?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.19.133
[linux/fpc-iii.git] / drivers / md / dm-mpath.c
blobc1ad84f3414cd11c8d08723adfbeeb82bb4a7193
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 unsigned long flags; /* Multipath state flags */
68
69 spinlock_t lock;
70 enum dm_queue_mode queue_mode;
71
72 struct pgpath *current_pgpath;
73 struct priority_group *current_pg;
74 struct priority_group *next_pg; /* Switch to this PG if set */
75
76 atomic_t nr_valid_paths; /* Total number of usable paths */
77 unsigned nr_priority_groups;
78 struct list_head priority_groups;
79
80 const char *hw_handler_name;
81 char *hw_handler_params;
82 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
83 unsigned pg_init_retries; /* Number of times to retry pg_init */
84 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
85 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
86 atomic_t pg_init_count; /* Number of times pg_init called */
87
88 struct mutex work_mutex;
89 struct work_struct trigger_event;
90 struct dm_target *ti;
91
92 struct work_struct process_queued_bios;
93 struct bio_list queued_bios;
94 };
95
96 /*
97 * Context information attached to each io we process.
98 */
99 struct dm_mpath_io {
100 struct pgpath *pgpath;
101 size_t nr_bytes;
102 };
103
104 typedef int (*action_fn) (struct pgpath *pgpath);
105
106 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
107 static void trigger_event(struct work_struct *work);
108 static void activate_or_offline_path(struct pgpath *pgpath);
109 static void activate_path_work(struct work_struct *work);
110 static void process_queued_bios(struct work_struct *work);
111
112 /*-----------------------------------------------
113 * Multipath state flags.
114 *-----------------------------------------------*/
115
116 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
117 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
118 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
119 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
120 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
121 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
122 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
123
124 /*-----------------------------------------------
125 * Allocation routines
126 *-----------------------------------------------*/
127
128 static struct pgpath *alloc_pgpath(void)
129 {
130 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
131
132 if (!pgpath)
133 return NULL;
134
135 pgpath->is_active = true;
136
137 return pgpath;
138 }
139
140 static void free_pgpath(struct pgpath *pgpath)
141 {
142 kfree(pgpath);
143 }
144
145 static struct priority_group *alloc_priority_group(void)
146 {
147 struct priority_group *pg;
148
149 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
150
151 if (pg)
152 INIT_LIST_HEAD(&pg->pgpaths);
153
154 return pg;
155 }
156
157 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
158 {
159 struct pgpath *pgpath, *tmp;
160
161 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
162 list_del(&pgpath->list);
163 dm_put_device(ti, pgpath->path.dev);
164 free_pgpath(pgpath);
165 }
166 }
167
168 static void free_priority_group(struct priority_group *pg,
169 struct dm_target *ti)
170 {
171 struct path_selector *ps = &pg->ps;
172
173 if (ps->type) {
174 ps->type->destroy(ps);
175 dm_put_path_selector(ps->type);
176 }
177
178 free_pgpaths(&pg->pgpaths, ti);
179 kfree(pg);
180 }
181
182 static struct multipath *alloc_multipath(struct dm_target *ti)
183 {
184 struct multipath *m;
185
186 m = kzalloc(sizeof(*m), GFP_KERNEL);
187 if (m) {
188 INIT_LIST_HEAD(&m->priority_groups);
189 spin_lock_init(&m->lock);
190 atomic_set(&m->nr_valid_paths, 0);
191 INIT_WORK(&m->trigger_event, trigger_event);
192 mutex_init(&m->work_mutex);
193
194 m->queue_mode = DM_TYPE_NONE;
195
196 m->ti = ti;
197 ti->private = m;
198 }
199
200 return m;
201 }
202
203 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
204 {
205 if (m->queue_mode == DM_TYPE_NONE) {
206 /*
207 * Default to request-based.
208 */
209 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
210 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
211 else
212 m->queue_mode = DM_TYPE_REQUEST_BASED;
213
214 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
215 INIT_WORK(&m->process_queued_bios, process_queued_bios);
216 /*
217 * bio-based doesn't support any direct scsi_dh management;
218 * it just discovers if a scsi_dh is attached.
219 */
220 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
221 }
222
223 dm_table_set_type(ti->table, m->queue_mode);
224
225 /*
226 * Init fields that are only used when a scsi_dh is attached
227 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
228 */
229 set_bit(MPATHF_QUEUE_IO, &m->flags);
230 atomic_set(&m->pg_init_in_progress, 0);
231 atomic_set(&m->pg_init_count, 0);
232 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
233 init_waitqueue_head(&m->pg_init_wait);
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 mutex_destroy(&m->work_mutex);
250 kfree(m);
251 }
252
253 static struct dm_mpath_io *get_mpio(union map_info *info)
254 {
255 return info->ptr;
256 }
257
258 static size_t multipath_per_bio_data_size(void)
259 {
260 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
261 }
262
263 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
264 {
265 return dm_per_bio_data(bio, multipath_per_bio_data_size());
266 }
267
268 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
269 {
270 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
271 void *bio_details = mpio + 1;
272 return bio_details;
273 }
274
275 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
276 {
277 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
278 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
279
280 mpio->nr_bytes = bio->bi_iter.bi_size;
281 mpio->pgpath = NULL;
282 *mpio_p = mpio;
283
284 dm_bio_record(bio_details, bio);
285 }
286
287 /*-----------------------------------------------
288 * Path selection
289 *-----------------------------------------------*/
290
291 static int __pg_init_all_paths(struct multipath *m)
292 {
293 struct pgpath *pgpath;
294 unsigned long pg_init_delay = 0;
295
296 lockdep_assert_held(&m->lock);
297
298 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
299 return 0;
300
301 atomic_inc(&m->pg_init_count);
302 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
303
304 /* Check here to reset pg_init_required */
305 if (!m->current_pg)
306 return 0;
307
308 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
309 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
310 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
311 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
312 /* Skip failed paths */
313 if (!pgpath->is_active)
314 continue;
315 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
316 pg_init_delay))
317 atomic_inc(&m->pg_init_in_progress);
318 }
319 return atomic_read(&m->pg_init_in_progress);
320 }
321
322 static int pg_init_all_paths(struct multipath *m)
323 {
324 int ret;
325 unsigned long flags;
326
327 spin_lock_irqsave(&m->lock, flags);
328 ret = __pg_init_all_paths(m);
329 spin_unlock_irqrestore(&m->lock, flags);
330
331 return ret;
332 }
333
334 static void __switch_pg(struct multipath *m, struct priority_group *pg)
335 {
336 m->current_pg = pg;
337
338 /* Must we initialise the PG first, and queue I/O till it's ready? */
339 if (m->hw_handler_name) {
340 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
341 set_bit(MPATHF_QUEUE_IO, &m->flags);
342 } else {
343 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
344 clear_bit(MPATHF_QUEUE_IO, &m->flags);
345 }
346
347 atomic_set(&m->pg_init_count, 0);
348 }
349
350 static struct pgpath *choose_path_in_pg(struct multipath *m,
351 struct priority_group *pg,
352 size_t nr_bytes)
353 {
354 unsigned long flags;
355 struct dm_path *path;
356 struct pgpath *pgpath;
357
358 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
359 if (!path)
360 return ERR_PTR(-ENXIO);
361
362 pgpath = path_to_pgpath(path);
363
364 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
365 /* Only update current_pgpath if pg changed */
366 spin_lock_irqsave(&m->lock, flags);
367 m->current_pgpath = pgpath;
368 __switch_pg(m, pg);
369 spin_unlock_irqrestore(&m->lock, flags);
370 }
371
372 return pgpath;
373 }
374
375 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
376 {
377 unsigned long flags;
378 struct priority_group *pg;
379 struct pgpath *pgpath;
380 unsigned bypassed = 1;
381
382 if (!atomic_read(&m->nr_valid_paths)) {
383 clear_bit(MPATHF_QUEUE_IO, &m->flags);
384 goto failed;
385 }
386
387 /* Were we instructed to switch PG? */
388 if (READ_ONCE(m->next_pg)) {
389 spin_lock_irqsave(&m->lock, flags);
390 pg = m->next_pg;
391 if (!pg) {
392 spin_unlock_irqrestore(&m->lock, flags);
393 goto check_current_pg;
394 }
395 m->next_pg = NULL;
396 spin_unlock_irqrestore(&m->lock, flags);
397 pgpath = choose_path_in_pg(m, pg, nr_bytes);
398 if (!IS_ERR_OR_NULL(pgpath))
399 return pgpath;
400 }
401
402 /* Don't change PG until it has no remaining paths */
403 check_current_pg:
404 pg = READ_ONCE(m->current_pg);
405 if (pg) {
406 pgpath = choose_path_in_pg(m, pg, nr_bytes);
407 if (!IS_ERR_OR_NULL(pgpath))
408 return pgpath;
409 }
410
411 /*
412 * Loop through priority groups until we find a valid path.
413 * First time we skip PGs marked 'bypassed'.
414 * Second time we only try the ones we skipped, but set
415 * pg_init_delay_retry so we do not hammer controllers.
416 */
417 do {
418 list_for_each_entry(pg, &m->priority_groups, list) {
419 if (pg->bypassed == !!bypassed)
420 continue;
421 pgpath = choose_path_in_pg(m, pg, nr_bytes);
422 if (!IS_ERR_OR_NULL(pgpath)) {
423 if (!bypassed)
424 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
425 return pgpath;
426 }
427 }
428 } while (bypassed--);
429
430 failed:
431 spin_lock_irqsave(&m->lock, flags);
432 m->current_pgpath = NULL;
433 m->current_pg = NULL;
434 spin_unlock_irqrestore(&m->lock, flags);
435
436 return NULL;
437 }
438
439 /*
440 * dm_report_EIO() is a macro instead of a function to make pr_debug()
441 * report the function name and line number of the function from which
442 * it has been invoked.
443 */
444 #define dm_report_EIO(m) \
445 do { \
446 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
447 \
448 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
449 dm_device_name(md), \
450 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
451 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
452 dm_noflush_suspending((m)->ti)); \
453 } while (0)
454
455 /*
456 * Check whether bios must be queued in the device-mapper core rather
457 * than here in the target.
458 *
459 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
460 * the same value then we are not between multipath_presuspend()
461 * and multipath_resume() calls and we have no need to check
462 * for the DMF_NOFLUSH_SUSPENDING flag.
463 */
464 static bool __must_push_back(struct multipath *m, unsigned long flags)
465 {
466 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
467 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
468 dm_noflush_suspending(m->ti));
469 }
470
471 /*
472 * Following functions use READ_ONCE to get atomic access to
473 * all m->flags to avoid taking spinlock
474 */
475 static bool must_push_back_rq(struct multipath *m)
476 {
477 unsigned long flags = READ_ONCE(m->flags);
478 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
479 }
480
481 static bool must_push_back_bio(struct multipath *m)
482 {
483 unsigned long flags = READ_ONCE(m->flags);
484 return __must_push_back(m, flags);
485 }
486
487 /*
488 * Map cloned requests (request-based multipath)
489 */
490 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
491 union map_info *map_context,
492 struct request **__clone)
493 {
494 struct multipath *m = ti->private;
495 size_t nr_bytes = blk_rq_bytes(rq);
496 struct pgpath *pgpath;
497 struct block_device *bdev;
498 struct dm_mpath_io *mpio = get_mpio(map_context);
499 struct request_queue *q;
500 struct request *clone;
501
502 /* Do we need to select a new pgpath? */
503 pgpath = READ_ONCE(m->current_pgpath);
504 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
505 pgpath = choose_pgpath(m, nr_bytes);
506
507 if (!pgpath) {
508 if (must_push_back_rq(m))
509 return DM_MAPIO_DELAY_REQUEUE;
510 dm_report_EIO(m); /* Failed */
511 return DM_MAPIO_KILL;
512 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
513 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
514 pg_init_all_paths(m);
515 return DM_MAPIO_DELAY_REQUEUE;
516 }
517
518 mpio->pgpath = pgpath;
519 mpio->nr_bytes = nr_bytes;
520
521 bdev = pgpath->path.dev->bdev;
522 q = bdev_get_queue(bdev);
523 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
524 BLK_MQ_REQ_NOWAIT);
525 if (IS_ERR(clone)) {
526 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
527 if (blk_queue_dying(q)) {
528 atomic_inc(&m->pg_init_in_progress);
529 activate_or_offline_path(pgpath);
530 return DM_MAPIO_DELAY_REQUEUE;
531 }
532
533 /*
534 * blk-mq's SCHED_RESTART can cover this requeue, so we
535 * needn't deal with it by DELAY_REQUEUE. More importantly,
536 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
537 * get the queue busy feedback (via BLK_STS_RESOURCE),
538 * otherwise I/O merging can suffer.
539 */
540 if (q->mq_ops)
541 return DM_MAPIO_REQUEUE;
542 else
543 return DM_MAPIO_DELAY_REQUEUE;
544 }
545 clone->bio = clone->biotail = NULL;
546 clone->rq_disk = bdev->bd_disk;
547 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
548 *__clone = clone;
549
550 if (pgpath->pg->ps.type->start_io)
551 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
552 &pgpath->path,
553 nr_bytes);
554 return DM_MAPIO_REMAPPED;
555 }
556
557 static void multipath_release_clone(struct request *clone,
558 union map_info *map_context)
559 {
560 if (unlikely(map_context)) {
561 /*
562 * non-NULL map_context means caller is still map
563 * method; must undo multipath_clone_and_map()
564 */
565 struct dm_mpath_io *mpio = get_mpio(map_context);
566 struct pgpath *pgpath = mpio->pgpath;
567
568 if (pgpath && pgpath->pg->ps.type->end_io)
569 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
570 &pgpath->path,
571 mpio->nr_bytes);
572 }
573
574 blk_put_request(clone);
575 }
576
577 /*
578 * Map cloned bios (bio-based multipath)
579 */
580
581 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
582 {
583 struct pgpath *pgpath;
584 unsigned long flags;
585 bool queue_io;
586
587 /* Do we need to select a new pgpath? */
588 pgpath = READ_ONCE(m->current_pgpath);
589 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
590 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
591
592 /* MPATHF_QUEUE_IO might have been cleared by choose_pgpath. */
593 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
594
595 if ((pgpath && queue_io) ||
596 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
597 /* Queue for the daemon to resubmit */
598 spin_lock_irqsave(&m->lock, flags);
599 bio_list_add(&m->queued_bios, bio);
600 spin_unlock_irqrestore(&m->lock, flags);
601
602 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
603 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
604 pg_init_all_paths(m);
605 else if (!queue_io)
606 queue_work(kmultipathd, &m->process_queued_bios);
607
608 return ERR_PTR(-EAGAIN);
609 }
610
611 return pgpath;
612 }
613
614 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
615 struct dm_mpath_io *mpio)
616 {
617 struct pgpath *pgpath = __map_bio(m, bio);
618
619 if (IS_ERR(pgpath))
620 return DM_MAPIO_SUBMITTED;
621
622 if (!pgpath) {
623 if (must_push_back_bio(m))
624 return DM_MAPIO_REQUEUE;
625 dm_report_EIO(m);
626 return DM_MAPIO_KILL;
627 }
628
629 mpio->pgpath = pgpath;
630
631 bio->bi_status = 0;
632 bio_set_dev(bio, pgpath->path.dev->bdev);
633 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
634
635 if (pgpath->pg->ps.type->start_io)
636 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
637 &pgpath->path,
638 mpio->nr_bytes);
639 return DM_MAPIO_REMAPPED;
640 }
641
642 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
643 {
644 struct multipath *m = ti->private;
645 struct dm_mpath_io *mpio = NULL;
646
647 multipath_init_per_bio_data(bio, &mpio);
648 return __multipath_map_bio(m, bio, mpio);
649 }
650
651 static void process_queued_io_list(struct multipath *m)
652 {
653 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
654 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
655 else if (m->queue_mode == DM_TYPE_BIO_BASED)
656 queue_work(kmultipathd, &m->process_queued_bios);
657 }
658
659 static void process_queued_bios(struct work_struct *work)
660 {
661 int r;
662 unsigned long flags;
663 struct bio *bio;
664 struct bio_list bios;
665 struct blk_plug plug;
666 struct multipath *m =
667 container_of(work, struct multipath, process_queued_bios);
668
669 bio_list_init(&bios);
670
671 spin_lock_irqsave(&m->lock, flags);
672
673 if (bio_list_empty(&m->queued_bios)) {
674 spin_unlock_irqrestore(&m->lock, flags);
675 return;
676 }
677
678 bio_list_merge(&bios, &m->queued_bios);
679 bio_list_init(&m->queued_bios);
680
681 spin_unlock_irqrestore(&m->lock, flags);
682
683 blk_start_plug(&plug);
684 while ((bio = bio_list_pop(&bios))) {
685 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
686 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
687 r = __multipath_map_bio(m, bio, mpio);
688 switch (r) {
689 case DM_MAPIO_KILL:
690 bio->bi_status = BLK_STS_IOERR;
691 bio_endio(bio);
692 break;
693 case DM_MAPIO_REQUEUE:
694 bio->bi_status = BLK_STS_DM_REQUEUE;
695 bio_endio(bio);
696 break;
697 case DM_MAPIO_REMAPPED:
698 generic_make_request(bio);
699 break;
700 case DM_MAPIO_SUBMITTED:
701 break;
702 default:
703 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
704 }
705 }
706 blk_finish_plug(&plug);
707 }
708
709 /*
710 * If we run out of usable paths, should we queue I/O or error it?
711 */
712 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
713 bool save_old_value)
714 {
715 unsigned long flags;
716
717 spin_lock_irqsave(&m->lock, flags);
718 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
719 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
720 (!save_old_value && queue_if_no_path));
721 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
722 spin_unlock_irqrestore(&m->lock, flags);
723
724 if (!queue_if_no_path) {
725 dm_table_run_md_queue_async(m->ti->table);
726 process_queued_io_list(m);
727 }
728
729 return 0;
730 }
731
732 /*
733 * An event is triggered whenever a path is taken out of use.
734 * Includes path failure and PG bypass.
735 */
736 static void trigger_event(struct work_struct *work)
737 {
738 struct multipath *m =
739 container_of(work, struct multipath, trigger_event);
740
741 dm_table_event(m->ti->table);
742 }
743
744 /*-----------------------------------------------------------------
745 * Constructor/argument parsing:
746 * <#multipath feature args> [<arg>]*
747 * <#hw_handler args> [hw_handler [<arg>]*]
748 * <#priority groups>
749 * <initial priority group>
750 * [<selector> <#selector args> [<arg>]*
751 * <#paths> <#per-path selector args>
752 * [<path> [<arg>]* ]+ ]+
753 *---------------------------------------------------------------*/
754 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
755 struct dm_target *ti)
756 {
757 int r;
758 struct path_selector_type *pst;
759 unsigned ps_argc;
760
761 static const struct dm_arg _args[] = {
762 {0, 1024, "invalid number of path selector args"},
763 };
764
765 pst = dm_get_path_selector(dm_shift_arg(as));
766 if (!pst) {
767 ti->error = "unknown path selector type";
768 return -EINVAL;
769 }
770
771 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
772 if (r) {
773 dm_put_path_selector(pst);
774 return -EINVAL;
775 }
776
777 r = pst->create(&pg->ps, ps_argc, as->argv);
778 if (r) {
779 dm_put_path_selector(pst);
780 ti->error = "path selector constructor failed";
781 return r;
782 }
783
784 pg->ps.type = pst;
785 dm_consume_args(as, ps_argc);
786
787 return 0;
788 }
789
790 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
791 const char **attached_handler_name, char **error)
792 {
793 struct request_queue *q = bdev_get_queue(bdev);
794 int r;
795
796 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
797 retain:
798 if (*attached_handler_name) {
799 /*
800 * Clear any hw_handler_params associated with a
801 * handler that isn't already attached.
802 */
803 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
804 kfree(m->hw_handler_params);
805 m->hw_handler_params = NULL;
806 }
807
808 /*
809 * Reset hw_handler_name to match the attached handler
810 *
811 * NB. This modifies the table line to show the actual
812 * handler instead of the original table passed in.
813 */
814 kfree(m->hw_handler_name);
815 m->hw_handler_name = *attached_handler_name;
816 *attached_handler_name = NULL;
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(bdev, b));
827 goto retain;
828 }
829 if (r < 0) {
830 *error = "error attaching hardware handler";
831 return r;
832 }
833
834 if (m->hw_handler_params) {
835 r = scsi_dh_set_params(q, m->hw_handler_params);
836 if (r < 0) {
837 *error = "unable to set hardware handler parameters";
838 return r;
839 }
840 }
841 }
842
843 return 0;
844 }
845
846 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
847 struct dm_target *ti)
848 {
849 int r;
850 struct pgpath *p;
851 struct multipath *m = ti->private;
852 struct request_queue *q;
853 const char *attached_handler_name = NULL;
854
855 /* we need at least a path arg */
856 if (as->argc < 1) {
857 ti->error = "no device given";
858 return ERR_PTR(-EINVAL);
859 }
860
861 p = alloc_pgpath();
862 if (!p)
863 return ERR_PTR(-ENOMEM);
864
865 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
866 &p->path.dev);
867 if (r) {
868 ti->error = "error getting device";
869 goto bad;
870 }
871
872 q = bdev_get_queue(p->path.dev->bdev);
873 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
874 if (attached_handler_name || m->hw_handler_name) {
875 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
876 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
877 kfree(attached_handler_name);
878 if (r) {
879 dm_put_device(ti, p->path.dev);
880 goto bad;
881 }
882 }
883
884 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
885 if (r) {
886 dm_put_device(ti, p->path.dev);
887 goto bad;
888 }
889
890 return p;
891 bad:
892 free_pgpath(p);
893 return ERR_PTR(r);
894 }
895
896 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
897 struct multipath *m)
898 {
899 static const struct dm_arg _args[] = {
900 {1, 1024, "invalid number of paths"},
901 {0, 1024, "invalid number of selector args"}
902 };
903
904 int r;
905 unsigned i, nr_selector_args, nr_args;
906 struct priority_group *pg;
907 struct dm_target *ti = m->ti;
908
909 if (as->argc < 2) {
910 as->argc = 0;
911 ti->error = "not enough priority group arguments";
912 return ERR_PTR(-EINVAL);
913 }
914
915 pg = alloc_priority_group();
916 if (!pg) {
917 ti->error = "couldn't allocate priority group";
918 return ERR_PTR(-ENOMEM);
919 }
920 pg->m = m;
921
922 r = parse_path_selector(as, pg, ti);
923 if (r)
924 goto bad;
925
926 /*
927 * read the paths
928 */
929 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
930 if (r)
931 goto bad;
932
933 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
934 if (r)
935 goto bad;
936
937 nr_args = 1 + nr_selector_args;
938 for (i = 0; i < pg->nr_pgpaths; i++) {
939 struct pgpath *pgpath;
940 struct dm_arg_set path_args;
941
942 if (as->argc < nr_args) {
943 ti->error = "not enough path parameters";
944 r = -EINVAL;
945 goto bad;
946 }
947
948 path_args.argc = nr_args;
949 path_args.argv = as->argv;
950
951 pgpath = parse_path(&path_args, &pg->ps, ti);
952 if (IS_ERR(pgpath)) {
953 r = PTR_ERR(pgpath);
954 goto bad;
955 }
956
957 pgpath->pg = pg;
958 list_add_tail(&pgpath->list, &pg->pgpaths);
959 dm_consume_args(as, nr_args);
960 }
961
962 return pg;
963
964 bad:
965 free_priority_group(pg, ti);
966 return ERR_PTR(r);
967 }
968
969 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
970 {
971 unsigned hw_argc;
972 int ret;
973 struct dm_target *ti = m->ti;
974
975 static const struct dm_arg _args[] = {
976 {0, 1024, "invalid number of hardware handler args"},
977 };
978
979 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
980 return -EINVAL;
981
982 if (!hw_argc)
983 return 0;
984
985 if (m->queue_mode == DM_TYPE_BIO_BASED) {
986 dm_consume_args(as, hw_argc);
987 DMERR("bio-based multipath doesn't allow hardware handler args");
988 return 0;
989 }
990
991 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
992 if (!m->hw_handler_name)
993 return -EINVAL;
994
995 if (hw_argc > 1) {
996 char *p;
997 int i, j, len = 4;
998
999 for (i = 0; i <= hw_argc - 2; i++)
1000 len += strlen(as->argv[i]) + 1;
1001 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1002 if (!p) {
1003 ti->error = "memory allocation failed";
1004 ret = -ENOMEM;
1005 goto fail;
1006 }
1007 j = sprintf(p, "%d", hw_argc - 1);
1008 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1009 j = sprintf(p, "%s", as->argv[i]);
1010 }
1011 dm_consume_args(as, hw_argc - 1);
1012
1013 return 0;
1014 fail:
1015 kfree(m->hw_handler_name);
1016 m->hw_handler_name = NULL;
1017 return ret;
1018 }
1019
1020 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1021 {
1022 int r;
1023 unsigned argc;
1024 struct dm_target *ti = m->ti;
1025 const char *arg_name;
1026
1027 static const struct dm_arg _args[] = {
1028 {0, 8, "invalid number of feature args"},
1029 {1, 50, "pg_init_retries must be between 1 and 50"},
1030 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1031 };
1032
1033 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1034 if (r)
1035 return -EINVAL;
1036
1037 if (!argc)
1038 return 0;
1039
1040 do {
1041 arg_name = dm_shift_arg(as);
1042 argc--;
1043
1044 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1045 r = queue_if_no_path(m, true, false);
1046 continue;
1047 }
1048
1049 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1050 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1051 continue;
1052 }
1053
1054 if (!strcasecmp(arg_name, "pg_init_retries") &&
1055 (argc >= 1)) {
1056 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1057 argc--;
1058 continue;
1059 }
1060
1061 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1062 (argc >= 1)) {
1063 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1064 argc--;
1065 continue;
1066 }
1067
1068 if (!strcasecmp(arg_name, "queue_mode") &&
1069 (argc >= 1)) {
1070 const char *queue_mode_name = dm_shift_arg(as);
1071
1072 if (!strcasecmp(queue_mode_name, "bio"))
1073 m->queue_mode = DM_TYPE_BIO_BASED;
1074 else if (!strcasecmp(queue_mode_name, "rq"))
1075 m->queue_mode = DM_TYPE_REQUEST_BASED;
1076 else if (!strcasecmp(queue_mode_name, "mq"))
1077 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1078 else {
1079 ti->error = "Unknown 'queue_mode' requested";
1080 r = -EINVAL;
1081 }
1082 argc--;
1083 continue;
1084 }
1085
1086 ti->error = "Unrecognised multipath feature request";
1087 r = -EINVAL;
1088 } while (argc && !r);
1089
1090 return r;
1091 }
1092
1093 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1094 {
1095 /* target arguments */
1096 static const struct dm_arg _args[] = {
1097 {0, 1024, "invalid number of priority groups"},
1098 {0, 1024, "invalid initial priority group number"},
1099 };
1100
1101 int r;
1102 struct multipath *m;
1103 struct dm_arg_set as;
1104 unsigned pg_count = 0;
1105 unsigned next_pg_num;
1106
1107 as.argc = argc;
1108 as.argv = argv;
1109
1110 m = alloc_multipath(ti);
1111 if (!m) {
1112 ti->error = "can't allocate multipath";
1113 return -EINVAL;
1114 }
1115
1116 r = parse_features(&as, m);
1117 if (r)
1118 goto bad;
1119
1120 r = alloc_multipath_stage2(ti, m);
1121 if (r)
1122 goto bad;
1123
1124 r = parse_hw_handler(&as, m);
1125 if (r)
1126 goto bad;
1127
1128 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1129 if (r)
1130 goto bad;
1131
1132 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1133 if (r)
1134 goto bad;
1135
1136 if ((!m->nr_priority_groups && next_pg_num) ||
1137 (m->nr_priority_groups && !next_pg_num)) {
1138 ti->error = "invalid initial priority group";
1139 r = -EINVAL;
1140 goto bad;
1141 }
1142
1143 /* parse the priority groups */
1144 while (as.argc) {
1145 struct priority_group *pg;
1146 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1147
1148 pg = parse_priority_group(&as, m);
1149 if (IS_ERR(pg)) {
1150 r = PTR_ERR(pg);
1151 goto bad;
1152 }
1153
1154 nr_valid_paths += pg->nr_pgpaths;
1155 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1156
1157 list_add_tail(&pg->list, &m->priority_groups);
1158 pg_count++;
1159 pg->pg_num = pg_count;
1160 if (!--next_pg_num)
1161 m->next_pg = pg;
1162 }
1163
1164 if (pg_count != m->nr_priority_groups) {
1165 ti->error = "priority group count mismatch";
1166 r = -EINVAL;
1167 goto bad;
1168 }
1169
1170 ti->num_flush_bios = 1;
1171 ti->num_discard_bios = 1;
1172 ti->num_write_same_bios = 1;
1173 ti->num_write_zeroes_bios = 1;
1174 if (m->queue_mode == DM_TYPE_BIO_BASED)
1175 ti->per_io_data_size = multipath_per_bio_data_size();
1176 else
1177 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1178
1179 return 0;
1180
1181 bad:
1182 free_multipath(m);
1183 return r;
1184 }
1185
1186 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1187 {
1188 DEFINE_WAIT(wait);
1189
1190 while (1) {
1191 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1192
1193 if (!atomic_read(&m->pg_init_in_progress))
1194 break;
1195
1196 io_schedule();
1197 }
1198 finish_wait(&m->pg_init_wait, &wait);
1199 }
1200
1201 static void flush_multipath_work(struct multipath *m)
1202 {
1203 if (m->hw_handler_name) {
1204 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1205 smp_mb__after_atomic();
1206
1207 flush_workqueue(kmpath_handlerd);
1208 multipath_wait_for_pg_init_completion(m);
1209
1210 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1211 smp_mb__after_atomic();
1212 }
1213
1214 flush_workqueue(kmultipathd);
1215 flush_work(&m->trigger_event);
1216 }
1217
1218 static void multipath_dtr(struct dm_target *ti)
1219 {
1220 struct multipath *m = ti->private;
1221
1222 flush_multipath_work(m);
1223 free_multipath(m);
1224 }
1225
1226 /*
1227 * Take a path out of use.
1228 */
1229 static int fail_path(struct pgpath *pgpath)
1230 {
1231 unsigned long flags;
1232 struct multipath *m = pgpath->pg->m;
1233
1234 spin_lock_irqsave(&m->lock, flags);
1235
1236 if (!pgpath->is_active)
1237 goto out;
1238
1239 DMWARN("Failing path %s.", pgpath->path.dev->name);
1240
1241 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1242 pgpath->is_active = false;
1243 pgpath->fail_count++;
1244
1245 atomic_dec(&m->nr_valid_paths);
1246
1247 if (pgpath == m->current_pgpath)
1248 m->current_pgpath = NULL;
1249
1250 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1251 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1252
1253 schedule_work(&m->trigger_event);
1254
1255 out:
1256 spin_unlock_irqrestore(&m->lock, flags);
1257
1258 return 0;
1259 }
1260
1261 /*
1262 * Reinstate a previously-failed path
1263 */
1264 static int reinstate_path(struct pgpath *pgpath)
1265 {
1266 int r = 0, run_queue = 0;
1267 unsigned long flags;
1268 struct multipath *m = pgpath->pg->m;
1269 unsigned nr_valid_paths;
1270
1271 spin_lock_irqsave(&m->lock, flags);
1272
1273 if (pgpath->is_active)
1274 goto out;
1275
1276 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1277
1278 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1279 if (r)
1280 goto out;
1281
1282 pgpath->is_active = true;
1283
1284 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1285 if (nr_valid_paths == 1) {
1286 m->current_pgpath = NULL;
1287 run_queue = 1;
1288 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1289 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1290 atomic_inc(&m->pg_init_in_progress);
1291 }
1292
1293 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1294 pgpath->path.dev->name, nr_valid_paths);
1295
1296 schedule_work(&m->trigger_event);
1297
1298 out:
1299 spin_unlock_irqrestore(&m->lock, flags);
1300 if (run_queue) {
1301 dm_table_run_md_queue_async(m->ti->table);
1302 process_queued_io_list(m);
1303 }
1304
1305 return r;
1306 }
1307
1308 /*
1309 * Fail or reinstate all paths that match the provided struct dm_dev.
1310 */
1311 static int action_dev(struct multipath *m, struct dm_dev *dev,
1312 action_fn action)
1313 {
1314 int r = -EINVAL;
1315 struct pgpath *pgpath;
1316 struct priority_group *pg;
1317
1318 list_for_each_entry(pg, &m->priority_groups, list) {
1319 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1320 if (pgpath->path.dev == dev)
1321 r = action(pgpath);
1322 }
1323 }
1324
1325 return r;
1326 }
1327
1328 /*
1329 * Temporarily try to avoid having to use the specified PG
1330 */
1331 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1332 bool bypassed)
1333 {
1334 unsigned long flags;
1335
1336 spin_lock_irqsave(&m->lock, flags);
1337
1338 pg->bypassed = bypassed;
1339 m->current_pgpath = NULL;
1340 m->current_pg = NULL;
1341
1342 spin_unlock_irqrestore(&m->lock, flags);
1343
1344 schedule_work(&m->trigger_event);
1345 }
1346
1347 /*
1348 * Switch to using the specified PG from the next I/O that gets mapped
1349 */
1350 static int switch_pg_num(struct multipath *m, const char *pgstr)
1351 {
1352 struct priority_group *pg;
1353 unsigned pgnum;
1354 unsigned long flags;
1355 char dummy;
1356
1357 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1358 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1359 DMWARN("invalid PG number supplied to switch_pg_num");
1360 return -EINVAL;
1361 }
1362
1363 spin_lock_irqsave(&m->lock, flags);
1364 list_for_each_entry(pg, &m->priority_groups, list) {
1365 pg->bypassed = false;
1366 if (--pgnum)
1367 continue;
1368
1369 m->current_pgpath = NULL;
1370 m->current_pg = NULL;
1371 m->next_pg = pg;
1372 }
1373 spin_unlock_irqrestore(&m->lock, flags);
1374
1375 schedule_work(&m->trigger_event);
1376 return 0;
1377 }
1378
1379 /*
1380 * Set/clear bypassed status of a PG.
1381 * PGs are numbered upwards from 1 in the order they were declared.
1382 */
1383 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1384 {
1385 struct priority_group *pg;
1386 unsigned pgnum;
1387 char dummy;
1388
1389 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1390 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1391 DMWARN("invalid PG number supplied to bypass_pg");
1392 return -EINVAL;
1393 }
1394
1395 list_for_each_entry(pg, &m->priority_groups, list) {
1396 if (!--pgnum)
1397 break;
1398 }
1399
1400 bypass_pg(m, pg, bypassed);
1401 return 0;
1402 }
1403
1404 /*
1405 * Should we retry pg_init immediately?
1406 */
1407 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1408 {
1409 unsigned long flags;
1410 bool limit_reached = false;
1411
1412 spin_lock_irqsave(&m->lock, flags);
1413
1414 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1415 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1416 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1417 else
1418 limit_reached = true;
1419
1420 spin_unlock_irqrestore(&m->lock, flags);
1421
1422 return limit_reached;
1423 }
1424
1425 static void pg_init_done(void *data, int errors)
1426 {
1427 struct pgpath *pgpath = data;
1428 struct priority_group *pg = pgpath->pg;
1429 struct multipath *m = pg->m;
1430 unsigned long flags;
1431 bool delay_retry = false;
1432
1433 /* device or driver problems */
1434 switch (errors) {
1435 case SCSI_DH_OK:
1436 break;
1437 case SCSI_DH_NOSYS:
1438 if (!m->hw_handler_name) {
1439 errors = 0;
1440 break;
1441 }
1442 DMERR("Could not failover the device: Handler scsi_dh_%s "
1443 "Error %d.", m->hw_handler_name, errors);
1444 /*
1445 * Fail path for now, so we do not ping pong
1446 */
1447 fail_path(pgpath);
1448 break;
1449 case SCSI_DH_DEV_TEMP_BUSY:
1450 /*
1451 * Probably doing something like FW upgrade on the
1452 * controller so try the other pg.
1453 */
1454 bypass_pg(m, pg, true);
1455 break;
1456 case SCSI_DH_RETRY:
1457 /* Wait before retrying. */
1458 delay_retry = 1;
1459 /* fall through */
1460 case SCSI_DH_IMM_RETRY:
1461 case SCSI_DH_RES_TEMP_UNAVAIL:
1462 if (pg_init_limit_reached(m, pgpath))
1463 fail_path(pgpath);
1464 errors = 0;
1465 break;
1466 case SCSI_DH_DEV_OFFLINED:
1467 default:
1468 /*
1469 * We probably do not want to fail the path for a device
1470 * error, but this is what the old dm did. In future
1471 * patches we can do more advanced handling.
1472 */
1473 fail_path(pgpath);
1474 }
1475
1476 spin_lock_irqsave(&m->lock, flags);
1477 if (errors) {
1478 if (pgpath == m->current_pgpath) {
1479 DMERR("Could not failover device. Error %d.", errors);
1480 m->current_pgpath = NULL;
1481 m->current_pg = NULL;
1482 }
1483 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1484 pg->bypassed = false;
1485
1486 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1487 /* Activations of other paths are still on going */
1488 goto out;
1489
1490 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1491 if (delay_retry)
1492 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1493 else
1494 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1495
1496 if (__pg_init_all_paths(m))
1497 goto out;
1498 }
1499 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1500
1501 process_queued_io_list(m);
1502
1503 /*
1504 * Wake up any thread waiting to suspend.
1505 */
1506 wake_up(&m->pg_init_wait);
1507
1508 out:
1509 spin_unlock_irqrestore(&m->lock, flags);
1510 }
1511
1512 static void activate_or_offline_path(struct pgpath *pgpath)
1513 {
1514 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1515
1516 if (pgpath->is_active && !blk_queue_dying(q))
1517 scsi_dh_activate(q, pg_init_done, pgpath);
1518 else
1519 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1520 }
1521
1522 static void activate_path_work(struct work_struct *work)
1523 {
1524 struct pgpath *pgpath =
1525 container_of(work, struct pgpath, activate_path.work);
1526
1527 activate_or_offline_path(pgpath);
1528 }
1529
1530 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1531 blk_status_t error, union map_info *map_context)
1532 {
1533 struct dm_mpath_io *mpio = get_mpio(map_context);
1534 struct pgpath *pgpath = mpio->pgpath;
1535 int r = DM_ENDIO_DONE;
1536
1537 /*
1538 * We don't queue any clone request inside the multipath target
1539 * during end I/O handling, since those clone requests don't have
1540 * bio clones. If we queue them inside the multipath target,
1541 * we need to make bio clones, that requires memory allocation.
1542 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1543 * don't have bio clones.)
1544 * Instead of queueing the clone request here, we queue the original
1545 * request into dm core, which will remake a clone request and
1546 * clone bios for it and resubmit it later.
1547 */
1548 if (error && blk_path_error(error)) {
1549 struct multipath *m = ti->private;
1550
1551 if (error == BLK_STS_RESOURCE)
1552 r = DM_ENDIO_DELAY_REQUEUE;
1553 else
1554 r = DM_ENDIO_REQUEUE;
1555
1556 if (pgpath)
1557 fail_path(pgpath);
1558
1559 if (atomic_read(&m->nr_valid_paths) == 0 &&
1560 !must_push_back_rq(m)) {
1561 if (error == BLK_STS_IOERR)
1562 dm_report_EIO(m);
1563 /* complete with the original error */
1564 r = DM_ENDIO_DONE;
1565 }
1566 }
1567
1568 if (pgpath) {
1569 struct path_selector *ps = &pgpath->pg->ps;
1570
1571 if (ps->type->end_io)
1572 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1573 }
1574
1575 return r;
1576 }
1577
1578 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1579 blk_status_t *error)
1580 {
1581 struct multipath *m = ti->private;
1582 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1583 struct pgpath *pgpath = mpio->pgpath;
1584 unsigned long flags;
1585 int r = DM_ENDIO_DONE;
1586
1587 if (!*error || !blk_path_error(*error))
1588 goto done;
1589
1590 if (pgpath)
1591 fail_path(pgpath);
1592
1593 if (atomic_read(&m->nr_valid_paths) == 0 &&
1594 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1595 if (must_push_back_bio(m)) {
1596 r = DM_ENDIO_REQUEUE;
1597 } else {
1598 dm_report_EIO(m);
1599 *error = BLK_STS_IOERR;
1600 }
1601 goto done;
1602 }
1603
1604 spin_lock_irqsave(&m->lock, flags);
1605 bio_list_add(&m->queued_bios, clone);
1606 spin_unlock_irqrestore(&m->lock, flags);
1607 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1608 queue_work(kmultipathd, &m->process_queued_bios);
1609
1610 r = DM_ENDIO_INCOMPLETE;
1611 done:
1612 if (pgpath) {
1613 struct path_selector *ps = &pgpath->pg->ps;
1614
1615 if (ps->type->end_io)
1616 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1617 }
1618
1619 return r;
1620 }
1621
1622 /*
1623 * Suspend can't complete until all the I/O is processed so if
1624 * the last path fails we must error any remaining I/O.
1625 * Note that if the freeze_bdev fails while suspending, the
1626 * queue_if_no_path state is lost - userspace should reset it.
1627 */
1628 static void multipath_presuspend(struct dm_target *ti)
1629 {
1630 struct multipath *m = ti->private;
1631
1632 queue_if_no_path(m, false, true);
1633 }
1634
1635 static void multipath_postsuspend(struct dm_target *ti)
1636 {
1637 struct multipath *m = ti->private;
1638
1639 mutex_lock(&m->work_mutex);
1640 flush_multipath_work(m);
1641 mutex_unlock(&m->work_mutex);
1642 }
1643
1644 /*
1645 * Restore the queue_if_no_path setting.
1646 */
1647 static void multipath_resume(struct dm_target *ti)
1648 {
1649 struct multipath *m = ti->private;
1650 unsigned long flags;
1651
1652 spin_lock_irqsave(&m->lock, flags);
1653 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1654 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1655 spin_unlock_irqrestore(&m->lock, flags);
1656 }
1657
1658 /*
1659 * Info output has the following format:
1660 * num_multipath_feature_args [multipath_feature_args]*
1661 * num_handler_status_args [handler_status_args]*
1662 * num_groups init_group_number
1663 * [A|D|E num_ps_status_args [ps_status_args]*
1664 * num_paths num_selector_args
1665 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1666 *
1667 * Table output has the following format (identical to the constructor string):
1668 * num_feature_args [features_args]*
1669 * num_handler_args hw_handler [hw_handler_args]*
1670 * num_groups init_group_number
1671 * [priority selector-name num_ps_args [ps_args]*
1672 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1673 */
1674 static void multipath_status(struct dm_target *ti, status_type_t type,
1675 unsigned status_flags, char *result, unsigned maxlen)
1676 {
1677 int sz = 0;
1678 unsigned long flags;
1679 struct multipath *m = ti->private;
1680 struct priority_group *pg;
1681 struct pgpath *p;
1682 unsigned pg_num;
1683 char state;
1684
1685 spin_lock_irqsave(&m->lock, flags);
1686
1687 /* Features */
1688 if (type == STATUSTYPE_INFO)
1689 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1690 atomic_read(&m->pg_init_count));
1691 else {
1692 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1693 (m->pg_init_retries > 0) * 2 +
1694 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1695 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1696 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1697
1698 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1699 DMEMIT("queue_if_no_path ");
1700 if (m->pg_init_retries)
1701 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1702 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1703 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1704 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1705 DMEMIT("retain_attached_hw_handler ");
1706 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1707 switch(m->queue_mode) {
1708 case DM_TYPE_BIO_BASED:
1709 DMEMIT("queue_mode bio ");
1710 break;
1711 case DM_TYPE_MQ_REQUEST_BASED:
1712 DMEMIT("queue_mode mq ");
1713 break;
1714 default:
1715 WARN_ON_ONCE(true);
1716 break;
1717 }
1718 }
1719 }
1720
1721 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1722 DMEMIT("0 ");
1723 else
1724 DMEMIT("1 %s ", m->hw_handler_name);
1725
1726 DMEMIT("%u ", m->nr_priority_groups);
1727
1728 if (m->next_pg)
1729 pg_num = m->next_pg->pg_num;
1730 else if (m->current_pg)
1731 pg_num = m->current_pg->pg_num;
1732 else
1733 pg_num = (m->nr_priority_groups ? 1 : 0);
1734
1735 DMEMIT("%u ", pg_num);
1736
1737 switch (type) {
1738 case STATUSTYPE_INFO:
1739 list_for_each_entry(pg, &m->priority_groups, list) {
1740 if (pg->bypassed)
1741 state = 'D'; /* Disabled */
1742 else if (pg == m->current_pg)
1743 state = 'A'; /* Currently Active */
1744 else
1745 state = 'E'; /* Enabled */
1746
1747 DMEMIT("%c ", state);
1748
1749 if (pg->ps.type->status)
1750 sz += pg->ps.type->status(&pg->ps, NULL, type,
1751 result + sz,
1752 maxlen - sz);
1753 else
1754 DMEMIT("0 ");
1755
1756 DMEMIT("%u %u ", pg->nr_pgpaths,
1757 pg->ps.type->info_args);
1758
1759 list_for_each_entry(p, &pg->pgpaths, list) {
1760 DMEMIT("%s %s %u ", p->path.dev->name,
1761 p->is_active ? "A" : "F",
1762 p->fail_count);
1763 if (pg->ps.type->status)
1764 sz += pg->ps.type->status(&pg->ps,
1765 &p->path, type, result + sz,
1766 maxlen - sz);
1767 }
1768 }
1769 break;
1770
1771 case STATUSTYPE_TABLE:
1772 list_for_each_entry(pg, &m->priority_groups, list) {
1773 DMEMIT("%s ", pg->ps.type->name);
1774
1775 if (pg->ps.type->status)
1776 sz += pg->ps.type->status(&pg->ps, NULL, type,
1777 result + sz,
1778 maxlen - sz);
1779 else
1780 DMEMIT("0 ");
1781
1782 DMEMIT("%u %u ", pg->nr_pgpaths,
1783 pg->ps.type->table_args);
1784
1785 list_for_each_entry(p, &pg->pgpaths, list) {
1786 DMEMIT("%s ", p->path.dev->name);
1787 if (pg->ps.type->status)
1788 sz += pg->ps.type->status(&pg->ps,
1789 &p->path, type, result + sz,
1790 maxlen - sz);
1791 }
1792 }
1793 break;
1794 }
1795
1796 spin_unlock_irqrestore(&m->lock, flags);
1797 }
1798
1799 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1800 char *result, unsigned maxlen)
1801 {
1802 int r = -EINVAL;
1803 struct dm_dev *dev;
1804 struct multipath *m = ti->private;
1805 action_fn action;
1806
1807 mutex_lock(&m->work_mutex);
1808
1809 if (dm_suspended(ti)) {
1810 r = -EBUSY;
1811 goto out;
1812 }
1813
1814 if (argc == 1) {
1815 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1816 r = queue_if_no_path(m, true, false);
1817 goto out;
1818 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1819 r = queue_if_no_path(m, false, false);
1820 goto out;
1821 }
1822 }
1823
1824 if (argc != 2) {
1825 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1826 goto out;
1827 }
1828
1829 if (!strcasecmp(argv[0], "disable_group")) {
1830 r = bypass_pg_num(m, argv[1], true);
1831 goto out;
1832 } else if (!strcasecmp(argv[0], "enable_group")) {
1833 r = bypass_pg_num(m, argv[1], false);
1834 goto out;
1835 } else if (!strcasecmp(argv[0], "switch_group")) {
1836 r = switch_pg_num(m, argv[1]);
1837 goto out;
1838 } else if (!strcasecmp(argv[0], "reinstate_path"))
1839 action = reinstate_path;
1840 else if (!strcasecmp(argv[0], "fail_path"))
1841 action = fail_path;
1842 else {
1843 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1844 goto out;
1845 }
1846
1847 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1848 if (r) {
1849 DMWARN("message: error getting device %s",
1850 argv[1]);
1851 goto out;
1852 }
1853
1854 r = action_dev(m, dev, action);
1855
1856 dm_put_device(ti, dev);
1857
1858 out:
1859 mutex_unlock(&m->work_mutex);
1860 return r;
1861 }
1862
1863 static int multipath_prepare_ioctl(struct dm_target *ti,
1864 struct block_device **bdev)
1865 {
1866 struct multipath *m = ti->private;
1867 struct pgpath *current_pgpath;
1868 int r;
1869
1870 current_pgpath = READ_ONCE(m->current_pgpath);
1871 if (!current_pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
1872 current_pgpath = choose_pgpath(m, 0);
1873
1874 if (current_pgpath) {
1875 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1876 *bdev = current_pgpath->path.dev->bdev;
1877 r = 0;
1878 } else {
1879 /* pg_init has not started or completed */
1880 r = -ENOTCONN;
1881 }
1882 } else {
1883 /* No path is available */
1884 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1885 r = -ENOTCONN;
1886 else
1887 r = -EIO;
1888 }
1889
1890 if (r == -ENOTCONN) {
1891 if (!READ_ONCE(m->current_pg)) {
1892 /* Path status changed, redo selection */
1893 (void) choose_pgpath(m, 0);
1894 }
1895 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1896 pg_init_all_paths(m);
1897 dm_table_run_md_queue_async(m->ti->table);
1898 process_queued_io_list(m);
1899 }
1900
1901 /*
1902 * Only pass ioctls through if the device sizes match exactly.
1903 */
1904 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1905 return 1;
1906 return r;
1907 }
1908
1909 static int multipath_iterate_devices(struct dm_target *ti,
1910 iterate_devices_callout_fn fn, void *data)
1911 {
1912 struct multipath *m = ti->private;
1913 struct priority_group *pg;
1914 struct pgpath *p;
1915 int ret = 0;
1916
1917 list_for_each_entry(pg, &m->priority_groups, list) {
1918 list_for_each_entry(p, &pg->pgpaths, list) {
1919 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1920 if (ret)
1921 goto out;
1922 }
1923 }
1924
1925 out:
1926 return ret;
1927 }
1928
1929 static int pgpath_busy(struct pgpath *pgpath)
1930 {
1931 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1932
1933 return blk_lld_busy(q);
1934 }
1935
1936 /*
1937 * We return "busy", only when we can map I/Os but underlying devices
1938 * are busy (so even if we map I/Os now, the I/Os will wait on
1939 * the underlying queue).
1940 * In other words, if we want to kill I/Os or queue them inside us
1941 * due to map unavailability, we don't return "busy". Otherwise,
1942 * dm core won't give us the I/Os and we can't do what we want.
1943 */
1944 static int multipath_busy(struct dm_target *ti)
1945 {
1946 bool busy = false, has_active = false;
1947 struct multipath *m = ti->private;
1948 struct priority_group *pg, *next_pg;
1949 struct pgpath *pgpath;
1950
1951 /* pg_init in progress */
1952 if (atomic_read(&m->pg_init_in_progress))
1953 return true;
1954
1955 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1956 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1957 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1958
1959 /* Guess which priority_group will be used at next mapping time */
1960 pg = READ_ONCE(m->current_pg);
1961 next_pg = READ_ONCE(m->next_pg);
1962 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1963 pg = next_pg;
1964
1965 if (!pg) {
1966 /*
1967 * We don't know which pg will be used at next mapping time.
1968 * We don't call choose_pgpath() here to avoid to trigger
1969 * pg_init just by busy checking.
1970 * So we don't know whether underlying devices we will be using
1971 * at next mapping time are busy or not. Just try mapping.
1972 */
1973 return busy;
1974 }
1975
1976 /*
1977 * If there is one non-busy active path at least, the path selector
1978 * will be able to select it. So we consider such a pg as not busy.
1979 */
1980 busy = true;
1981 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1982 if (pgpath->is_active) {
1983 has_active = true;
1984 if (!pgpath_busy(pgpath)) {
1985 busy = false;
1986 break;
1987 }
1988 }
1989 }
1990
1991 if (!has_active) {
1992 /*
1993 * No active path in this pg, so this pg won't be used and
1994 * the current_pg will be changed at next mapping time.
1995 * We need to try mapping to determine it.
1996 */
1997 busy = false;
1998 }
1999
2000 return busy;
2001 }
2002
2003 /*-----------------------------------------------------------------
2004 * Module setup
2005 *---------------------------------------------------------------*/
2006 static struct target_type multipath_target = {
2007 .name = "multipath",
2008 .version = {1, 13, 0},
2009 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2010 DM_TARGET_PASSES_INTEGRITY,
2011 .module = THIS_MODULE,
2012 .ctr = multipath_ctr,
2013 .dtr = multipath_dtr,
2014 .clone_and_map_rq = multipath_clone_and_map,
2015 .release_clone_rq = multipath_release_clone,
2016 .rq_end_io = multipath_end_io,
2017 .map = multipath_map_bio,
2018 .end_io = multipath_end_io_bio,
2019 .presuspend = multipath_presuspend,
2020 .postsuspend = multipath_postsuspend,
2021 .resume = multipath_resume,
2022 .status = multipath_status,
2023 .message = multipath_message,
2024 .prepare_ioctl = multipath_prepare_ioctl,
2025 .iterate_devices = multipath_iterate_devices,
2026 .busy = multipath_busy,
2027 };
2028
2029 static int __init dm_multipath_init(void)
2030 {
2031 int r;
2032
2033 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2034 if (!kmultipathd) {
2035 DMERR("failed to create workqueue kmpathd");
2036 r = -ENOMEM;
2037 goto bad_alloc_kmultipathd;
2038 }
2039
2040 /*
2041 * A separate workqueue is used to handle the device handlers
2042 * to avoid overloading existing workqueue. Overloading the
2043 * old workqueue would also create a bottleneck in the
2044 * path of the storage hardware device activation.
2045 */
2046 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2047 WQ_MEM_RECLAIM);
2048 if (!kmpath_handlerd) {
2049 DMERR("failed to create workqueue kmpath_handlerd");
2050 r = -ENOMEM;
2051 goto bad_alloc_kmpath_handlerd;
2052 }
2053
2054 r = dm_register_target(&multipath_target);
2055 if (r < 0) {
2056 DMERR("request-based register failed %d", r);
2057 r = -EINVAL;
2058 goto bad_register_target;
2059 }
2060
2061 return 0;
2062
2063 bad_register_target:
2064 destroy_workqueue(kmpath_handlerd);
2065 bad_alloc_kmpath_handlerd:
2066 destroy_workqueue(kmultipathd);
2067 bad_alloc_kmultipathd:
2068 return r;
2069 }
2070
2071 static void __exit dm_multipath_exit(void)
2072 {
2073 destroy_workqueue(kmpath_handlerd);
2074 destroy_workqueue(kmultipathd);
2075
2076 dm_unregister_target(&multipath_target);
2077 }
2078
2079 module_init(dm_multipath_init);
2080 module_exit(dm_multipath_exit);
2081
2082 MODULE_DESCRIPTION(DM_NAME " multipath target");
2083 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2084 MODULE_LICENSE("GPL");
Linux with added FPC-III support