]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/md/dm-mpath.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:


a3223c8fbaacc290c8457b49a20ea2a5f10419f4
[mirror_ubuntu-bionic-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(READ_ONCE(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 (READ_ONCE(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 = READ_ONCE(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 do { \
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 } while (0)
459
460 /*
461 * Check whether bios must be queued in the device-mapper core rather
462 * than here in the target.
463 *
464 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
465 * the same value then we are not between multipath_presuspend()
466 * and multipath_resume() calls and we have no need to check
467 * for the DMF_NOFLUSH_SUSPENDING flag.
468 */
469 static bool __must_push_back(struct multipath *m, unsigned long flags)
470 {
471 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
472 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
473 dm_noflush_suspending(m->ti));
474 }
475
476 /*
477 * Following functions use READ_ONCE to get atomic access to
478 * all m->flags to avoid taking spinlock
479 */
480 static bool must_push_back_rq(struct multipath *m)
481 {
482 unsigned long flags = READ_ONCE(m->flags);
483 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
484 }
485
486 static bool must_push_back_bio(struct multipath *m)
487 {
488 unsigned long flags = READ_ONCE(m->flags);
489 return __must_push_back(m, flags);
490 }
491
492 /*
493 * Map cloned requests (request-based multipath)
494 */
495 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
496 union map_info *map_context,
497 struct request **__clone)
498 {
499 struct multipath *m = ti->private;
500 size_t nr_bytes = blk_rq_bytes(rq);
501 struct pgpath *pgpath;
502 struct block_device *bdev;
503 struct dm_mpath_io *mpio = get_mpio(map_context);
504 struct request_queue *q;
505 struct request *clone;
506
507 /* Do we need to select a new pgpath? */
508 pgpath = READ_ONCE(m->current_pgpath);
509 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
510 pgpath = choose_pgpath(m, nr_bytes);
511
512 if (!pgpath) {
513 if (must_push_back_rq(m))
514 return DM_MAPIO_DELAY_REQUEUE;
515 dm_report_EIO(m); /* Failed */
516 return DM_MAPIO_KILL;
517 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
518 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
519 if (pg_init_all_paths(m))
520 return DM_MAPIO_DELAY_REQUEUE;
521 return DM_MAPIO_REQUEUE;
522 }
523
524 memset(mpio, 0, sizeof(*mpio));
525 mpio->pgpath = pgpath;
526 mpio->nr_bytes = nr_bytes;
527
528 bdev = pgpath->path.dev->bdev;
529 q = bdev_get_queue(bdev);
530 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
531 if (IS_ERR(clone)) {
532 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
533 bool queue_dying = blk_queue_dying(q);
534 if (queue_dying) {
535 atomic_inc(&m->pg_init_in_progress);
536 activate_or_offline_path(pgpath);
537 return DM_MAPIO_DELAY_REQUEUE;
538 }
539
540 /*
541 * blk-mq's SCHED_RESTART can cover this requeue, so we
542 * needn't deal with it by DELAY_REQUEUE. More importantly,
543 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
544 * get the queue busy feedback (via BLK_STS_RESOURCE),
545 * otherwise I/O merging can suffer.
546 */
547 if (q->mq_ops)
548 return DM_MAPIO_REQUEUE;
549 else
550 return DM_MAPIO_DELAY_REQUEUE;
551 }
552 clone->bio = clone->biotail = NULL;
553 clone->rq_disk = bdev->bd_disk;
554 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
555 *__clone = clone;
556
557 if (pgpath->pg->ps.type->start_io)
558 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
559 &pgpath->path,
560 nr_bytes);
561 return DM_MAPIO_REMAPPED;
562 }
563
564 static void multipath_release_clone(struct request *clone)
565 {
566 blk_put_request(clone);
567 }
568
569 /*
570 * Map cloned bios (bio-based multipath)
571 */
572 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
573 {
574 size_t nr_bytes = bio->bi_iter.bi_size;
575 struct pgpath *pgpath;
576 unsigned long flags;
577 bool queue_io;
578
579 /* Do we need to select a new pgpath? */
580 pgpath = READ_ONCE(m->current_pgpath);
581 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
582 if (!pgpath || !queue_io)
583 pgpath = choose_pgpath(m, nr_bytes);
584
585 if ((pgpath && queue_io) ||
586 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
587 /* Queue for the daemon to resubmit */
588 spin_lock_irqsave(&m->lock, flags);
589 bio_list_add(&m->queued_bios, bio);
590 spin_unlock_irqrestore(&m->lock, flags);
591 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
592 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
593 pg_init_all_paths(m);
594 else if (!queue_io)
595 queue_work(kmultipathd, &m->process_queued_bios);
596 return DM_MAPIO_SUBMITTED;
597 }
598
599 if (!pgpath) {
600 if (must_push_back_bio(m))
601 return DM_MAPIO_REQUEUE;
602 dm_report_EIO(m);
603 return DM_MAPIO_KILL;
604 }
605
606 mpio->pgpath = pgpath;
607 mpio->nr_bytes = nr_bytes;
608
609 bio->bi_status = 0;
610 bio_set_dev(bio, pgpath->path.dev->bdev);
611 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
612
613 if (pgpath->pg->ps.type->start_io)
614 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
615 &pgpath->path,
616 nr_bytes);
617 return DM_MAPIO_REMAPPED;
618 }
619
620 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
621 {
622 struct multipath *m = ti->private;
623 struct dm_mpath_io *mpio = NULL;
624
625 multipath_init_per_bio_data(bio, &mpio, NULL);
626
627 return __multipath_map_bio(m, bio, mpio);
628 }
629
630 static void process_queued_io_list(struct multipath *m)
631 {
632 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
633 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
634 else if (m->queue_mode == DM_TYPE_BIO_BASED)
635 queue_work(kmultipathd, &m->process_queued_bios);
636 }
637
638 static void process_queued_bios(struct work_struct *work)
639 {
640 int r;
641 unsigned long flags;
642 struct bio *bio;
643 struct bio_list bios;
644 struct blk_plug plug;
645 struct multipath *m =
646 container_of(work, struct multipath, process_queued_bios);
647
648 bio_list_init(&bios);
649
650 spin_lock_irqsave(&m->lock, flags);
651
652 if (bio_list_empty(&m->queued_bios)) {
653 spin_unlock_irqrestore(&m->lock, flags);
654 return;
655 }
656
657 bio_list_merge(&bios, &m->queued_bios);
658 bio_list_init(&m->queued_bios);
659
660 spin_unlock_irqrestore(&m->lock, flags);
661
662 blk_start_plug(&plug);
663 while ((bio = bio_list_pop(&bios))) {
664 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
665 switch (r) {
666 case DM_MAPIO_KILL:
667 bio->bi_status = BLK_STS_IOERR;
668 bio_endio(bio);
669 break;
670 case DM_MAPIO_REQUEUE:
671 bio->bi_status = BLK_STS_DM_REQUEUE;
672 bio_endio(bio);
673 break;
674 case DM_MAPIO_REMAPPED:
675 generic_make_request(bio);
676 break;
677 case 0:
678 break;
679 default:
680 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
681 }
682 }
683 blk_finish_plug(&plug);
684 }
685
686 /*
687 * If we run out of usable paths, should we queue I/O or error it?
688 */
689 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
690 bool save_old_value)
691 {
692 unsigned long flags;
693
694 spin_lock_irqsave(&m->lock, flags);
695 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
696 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
697 (!save_old_value && queue_if_no_path));
698 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
699 spin_unlock_irqrestore(&m->lock, flags);
700
701 if (!queue_if_no_path) {
702 dm_table_run_md_queue_async(m->ti->table);
703 process_queued_io_list(m);
704 }
705
706 return 0;
707 }
708
709 /*
710 * An event is triggered whenever a path is taken out of use.
711 * Includes path failure and PG bypass.
712 */
713 static void trigger_event(struct work_struct *work)
714 {
715 struct multipath *m =
716 container_of(work, struct multipath, trigger_event);
717
718 dm_table_event(m->ti->table);
719 }
720
721 /*-----------------------------------------------------------------
722 * Constructor/argument parsing:
723 * <#multipath feature args> [<arg>]*
724 * <#hw_handler args> [hw_handler [<arg>]*]
725 * <#priority groups>
726 * <initial priority group>
727 * [<selector> <#selector args> [<arg>]*
728 * <#paths> <#per-path selector args>
729 * [<path> [<arg>]* ]+ ]+
730 *---------------------------------------------------------------*/
731 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
732 struct dm_target *ti)
733 {
734 int r;
735 struct path_selector_type *pst;
736 unsigned ps_argc;
737
738 static const struct dm_arg _args[] = {
739 {0, 1024, "invalid number of path selector args"},
740 };
741
742 pst = dm_get_path_selector(dm_shift_arg(as));
743 if (!pst) {
744 ti->error = "unknown path selector type";
745 return -EINVAL;
746 }
747
748 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
749 if (r) {
750 dm_put_path_selector(pst);
751 return -EINVAL;
752 }
753
754 r = pst->create(&pg->ps, ps_argc, as->argv);
755 if (r) {
756 dm_put_path_selector(pst);
757 ti->error = "path selector constructor failed";
758 return r;
759 }
760
761 pg->ps.type = pst;
762 dm_consume_args(as, ps_argc);
763
764 return 0;
765 }
766
767 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
768 struct dm_target *ti)
769 {
770 int r;
771 struct pgpath *p;
772 struct multipath *m = ti->private;
773 struct request_queue *q = NULL;
774 const char *attached_handler_name;
775
776 /* we need at least a path arg */
777 if (as->argc < 1) {
778 ti->error = "no device given";
779 return ERR_PTR(-EINVAL);
780 }
781
782 p = alloc_pgpath();
783 if (!p)
784 return ERR_PTR(-ENOMEM);
785
786 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
787 &p->path.dev);
788 if (r) {
789 ti->error = "error getting device";
790 goto bad;
791 }
792
793 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
794 q = bdev_get_queue(p->path.dev->bdev);
795
796 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
797 retain:
798 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
799 if (attached_handler_name) {
800 /*
801 * Clear any hw_handler_params associated with a
802 * handler that isn't already attached.
803 */
804 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
805 kfree(m->hw_handler_params);
806 m->hw_handler_params = NULL;
807 }
808
809 /*
810 * Reset hw_handler_name to match the attached handler
811 *
812 * NB. This modifies the table line to show the actual
813 * handler instead of the original table passed in.
814 */
815 kfree(m->hw_handler_name);
816 m->hw_handler_name = attached_handler_name;
817 }
818 }
819
820 if (m->hw_handler_name) {
821 r = scsi_dh_attach(q, m->hw_handler_name);
822 if (r == -EBUSY) {
823 char b[BDEVNAME_SIZE];
824
825 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
826 bdevname(p->path.dev->bdev, b));
827 goto retain;
828 }
829 if (r < 0) {
830 ti->error = "error attaching hardware handler";
831 dm_put_device(ti, p->path.dev);
832 goto bad;
833 }
834
835 if (m->hw_handler_params) {
836 r = scsi_dh_set_params(q, m->hw_handler_params);
837 if (r < 0) {
838 ti->error = "unable to set hardware "
839 "handler parameters";
840 dm_put_device(ti, p->path.dev);
841 goto bad;
842 }
843 }
844 }
845
846 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
847 if (r) {
848 dm_put_device(ti, p->path.dev);
849 goto bad;
850 }
851
852 return p;
853
854 bad:
855 free_pgpath(p);
856 return ERR_PTR(r);
857 }
858
859 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
860 struct multipath *m)
861 {
862 static const struct dm_arg _args[] = {
863 {1, 1024, "invalid number of paths"},
864 {0, 1024, "invalid number of selector args"}
865 };
866
867 int r;
868 unsigned i, nr_selector_args, nr_args;
869 struct priority_group *pg;
870 struct dm_target *ti = m->ti;
871
872 if (as->argc < 2) {
873 as->argc = 0;
874 ti->error = "not enough priority group arguments";
875 return ERR_PTR(-EINVAL);
876 }
877
878 pg = alloc_priority_group();
879 if (!pg) {
880 ti->error = "couldn't allocate priority group";
881 return ERR_PTR(-ENOMEM);
882 }
883 pg->m = m;
884
885 r = parse_path_selector(as, pg, ti);
886 if (r)
887 goto bad;
888
889 /*
890 * read the paths
891 */
892 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
893 if (r)
894 goto bad;
895
896 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
897 if (r)
898 goto bad;
899
900 nr_args = 1 + nr_selector_args;
901 for (i = 0; i < pg->nr_pgpaths; i++) {
902 struct pgpath *pgpath;
903 struct dm_arg_set path_args;
904
905 if (as->argc < nr_args) {
906 ti->error = "not enough path parameters";
907 r = -EINVAL;
908 goto bad;
909 }
910
911 path_args.argc = nr_args;
912 path_args.argv = as->argv;
913
914 pgpath = parse_path(&path_args, &pg->ps, ti);
915 if (IS_ERR(pgpath)) {
916 r = PTR_ERR(pgpath);
917 goto bad;
918 }
919
920 pgpath->pg = pg;
921 list_add_tail(&pgpath->list, &pg->pgpaths);
922 dm_consume_args(as, nr_args);
923 }
924
925 return pg;
926
927 bad:
928 free_priority_group(pg, ti);
929 return ERR_PTR(r);
930 }
931
932 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
933 {
934 unsigned hw_argc;
935 int ret;
936 struct dm_target *ti = m->ti;
937
938 static const struct dm_arg _args[] = {
939 {0, 1024, "invalid number of hardware handler args"},
940 };
941
942 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
943 return -EINVAL;
944
945 if (!hw_argc)
946 return 0;
947
948 if (m->queue_mode == DM_TYPE_BIO_BASED) {
949 dm_consume_args(as, hw_argc);
950 DMERR("bio-based multipath doesn't allow hardware handler args");
951 return 0;
952 }
953
954 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
955 if (!m->hw_handler_name)
956 return -EINVAL;
957
958 if (hw_argc > 1) {
959 char *p;
960 int i, j, len = 4;
961
962 for (i = 0; i <= hw_argc - 2; i++)
963 len += strlen(as->argv[i]) + 1;
964 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
965 if (!p) {
966 ti->error = "memory allocation failed";
967 ret = -ENOMEM;
968 goto fail;
969 }
970 j = sprintf(p, "%d", hw_argc - 1);
971 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
972 j = sprintf(p, "%s", as->argv[i]);
973 }
974 dm_consume_args(as, hw_argc - 1);
975
976 return 0;
977 fail:
978 kfree(m->hw_handler_name);
979 m->hw_handler_name = NULL;
980 return ret;
981 }
982
983 static int parse_features(struct dm_arg_set *as, struct multipath *m)
984 {
985 int r;
986 unsigned argc;
987 struct dm_target *ti = m->ti;
988 const char *arg_name;
989
990 static const struct dm_arg _args[] = {
991 {0, 8, "invalid number of feature args"},
992 {1, 50, "pg_init_retries must be between 1 and 50"},
993 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
994 };
995
996 r = dm_read_arg_group(_args, as, &argc, &ti->error);
997 if (r)
998 return -EINVAL;
999
1000 if (!argc)
1001 return 0;
1002
1003 do {
1004 arg_name = dm_shift_arg(as);
1005 argc--;
1006
1007 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1008 r = queue_if_no_path(m, true, false);
1009 continue;
1010 }
1011
1012 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1013 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1014 continue;
1015 }
1016
1017 if (!strcasecmp(arg_name, "pg_init_retries") &&
1018 (argc >= 1)) {
1019 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1020 argc--;
1021 continue;
1022 }
1023
1024 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1025 (argc >= 1)) {
1026 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1027 argc--;
1028 continue;
1029 }
1030
1031 if (!strcasecmp(arg_name, "queue_mode") &&
1032 (argc >= 1)) {
1033 const char *queue_mode_name = dm_shift_arg(as);
1034
1035 if (!strcasecmp(queue_mode_name, "bio"))
1036 m->queue_mode = DM_TYPE_BIO_BASED;
1037 else if (!strcasecmp(queue_mode_name, "rq"))
1038 m->queue_mode = DM_TYPE_REQUEST_BASED;
1039 else if (!strcasecmp(queue_mode_name, "mq"))
1040 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1041 else {
1042 ti->error = "Unknown 'queue_mode' requested";
1043 r = -EINVAL;
1044 }
1045 argc--;
1046 continue;
1047 }
1048
1049 ti->error = "Unrecognised multipath feature request";
1050 r = -EINVAL;
1051 } while (argc && !r);
1052
1053 return r;
1054 }
1055
1056 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1057 {
1058 /* target arguments */
1059 static const struct dm_arg _args[] = {
1060 {0, 1024, "invalid number of priority groups"},
1061 {0, 1024, "invalid initial priority group number"},
1062 };
1063
1064 int r;
1065 struct multipath *m;
1066 struct dm_arg_set as;
1067 unsigned pg_count = 0;
1068 unsigned next_pg_num;
1069
1070 as.argc = argc;
1071 as.argv = argv;
1072
1073 m = alloc_multipath(ti);
1074 if (!m) {
1075 ti->error = "can't allocate multipath";
1076 return -EINVAL;
1077 }
1078
1079 r = parse_features(&as, m);
1080 if (r)
1081 goto bad;
1082
1083 r = alloc_multipath_stage2(ti, m);
1084 if (r)
1085 goto bad;
1086
1087 r = parse_hw_handler(&as, m);
1088 if (r)
1089 goto bad;
1090
1091 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1092 if (r)
1093 goto bad;
1094
1095 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1096 if (r)
1097 goto bad;
1098
1099 if ((!m->nr_priority_groups && next_pg_num) ||
1100 (m->nr_priority_groups && !next_pg_num)) {
1101 ti->error = "invalid initial priority group";
1102 r = -EINVAL;
1103 goto bad;
1104 }
1105
1106 /* parse the priority groups */
1107 while (as.argc) {
1108 struct priority_group *pg;
1109 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1110
1111 pg = parse_priority_group(&as, m);
1112 if (IS_ERR(pg)) {
1113 r = PTR_ERR(pg);
1114 goto bad;
1115 }
1116
1117 nr_valid_paths += pg->nr_pgpaths;
1118 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1119
1120 list_add_tail(&pg->list, &m->priority_groups);
1121 pg_count++;
1122 pg->pg_num = pg_count;
1123 if (!--next_pg_num)
1124 m->next_pg = pg;
1125 }
1126
1127 if (pg_count != m->nr_priority_groups) {
1128 ti->error = "priority group count mismatch";
1129 r = -EINVAL;
1130 goto bad;
1131 }
1132
1133 ti->num_flush_bios = 1;
1134 ti->num_discard_bios = 1;
1135 ti->num_write_same_bios = 1;
1136 ti->num_write_zeroes_bios = 1;
1137 if (m->queue_mode == DM_TYPE_BIO_BASED)
1138 ti->per_io_data_size = multipath_per_bio_data_size();
1139 else
1140 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1141
1142 return 0;
1143
1144 bad:
1145 free_multipath(m);
1146 return r;
1147 }
1148
1149 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1150 {
1151 DEFINE_WAIT(wait);
1152
1153 while (1) {
1154 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1155
1156 if (!atomic_read(&m->pg_init_in_progress))
1157 break;
1158
1159 io_schedule();
1160 }
1161 finish_wait(&m->pg_init_wait, &wait);
1162 }
1163
1164 static void flush_multipath_work(struct multipath *m)
1165 {
1166 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1167 smp_mb__after_atomic();
1168
1169 flush_workqueue(kmpath_handlerd);
1170 multipath_wait_for_pg_init_completion(m);
1171 flush_workqueue(kmultipathd);
1172 flush_work(&m->trigger_event);
1173
1174 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1175 smp_mb__after_atomic();
1176 }
1177
1178 static void multipath_dtr(struct dm_target *ti)
1179 {
1180 struct multipath *m = ti->private;
1181
1182 flush_multipath_work(m);
1183 free_multipath(m);
1184 }
1185
1186 /*
1187 * Take a path out of use.
1188 */
1189 static int fail_path(struct pgpath *pgpath)
1190 {
1191 unsigned long flags;
1192 struct multipath *m = pgpath->pg->m;
1193
1194 spin_lock_irqsave(&m->lock, flags);
1195
1196 if (!pgpath->is_active)
1197 goto out;
1198
1199 DMWARN("Failing path %s.", pgpath->path.dev->name);
1200
1201 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1202 pgpath->is_active = false;
1203 pgpath->fail_count++;
1204
1205 atomic_dec(&m->nr_valid_paths);
1206
1207 if (pgpath == m->current_pgpath)
1208 m->current_pgpath = NULL;
1209
1210 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1211 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1212
1213 schedule_work(&m->trigger_event);
1214
1215 out:
1216 spin_unlock_irqrestore(&m->lock, flags);
1217
1218 return 0;
1219 }
1220
1221 /*
1222 * Reinstate a previously-failed path
1223 */
1224 static int reinstate_path(struct pgpath *pgpath)
1225 {
1226 int r = 0, run_queue = 0;
1227 unsigned long flags;
1228 struct multipath *m = pgpath->pg->m;
1229 unsigned nr_valid_paths;
1230
1231 spin_lock_irqsave(&m->lock, flags);
1232
1233 if (pgpath->is_active)
1234 goto out;
1235
1236 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1237
1238 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1239 if (r)
1240 goto out;
1241
1242 pgpath->is_active = true;
1243
1244 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1245 if (nr_valid_paths == 1) {
1246 m->current_pgpath = NULL;
1247 run_queue = 1;
1248 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1249 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1250 atomic_inc(&m->pg_init_in_progress);
1251 }
1252
1253 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1254 pgpath->path.dev->name, nr_valid_paths);
1255
1256 schedule_work(&m->trigger_event);
1257
1258 out:
1259 spin_unlock_irqrestore(&m->lock, flags);
1260 if (run_queue) {
1261 dm_table_run_md_queue_async(m->ti->table);
1262 process_queued_io_list(m);
1263 }
1264
1265 return r;
1266 }
1267
1268 /*
1269 * Fail or reinstate all paths that match the provided struct dm_dev.
1270 */
1271 static int action_dev(struct multipath *m, struct dm_dev *dev,
1272 action_fn action)
1273 {
1274 int r = -EINVAL;
1275 struct pgpath *pgpath;
1276 struct priority_group *pg;
1277
1278 list_for_each_entry(pg, &m->priority_groups, list) {
1279 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1280 if (pgpath->path.dev == dev)
1281 r = action(pgpath);
1282 }
1283 }
1284
1285 return r;
1286 }
1287
1288 /*
1289 * Temporarily try to avoid having to use the specified PG
1290 */
1291 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1292 bool bypassed)
1293 {
1294 unsigned long flags;
1295
1296 spin_lock_irqsave(&m->lock, flags);
1297
1298 pg->bypassed = bypassed;
1299 m->current_pgpath = NULL;
1300 m->current_pg = NULL;
1301
1302 spin_unlock_irqrestore(&m->lock, flags);
1303
1304 schedule_work(&m->trigger_event);
1305 }
1306
1307 /*
1308 * Switch to using the specified PG from the next I/O that gets mapped
1309 */
1310 static int switch_pg_num(struct multipath *m, const char *pgstr)
1311 {
1312 struct priority_group *pg;
1313 unsigned pgnum;
1314 unsigned long flags;
1315 char dummy;
1316
1317 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1318 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1319 DMWARN("invalid PG number supplied to switch_pg_num");
1320 return -EINVAL;
1321 }
1322
1323 spin_lock_irqsave(&m->lock, flags);
1324 list_for_each_entry(pg, &m->priority_groups, list) {
1325 pg->bypassed = false;
1326 if (--pgnum)
1327 continue;
1328
1329 m->current_pgpath = NULL;
1330 m->current_pg = NULL;
1331 m->next_pg = pg;
1332 }
1333 spin_unlock_irqrestore(&m->lock, flags);
1334
1335 schedule_work(&m->trigger_event);
1336 return 0;
1337 }
1338
1339 /*
1340 * Set/clear bypassed status of a PG.
1341 * PGs are numbered upwards from 1 in the order they were declared.
1342 */
1343 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1344 {
1345 struct priority_group *pg;
1346 unsigned pgnum;
1347 char dummy;
1348
1349 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1350 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1351 DMWARN("invalid PG number supplied to bypass_pg");
1352 return -EINVAL;
1353 }
1354
1355 list_for_each_entry(pg, &m->priority_groups, list) {
1356 if (!--pgnum)
1357 break;
1358 }
1359
1360 bypass_pg(m, pg, bypassed);
1361 return 0;
1362 }
1363
1364 /*
1365 * Should we retry pg_init immediately?
1366 */
1367 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1368 {
1369 unsigned long flags;
1370 bool limit_reached = false;
1371
1372 spin_lock_irqsave(&m->lock, flags);
1373
1374 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1375 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1376 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1377 else
1378 limit_reached = true;
1379
1380 spin_unlock_irqrestore(&m->lock, flags);
1381
1382 return limit_reached;
1383 }
1384
1385 static void pg_init_done(void *data, int errors)
1386 {
1387 struct pgpath *pgpath = data;
1388 struct priority_group *pg = pgpath->pg;
1389 struct multipath *m = pg->m;
1390 unsigned long flags;
1391 bool delay_retry = false;
1392
1393 /* device or driver problems */
1394 switch (errors) {
1395 case SCSI_DH_OK:
1396 break;
1397 case SCSI_DH_NOSYS:
1398 if (!m->hw_handler_name) {
1399 errors = 0;
1400 break;
1401 }
1402 DMERR("Could not failover the device: Handler scsi_dh_%s "
1403 "Error %d.", m->hw_handler_name, errors);
1404 /*
1405 * Fail path for now, so we do not ping pong
1406 */
1407 fail_path(pgpath);
1408 break;
1409 case SCSI_DH_DEV_TEMP_BUSY:
1410 /*
1411 * Probably doing something like FW upgrade on the
1412 * controller so try the other pg.
1413 */
1414 bypass_pg(m, pg, true);
1415 break;
1416 case SCSI_DH_RETRY:
1417 /* Wait before retrying. */
1418 delay_retry = 1;
1419 /* fall through */
1420 case SCSI_DH_IMM_RETRY:
1421 case SCSI_DH_RES_TEMP_UNAVAIL:
1422 if (pg_init_limit_reached(m, pgpath))
1423 fail_path(pgpath);
1424 errors = 0;
1425 break;
1426 case SCSI_DH_DEV_OFFLINED:
1427 default:
1428 /*
1429 * We probably do not want to fail the path for a device
1430 * error, but this is what the old dm did. In future
1431 * patches we can do more advanced handling.
1432 */
1433 fail_path(pgpath);
1434 }
1435
1436 spin_lock_irqsave(&m->lock, flags);
1437 if (errors) {
1438 if (pgpath == m->current_pgpath) {
1439 DMERR("Could not failover device. Error %d.", errors);
1440 m->current_pgpath = NULL;
1441 m->current_pg = NULL;
1442 }
1443 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1444 pg->bypassed = false;
1445
1446 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1447 /* Activations of other paths are still on going */
1448 goto out;
1449
1450 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1451 if (delay_retry)
1452 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1453 else
1454 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1455
1456 if (__pg_init_all_paths(m))
1457 goto out;
1458 }
1459 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1460
1461 process_queued_io_list(m);
1462
1463 /*
1464 * Wake up any thread waiting to suspend.
1465 */
1466 wake_up(&m->pg_init_wait);
1467
1468 out:
1469 spin_unlock_irqrestore(&m->lock, flags);
1470 }
1471
1472 static void activate_or_offline_path(struct pgpath *pgpath)
1473 {
1474 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1475
1476 if (pgpath->is_active && !blk_queue_dying(q))
1477 scsi_dh_activate(q, pg_init_done, pgpath);
1478 else
1479 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1480 }
1481
1482 static void activate_path_work(struct work_struct *work)
1483 {
1484 struct pgpath *pgpath =
1485 container_of(work, struct pgpath, activate_path.work);
1486
1487 activate_or_offline_path(pgpath);
1488 }
1489
1490 static int noretry_error(blk_status_t error)
1491 {
1492 switch (error) {
1493 case BLK_STS_NOTSUPP:
1494 case BLK_STS_NOSPC:
1495 case BLK_STS_TARGET:
1496 case BLK_STS_NEXUS:
1497 case BLK_STS_MEDIUM:
1498 return 1;
1499 }
1500
1501 /* Anything else could be a path failure, so should be retried */
1502 return 0;
1503 }
1504
1505 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1506 blk_status_t error, union map_info *map_context)
1507 {
1508 struct dm_mpath_io *mpio = get_mpio(map_context);
1509 struct pgpath *pgpath = mpio->pgpath;
1510 int r = DM_ENDIO_DONE;
1511
1512 /*
1513 * We don't queue any clone request inside the multipath target
1514 * during end I/O handling, since those clone requests don't have
1515 * bio clones. If we queue them inside the multipath target,
1516 * we need to make bio clones, that requires memory allocation.
1517 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1518 * don't have bio clones.)
1519 * Instead of queueing the clone request here, we queue the original
1520 * request into dm core, which will remake a clone request and
1521 * clone bios for it and resubmit it later.
1522 */
1523 if (error && !noretry_error(error)) {
1524 struct multipath *m = ti->private;
1525
1526 r = DM_ENDIO_REQUEUE;
1527
1528 if (pgpath)
1529 fail_path(pgpath);
1530
1531 if (atomic_read(&m->nr_valid_paths) == 0 &&
1532 !must_push_back_rq(m)) {
1533 if (error == BLK_STS_IOERR)
1534 dm_report_EIO(m);
1535 /* complete with the original error */
1536 r = DM_ENDIO_DONE;
1537 }
1538 }
1539
1540 if (pgpath) {
1541 struct path_selector *ps = &pgpath->pg->ps;
1542
1543 if (ps->type->end_io)
1544 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1545 }
1546
1547 return r;
1548 }
1549
1550 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1551 blk_status_t *error)
1552 {
1553 struct multipath *m = ti->private;
1554 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1555 struct pgpath *pgpath = mpio->pgpath;
1556 unsigned long flags;
1557 int r = DM_ENDIO_DONE;
1558
1559 if (!*error || noretry_error(*error))
1560 goto done;
1561
1562 if (pgpath)
1563 fail_path(pgpath);
1564
1565 if (atomic_read(&m->nr_valid_paths) == 0 &&
1566 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1567 if (must_push_back_bio(m)) {
1568 r = DM_ENDIO_REQUEUE;
1569 } else {
1570 dm_report_EIO(m);
1571 *error = BLK_STS_IOERR;
1572 }
1573 goto done;
1574 }
1575
1576 /* Queue for the daemon to resubmit */
1577 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1578
1579 spin_lock_irqsave(&m->lock, flags);
1580 bio_list_add(&m->queued_bios, clone);
1581 spin_unlock_irqrestore(&m->lock, flags);
1582 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1583 queue_work(kmultipathd, &m->process_queued_bios);
1584
1585 r = DM_ENDIO_INCOMPLETE;
1586 done:
1587 if (pgpath) {
1588 struct path_selector *ps = &pgpath->pg->ps;
1589
1590 if (ps->type->end_io)
1591 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1592 }
1593
1594 return r;
1595 }
1596
1597 /*
1598 * Suspend can't complete until all the I/O is processed so if
1599 * the last path fails we must error any remaining I/O.
1600 * Note that if the freeze_bdev fails while suspending, the
1601 * queue_if_no_path state is lost - userspace should reset it.
1602 */
1603 static void multipath_presuspend(struct dm_target *ti)
1604 {
1605 struct multipath *m = ti->private;
1606
1607 queue_if_no_path(m, false, true);
1608 }
1609
1610 static void multipath_postsuspend(struct dm_target *ti)
1611 {
1612 struct multipath *m = ti->private;
1613
1614 mutex_lock(&m->work_mutex);
1615 flush_multipath_work(m);
1616 mutex_unlock(&m->work_mutex);
1617 }
1618
1619 /*
1620 * Restore the queue_if_no_path setting.
1621 */
1622 static void multipath_resume(struct dm_target *ti)
1623 {
1624 struct multipath *m = ti->private;
1625 unsigned long flags;
1626
1627 spin_lock_irqsave(&m->lock, flags);
1628 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1629 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1630 spin_unlock_irqrestore(&m->lock, flags);
1631 }
1632
1633 /*
1634 * Info output has the following format:
1635 * num_multipath_feature_args [multipath_feature_args]*
1636 * num_handler_status_args [handler_status_args]*
1637 * num_groups init_group_number
1638 * [A|D|E num_ps_status_args [ps_status_args]*
1639 * num_paths num_selector_args
1640 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1641 *
1642 * Table output has the following format (identical to the constructor string):
1643 * num_feature_args [features_args]*
1644 * num_handler_args hw_handler [hw_handler_args]*
1645 * num_groups init_group_number
1646 * [priority selector-name num_ps_args [ps_args]*
1647 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1648 */
1649 static void multipath_status(struct dm_target *ti, status_type_t type,
1650 unsigned status_flags, char *result, unsigned maxlen)
1651 {
1652 int sz = 0;
1653 unsigned long flags;
1654 struct multipath *m = ti->private;
1655 struct priority_group *pg;
1656 struct pgpath *p;
1657 unsigned pg_num;
1658 char state;
1659
1660 spin_lock_irqsave(&m->lock, flags);
1661
1662 /* Features */
1663 if (type == STATUSTYPE_INFO)
1664 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1665 atomic_read(&m->pg_init_count));
1666 else {
1667 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1668 (m->pg_init_retries > 0) * 2 +
1669 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1670 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1671 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1672
1673 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1674 DMEMIT("queue_if_no_path ");
1675 if (m->pg_init_retries)
1676 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1677 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1678 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1679 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1680 DMEMIT("retain_attached_hw_handler ");
1681 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1682 switch(m->queue_mode) {
1683 case DM_TYPE_BIO_BASED:
1684 DMEMIT("queue_mode bio ");
1685 break;
1686 case DM_TYPE_MQ_REQUEST_BASED:
1687 DMEMIT("queue_mode mq ");
1688 break;
1689 default:
1690 WARN_ON_ONCE(true);
1691 break;
1692 }
1693 }
1694 }
1695
1696 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1697 DMEMIT("0 ");
1698 else
1699 DMEMIT("1 %s ", m->hw_handler_name);
1700
1701 DMEMIT("%u ", m->nr_priority_groups);
1702
1703 if (m->next_pg)
1704 pg_num = m->next_pg->pg_num;
1705 else if (m->current_pg)
1706 pg_num = m->current_pg->pg_num;
1707 else
1708 pg_num = (m->nr_priority_groups ? 1 : 0);
1709
1710 DMEMIT("%u ", pg_num);
1711
1712 switch (type) {
1713 case STATUSTYPE_INFO:
1714 list_for_each_entry(pg, &m->priority_groups, list) {
1715 if (pg->bypassed)
1716 state = 'D'; /* Disabled */
1717 else if (pg == m->current_pg)
1718 state = 'A'; /* Currently Active */
1719 else
1720 state = 'E'; /* Enabled */
1721
1722 DMEMIT("%c ", state);
1723
1724 if (pg->ps.type->status)
1725 sz += pg->ps.type->status(&pg->ps, NULL, type,
1726 result + sz,
1727 maxlen - sz);
1728 else
1729 DMEMIT("0 ");
1730
1731 DMEMIT("%u %u ", pg->nr_pgpaths,
1732 pg->ps.type->info_args);
1733
1734 list_for_each_entry(p, &pg->pgpaths, list) {
1735 DMEMIT("%s %s %u ", p->path.dev->name,
1736 p->is_active ? "A" : "F",
1737 p->fail_count);
1738 if (pg->ps.type->status)
1739 sz += pg->ps.type->status(&pg->ps,
1740 &p->path, type, result + sz,
1741 maxlen - sz);
1742 }
1743 }
1744 break;
1745
1746 case STATUSTYPE_TABLE:
1747 list_for_each_entry(pg, &m->priority_groups, list) {
1748 DMEMIT("%s ", pg->ps.type->name);
1749
1750 if (pg->ps.type->status)
1751 sz += pg->ps.type->status(&pg->ps, NULL, type,
1752 result + sz,
1753 maxlen - sz);
1754 else
1755 DMEMIT("0 ");
1756
1757 DMEMIT("%u %u ", pg->nr_pgpaths,
1758 pg->ps.type->table_args);
1759
1760 list_for_each_entry(p, &pg->pgpaths, list) {
1761 DMEMIT("%s ", p->path.dev->name);
1762 if (pg->ps.type->status)
1763 sz += pg->ps.type->status(&pg->ps,
1764 &p->path, type, result + sz,
1765 maxlen - sz);
1766 }
1767 }
1768 break;
1769 }
1770
1771 spin_unlock_irqrestore(&m->lock, flags);
1772 }
1773
1774 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1775 {
1776 int r = -EINVAL;
1777 struct dm_dev *dev;
1778 struct multipath *m = ti->private;
1779 action_fn action;
1780
1781 mutex_lock(&m->work_mutex);
1782
1783 if (dm_suspended(ti)) {
1784 r = -EBUSY;
1785 goto out;
1786 }
1787
1788 if (argc == 1) {
1789 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1790 r = queue_if_no_path(m, true, false);
1791 goto out;
1792 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1793 r = queue_if_no_path(m, false, false);
1794 goto out;
1795 }
1796 }
1797
1798 if (argc != 2) {
1799 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1800 goto out;
1801 }
1802
1803 if (!strcasecmp(argv[0], "disable_group")) {
1804 r = bypass_pg_num(m, argv[1], true);
1805 goto out;
1806 } else if (!strcasecmp(argv[0], "enable_group")) {
1807 r = bypass_pg_num(m, argv[1], false);
1808 goto out;
1809 } else if (!strcasecmp(argv[0], "switch_group")) {
1810 r = switch_pg_num(m, argv[1]);
1811 goto out;
1812 } else if (!strcasecmp(argv[0], "reinstate_path"))
1813 action = reinstate_path;
1814 else if (!strcasecmp(argv[0], "fail_path"))
1815 action = fail_path;
1816 else {
1817 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1818 goto out;
1819 }
1820
1821 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1822 if (r) {
1823 DMWARN("message: error getting device %s",
1824 argv[1]);
1825 goto out;
1826 }
1827
1828 r = action_dev(m, dev, action);
1829
1830 dm_put_device(ti, dev);
1831
1832 out:
1833 mutex_unlock(&m->work_mutex);
1834 return r;
1835 }
1836
1837 static int multipath_prepare_ioctl(struct dm_target *ti,
1838 struct block_device **bdev, fmode_t *mode)
1839 {
1840 struct multipath *m = ti->private;
1841 struct pgpath *current_pgpath;
1842 int r;
1843
1844 current_pgpath = READ_ONCE(m->current_pgpath);
1845 if (!current_pgpath)
1846 current_pgpath = choose_pgpath(m, 0);
1847
1848 if (current_pgpath) {
1849 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1850 *bdev = current_pgpath->path.dev->bdev;
1851 *mode = current_pgpath->path.dev->mode;
1852 r = 0;
1853 } else {
1854 /* pg_init has not started or completed */
1855 r = -ENOTCONN;
1856 }
1857 } else {
1858 /* No path is available */
1859 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1860 r = -ENOTCONN;
1861 else
1862 r = -EIO;
1863 }
1864
1865 if (r == -ENOTCONN) {
1866 if (!READ_ONCE(m->current_pg)) {
1867 /* Path status changed, redo selection */
1868 (void) choose_pgpath(m, 0);
1869 }
1870 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1871 pg_init_all_paths(m);
1872 dm_table_run_md_queue_async(m->ti->table);
1873 process_queued_io_list(m);
1874 }
1875
1876 /*
1877 * Only pass ioctls through if the device sizes match exactly.
1878 */
1879 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1880 return 1;
1881 return r;
1882 }
1883
1884 static int multipath_iterate_devices(struct dm_target *ti,
1885 iterate_devices_callout_fn fn, void *data)
1886 {
1887 struct multipath *m = ti->private;
1888 struct priority_group *pg;
1889 struct pgpath *p;
1890 int ret = 0;
1891
1892 list_for_each_entry(pg, &m->priority_groups, list) {
1893 list_for_each_entry(p, &pg->pgpaths, list) {
1894 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1895 if (ret)
1896 goto out;
1897 }
1898 }
1899
1900 out:
1901 return ret;
1902 }
1903
1904 static int pgpath_busy(struct pgpath *pgpath)
1905 {
1906 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1907
1908 return blk_lld_busy(q);
1909 }
1910
1911 /*
1912 * We return "busy", only when we can map I/Os but underlying devices
1913 * are busy (so even if we map I/Os now, the I/Os will wait on
1914 * the underlying queue).
1915 * In other words, if we want to kill I/Os or queue them inside us
1916 * due to map unavailability, we don't return "busy". Otherwise,
1917 * dm core won't give us the I/Os and we can't do what we want.
1918 */
1919 static int multipath_busy(struct dm_target *ti)
1920 {
1921 bool busy = false, has_active = false;
1922 struct multipath *m = ti->private;
1923 struct priority_group *pg, *next_pg;
1924 struct pgpath *pgpath;
1925
1926 /* pg_init in progress */
1927 if (atomic_read(&m->pg_init_in_progress))
1928 return true;
1929
1930 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1931 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1932 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1933
1934 /* Guess which priority_group will be used at next mapping time */
1935 pg = READ_ONCE(m->current_pg);
1936 next_pg = READ_ONCE(m->next_pg);
1937 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1938 pg = next_pg;
1939
1940 if (!pg) {
1941 /*
1942 * We don't know which pg will be used at next mapping time.
1943 * We don't call choose_pgpath() here to avoid to trigger
1944 * pg_init just by busy checking.
1945 * So we don't know whether underlying devices we will be using
1946 * at next mapping time are busy or not. Just try mapping.
1947 */
1948 return busy;
1949 }
1950
1951 /*
1952 * If there is one non-busy active path at least, the path selector
1953 * will be able to select it. So we consider such a pg as not busy.
1954 */
1955 busy = true;
1956 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1957 if (pgpath->is_active) {
1958 has_active = true;
1959 if (!pgpath_busy(pgpath)) {
1960 busy = false;
1961 break;
1962 }
1963 }
1964 }
1965
1966 if (!has_active) {
1967 /*
1968 * No active path in this pg, so this pg won't be used and
1969 * the current_pg will be changed at next mapping time.
1970 * We need to try mapping to determine it.
1971 */
1972 busy = false;
1973 }
1974
1975 return busy;
1976 }
1977
1978 /*-----------------------------------------------------------------
1979 * Module setup
1980 *---------------------------------------------------------------*/
1981 static struct target_type multipath_target = {
1982 .name = "multipath",
1983 .version = {1, 13, 0},
1984 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
1985 DM_TARGET_PASSES_INTEGRITY,
1986 .module = THIS_MODULE,
1987 .ctr = multipath_ctr,
1988 .dtr = multipath_dtr,
1989 .clone_and_map_rq = multipath_clone_and_map,
1990 .release_clone_rq = multipath_release_clone,
1991 .rq_end_io = multipath_end_io,
1992 .map = multipath_map_bio,
1993 .end_io = multipath_end_io_bio,
1994 .presuspend = multipath_presuspend,
1995 .postsuspend = multipath_postsuspend,
1996 .resume = multipath_resume,
1997 .status = multipath_status,
1998 .message = multipath_message,
1999 .prepare_ioctl = multipath_prepare_ioctl,
2000 .iterate_devices = multipath_iterate_devices,
2001 .busy = multipath_busy,
2002 };
2003
2004 static int __init dm_multipath_init(void)
2005 {
2006 int r;
2007
2008 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2009 if (!kmultipathd) {
2010 DMERR("failed to create workqueue kmpathd");
2011 r = -ENOMEM;
2012 goto bad_alloc_kmultipathd;
2013 }
2014
2015 /*
2016 * A separate workqueue is used to handle the device handlers
2017 * to avoid overloading existing workqueue. Overloading the
2018 * old workqueue would also create a bottleneck in the
2019 * path of the storage hardware device activation.
2020 */
2021 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2022 WQ_MEM_RECLAIM);
2023 if (!kmpath_handlerd) {
2024 DMERR("failed to create workqueue kmpath_handlerd");
2025 r = -ENOMEM;
2026 goto bad_alloc_kmpath_handlerd;
2027 }
2028
2029 r = dm_register_target(&multipath_target);
2030 if (r < 0) {
2031 DMERR("request-based register failed %d", r);
2032 r = -EINVAL;
2033 goto bad_register_target;
2034 }
2035
2036 return 0;
2037
2038 bad_register_target:
2039 destroy_workqueue(kmpath_handlerd);
2040 bad_alloc_kmpath_handlerd:
2041 destroy_workqueue(kmultipathd);
2042 bad_alloc_kmultipathd:
2043 return r;
2044 }
2045
2046 static void __exit dm_multipath_exit(void)
2047 {
2048 destroy_workqueue(kmpath_handlerd);
2049 destroy_workqueue(kmultipathd);
2050
2051 dm_unregister_target(&multipath_target);
2052 }
2053
2054 module_init(dm_multipath_init);
2055 module_exit(dm_multipath_exit);
2056
2057 MODULE_DESCRIPTION(DM_NAME " multipath target");
2058 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2059 MODULE_LICENSE("GPL");
ubuntu-bionic-kernel mirror