?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
staging: rtl8188eu: rename HalSetBrateCfg() - style
[linux/fpc-iii.git] / drivers / md / dm-mpath.c
blobd94ba6f72ff59e3723cc67337ff6bca21a472fda
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 {
559 blk_put_request(clone);
560 }
561
562 /*
563 * Map cloned bios (bio-based multipath)
564 */
565
566 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
567 {
568 struct pgpath *pgpath;
569 unsigned long flags;
570 bool queue_io;
571
572 /* Do we need to select a new pgpath? */
573 pgpath = READ_ONCE(m->current_pgpath);
574 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
575 if (!pgpath || !queue_io)
576 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
577
578 if ((pgpath && queue_io) ||
579 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
580 /* Queue for the daemon to resubmit */
581 spin_lock_irqsave(&m->lock, flags);
582 bio_list_add(&m->queued_bios, bio);
583 spin_unlock_irqrestore(&m->lock, flags);
584
585 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
586 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
587 pg_init_all_paths(m);
588 else if (!queue_io)
589 queue_work(kmultipathd, &m->process_queued_bios);
590
591 return ERR_PTR(-EAGAIN);
592 }
593
594 return pgpath;
595 }
596
597 static struct pgpath *__map_bio_fast(struct multipath *m, struct bio *bio)
598 {
599 struct pgpath *pgpath;
600 unsigned long flags;
601
602 /* Do we need to select a new pgpath? */
603 /*
604 * FIXME: currently only switching path if no path (due to failure, etc)
605 * - which negates the point of using a path selector
606 */
607 pgpath = READ_ONCE(m->current_pgpath);
608 if (!pgpath)
609 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
610
611 if (!pgpath) {
612 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
613 /* Queue for the daemon to resubmit */
614 spin_lock_irqsave(&m->lock, flags);
615 bio_list_add(&m->queued_bios, bio);
616 spin_unlock_irqrestore(&m->lock, flags);
617 queue_work(kmultipathd, &m->process_queued_bios);
618
619 return ERR_PTR(-EAGAIN);
620 }
621 return NULL;
622 }
623
624 return pgpath;
625 }
626
627 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
628 struct dm_mpath_io *mpio)
629 {
630 struct pgpath *pgpath;
631
632 if (!m->hw_handler_name)
633 pgpath = __map_bio_fast(m, bio);
634 else
635 pgpath = __map_bio(m, bio);
636
637 if (IS_ERR(pgpath))
638 return DM_MAPIO_SUBMITTED;
639
640 if (!pgpath) {
641 if (must_push_back_bio(m))
642 return DM_MAPIO_REQUEUE;
643 dm_report_EIO(m);
644 return DM_MAPIO_KILL;
645 }
646
647 mpio->pgpath = pgpath;
648
649 bio->bi_status = 0;
650 bio_set_dev(bio, pgpath->path.dev->bdev);
651 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
652
653 if (pgpath->pg->ps.type->start_io)
654 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
655 &pgpath->path,
656 mpio->nr_bytes);
657 return DM_MAPIO_REMAPPED;
658 }
659
660 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
661 {
662 struct multipath *m = ti->private;
663 struct dm_mpath_io *mpio = NULL;
664
665 multipath_init_per_bio_data(bio, &mpio);
666 return __multipath_map_bio(m, bio, mpio);
667 }
668
669 static void process_queued_io_list(struct multipath *m)
670 {
671 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
672 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
673 else if (m->queue_mode == DM_TYPE_BIO_BASED)
674 queue_work(kmultipathd, &m->process_queued_bios);
675 }
676
677 static void process_queued_bios(struct work_struct *work)
678 {
679 int r;
680 unsigned long flags;
681 struct bio *bio;
682 struct bio_list bios;
683 struct blk_plug plug;
684 struct multipath *m =
685 container_of(work, struct multipath, process_queued_bios);
686
687 bio_list_init(&bios);
688
689 spin_lock_irqsave(&m->lock, flags);
690
691 if (bio_list_empty(&m->queued_bios)) {
692 spin_unlock_irqrestore(&m->lock, flags);
693 return;
694 }
695
696 bio_list_merge(&bios, &m->queued_bios);
697 bio_list_init(&m->queued_bios);
698
699 spin_unlock_irqrestore(&m->lock, flags);
700
701 blk_start_plug(&plug);
702 while ((bio = bio_list_pop(&bios))) {
703 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
704 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
705 r = __multipath_map_bio(m, bio, mpio);
706 switch (r) {
707 case DM_MAPIO_KILL:
708 bio->bi_status = BLK_STS_IOERR;
709 bio_endio(bio);
710 break;
711 case DM_MAPIO_REQUEUE:
712 bio->bi_status = BLK_STS_DM_REQUEUE;
713 bio_endio(bio);
714 break;
715 case DM_MAPIO_REMAPPED:
716 generic_make_request(bio);
717 break;
718 case DM_MAPIO_SUBMITTED:
719 break;
720 default:
721 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
722 }
723 }
724 blk_finish_plug(&plug);
725 }
726
727 /*
728 * If we run out of usable paths, should we queue I/O or error it?
729 */
730 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
731 bool save_old_value)
732 {
733 unsigned long flags;
734
735 spin_lock_irqsave(&m->lock, flags);
736 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
737 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
738 (!save_old_value && queue_if_no_path));
739 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
740 spin_unlock_irqrestore(&m->lock, flags);
741
742 if (!queue_if_no_path) {
743 dm_table_run_md_queue_async(m->ti->table);
744 process_queued_io_list(m);
745 }
746
747 return 0;
748 }
749
750 /*
751 * An event is triggered whenever a path is taken out of use.
752 * Includes path failure and PG bypass.
753 */
754 static void trigger_event(struct work_struct *work)
755 {
756 struct multipath *m =
757 container_of(work, struct multipath, trigger_event);
758
759 dm_table_event(m->ti->table);
760 }
761
762 /*-----------------------------------------------------------------
763 * Constructor/argument parsing:
764 * <#multipath feature args> [<arg>]*
765 * <#hw_handler args> [hw_handler [<arg>]*]
766 * <#priority groups>
767 * <initial priority group>
768 * [<selector> <#selector args> [<arg>]*
769 * <#paths> <#per-path selector args>
770 * [<path> [<arg>]* ]+ ]+
771 *---------------------------------------------------------------*/
772 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
773 struct dm_target *ti)
774 {
775 int r;
776 struct path_selector_type *pst;
777 unsigned ps_argc;
778
779 static const struct dm_arg _args[] = {
780 {0, 1024, "invalid number of path selector args"},
781 };
782
783 pst = dm_get_path_selector(dm_shift_arg(as));
784 if (!pst) {
785 ti->error = "unknown path selector type";
786 return -EINVAL;
787 }
788
789 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
790 if (r) {
791 dm_put_path_selector(pst);
792 return -EINVAL;
793 }
794
795 r = pst->create(&pg->ps, ps_argc, as->argv);
796 if (r) {
797 dm_put_path_selector(pst);
798 ti->error = "path selector constructor failed";
799 return r;
800 }
801
802 pg->ps.type = pst;
803 dm_consume_args(as, ps_argc);
804
805 return 0;
806 }
807
808 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
809 const char *attached_handler_name, char **error)
810 {
811 struct request_queue *q = bdev_get_queue(bdev);
812 int r;
813
814 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
815 retain:
816 if (attached_handler_name) {
817 /*
818 * Clear any hw_handler_params associated with a
819 * handler that isn't already attached.
820 */
821 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
822 kfree(m->hw_handler_params);
823 m->hw_handler_params = NULL;
824 }
825
826 /*
827 * Reset hw_handler_name to match the attached handler
828 *
829 * NB. This modifies the table line to show the actual
830 * handler instead of the original table passed in.
831 */
832 kfree(m->hw_handler_name);
833 m->hw_handler_name = attached_handler_name;
834 }
835 }
836
837 if (m->hw_handler_name) {
838 r = scsi_dh_attach(q, m->hw_handler_name);
839 if (r == -EBUSY) {
840 char b[BDEVNAME_SIZE];
841
842 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
843 bdevname(bdev, b));
844 goto retain;
845 }
846 if (r < 0) {
847 *error = "error attaching hardware handler";
848 return r;
849 }
850
851 if (m->hw_handler_params) {
852 r = scsi_dh_set_params(q, m->hw_handler_params);
853 if (r < 0) {
854 *error = "unable to set hardware handler parameters";
855 return r;
856 }
857 }
858 }
859
860 return 0;
861 }
862
863 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
864 struct dm_target *ti)
865 {
866 int r;
867 struct pgpath *p;
868 struct multipath *m = ti->private;
869 struct request_queue *q;
870 const char *attached_handler_name;
871
872 /* we need at least a path arg */
873 if (as->argc < 1) {
874 ti->error = "no device given";
875 return ERR_PTR(-EINVAL);
876 }
877
878 p = alloc_pgpath();
879 if (!p)
880 return ERR_PTR(-ENOMEM);
881
882 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
883 &p->path.dev);
884 if (r) {
885 ti->error = "error getting device";
886 goto bad;
887 }
888
889 q = bdev_get_queue(p->path.dev->bdev);
890 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
891 if (attached_handler_name || m->hw_handler_name) {
892 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
893 r = setup_scsi_dh(p->path.dev->bdev, m, attached_handler_name, &ti->error);
894 if (r) {
895 dm_put_device(ti, p->path.dev);
896 goto bad;
897 }
898 }
899
900 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
901 if (r) {
902 dm_put_device(ti, p->path.dev);
903 goto bad;
904 }
905
906 return p;
907 bad:
908 free_pgpath(p);
909 return ERR_PTR(r);
910 }
911
912 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
913 struct multipath *m)
914 {
915 static const struct dm_arg _args[] = {
916 {1, 1024, "invalid number of paths"},
917 {0, 1024, "invalid number of selector args"}
918 };
919
920 int r;
921 unsigned i, nr_selector_args, nr_args;
922 struct priority_group *pg;
923 struct dm_target *ti = m->ti;
924
925 if (as->argc < 2) {
926 as->argc = 0;
927 ti->error = "not enough priority group arguments";
928 return ERR_PTR(-EINVAL);
929 }
930
931 pg = alloc_priority_group();
932 if (!pg) {
933 ti->error = "couldn't allocate priority group";
934 return ERR_PTR(-ENOMEM);
935 }
936 pg->m = m;
937
938 r = parse_path_selector(as, pg, ti);
939 if (r)
940 goto bad;
941
942 /*
943 * read the paths
944 */
945 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
946 if (r)
947 goto bad;
948
949 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
950 if (r)
951 goto bad;
952
953 nr_args = 1 + nr_selector_args;
954 for (i = 0; i < pg->nr_pgpaths; i++) {
955 struct pgpath *pgpath;
956 struct dm_arg_set path_args;
957
958 if (as->argc < nr_args) {
959 ti->error = "not enough path parameters";
960 r = -EINVAL;
961 goto bad;
962 }
963
964 path_args.argc = nr_args;
965 path_args.argv = as->argv;
966
967 pgpath = parse_path(&path_args, &pg->ps, ti);
968 if (IS_ERR(pgpath)) {
969 r = PTR_ERR(pgpath);
970 goto bad;
971 }
972
973 pgpath->pg = pg;
974 list_add_tail(&pgpath->list, &pg->pgpaths);
975 dm_consume_args(as, nr_args);
976 }
977
978 return pg;
979
980 bad:
981 free_priority_group(pg, ti);
982 return ERR_PTR(r);
983 }
984
985 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
986 {
987 unsigned hw_argc;
988 int ret;
989 struct dm_target *ti = m->ti;
990
991 static const struct dm_arg _args[] = {
992 {0, 1024, "invalid number of hardware handler args"},
993 };
994
995 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
996 return -EINVAL;
997
998 if (!hw_argc)
999 return 0;
1000
1001 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1002 dm_consume_args(as, hw_argc);
1003 DMERR("bio-based multipath doesn't allow hardware handler args");
1004 return 0;
1005 }
1006
1007 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1008 if (!m->hw_handler_name)
1009 return -EINVAL;
1010
1011 if (hw_argc > 1) {
1012 char *p;
1013 int i, j, len = 4;
1014
1015 for (i = 0; i <= hw_argc - 2; i++)
1016 len += strlen(as->argv[i]) + 1;
1017 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1018 if (!p) {
1019 ti->error = "memory allocation failed";
1020 ret = -ENOMEM;
1021 goto fail;
1022 }
1023 j = sprintf(p, "%d", hw_argc - 1);
1024 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1025 j = sprintf(p, "%s", as->argv[i]);
1026 }
1027 dm_consume_args(as, hw_argc - 1);
1028
1029 return 0;
1030 fail:
1031 kfree(m->hw_handler_name);
1032 m->hw_handler_name = NULL;
1033 return ret;
1034 }
1035
1036 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1037 {
1038 int r;
1039 unsigned argc;
1040 struct dm_target *ti = m->ti;
1041 const char *arg_name;
1042
1043 static const struct dm_arg _args[] = {
1044 {0, 8, "invalid number of feature args"},
1045 {1, 50, "pg_init_retries must be between 1 and 50"},
1046 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1047 };
1048
1049 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1050 if (r)
1051 return -EINVAL;
1052
1053 if (!argc)
1054 return 0;
1055
1056 do {
1057 arg_name = dm_shift_arg(as);
1058 argc--;
1059
1060 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1061 r = queue_if_no_path(m, true, false);
1062 continue;
1063 }
1064
1065 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1066 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1067 continue;
1068 }
1069
1070 if (!strcasecmp(arg_name, "pg_init_retries") &&
1071 (argc >= 1)) {
1072 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1073 argc--;
1074 continue;
1075 }
1076
1077 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1078 (argc >= 1)) {
1079 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1080 argc--;
1081 continue;
1082 }
1083
1084 if (!strcasecmp(arg_name, "queue_mode") &&
1085 (argc >= 1)) {
1086 const char *queue_mode_name = dm_shift_arg(as);
1087
1088 if (!strcasecmp(queue_mode_name, "bio"))
1089 m->queue_mode = DM_TYPE_BIO_BASED;
1090 else if (!strcasecmp(queue_mode_name, "rq"))
1091 m->queue_mode = DM_TYPE_REQUEST_BASED;
1092 else if (!strcasecmp(queue_mode_name, "mq"))
1093 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1094 else {
1095 ti->error = "Unknown 'queue_mode' requested";
1096 r = -EINVAL;
1097 }
1098 argc--;
1099 continue;
1100 }
1101
1102 ti->error = "Unrecognised multipath feature request";
1103 r = -EINVAL;
1104 } while (argc && !r);
1105
1106 return r;
1107 }
1108
1109 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1110 {
1111 /* target arguments */
1112 static const struct dm_arg _args[] = {
1113 {0, 1024, "invalid number of priority groups"},
1114 {0, 1024, "invalid initial priority group number"},
1115 };
1116
1117 int r;
1118 struct multipath *m;
1119 struct dm_arg_set as;
1120 unsigned pg_count = 0;
1121 unsigned next_pg_num;
1122
1123 as.argc = argc;
1124 as.argv = argv;
1125
1126 m = alloc_multipath(ti);
1127 if (!m) {
1128 ti->error = "can't allocate multipath";
1129 return -EINVAL;
1130 }
1131
1132 r = parse_features(&as, m);
1133 if (r)
1134 goto bad;
1135
1136 r = alloc_multipath_stage2(ti, m);
1137 if (r)
1138 goto bad;
1139
1140 r = parse_hw_handler(&as, m);
1141 if (r)
1142 goto bad;
1143
1144 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1145 if (r)
1146 goto bad;
1147
1148 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1149 if (r)
1150 goto bad;
1151
1152 if ((!m->nr_priority_groups && next_pg_num) ||
1153 (m->nr_priority_groups && !next_pg_num)) {
1154 ti->error = "invalid initial priority group";
1155 r = -EINVAL;
1156 goto bad;
1157 }
1158
1159 /* parse the priority groups */
1160 while (as.argc) {
1161 struct priority_group *pg;
1162 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1163
1164 pg = parse_priority_group(&as, m);
1165 if (IS_ERR(pg)) {
1166 r = PTR_ERR(pg);
1167 goto bad;
1168 }
1169
1170 nr_valid_paths += pg->nr_pgpaths;
1171 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1172
1173 list_add_tail(&pg->list, &m->priority_groups);
1174 pg_count++;
1175 pg->pg_num = pg_count;
1176 if (!--next_pg_num)
1177 m->next_pg = pg;
1178 }
1179
1180 if (pg_count != m->nr_priority_groups) {
1181 ti->error = "priority group count mismatch";
1182 r = -EINVAL;
1183 goto bad;
1184 }
1185
1186 ti->num_flush_bios = 1;
1187 ti->num_discard_bios = 1;
1188 ti->num_write_same_bios = 1;
1189 ti->num_write_zeroes_bios = 1;
1190 if (m->queue_mode == DM_TYPE_BIO_BASED)
1191 ti->per_io_data_size = multipath_per_bio_data_size();
1192 else
1193 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1194
1195 return 0;
1196
1197 bad:
1198 free_multipath(m);
1199 return r;
1200 }
1201
1202 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1203 {
1204 DEFINE_WAIT(wait);
1205
1206 while (1) {
1207 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1208
1209 if (!atomic_read(&m->pg_init_in_progress))
1210 break;
1211
1212 io_schedule();
1213 }
1214 finish_wait(&m->pg_init_wait, &wait);
1215 }
1216
1217 static void flush_multipath_work(struct multipath *m)
1218 {
1219 if (m->hw_handler_name) {
1220 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1221 smp_mb__after_atomic();
1222
1223 flush_workqueue(kmpath_handlerd);
1224 multipath_wait_for_pg_init_completion(m);
1225
1226 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1227 smp_mb__after_atomic();
1228 }
1229
1230 flush_workqueue(kmultipathd);
1231 flush_work(&m->trigger_event);
1232 }
1233
1234 static void multipath_dtr(struct dm_target *ti)
1235 {
1236 struct multipath *m = ti->private;
1237
1238 flush_multipath_work(m);
1239 free_multipath(m);
1240 }
1241
1242 /*
1243 * Take a path out of use.
1244 */
1245 static int fail_path(struct pgpath *pgpath)
1246 {
1247 unsigned long flags;
1248 struct multipath *m = pgpath->pg->m;
1249
1250 spin_lock_irqsave(&m->lock, flags);
1251
1252 if (!pgpath->is_active)
1253 goto out;
1254
1255 DMWARN("Failing path %s.", pgpath->path.dev->name);
1256
1257 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1258 pgpath->is_active = false;
1259 pgpath->fail_count++;
1260
1261 atomic_dec(&m->nr_valid_paths);
1262
1263 if (pgpath == m->current_pgpath)
1264 m->current_pgpath = NULL;
1265
1266 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1267 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1268
1269 schedule_work(&m->trigger_event);
1270
1271 out:
1272 spin_unlock_irqrestore(&m->lock, flags);
1273
1274 return 0;
1275 }
1276
1277 /*
1278 * Reinstate a previously-failed path
1279 */
1280 static int reinstate_path(struct pgpath *pgpath)
1281 {
1282 int r = 0, run_queue = 0;
1283 unsigned long flags;
1284 struct multipath *m = pgpath->pg->m;
1285 unsigned nr_valid_paths;
1286
1287 spin_lock_irqsave(&m->lock, flags);
1288
1289 if (pgpath->is_active)
1290 goto out;
1291
1292 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1293
1294 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1295 if (r)
1296 goto out;
1297
1298 pgpath->is_active = true;
1299
1300 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1301 if (nr_valid_paths == 1) {
1302 m->current_pgpath = NULL;
1303 run_queue = 1;
1304 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1305 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1306 atomic_inc(&m->pg_init_in_progress);
1307 }
1308
1309 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1310 pgpath->path.dev->name, nr_valid_paths);
1311
1312 schedule_work(&m->trigger_event);
1313
1314 out:
1315 spin_unlock_irqrestore(&m->lock, flags);
1316 if (run_queue) {
1317 dm_table_run_md_queue_async(m->ti->table);
1318 process_queued_io_list(m);
1319 }
1320
1321 return r;
1322 }
1323
1324 /*
1325 * Fail or reinstate all paths that match the provided struct dm_dev.
1326 */
1327 static int action_dev(struct multipath *m, struct dm_dev *dev,
1328 action_fn action)
1329 {
1330 int r = -EINVAL;
1331 struct pgpath *pgpath;
1332 struct priority_group *pg;
1333
1334 list_for_each_entry(pg, &m->priority_groups, list) {
1335 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1336 if (pgpath->path.dev == dev)
1337 r = action(pgpath);
1338 }
1339 }
1340
1341 return r;
1342 }
1343
1344 /*
1345 * Temporarily try to avoid having to use the specified PG
1346 */
1347 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1348 bool bypassed)
1349 {
1350 unsigned long flags;
1351
1352 spin_lock_irqsave(&m->lock, flags);
1353
1354 pg->bypassed = bypassed;
1355 m->current_pgpath = NULL;
1356 m->current_pg = NULL;
1357
1358 spin_unlock_irqrestore(&m->lock, flags);
1359
1360 schedule_work(&m->trigger_event);
1361 }
1362
1363 /*
1364 * Switch to using the specified PG from the next I/O that gets mapped
1365 */
1366 static int switch_pg_num(struct multipath *m, const char *pgstr)
1367 {
1368 struct priority_group *pg;
1369 unsigned pgnum;
1370 unsigned long flags;
1371 char dummy;
1372
1373 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1374 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1375 DMWARN("invalid PG number supplied to switch_pg_num");
1376 return -EINVAL;
1377 }
1378
1379 spin_lock_irqsave(&m->lock, flags);
1380 list_for_each_entry(pg, &m->priority_groups, list) {
1381 pg->bypassed = false;
1382 if (--pgnum)
1383 continue;
1384
1385 m->current_pgpath = NULL;
1386 m->current_pg = NULL;
1387 m->next_pg = pg;
1388 }
1389 spin_unlock_irqrestore(&m->lock, flags);
1390
1391 schedule_work(&m->trigger_event);
1392 return 0;
1393 }
1394
1395 /*
1396 * Set/clear bypassed status of a PG.
1397 * PGs are numbered upwards from 1 in the order they were declared.
1398 */
1399 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1400 {
1401 struct priority_group *pg;
1402 unsigned pgnum;
1403 char dummy;
1404
1405 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1406 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1407 DMWARN("invalid PG number supplied to bypass_pg");
1408 return -EINVAL;
1409 }
1410
1411 list_for_each_entry(pg, &m->priority_groups, list) {
1412 if (!--pgnum)
1413 break;
1414 }
1415
1416 bypass_pg(m, pg, bypassed);
1417 return 0;
1418 }
1419
1420 /*
1421 * Should we retry pg_init immediately?
1422 */
1423 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1424 {
1425 unsigned long flags;
1426 bool limit_reached = false;
1427
1428 spin_lock_irqsave(&m->lock, flags);
1429
1430 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1431 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1432 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1433 else
1434 limit_reached = true;
1435
1436 spin_unlock_irqrestore(&m->lock, flags);
1437
1438 return limit_reached;
1439 }
1440
1441 static void pg_init_done(void *data, int errors)
1442 {
1443 struct pgpath *pgpath = data;
1444 struct priority_group *pg = pgpath->pg;
1445 struct multipath *m = pg->m;
1446 unsigned long flags;
1447 bool delay_retry = false;
1448
1449 /* device or driver problems */
1450 switch (errors) {
1451 case SCSI_DH_OK:
1452 break;
1453 case SCSI_DH_NOSYS:
1454 if (!m->hw_handler_name) {
1455 errors = 0;
1456 break;
1457 }
1458 DMERR("Could not failover the device: Handler scsi_dh_%s "
1459 "Error %d.", m->hw_handler_name, errors);
1460 /*
1461 * Fail path for now, so we do not ping pong
1462 */
1463 fail_path(pgpath);
1464 break;
1465 case SCSI_DH_DEV_TEMP_BUSY:
1466 /*
1467 * Probably doing something like FW upgrade on the
1468 * controller so try the other pg.
1469 */
1470 bypass_pg(m, pg, true);
1471 break;
1472 case SCSI_DH_RETRY:
1473 /* Wait before retrying. */
1474 delay_retry = 1;
1475 /* fall through */
1476 case SCSI_DH_IMM_RETRY:
1477 case SCSI_DH_RES_TEMP_UNAVAIL:
1478 if (pg_init_limit_reached(m, pgpath))
1479 fail_path(pgpath);
1480 errors = 0;
1481 break;
1482 case SCSI_DH_DEV_OFFLINED:
1483 default:
1484 /*
1485 * We probably do not want to fail the path for a device
1486 * error, but this is what the old dm did. In future
1487 * patches we can do more advanced handling.
1488 */
1489 fail_path(pgpath);
1490 }
1491
1492 spin_lock_irqsave(&m->lock, flags);
1493 if (errors) {
1494 if (pgpath == m->current_pgpath) {
1495 DMERR("Could not failover device. Error %d.", errors);
1496 m->current_pgpath = NULL;
1497 m->current_pg = NULL;
1498 }
1499 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1500 pg->bypassed = false;
1501
1502 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1503 /* Activations of other paths are still on going */
1504 goto out;
1505
1506 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1507 if (delay_retry)
1508 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1509 else
1510 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1511
1512 if (__pg_init_all_paths(m))
1513 goto out;
1514 }
1515 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1516
1517 process_queued_io_list(m);
1518
1519 /*
1520 * Wake up any thread waiting to suspend.
1521 */
1522 wake_up(&m->pg_init_wait);
1523
1524 out:
1525 spin_unlock_irqrestore(&m->lock, flags);
1526 }
1527
1528 static void activate_or_offline_path(struct pgpath *pgpath)
1529 {
1530 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1531
1532 if (pgpath->is_active && !blk_queue_dying(q))
1533 scsi_dh_activate(q, pg_init_done, pgpath);
1534 else
1535 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1536 }
1537
1538 static void activate_path_work(struct work_struct *work)
1539 {
1540 struct pgpath *pgpath =
1541 container_of(work, struct pgpath, activate_path.work);
1542
1543 activate_or_offline_path(pgpath);
1544 }
1545
1546 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1547 blk_status_t error, union map_info *map_context)
1548 {
1549 struct dm_mpath_io *mpio = get_mpio(map_context);
1550 struct pgpath *pgpath = mpio->pgpath;
1551 int r = DM_ENDIO_DONE;
1552
1553 /*
1554 * We don't queue any clone request inside the multipath target
1555 * during end I/O handling, since those clone requests don't have
1556 * bio clones. If we queue them inside the multipath target,
1557 * we need to make bio clones, that requires memory allocation.
1558 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1559 * don't have bio clones.)
1560 * Instead of queueing the clone request here, we queue the original
1561 * request into dm core, which will remake a clone request and
1562 * clone bios for it and resubmit it later.
1563 */
1564 if (error && blk_path_error(error)) {
1565 struct multipath *m = ti->private;
1566
1567 if (error == BLK_STS_RESOURCE)
1568 r = DM_ENDIO_DELAY_REQUEUE;
1569 else
1570 r = DM_ENDIO_REQUEUE;
1571
1572 if (pgpath)
1573 fail_path(pgpath);
1574
1575 if (atomic_read(&m->nr_valid_paths) == 0 &&
1576 !must_push_back_rq(m)) {
1577 if (error == BLK_STS_IOERR)
1578 dm_report_EIO(m);
1579 /* complete with the original error */
1580 r = DM_ENDIO_DONE;
1581 }
1582 }
1583
1584 if (pgpath) {
1585 struct path_selector *ps = &pgpath->pg->ps;
1586
1587 if (ps->type->end_io)
1588 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1589 }
1590
1591 return r;
1592 }
1593
1594 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1595 blk_status_t *error)
1596 {
1597 struct multipath *m = ti->private;
1598 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1599 struct pgpath *pgpath = mpio->pgpath;
1600 unsigned long flags;
1601 int r = DM_ENDIO_DONE;
1602
1603 if (!*error || !blk_path_error(*error))
1604 goto done;
1605
1606 if (pgpath)
1607 fail_path(pgpath);
1608
1609 if (atomic_read(&m->nr_valid_paths) == 0 &&
1610 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1611 if (must_push_back_bio(m)) {
1612 r = DM_ENDIO_REQUEUE;
1613 } else {
1614 dm_report_EIO(m);
1615 *error = BLK_STS_IOERR;
1616 }
1617 goto done;
1618 }
1619
1620 spin_lock_irqsave(&m->lock, flags);
1621 bio_list_add(&m->queued_bios, clone);
1622 spin_unlock_irqrestore(&m->lock, flags);
1623 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1624 queue_work(kmultipathd, &m->process_queued_bios);
1625
1626 r = DM_ENDIO_INCOMPLETE;
1627 done:
1628 if (pgpath) {
1629 struct path_selector *ps = &pgpath->pg->ps;
1630
1631 if (ps->type->end_io)
1632 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1633 }
1634
1635 return r;
1636 }
1637
1638 /*
1639 * Suspend can't complete until all the I/O is processed so if
1640 * the last path fails we must error any remaining I/O.
1641 * Note that if the freeze_bdev fails while suspending, the
1642 * queue_if_no_path state is lost - userspace should reset it.
1643 */
1644 static void multipath_presuspend(struct dm_target *ti)
1645 {
1646 struct multipath *m = ti->private;
1647
1648 queue_if_no_path(m, false, true);
1649 }
1650
1651 static void multipath_postsuspend(struct dm_target *ti)
1652 {
1653 struct multipath *m = ti->private;
1654
1655 mutex_lock(&m->work_mutex);
1656 flush_multipath_work(m);
1657 mutex_unlock(&m->work_mutex);
1658 }
1659
1660 /*
1661 * Restore the queue_if_no_path setting.
1662 */
1663 static void multipath_resume(struct dm_target *ti)
1664 {
1665 struct multipath *m = ti->private;
1666 unsigned long flags;
1667
1668 spin_lock_irqsave(&m->lock, flags);
1669 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1670 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1671 spin_unlock_irqrestore(&m->lock, flags);
1672 }
1673
1674 /*
1675 * Info output has the following format:
1676 * num_multipath_feature_args [multipath_feature_args]*
1677 * num_handler_status_args [handler_status_args]*
1678 * num_groups init_group_number
1679 * [A|D|E num_ps_status_args [ps_status_args]*
1680 * num_paths num_selector_args
1681 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1682 *
1683 * Table output has the following format (identical to the constructor string):
1684 * num_feature_args [features_args]*
1685 * num_handler_args hw_handler [hw_handler_args]*
1686 * num_groups init_group_number
1687 * [priority selector-name num_ps_args [ps_args]*
1688 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1689 */
1690 static void multipath_status(struct dm_target *ti, status_type_t type,
1691 unsigned status_flags, char *result, unsigned maxlen)
1692 {
1693 int sz = 0;
1694 unsigned long flags;
1695 struct multipath *m = ti->private;
1696 struct priority_group *pg;
1697 struct pgpath *p;
1698 unsigned pg_num;
1699 char state;
1700
1701 spin_lock_irqsave(&m->lock, flags);
1702
1703 /* Features */
1704 if (type == STATUSTYPE_INFO)
1705 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1706 atomic_read(&m->pg_init_count));
1707 else {
1708 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1709 (m->pg_init_retries > 0) * 2 +
1710 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1711 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1712 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1713
1714 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1715 DMEMIT("queue_if_no_path ");
1716 if (m->pg_init_retries)
1717 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1718 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1719 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1720 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1721 DMEMIT("retain_attached_hw_handler ");
1722 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1723 switch(m->queue_mode) {
1724 case DM_TYPE_BIO_BASED:
1725 DMEMIT("queue_mode bio ");
1726 break;
1727 case DM_TYPE_MQ_REQUEST_BASED:
1728 DMEMIT("queue_mode mq ");
1729 break;
1730 default:
1731 WARN_ON_ONCE(true);
1732 break;
1733 }
1734 }
1735 }
1736
1737 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1738 DMEMIT("0 ");
1739 else
1740 DMEMIT("1 %s ", m->hw_handler_name);
1741
1742 DMEMIT("%u ", m->nr_priority_groups);
1743
1744 if (m->next_pg)
1745 pg_num = m->next_pg->pg_num;
1746 else if (m->current_pg)
1747 pg_num = m->current_pg->pg_num;
1748 else
1749 pg_num = (m->nr_priority_groups ? 1 : 0);
1750
1751 DMEMIT("%u ", pg_num);
1752
1753 switch (type) {
1754 case STATUSTYPE_INFO:
1755 list_for_each_entry(pg, &m->priority_groups, list) {
1756 if (pg->bypassed)
1757 state = 'D'; /* Disabled */
1758 else if (pg == m->current_pg)
1759 state = 'A'; /* Currently Active */
1760 else
1761 state = 'E'; /* Enabled */
1762
1763 DMEMIT("%c ", state);
1764
1765 if (pg->ps.type->status)
1766 sz += pg->ps.type->status(&pg->ps, NULL, type,
1767 result + sz,
1768 maxlen - sz);
1769 else
1770 DMEMIT("0 ");
1771
1772 DMEMIT("%u %u ", pg->nr_pgpaths,
1773 pg->ps.type->info_args);
1774
1775 list_for_each_entry(p, &pg->pgpaths, list) {
1776 DMEMIT("%s %s %u ", p->path.dev->name,
1777 p->is_active ? "A" : "F",
1778 p->fail_count);
1779 if (pg->ps.type->status)
1780 sz += pg->ps.type->status(&pg->ps,
1781 &p->path, type, result + sz,
1782 maxlen - sz);
1783 }
1784 }
1785 break;
1786
1787 case STATUSTYPE_TABLE:
1788 list_for_each_entry(pg, &m->priority_groups, list) {
1789 DMEMIT("%s ", pg->ps.type->name);
1790
1791 if (pg->ps.type->status)
1792 sz += pg->ps.type->status(&pg->ps, NULL, type,
1793 result + sz,
1794 maxlen - sz);
1795 else
1796 DMEMIT("0 ");
1797
1798 DMEMIT("%u %u ", pg->nr_pgpaths,
1799 pg->ps.type->table_args);
1800
1801 list_for_each_entry(p, &pg->pgpaths, list) {
1802 DMEMIT("%s ", p->path.dev->name);
1803 if (pg->ps.type->status)
1804 sz += pg->ps.type->status(&pg->ps,
1805 &p->path, type, result + sz,
1806 maxlen - sz);
1807 }
1808 }
1809 break;
1810 }
1811
1812 spin_unlock_irqrestore(&m->lock, flags);
1813 }
1814
1815 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1816 char *result, unsigned maxlen)
1817 {
1818 int r = -EINVAL;
1819 struct dm_dev *dev;
1820 struct multipath *m = ti->private;
1821 action_fn action;
1822
1823 mutex_lock(&m->work_mutex);
1824
1825 if (dm_suspended(ti)) {
1826 r = -EBUSY;
1827 goto out;
1828 }
1829
1830 if (argc == 1) {
1831 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1832 r = queue_if_no_path(m, true, false);
1833 goto out;
1834 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1835 r = queue_if_no_path(m, false, false);
1836 goto out;
1837 }
1838 }
1839
1840 if (argc != 2) {
1841 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1842 goto out;
1843 }
1844
1845 if (!strcasecmp(argv[0], "disable_group")) {
1846 r = bypass_pg_num(m, argv[1], true);
1847 goto out;
1848 } else if (!strcasecmp(argv[0], "enable_group")) {
1849 r = bypass_pg_num(m, argv[1], false);
1850 goto out;
1851 } else if (!strcasecmp(argv[0], "switch_group")) {
1852 r = switch_pg_num(m, argv[1]);
1853 goto out;
1854 } else if (!strcasecmp(argv[0], "reinstate_path"))
1855 action = reinstate_path;
1856 else if (!strcasecmp(argv[0], "fail_path"))
1857 action = fail_path;
1858 else {
1859 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1860 goto out;
1861 }
1862
1863 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1864 if (r) {
1865 DMWARN("message: error getting device %s",
1866 argv[1]);
1867 goto out;
1868 }
1869
1870 r = action_dev(m, dev, action);
1871
1872 dm_put_device(ti, dev);
1873
1874 out:
1875 mutex_unlock(&m->work_mutex);
1876 return r;
1877 }
1878
1879 static int multipath_prepare_ioctl(struct dm_target *ti,
1880 struct block_device **bdev)
1881 {
1882 struct multipath *m = ti->private;
1883 struct pgpath *current_pgpath;
1884 int r;
1885
1886 current_pgpath = READ_ONCE(m->current_pgpath);
1887 if (!current_pgpath)
1888 current_pgpath = choose_pgpath(m, 0);
1889
1890 if (current_pgpath) {
1891 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1892 *bdev = current_pgpath->path.dev->bdev;
1893 r = 0;
1894 } else {
1895 /* pg_init has not started or completed */
1896 r = -ENOTCONN;
1897 }
1898 } else {
1899 /* No path is available */
1900 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1901 r = -ENOTCONN;
1902 else
1903 r = -EIO;
1904 }
1905
1906 if (r == -ENOTCONN) {
1907 if (!READ_ONCE(m->current_pg)) {
1908 /* Path status changed, redo selection */
1909 (void) choose_pgpath(m, 0);
1910 }
1911 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1912 pg_init_all_paths(m);
1913 dm_table_run_md_queue_async(m->ti->table);
1914 process_queued_io_list(m);
1915 }
1916
1917 /*
1918 * Only pass ioctls through if the device sizes match exactly.
1919 */
1920 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1921 return 1;
1922 return r;
1923 }
1924
1925 static int multipath_iterate_devices(struct dm_target *ti,
1926 iterate_devices_callout_fn fn, void *data)
1927 {
1928 struct multipath *m = ti->private;
1929 struct priority_group *pg;
1930 struct pgpath *p;
1931 int ret = 0;
1932
1933 list_for_each_entry(pg, &m->priority_groups, list) {
1934 list_for_each_entry(p, &pg->pgpaths, list) {
1935 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1936 if (ret)
1937 goto out;
1938 }
1939 }
1940
1941 out:
1942 return ret;
1943 }
1944
1945 static int pgpath_busy(struct pgpath *pgpath)
1946 {
1947 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1948
1949 return blk_lld_busy(q);
1950 }
1951
1952 /*
1953 * We return "busy", only when we can map I/Os but underlying devices
1954 * are busy (so even if we map I/Os now, the I/Os will wait on
1955 * the underlying queue).
1956 * In other words, if we want to kill I/Os or queue them inside us
1957 * due to map unavailability, we don't return "busy". Otherwise,
1958 * dm core won't give us the I/Os and we can't do what we want.
1959 */
1960 static int multipath_busy(struct dm_target *ti)
1961 {
1962 bool busy = false, has_active = false;
1963 struct multipath *m = ti->private;
1964 struct priority_group *pg, *next_pg;
1965 struct pgpath *pgpath;
1966
1967 /* pg_init in progress */
1968 if (atomic_read(&m->pg_init_in_progress))
1969 return true;
1970
1971 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1972 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1973 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1974
1975 /* Guess which priority_group will be used at next mapping time */
1976 pg = READ_ONCE(m->current_pg);
1977 next_pg = READ_ONCE(m->next_pg);
1978 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1979 pg = next_pg;
1980
1981 if (!pg) {
1982 /*
1983 * We don't know which pg will be used at next mapping time.
1984 * We don't call choose_pgpath() here to avoid to trigger
1985 * pg_init just by busy checking.
1986 * So we don't know whether underlying devices we will be using
1987 * at next mapping time are busy or not. Just try mapping.
1988 */
1989 return busy;
1990 }
1991
1992 /*
1993 * If there is one non-busy active path at least, the path selector
1994 * will be able to select it. So we consider such a pg as not busy.
1995 */
1996 busy = true;
1997 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1998 if (pgpath->is_active) {
1999 has_active = true;
2000 if (!pgpath_busy(pgpath)) {
2001 busy = false;
2002 break;
2003 }
2004 }
2005 }
2006
2007 if (!has_active) {
2008 /*
2009 * No active path in this pg, so this pg won't be used and
2010 * the current_pg will be changed at next mapping time.
2011 * We need to try mapping to determine it.
2012 */
2013 busy = false;
2014 }
2015
2016 return busy;
2017 }
2018
2019 /*-----------------------------------------------------------------
2020 * Module setup
2021 *---------------------------------------------------------------*/
2022 static struct target_type multipath_target = {
2023 .name = "multipath",
2024 .version = {1, 13, 0},
2025 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2026 DM_TARGET_PASSES_INTEGRITY,
2027 .module = THIS_MODULE,
2028 .ctr = multipath_ctr,
2029 .dtr = multipath_dtr,
2030 .clone_and_map_rq = multipath_clone_and_map,
2031 .release_clone_rq = multipath_release_clone,
2032 .rq_end_io = multipath_end_io,
2033 .map = multipath_map_bio,
2034 .end_io = multipath_end_io_bio,
2035 .presuspend = multipath_presuspend,
2036 .postsuspend = multipath_postsuspend,
2037 .resume = multipath_resume,
2038 .status = multipath_status,
2039 .message = multipath_message,
2040 .prepare_ioctl = multipath_prepare_ioctl,
2041 .iterate_devices = multipath_iterate_devices,
2042 .busy = multipath_busy,
2043 };
2044
2045 static int __init dm_multipath_init(void)
2046 {
2047 int r;
2048
2049 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2050 if (!kmultipathd) {
2051 DMERR("failed to create workqueue kmpathd");
2052 r = -ENOMEM;
2053 goto bad_alloc_kmultipathd;
2054 }
2055
2056 /*
2057 * A separate workqueue is used to handle the device handlers
2058 * to avoid overloading existing workqueue. Overloading the
2059 * old workqueue would also create a bottleneck in the
2060 * path of the storage hardware device activation.
2061 */
2062 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2063 WQ_MEM_RECLAIM);
2064 if (!kmpath_handlerd) {
2065 DMERR("failed to create workqueue kmpath_handlerd");
2066 r = -ENOMEM;
2067 goto bad_alloc_kmpath_handlerd;
2068 }
2069
2070 r = dm_register_target(&multipath_target);
2071 if (r < 0) {
2072 DMERR("request-based register failed %d", r);
2073 r = -EINVAL;
2074 goto bad_register_target;
2075 }
2076
2077 return 0;
2078
2079 bad_register_target:
2080 destroy_workqueue(kmpath_handlerd);
2081 bad_alloc_kmpath_handlerd:
2082 destroy_workqueue(kmultipathd);
2083 bad_alloc_kmultipathd:
2084 return r;
2085 }
2086
2087 static void __exit dm_multipath_exit(void)
2088 {
2089 destroy_workqueue(kmpath_handlerd);
2090 destroy_workqueue(kmultipathd);
2091
2092 dm_unregister_target(&multipath_target);
2093 }
2094
2095 module_init(dm_multipath_init);
2096 module_exit(dm_multipath_exit);
2097
2098 MODULE_DESCRIPTION(DM_NAME " multipath target");
2099 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2100 MODULE_LICENSE("GPL");
Linux with added FPC-III support