_BASED;
992 else if (!strcasecmp(queue_mode_name, "rq"))
993 m->queue_mode = DM_TYPE_REQUEST_BASED;
994 else if (!strcasecmp(queue_mode_name, "mq"))
995 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
996 else {
997 ti->error = "Unknown 'queue_mode' requested";
998 r = -EINVAL;
999 }
1000 argc--;
1001 continue;
1002 }
1003
1004 ti->error = "Unrecognised multipath feature request";
1005 r = -EINVAL;
1006 } while (argc && !r);
1007
1008 return r;
1009 }
1010
1011 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1012 {
1013 /* target arguments */
1014 static struct dm_arg _args[] = {
1015 {0, 1024, "invalid number of priority groups"},
1016 {0, 1024, "invalid initial priority group number"},
1017 };
1018
1019 int r;
1020 struct multipath *m;
1021 struct dm_arg_set as;
1022 unsigned pg_count = 0;
1023 unsigned next_pg_num;
1024
1025 as.argc = argc;
1026 as.argv = argv;
1027
1028 m = alloc_multipath(ti);
1029 if (!m) {
1030 ti->error = "can't allocate multipath";
1031 return -EINVAL;
1032 }
1033
1034 r = parse_features(&as, m);
1035 if (r)
1036 goto bad;
1037
1038 r = alloc_multipath_stage2(ti, m);
1039 if (r)
1040 goto bad;
1041
1042 r = parse_hw_handler(&as, m);
1043 if (r)
1044 goto bad;
1045
1046 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1047 if (r)
1048 goto bad;
1049
1050 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1051 if (r)
1052 goto bad;
1053
1054 if ((!m->nr_priority_groups && next_pg_num) ||
1055 (m->nr_priority_groups && !next_pg_num)) {
1056 ti->error = "invalid initial priority group";
1057 r = -EINVAL;
1058 goto bad;
1059 }
1060
1061 /* parse the priority groups */
1062 while (as.argc) {
1063 struct priority_group *pg;
1064 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1065
1066 pg = parse_priority_group(&as, m);
1067 if (IS_ERR(pg)) {
1068 r = PTR_ERR(pg);
1069 goto bad;
1070 }
1071
1072 nr_valid_paths += pg->nr_pgpaths;
1073 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1074
1075 list_add_tail(&pg->list, &m->priority_groups);
1076 pg_count++;
1077 pg->pg_num = pg_count;
1078 if (!--next_pg_num)
1079 m->next_pg = pg;
1080 }
1081
1082 if (pg_count != m->nr_priority_groups) {
1083 ti->error = "priority group count mismatch";
1084 r = -EINVAL;
1085 goto bad;
1086 }
1087
1088 ti->num_flush_bios = 1;
1089 ti->num_discard_bios = 1;
1090 ti->num_write_same_bios = 1;
1091 if (m->queue_mode == DM_TYPE_BIO_BASED)
1092 ti->per_io_data_size = multipath_per_bio_data_size();
1093 else
1094 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1095
1096 return 0;
1097
1098 bad:
1099 free_multipath(m);
1100 return r;
1101 }
1102
1103 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1104 {
1105 DEFINE_WAIT(wait);
1106
1107 while (1) {
1108 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1109
1110 if (!atomic_read(&m->pg_init_in_progress))
1111 break;
1112
1113 io_schedule();
1114 }
1115 finish_wait(&m->pg_init_wait, &wait);
1116 }
1117
1118 static void flush_multipath_work(struct multipath *m)
1119 {
1120 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1121 smp_mb__after_atomic();
1122
1123 flush_workqueue(kmpath_handlerd);
1124 multipath_wait_for_pg_init_completion(m);
1125 flush_workqueue(kmultipathd);
1126 flush_work(&m->trigger_event);
1127
1128 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1129 smp_mb__after_atomic();
1130 }
1131
1132 static void multipath_dtr(struct dm_target *ti)
1133 {
1134 struct multipath *m = ti->private;
1135
1136 flush_multipath_work(m);
1137 free_multipath(m);
1138 }
1139
1140 /*
1141 * Take a path out of use.
1142 */
1143 static int fail_path(struct pgpath *pgpath)
1144 {
1145 unsigned long flags;
1146 struct multipath *m = pgpath->pg->m;
1147
1148 spin_lock_irqsave(&m->lock, flags);
1149
1150 if (!pgpath->is_active)
1151 goto out;
1152
1153 DMWARN("Failing path %s.", pgpath->path.dev->name);
1154
1155 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1156 pgpath->is_active = false;
1157 pgpath->fail_count++;
1158
1159 atomic_dec(&m->nr_valid_paths);
1160
1161 if (pgpath == m->current_pgpath)
1162 m->current_pgpath = NULL;
1163
1164 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1165 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1166
1167 schedule_work(&m->trigger_event);
1168
1169 out:
1170 spin_unlock_irqrestore(&m->lock, flags);
1171
1172 return 0;
1173 }
1174
1175 /*
1176 * Reinstate a previously-failed path
1177 */
1178 static int reinstate_path(struct pgpath *pgpath)
1179 {
1180 int r = 0, run_queue = 0;
1181 unsigned long flags;
1182 struct multipath *m = pgpath->pg->m;
1183 unsigned nr_valid_paths;
1184
1185 spin_lock_irqsave(&m->lock, flags);
1186
1187 if (pgpath->is_active)
1188 goto out;
1189
1190 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1191
1192 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1193 if (r)
1194 goto out;
1195
1196 pgpath->is_active = true;
1197
1198 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1199 if (nr_valid_paths == 1) {
1200 m->current_pgpath = NULL;
1201 run_queue = 1;
1202 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1203 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1204 atomic_inc(&m->pg_init_in_progress);
1205 }
1206
1207 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1208 pgpath->path.dev->name, nr_valid_paths);
1209
1210 schedule_work(&m->trigger_event);
1211
1212 out:
1213 spin_unlock_irqrestore(&m->lock, flags);
1214 if (run_queue) {
1215 dm_table_run_md_queue_async(m->ti->table);
1216 process_queued_io_list(m);
1217 }
1218
1219 return r;
1220 }
1221
1222 /*
1223 * Fail or reinstate all paths that match the provided struct dm_dev.
1224 */
1225 static int action_dev(struct multipath *m, struct dm_dev *dev,
1226 action_fn action)
1227 {
1228 int r = -EINVAL;
1229 struct pgpath *pgpath;
1230 struct priority_group *pg;
1231
1232 list_for_each_entry(pg, &m->priority_groups, list) {
1233 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1234 if (pgpath->path.dev == dev)
1235 r = action(pgpath);
1236 }
1237 }
1238
1239 return r;
1240 }
1241
1242 /*
1243 * Temporarily try to avoid having to use the specified PG
1244 */
1245 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1246 bool bypassed)
1247 {
1248 unsigned long flags;
1249
1250 spin_lock_irqsave(&m->lock, flags);
1251
1252 pg->bypassed = bypassed;
1253 m->current_pgpath = NULL;
1254 m->current_pg = NULL;
1255
1256 spin_unlock_irqrestore(&m->lock, flags);
1257
1258 schedule_work(&m->trigger_event);
1259 }
1260
1261 /*
1262 * Switch to using the specified PG from the next I/O that gets mapped
1263 */
1264 static int switch_pg_num(struct multipath *m, const char *pgstr)
1265 {
1266 struct priority_group *pg;
1267 unsigned pgnum;
1268 unsigned long flags;
1269 char dummy;
1270
1271 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1272 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1273 DMWARN("invalid PG number supplied to switch_pg_num");
1274 return -EINVAL;
1275 }
1276
1277 spin_lock_irqsave(&m->lock, flags);
1278 list_for_each_entry(pg, &m->priority_groups, list) {
1279 pg->bypassed = false;
1280 if (--pgnum)
1281 continue;
1282
1283 m->current_pgpath = NULL;
1284 m->current_pg = NULL;
1285 m->next_pg = pg;
1286 }
1287 spin_unlock_irqrestore(&m->lock, flags);
1288
1289 schedule_work(&m->trigger_event);
1290 return 0;
1291 }
1292
1293 /*
1294 * Set/clear bypassed status of a PG.
1295 * PGs are numbered upwards from 1 in the order they were declared.
1296 */
1297 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1298 {
1299 struct priority_group *pg;
1300 unsigned pgnum;
1301 char dummy;
1302
1303 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1304 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1305 DMWARN("invalid PG number supplied to bypass_pg");
1306 return -EINVAL;
1307 }
1308
1309 list_for_each_entry(pg, &m->priority_groups, list) {
1310 if (!--pgnum)
1311 break;
1312 }
1313
1314 bypass_pg(m, pg, bypassed);
1315 return 0;
1316 }
1317
1318 /*
1319 * Should we retry pg_init immediately?
1320 */
1321 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1322 {
1323 unsigned long flags;
1324 bool limit_reached = false;
1325
1326 spin_lock_irqsave(&m->lock, flags);
1327
1328 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1329 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1330 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1331 else
1332 limit_reached = true;
1333
1334 spin_unlock_irqrestore(&m->lock, flags);
1335
1336 return limit_reached;
1337 }
1338
1339 static void pg_init_done(void *data, int errors)
1340 {
1341 struct pgpath *pgpath = data;
1342 struct priority_group *pg = pgpath->pg;
1343 struct multipath *m = pg->m;
1344 unsigned long flags;
1345 bool delay_retry = false;
1346
1347 /* device or driver problems */
1348 switch (errors) {
1349 case SCSI_DH_OK:
1350 break;
1351 case SCSI_DH_NOSYS:
1352 if (!m->hw_handler_name) {
1353 errors = 0;
1354 break;
1355 }
1356 DMERR("Could not failover the device: Handler scsi_dh_%s "
1357 "Error %d.", m->hw_handler_name, errors);
1358 /*
1359 * Fail path for now, so we do not ping pong
1360 */
1361 fail_path(pgpath);
1362 break;
1363 case SCSI_DH_DEV_TEMP_BUSY:
1364 /*
1365 * Probably doing something like FW upgrade on the
1366 * controller so try the other pg.
1367 */
1368 bypass_pg(m, pg, true);
1369 break;
1370 case SCSI_DH_RETRY:
1371 /* Wait before retrying. */
1372 delay_retry = 1;
1373 case SCSI_DH_IMM_RETRY:
1374 case SCSI_DH_RES_TEMP_UNAVAIL:
1375 if (pg_init_limit_reached(m, pgpath))
1376 fail_path(pgpath);
1377 errors = 0;
1378 break;
1379 case SCSI_DH_DEV_OFFLINED:
1380 default:
1381 /*
1382 * We probably do not want to fail the path for a device
1383 * error, but this is what the old dm did. In future
1384 * patches we can do more advanced handling.
1385 */
1386 fail_path(pgpath);
1387 }
1388
1389 spin_lock_irqsave(&m->lock, flags);
1390 if (errors) {
1391 if (pgpath == m->current_pgpath) {
1392 DMERR("Could not failover device. Error %d.", errors);
1393 m->current_pgpath = NULL;
1394 m->current_pg = NULL;
1395 }
1396 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1397 pg->bypassed = false;
1398
1399 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1400 /* Activations of other paths are still on going */
1401 goto out;
1402
1403 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1404 if (delay_retry)
1405 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1406 else
1407 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1408
1409 if (__pg_init_all_paths(m))
1410 goto out;
1411 }
1412 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1413
1414 process_queued_io_list(m);
1415
1416 /*
1417 * Wake up any thread waiting to suspend.
1418 */
1419 wake_up(&m->pg_init_wait);
1420
1421 out:
1422 spin_unlock_irqrestore(&m->lock, flags);
1423 }
1424
1425 static void activate_or_offline_path(struct pgpath *pgpath)
1426 {
1427 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1428
1429 if (pgpath->is_active && !blk_queue_dying(q))
1430 scsi_dh_activate(q, pg_init_done, pgpath);
1431 else
1432 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1433 }
1434
1435 static void activate_path_work(struct work_struct *work)
1436 {
1437 struct pgpath *pgpath =
1438 container_of(work, struct pgpath, activate_path.work);
1439
1440 activate_or_offline_path(pgpath);
1441 }
1442
1443 static int noretry_error(int error)
1444 {
1445 switch (error) {
1446 case -EBADE:
1447 /*
1448 * EBADE signals an reservation conflict.
1449 * We shouldn't fail the path here as we can communicate with
1450 * the target. We should failover to the next path, but in
1451 * doing so we might be causing a ping-pong between paths.
1452 * So just return the reservation conflict error.
1453 */
1454 case -EOPNOTSUPP:
1455 case -EREMOTEIO:
1456 case -EILSEQ:
1457 case -ENODATA:
1458 case -ENOSPC:
1459 return 1;
1460 }
1461
1462 /* Anything else could be a path failure, so should be retried */
1463 return 0;
1464 }
1465
1466 /*
1467 * end_io handling
1468 */
1469 static int do_end_io(struct multipath *m, struct request *clone,
1470 int error, struct dm_mpath_io *mpio)
1471 {
1472 /*
1473 * We don't queue any clone request inside the multipath target
1474 * during end I/O handling, since those clone requests don't have
1475 * bio clones. If we queue them inside the multipath target,
1476 * we need to make bio clones, that requires memory allocation.
1477 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1478 * don't have bio clones.)
1479 * Instead of queueing the clone request here, we queue the original
1480 * request into dm core, which will remake a clone request and
1481 * clone bios for it and resubmit it later.
1482 */
1483 int r = DM_ENDIO_REQUEUE;
1484
1485 if (!error && !clone->errors)
1486 return 0; /* I/O complete */
1487
1488 if (noretry_error(error))
1489 return error;
1490
1491 if (mpio->pgpath)
1492 fail_path(mpio->pgpath);
1493
1494 if (atomic_read(&m->nr_valid_paths) == 0 &&
1495 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1496 r = dm_report_EIO(m);
1497
1498 return r;
1499 }
1500
1501 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1502 int error, union map_info *map_context)
1503 {
1504 struct multipath *m = ti->private;
1505 struct dm_mpath_io *mpio = get_mpio(map_context);
1506 struct pgpath *pgpath;
1507 struct path_selector *ps;
1508 int r;
1509
1510 BUG_ON(!mpio);
1511
1512 r = do_end_io(m, clone, error, mpio);
1513 pgpath = mpio->pgpath;
1514 if (pgpath) {
1515 ps = &pgpath->pg->ps;
1516 if (ps->type->end_io)
1517 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1518 }
1519
1520 return r;
1521 }
1522
1523 static int do_end_io_bio(struct multipath *m, struct bio *clone,
1524 int error, struct dm_mpath_io *mpio)
1525 {
1526 unsigned long flags;
1527
1528 if (!error)
1529 return 0; /* I/O complete */
1530
1531 if (noretry_error(error))
1532 return error;
1533
1534 if (mpio->pgpath)
1535 fail_path(mpio->pgpath);
1536
1537 if (atomic_read(&m->nr_valid_paths) == 0 &&
1538 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1539 return dm_report_EIO(m);
1540
1541 /* Queue for the daemon to resubmit */
1542 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1543
1544 spin_lock_irqsave(&m->lock, flags);
1545 bio_list_add(&m->queued_bios, clone);
1546 spin_unlock_irqrestore(&m->lock, flags);
1547 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1548 queue_work(kmultipathd, &m->process_queued_bios);
1549
1550 return DM_ENDIO_INCOMPLETE;
1551 }
1552
1553 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1554 {
1555 struct multipath *m = ti->private;
1556 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1557 struct pgpath *pgpath;
1558 struct path_selector *ps;
1559 int r;
1560
1561 BUG_ON(!mpio);
1562
1563 r = do_end_io_bio(m, clone, error, mpio);
1564 pgpath = mpio->pgpath;
1565 if (pgpath) {
1566 ps = &pgpath->pg->ps;
1567 if (ps->type->end_io)
1568 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1569 }
1570
1571 return r;
1572 }
1573
1574 /*
1575 * Suspend can't complete until all the I/O is processed so if
1576 * the last path fails we must error any remaining I/O.
1577 * Note that if the freeze_bdev fails while suspending, the
1578 * queue_if_no_path state is lost - userspace should reset it.
1579 */
1580 static void multipath_presuspend(struct dm_target *ti)
1581 {
1582 struct multipath *m = ti->private;
1583
1584 queue_if_no_path(m, false, true);
1585 }
1586
1587 static void multipath_postsuspend(struct dm_target *ti)
1588 {
1589 struct multipath *m = ti->private;
1590
1591 mutex_lock(&m->work_mutex);
1592 flush_multipath_work(m);
1593 mutex_unlock(&m->work_mutex);
1594 }
1595
1596 /*
1597 * Restore the queue_if_no_path setting.
1598 */
1599 static void multipath_resume(struct dm_target *ti)
1600 {
1601 struct multipath *m = ti->private;
1602 unsigned long flags;
1603
1604 spin_lock_irqsave(&m->lock, flags);
1605 assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1606 MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1607 spin_unlock_irqrestore(&m->lock, flags);
1608 }
1609
1610 /*
1611 * Info output has the following format:
1612 * num_multipath_feature_args [multipath_feature_args]*
1613 * num_handler_status_args [handler_status_args]*
1614 * num_groups init_group_number
1615 * [A|D|E num_ps_status_args [ps_status_args]*
1616 * num_paths num_selector_args
1617 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1618 *
1619 * Table output has the following format (identical to the constructor string):
1620 * num_feature_args [features_args]*
1621 * num_handler_args hw_handler [hw_handler_args]*
1622 * num_groups init_group_number
1623 * [priority selector-name num_ps_args [ps_args]*
1624 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1625 */
1626 static void multipath_status(struct dm_target *ti, status_type_t type,
1627 unsigned status_flags, char *result, unsigned maxlen)
1628 {
1629 int sz = 0;
1630 unsigned long flags;
1631 struct multipath *m = ti->private;
1632 struct priority_group *pg;
1633 struct pgpath *p;
1634 unsigned pg_num;
1635 char state;
1636
1637 spin_lock_irqsave(&m->lock, flags);
1638
1639 /* Features */
1640 if (type == STATUSTYPE_INFO)
1641 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1642 atomic_read(&m->pg_init_count));
1643 else {
1644 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1645 (m->pg_init_retries > 0) * 2 +
1646 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1647 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1648 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1649
1650 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1651 DMEMIT("queue_if_no_path ");
1652 if (m->pg_init_retries)
1653 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1654 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1655 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1656 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1657 DMEMIT("retain_attached_hw_handler ");
1658 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1659 switch(m->queue_mode) {
1660 case DM_TYPE_BIO_BASED:
1661 DMEMIT("queue_mode bio ");
1662 break;
1663 case DM_TYPE_MQ_REQUEST_BASED:
1664 DMEMIT("queue_mode mq ");
1665 break;
1666 default:
1667 WARN_ON_ONCE(true);
1668 break;
1669 }
1670 }
1671 }
1672
1673 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1674 DMEMIT("0 ");
1675 else
1676 DMEMIT("1 %s ", m->hw_handler_name);
1677
1678 DMEMIT("%u ", m->nr_priority_groups);
1679
1680 if (m->next_pg)
1681 pg_num = m->next_pg->pg_num;
1682 else if (m->current_pg)
1683 pg_num = m->current_pg->pg_num;
1684 else
1685 pg_num = (m->nr_priority_groups ? 1 : 0);
1686
1687 DMEMIT("%u ", pg_num);
1688
1689 switch (type) {
1690 case STATUSTYPE_INFO:
1691 list_for_each_entry(pg, &m->priority_groups, list) {
1692 if (pg->bypassed)
1693 state = 'D'; /* Disabled */
1694 else if (pg == m->current_pg)
1695 state = 'A'; /* Currently Active */
1696 else
1697 state = 'E'; /* Enabled */
1698
1699 DMEMIT("%c ", state);
1700
1701 if (pg->ps.type->status)
1702 sz += pg->ps.type->status(&pg->ps, NULL, type,
1703 result + sz,
1704 maxlen - sz);
1705 else
1706 DMEMIT("0 ");
1707
1708 DMEMIT("%u %u ", pg->nr_pgpaths,
1709 pg->ps.type->info_args);
1710
1711 list_for_each_entry(p, &pg->pgpaths, list) {
1712 DMEMIT("%s %s %u ", p->path.dev->name,
1713 p->is_active ? "A" : "F",
1714 p->fail_count);
1715 if (pg->ps.type->status)
1716 sz += pg->ps.type->status(&pg->ps,
1717 &p->path, type, result + sz,
1718 maxlen - sz);
1719 }
1720 }
1721 break;
1722
1723 case STATUSTYPE_TABLE:
1724 list_for_each_entry(pg, &m->priority_groups, list) {
1725 DMEMIT("%s ", pg->ps.type->name);
1726
1727 if (pg->ps.type->status)
1728 sz += pg->ps.type->status(&pg->ps, NULL, type,
1729 result + sz,
1730 maxlen - sz);
1731 else
1732 DMEMIT("0 ");
1733
1734 DMEMIT("%u %u ", pg->nr_pgpaths,
1735 pg->ps.type->table_args);
1736
1737 list_for_each_entry(p, &pg->pgpaths, list) {
1738 DMEMIT("%s ", p->path.dev->name);
1739 if (pg->ps.type->status)
1740 sz += pg->ps.type->status(&pg->ps,
1741 &p->path, type, result + sz,
1742 maxlen - sz);
1743 }
1744 }
1745 break;
1746 }
1747
1748 spin_unlock_irqrestore(&m->lock, flags);
1749 }
1750
1751 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1752 {
1753 int r = -EINVAL;
1754 struct dm_dev *dev;
1755 struct multipath *m = ti->private;
1756 action_fn action;
1757
1758 mutex_lock(&m->work_mutex);
1759
1760 if (dm_suspended(ti)) {
1761 r = -EBUSY;
1762 goto out;
1763 }
1764
1765 if (argc == 1) {
1766 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1767 r = queue_if_no_path(m, true, false);
1768 goto out;
1769 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1770 r = queue_if_no_path(m, false, false);
1771 goto out;
1772 }
1773 }
1774
1775 if (argc != 2) {
1776 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1777 goto out;
1778 }
1779
1780 if (!strcasecmp(argv[0], "disable_group")) {
1781 r = bypass_pg_num(m, argv[1], true);
1782 goto out;
1783 } else if (!strcasecmp(argv[0], "enable_group")) {
1784 r = bypass_pg_num(m, argv[1], false);
1785 goto out;
1786 } else if (!strcasecmp(argv[0], "switch_group")) {
1787 r = switch_pg_num(m, argv[1]);
1788 goto out;
1789 } else if (!strcasecmp(argv[0], "reinstate_path"))
1790 action = reinstate_path;
1791 else if (!strcasecmp(argv[0], "fail_path"))
1792 action = fail_path;
1793 else {
1794 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1795 goto out;
1796 }
1797
1798 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1799 if (r) {
1800 DMWARN("message: error getting device %s",
1801 argv[1]);
1802 goto out;
1803 }
1804
1805 r = action_dev(m, dev, action);
1806
1807 dm_put_device(ti, dev);
1808
1809 out:
1810 mutex_unlock(&m->work_mutex);
1811 return r;
1812 }
1813
1814 static int multipath_prepare_ioctl(struct dm_target *ti,
1815 struct block_device **bdev, fmode_t *mode)
1816 {
1817 struct multipath *m = ti->private;
1818 struct pgpath *current_pgpath;
1819 int r;
1820
1821 current_pgpath = lockless_dereference(m->current_pgpath);
1822 if (!current_pgpath)
1823 current_pgpath = choose_pgpath(m, 0);
1824
1825 if (current_pgpath) {
1826 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1827 *bdev = current_pgpath->path.dev->bdev;
1828 *mode = current_pgpath->path.dev->mode;
1829 r = 0;
1830 } else {
1831 /* pg_init has not started or completed */
1832 r = -ENOTCONN;
1833 }
1834 } else {
1835 /* No path is available */
1836 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1837 r = -ENOTCONN;
1838 else
1839 r = -EIO;
1840 }
1841
1842 if (r == -ENOTCONN) {
1843 if (!lockless_dereference(m->current_pg)) {
1844 /* Path status changed, redo selection */
1845 (void) choose_pgpath(m, 0);
1846 }
1847 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1848 pg_init_all_paths(m);
1849 dm_table_run_md_queue_async(m->ti->table);
1850 process_queued_io_list(m);
1851 }
1852
1853 /*
1854 * Only pass ioctls through if the device sizes match exactly.
1855 */
1856 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1857 return 1;
1858 return r;
1859 }
1860
1861 static int multipath_iterate_devices(struct dm_target *ti,
1862 iterate_devices_callout_fn fn, void *data)
1863 {
1864 struct multipath *m = ti->private;
1865 struct priority_group *pg;
1866 struct pgpath *p;
1867 int ret = 0;
1868
1869 list_for_each_entry(pg, &m->priority_groups, list) {
1870 list_for_each_entry(p, &pg->pgpaths, list) {
1871 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1872 if (ret)
1873 goto out;
1874 }
1875 }
1876
1877 out:
1878 return ret;
1879 }
1880
1881 static int pgpath_busy(struct pgpath *pgpath)
1882 {
1883 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1884
1885 return blk_lld_busy(q);
1886 }
1887
1888 /*
1889 * We return "busy", only when we can map I/Os but underlying devices
1890 * are busy (so even if we map I/Os now, the I/Os will wait on
1891 * the underlying queue).
1892 * In other words, if we want to kill I/Os or queue them inside us
1893 * due to map unavailability, we don't return "busy". Otherwise,
1894 * dm core won't give us the I/Os and we can't do what we want.
1895 */
1896 static int multipath_busy(struct dm_target *ti)
1897 {
1898 bool busy = false, has_active = false;
1899 struct multipath *m = ti->private;
1900 struct priority_group *pg, *next_pg;
1901 struct pgpath *pgpath;
1902
1903 /* pg_init in progress */
1904 if (atomic_read(&m->pg_init_in_progress))
1905 return true;
1906
1907 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1908 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1909 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1910
1911 /* Guess which priority_group will be used at next mapping time */
1912 pg = lockless_dereference(m->current_pg);
1913 next_pg = lockless_dereference(m->next_pg);
1914 if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1915 pg = next_pg;
1916
1917 if (!pg) {
1918 /*
1919 * We don't know which pg will be used at next mapping time.
1920 * We don't call choose_pgpath() here to avoid to trigger
1921 * pg_init just by busy checking.
1922 * So we don't know whether underlying devices we will be using
1923 * at next mapping time are busy or not. Just try mapping.
1924 */
1925 return busy;
1926 }
1927
1928 /*
1929 * If there is one non-busy active path at least, the path selector
1930 * will be able to select it. So we consider such a pg as not busy.
1931 */
1932 busy = true;
1933 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1934 if (pgpath->is_active) {
1935 has_active = true;
1936 if (!pgpath_busy(pgpath)) {
1937 busy = false;
1938 break;
1939 }
1940 }
1941 }
1942
1943 if (!has_active) {
1944 /*
1945 * No active path in this pg, so this pg won't be used and
1946 * the current_pg will be changed at next mapping time.
1947 * We need to try mapping to determine it.
1948 */
1949 busy = false;
1950 }
1951
1952 return busy;
1953 }
1954
1955 /*-----------------------------------------------------------------
1956 * Module setup
1957 *---------------------------------------------------------------*/
1958 static struct target_type multipath_target = {
1959 .name = "multipath",
1960 .version = {1, 12, 0},
1961 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1962 .module = THIS_MODULE,
1963 .ctr = multipath_ctr,
1964 .dtr = multipath_dtr,
1965 .clone_and_map_rq = multipath_clone_and_map,
1966 .release_clone_rq = multipath_release_clone,
1967 .rq_end_io = multipath_end_io,
1968 .map = multipath_map_bio,
1969 .end_io = multipath_end_io_bio,
1970 .presuspend = multipath_presuspend,
1971 .postsuspend = multipath_postsuspend,
1972 .resume = multipath_resume,
1973 .status = multipath_status,
1974 .message = multipath_message,
1975 .prepare_ioctl = multipath_prepare_ioctl,
1976 .iterate_devices = multipath_iterate_devices,
1977 .busy = multipath_busy,
1978 };
1979
1980 static int __init dm_multipath_init(void)
1981 {
1982 int r;
1983
1984 r = dm_register_target(&multipath_target);
1985 if (r < 0) {
1986 DMERR("request-based register failed %d", r);
1987 r = -EINVAL;
1988 goto bad_register_target;
1989 }
1990
1991 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1992 if (!kmultipathd) {
1993 DMERR("failed to create workqueue kmpathd");
1994 r = -ENOMEM;
1995 goto bad_alloc_kmultipathd;
1996 }
1997
1998 /*
1999 * A separate workqueue is used to handle the device handlers
2000 * to avoid overloading existing workqueue. Overloading the
2001 * old workqueue would also create a bottleneck in the
2002 * path of the storage hardware device activation.
2003 */
2004 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2005 WQ_MEM_RECLAIM);
2006 if (!kmpath_handlerd) {
2007 DMERR("failed to create workqueue kmpath_handlerd");
2008 r = -ENOMEM;
2009 goto bad_alloc_kmpath_handlerd;
2010 }
2011
2012 return 0;
2013
2014 bad_alloc_kmpath_handlerd:
2015 destroy_workqueue(kmultipathd);
2016 bad_alloc_kmultipathd:
2017 dm_unregister_target(&multipath_target);
2018 bad_register_target:
2019 return r;
2020 }
2021
2022 static void __exit dm_multipath_exit(void)
2023 {
2024 destroy_workqueue(kmpath_handlerd);
2025 destroy_workqueue(kmultipathd);
2026
2027 dm_unregister_target(&multipath_target);
2028 }
2029
2030 module_init(dm_multipath_init);
2031 module_exit(dm_multipath_exit);
2032
2033 MODULE_DESCRIPTION(DM_NAME " multipath target");
2034 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2035 MODULE_LICENSE("GPL");
Ubuntu Artful kernel mirror