xref: /openbmc/linux/drivers/md/dm.c (revision b34e08d5)
1 /*
2  * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3  * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include "dm.h"
9 #include "dm-uevent.h"
10 
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/mutex.h>
14 #include <linux/moduleparam.h>
15 #include <linux/blkpg.h>
16 #include <linux/bio.h>
17 #include <linux/mempool.h>
18 #include <linux/slab.h>
19 #include <linux/idr.h>
20 #include <linux/hdreg.h>
21 #include <linux/delay.h>
22 
23 #include <trace/events/block.h>
24 
25 #define DM_MSG_PREFIX "core"
26 
27 #ifdef CONFIG_PRINTK
28 /*
29  * ratelimit state to be used in DMXXX_LIMIT().
30  */
31 DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
32 		       DEFAULT_RATELIMIT_INTERVAL,
33 		       DEFAULT_RATELIMIT_BURST);
34 EXPORT_SYMBOL(dm_ratelimit_state);
35 #endif
36 
37 /*
38  * Cookies are numeric values sent with CHANGE and REMOVE
39  * uevents while resuming, removing or renaming the device.
40  */
41 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
42 #define DM_COOKIE_LENGTH 24
43 
44 static const char *_name = DM_NAME;
45 
46 static unsigned int major = 0;
47 static unsigned int _major = 0;
48 
49 static DEFINE_IDR(_minor_idr);
50 
51 static DEFINE_SPINLOCK(_minor_lock);
52 
53 static void do_deferred_remove(struct work_struct *w);
54 
55 static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
56 
57 /*
58  * For bio-based dm.
59  * One of these is allocated per bio.
60  */
61 struct dm_io {
62 	struct mapped_device *md;
63 	int error;
64 	atomic_t io_count;
65 	struct bio *bio;
66 	unsigned long start_time;
67 	spinlock_t endio_lock;
68 	struct dm_stats_aux stats_aux;
69 };
70 
71 /*
72  * For request-based dm.
73  * One of these is allocated per request.
74  */
75 struct dm_rq_target_io {
76 	struct mapped_device *md;
77 	struct dm_target *ti;
78 	struct request *orig, clone;
79 	int error;
80 	union map_info info;
81 };
82 
83 /*
84  * For request-based dm - the bio clones we allocate are embedded in these
85  * structs.
86  *
87  * We allocate these with bio_alloc_bioset, using the front_pad parameter when
88  * the bioset is created - this means the bio has to come at the end of the
89  * struct.
90  */
91 struct dm_rq_clone_bio_info {
92 	struct bio *orig;
93 	struct dm_rq_target_io *tio;
94 	struct bio clone;
95 };
96 
97 union map_info *dm_get_rq_mapinfo(struct request *rq)
98 {
99 	if (rq && rq->end_io_data)
100 		return &((struct dm_rq_target_io *)rq->end_io_data)->info;
101 	return NULL;
102 }
103 EXPORT_SYMBOL_GPL(dm_get_rq_mapinfo);
104 
105 #define MINOR_ALLOCED ((void *)-1)
106 
107 /*
108  * Bits for the md->flags field.
109  */
110 #define DMF_BLOCK_IO_FOR_SUSPEND 0
111 #define DMF_SUSPENDED 1
112 #define DMF_FROZEN 2
113 #define DMF_FREEING 3
114 #define DMF_DELETING 4
115 #define DMF_NOFLUSH_SUSPENDING 5
116 #define DMF_MERGE_IS_OPTIONAL 6
117 #define DMF_DEFERRED_REMOVE 7
118 
119 /*
120  * A dummy definition to make RCU happy.
121  * struct dm_table should never be dereferenced in this file.
122  */
123 struct dm_table {
124 	int undefined__;
125 };
126 
127 /*
128  * Work processed by per-device workqueue.
129  */
130 struct mapped_device {
131 	struct srcu_struct io_barrier;
132 	struct mutex suspend_lock;
133 	atomic_t holders;
134 	atomic_t open_count;
135 
136 	/*
137 	 * The current mapping.
138 	 * Use dm_get_live_table{_fast} or take suspend_lock for
139 	 * dereference.
140 	 */
141 	struct dm_table *map;
142 
143 	unsigned long flags;
144 
145 	struct request_queue *queue;
146 	unsigned type;
147 	/* Protect queue and type against concurrent access. */
148 	struct mutex type_lock;
149 
150 	struct target_type *immutable_target_type;
151 
152 	struct gendisk *disk;
153 	char name[16];
154 
155 	void *interface_ptr;
156 
157 	/*
158 	 * A list of ios that arrived while we were suspended.
159 	 */
160 	atomic_t pending[2];
161 	wait_queue_head_t wait;
162 	struct work_struct work;
163 	struct bio_list deferred;
164 	spinlock_t deferred_lock;
165 
166 	/*
167 	 * Processing queue (flush)
168 	 */
169 	struct workqueue_struct *wq;
170 
171 	/*
172 	 * io objects are allocated from here.
173 	 */
174 	mempool_t *io_pool;
175 
176 	struct bio_set *bs;
177 
178 	/*
179 	 * Event handling.
180 	 */
181 	atomic_t event_nr;
182 	wait_queue_head_t eventq;
183 	atomic_t uevent_seq;
184 	struct list_head uevent_list;
185 	spinlock_t uevent_lock; /* Protect access to uevent_list */
186 
187 	/*
188 	 * freeze/thaw support require holding onto a super block
189 	 */
190 	struct super_block *frozen_sb;
191 	struct block_device *bdev;
192 
193 	/* forced geometry settings */
194 	struct hd_geometry geometry;
195 
196 	/* kobject and completion */
197 	struct dm_kobject_holder kobj_holder;
198 
199 	/* zero-length flush that will be cloned and submitted to targets */
200 	struct bio flush_bio;
201 
202 	struct dm_stats stats;
203 };
204 
205 /*
206  * For mempools pre-allocation at the table loading time.
207  */
208 struct dm_md_mempools {
209 	mempool_t *io_pool;
210 	struct bio_set *bs;
211 };
212 
213 #define RESERVED_BIO_BASED_IOS		16
214 #define RESERVED_REQUEST_BASED_IOS	256
215 #define RESERVED_MAX_IOS		1024
216 static struct kmem_cache *_io_cache;
217 static struct kmem_cache *_rq_tio_cache;
218 
219 /*
220  * Bio-based DM's mempools' reserved IOs set by the user.
221  */
222 static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
223 
224 /*
225  * Request-based DM's mempools' reserved IOs set by the user.
226  */
227 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
228 
229 static unsigned __dm_get_reserved_ios(unsigned *reserved_ios,
230 				      unsigned def, unsigned max)
231 {
232 	unsigned ios = ACCESS_ONCE(*reserved_ios);
233 	unsigned modified_ios = 0;
234 
235 	if (!ios)
236 		modified_ios = def;
237 	else if (ios > max)
238 		modified_ios = max;
239 
240 	if (modified_ios) {
241 		(void)cmpxchg(reserved_ios, ios, modified_ios);
242 		ios = modified_ios;
243 	}
244 
245 	return ios;
246 }
247 
248 unsigned dm_get_reserved_bio_based_ios(void)
249 {
250 	return __dm_get_reserved_ios(&reserved_bio_based_ios,
251 				     RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
252 }
253 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
254 
255 unsigned dm_get_reserved_rq_based_ios(void)
256 {
257 	return __dm_get_reserved_ios(&reserved_rq_based_ios,
258 				     RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
259 }
260 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
261 
262 static int __init local_init(void)
263 {
264 	int r = -ENOMEM;
265 
266 	/* allocate a slab for the dm_ios */
267 	_io_cache = KMEM_CACHE(dm_io, 0);
268 	if (!_io_cache)
269 		return r;
270 
271 	_rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
272 	if (!_rq_tio_cache)
273 		goto out_free_io_cache;
274 
275 	r = dm_uevent_init();
276 	if (r)
277 		goto out_free_rq_tio_cache;
278 
279 	_major = major;
280 	r = register_blkdev(_major, _name);
281 	if (r < 0)
282 		goto out_uevent_exit;
283 
284 	if (!_major)
285 		_major = r;
286 
287 	return 0;
288 
289 out_uevent_exit:
290 	dm_uevent_exit();
291 out_free_rq_tio_cache:
292 	kmem_cache_destroy(_rq_tio_cache);
293 out_free_io_cache:
294 	kmem_cache_destroy(_io_cache);
295 
296 	return r;
297 }
298 
299 static void local_exit(void)
300 {
301 	flush_scheduled_work();
302 
303 	kmem_cache_destroy(_rq_tio_cache);
304 	kmem_cache_destroy(_io_cache);
305 	unregister_blkdev(_major, _name);
306 	dm_uevent_exit();
307 
308 	_major = 0;
309 
310 	DMINFO("cleaned up");
311 }
312 
313 static int (*_inits[])(void) __initdata = {
314 	local_init,
315 	dm_target_init,
316 	dm_linear_init,
317 	dm_stripe_init,
318 	dm_io_init,
319 	dm_kcopyd_init,
320 	dm_interface_init,
321 	dm_statistics_init,
322 };
323 
324 static void (*_exits[])(void) = {
325 	local_exit,
326 	dm_target_exit,
327 	dm_linear_exit,
328 	dm_stripe_exit,
329 	dm_io_exit,
330 	dm_kcopyd_exit,
331 	dm_interface_exit,
332 	dm_statistics_exit,
333 };
334 
335 static int __init dm_init(void)
336 {
337 	const int count = ARRAY_SIZE(_inits);
338 
339 	int r, i;
340 
341 	for (i = 0; i < count; i++) {
342 		r = _inits[i]();
343 		if (r)
344 			goto bad;
345 	}
346 
347 	return 0;
348 
349       bad:
350 	while (i--)
351 		_exits[i]();
352 
353 	return r;
354 }
355 
356 static void __exit dm_exit(void)
357 {
358 	int i = ARRAY_SIZE(_exits);
359 
360 	while (i--)
361 		_exits[i]();
362 
363 	/*
364 	 * Should be empty by this point.
365 	 */
366 	idr_destroy(&_minor_idr);
367 }
368 
369 /*
370  * Block device functions
371  */
372 int dm_deleting_md(struct mapped_device *md)
373 {
374 	return test_bit(DMF_DELETING, &md->flags);
375 }
376 
377 static int dm_blk_open(struct block_device *bdev, fmode_t mode)
378 {
379 	struct mapped_device *md;
380 
381 	spin_lock(&_minor_lock);
382 
383 	md = bdev->bd_disk->private_data;
384 	if (!md)
385 		goto out;
386 
387 	if (test_bit(DMF_FREEING, &md->flags) ||
388 	    dm_deleting_md(md)) {
389 		md = NULL;
390 		goto out;
391 	}
392 
393 	dm_get(md);
394 	atomic_inc(&md->open_count);
395 
396 out:
397 	spin_unlock(&_minor_lock);
398 
399 	return md ? 0 : -ENXIO;
400 }
401 
402 static void dm_blk_close(struct gendisk *disk, fmode_t mode)
403 {
404 	struct mapped_device *md = disk->private_data;
405 
406 	spin_lock(&_minor_lock);
407 
408 	if (atomic_dec_and_test(&md->open_count) &&
409 	    (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
410 		schedule_work(&deferred_remove_work);
411 
412 	dm_put(md);
413 
414 	spin_unlock(&_minor_lock);
415 }
416 
417 int dm_open_count(struct mapped_device *md)
418 {
419 	return atomic_read(&md->open_count);
420 }
421 
422 /*
423  * Guarantees nothing is using the device before it's deleted.
424  */
425 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
426 {
427 	int r = 0;
428 
429 	spin_lock(&_minor_lock);
430 
431 	if (dm_open_count(md)) {
432 		r = -EBUSY;
433 		if (mark_deferred)
434 			set_bit(DMF_DEFERRED_REMOVE, &md->flags);
435 	} else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
436 		r = -EEXIST;
437 	else
438 		set_bit(DMF_DELETING, &md->flags);
439 
440 	spin_unlock(&_minor_lock);
441 
442 	return r;
443 }
444 
445 int dm_cancel_deferred_remove(struct mapped_device *md)
446 {
447 	int r = 0;
448 
449 	spin_lock(&_minor_lock);
450 
451 	if (test_bit(DMF_DELETING, &md->flags))
452 		r = -EBUSY;
453 	else
454 		clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
455 
456 	spin_unlock(&_minor_lock);
457 
458 	return r;
459 }
460 
461 static void do_deferred_remove(struct work_struct *w)
462 {
463 	dm_deferred_remove();
464 }
465 
466 sector_t dm_get_size(struct mapped_device *md)
467 {
468 	return get_capacity(md->disk);
469 }
470 
471 struct request_queue *dm_get_md_queue(struct mapped_device *md)
472 {
473 	return md->queue;
474 }
475 
476 struct dm_stats *dm_get_stats(struct mapped_device *md)
477 {
478 	return &md->stats;
479 }
480 
481 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
482 {
483 	struct mapped_device *md = bdev->bd_disk->private_data;
484 
485 	return dm_get_geometry(md, geo);
486 }
487 
488 static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
489 			unsigned int cmd, unsigned long arg)
490 {
491 	struct mapped_device *md = bdev->bd_disk->private_data;
492 	int srcu_idx;
493 	struct dm_table *map;
494 	struct dm_target *tgt;
495 	int r = -ENOTTY;
496 
497 retry:
498 	map = dm_get_live_table(md, &srcu_idx);
499 
500 	if (!map || !dm_table_get_size(map))
501 		goto out;
502 
503 	/* We only support devices that have a single target */
504 	if (dm_table_get_num_targets(map) != 1)
505 		goto out;
506 
507 	tgt = dm_table_get_target(map, 0);
508 
509 	if (dm_suspended_md(md)) {
510 		r = -EAGAIN;
511 		goto out;
512 	}
513 
514 	if (tgt->type->ioctl)
515 		r = tgt->type->ioctl(tgt, cmd, arg);
516 
517 out:
518 	dm_put_live_table(md, srcu_idx);
519 
520 	if (r == -ENOTCONN) {
521 		msleep(10);
522 		goto retry;
523 	}
524 
525 	return r;
526 }
527 
528 static struct dm_io *alloc_io(struct mapped_device *md)
529 {
530 	return mempool_alloc(md->io_pool, GFP_NOIO);
531 }
532 
533 static void free_io(struct mapped_device *md, struct dm_io *io)
534 {
535 	mempool_free(io, md->io_pool);
536 }
537 
538 static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
539 {
540 	bio_put(&tio->clone);
541 }
542 
543 static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
544 					    gfp_t gfp_mask)
545 {
546 	return mempool_alloc(md->io_pool, gfp_mask);
547 }
548 
549 static void free_rq_tio(struct dm_rq_target_io *tio)
550 {
551 	mempool_free(tio, tio->md->io_pool);
552 }
553 
554 static int md_in_flight(struct mapped_device *md)
555 {
556 	return atomic_read(&md->pending[READ]) +
557 	       atomic_read(&md->pending[WRITE]);
558 }
559 
560 static void start_io_acct(struct dm_io *io)
561 {
562 	struct mapped_device *md = io->md;
563 	struct bio *bio = io->bio;
564 	int cpu;
565 	int rw = bio_data_dir(bio);
566 
567 	io->start_time = jiffies;
568 
569 	cpu = part_stat_lock();
570 	part_round_stats(cpu, &dm_disk(md)->part0);
571 	part_stat_unlock();
572 	atomic_set(&dm_disk(md)->part0.in_flight[rw],
573 		atomic_inc_return(&md->pending[rw]));
574 
575 	if (unlikely(dm_stats_used(&md->stats)))
576 		dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
577 				    bio_sectors(bio), false, 0, &io->stats_aux);
578 }
579 
580 static void end_io_acct(struct dm_io *io)
581 {
582 	struct mapped_device *md = io->md;
583 	struct bio *bio = io->bio;
584 	unsigned long duration = jiffies - io->start_time;
585 	int pending, cpu;
586 	int rw = bio_data_dir(bio);
587 
588 	cpu = part_stat_lock();
589 	part_round_stats(cpu, &dm_disk(md)->part0);
590 	part_stat_add(cpu, &dm_disk(md)->part0, ticks[rw], duration);
591 	part_stat_unlock();
592 
593 	if (unlikely(dm_stats_used(&md->stats)))
594 		dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
595 				    bio_sectors(bio), true, duration, &io->stats_aux);
596 
597 	/*
598 	 * After this is decremented the bio must not be touched if it is
599 	 * a flush.
600 	 */
601 	pending = atomic_dec_return(&md->pending[rw]);
602 	atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
603 	pending += atomic_read(&md->pending[rw^0x1]);
604 
605 	/* nudge anyone waiting on suspend queue */
606 	if (!pending)
607 		wake_up(&md->wait);
608 }
609 
610 /*
611  * Add the bio to the list of deferred io.
612  */
613 static void queue_io(struct mapped_device *md, struct bio *bio)
614 {
615 	unsigned long flags;
616 
617 	spin_lock_irqsave(&md->deferred_lock, flags);
618 	bio_list_add(&md->deferred, bio);
619 	spin_unlock_irqrestore(&md->deferred_lock, flags);
620 	queue_work(md->wq, &md->work);
621 }
622 
623 /*
624  * Everyone (including functions in this file), should use this
625  * function to access the md->map field, and make sure they call
626  * dm_put_live_table() when finished.
627  */
628 struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
629 {
630 	*srcu_idx = srcu_read_lock(&md->io_barrier);
631 
632 	return srcu_dereference(md->map, &md->io_barrier);
633 }
634 
635 void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
636 {
637 	srcu_read_unlock(&md->io_barrier, srcu_idx);
638 }
639 
640 void dm_sync_table(struct mapped_device *md)
641 {
642 	synchronize_srcu(&md->io_barrier);
643 	synchronize_rcu_expedited();
644 }
645 
646 /*
647  * A fast alternative to dm_get_live_table/dm_put_live_table.
648  * The caller must not block between these two functions.
649  */
650 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
651 {
652 	rcu_read_lock();
653 	return rcu_dereference(md->map);
654 }
655 
656 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
657 {
658 	rcu_read_unlock();
659 }
660 
661 /*
662  * Get the geometry associated with a dm device
663  */
664 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
665 {
666 	*geo = md->geometry;
667 
668 	return 0;
669 }
670 
671 /*
672  * Set the geometry of a device.
673  */
674 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
675 {
676 	sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
677 
678 	if (geo->start > sz) {
679 		DMWARN("Start sector is beyond the geometry limits.");
680 		return -EINVAL;
681 	}
682 
683 	md->geometry = *geo;
684 
685 	return 0;
686 }
687 
688 /*-----------------------------------------------------------------
689  * CRUD START:
690  *   A more elegant soln is in the works that uses the queue
691  *   merge fn, unfortunately there are a couple of changes to
692  *   the block layer that I want to make for this.  So in the
693  *   interests of getting something for people to use I give
694  *   you this clearly demarcated crap.
695  *---------------------------------------------------------------*/
696 
697 static int __noflush_suspending(struct mapped_device *md)
698 {
699 	return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
700 }
701 
702 /*
703  * Decrements the number of outstanding ios that a bio has been
704  * cloned into, completing the original io if necc.
705  */
706 static void dec_pending(struct dm_io *io, int error)
707 {
708 	unsigned long flags;
709 	int io_error;
710 	struct bio *bio;
711 	struct mapped_device *md = io->md;
712 
713 	/* Push-back supersedes any I/O errors */
714 	if (unlikely(error)) {
715 		spin_lock_irqsave(&io->endio_lock, flags);
716 		if (!(io->error > 0 && __noflush_suspending(md)))
717 			io->error = error;
718 		spin_unlock_irqrestore(&io->endio_lock, flags);
719 	}
720 
721 	if (atomic_dec_and_test(&io->io_count)) {
722 		if (io->error == DM_ENDIO_REQUEUE) {
723 			/*
724 			 * Target requested pushing back the I/O.
725 			 */
726 			spin_lock_irqsave(&md->deferred_lock, flags);
727 			if (__noflush_suspending(md))
728 				bio_list_add_head(&md->deferred, io->bio);
729 			else
730 				/* noflush suspend was interrupted. */
731 				io->error = -EIO;
732 			spin_unlock_irqrestore(&md->deferred_lock, flags);
733 		}
734 
735 		io_error = io->error;
736 		bio = io->bio;
737 		end_io_acct(io);
738 		free_io(md, io);
739 
740 		if (io_error == DM_ENDIO_REQUEUE)
741 			return;
742 
743 		if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
744 			/*
745 			 * Preflush done for flush with data, reissue
746 			 * without REQ_FLUSH.
747 			 */
748 			bio->bi_rw &= ~REQ_FLUSH;
749 			queue_io(md, bio);
750 		} else {
751 			/* done with normal IO or empty flush */
752 			trace_block_bio_complete(md->queue, bio, io_error);
753 			bio_endio(bio, io_error);
754 		}
755 	}
756 }
757 
758 static void clone_endio(struct bio *bio, int error)
759 {
760 	int r = 0;
761 	struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
762 	struct dm_io *io = tio->io;
763 	struct mapped_device *md = tio->io->md;
764 	dm_endio_fn endio = tio->ti->type->end_io;
765 
766 	if (!bio_flagged(bio, BIO_UPTODATE) && !error)
767 		error = -EIO;
768 
769 	if (endio) {
770 		r = endio(tio->ti, bio, error);
771 		if (r < 0 || r == DM_ENDIO_REQUEUE)
772 			/*
773 			 * error and requeue request are handled
774 			 * in dec_pending().
775 			 */
776 			error = r;
777 		else if (r == DM_ENDIO_INCOMPLETE)
778 			/* The target will handle the io */
779 			return;
780 		else if (r) {
781 			DMWARN("unimplemented target endio return value: %d", r);
782 			BUG();
783 		}
784 	}
785 
786 	free_tio(md, tio);
787 	dec_pending(io, error);
788 }
789 
790 /*
791  * Partial completion handling for request-based dm
792  */
793 static void end_clone_bio(struct bio *clone, int error)
794 {
795 	struct dm_rq_clone_bio_info *info =
796 		container_of(clone, struct dm_rq_clone_bio_info, clone);
797 	struct dm_rq_target_io *tio = info->tio;
798 	struct bio *bio = info->orig;
799 	unsigned int nr_bytes = info->orig->bi_iter.bi_size;
800 
801 	bio_put(clone);
802 
803 	if (tio->error)
804 		/*
805 		 * An error has already been detected on the request.
806 		 * Once error occurred, just let clone->end_io() handle
807 		 * the remainder.
808 		 */
809 		return;
810 	else if (error) {
811 		/*
812 		 * Don't notice the error to the upper layer yet.
813 		 * The error handling decision is made by the target driver,
814 		 * when the request is completed.
815 		 */
816 		tio->error = error;
817 		return;
818 	}
819 
820 	/*
821 	 * I/O for the bio successfully completed.
822 	 * Notice the data completion to the upper layer.
823 	 */
824 
825 	/*
826 	 * bios are processed from the head of the list.
827 	 * So the completing bio should always be rq->bio.
828 	 * If it's not, something wrong is happening.
829 	 */
830 	if (tio->orig->bio != bio)
831 		DMERR("bio completion is going in the middle of the request");
832 
833 	/*
834 	 * Update the original request.
835 	 * Do not use blk_end_request() here, because it may complete
836 	 * the original request before the clone, and break the ordering.
837 	 */
838 	blk_update_request(tio->orig, 0, nr_bytes);
839 }
840 
841 /*
842  * Don't touch any member of the md after calling this function because
843  * the md may be freed in dm_put() at the end of this function.
844  * Or do dm_get() before calling this function and dm_put() later.
845  */
846 static void rq_completed(struct mapped_device *md, int rw, int run_queue)
847 {
848 	atomic_dec(&md->pending[rw]);
849 
850 	/* nudge anyone waiting on suspend queue */
851 	if (!md_in_flight(md))
852 		wake_up(&md->wait);
853 
854 	/*
855 	 * Run this off this callpath, as drivers could invoke end_io while
856 	 * inside their request_fn (and holding the queue lock). Calling
857 	 * back into ->request_fn() could deadlock attempting to grab the
858 	 * queue lock again.
859 	 */
860 	if (run_queue)
861 		blk_run_queue_async(md->queue);
862 
863 	/*
864 	 * dm_put() must be at the end of this function. See the comment above
865 	 */
866 	dm_put(md);
867 }
868 
869 static void free_rq_clone(struct request *clone)
870 {
871 	struct dm_rq_target_io *tio = clone->end_io_data;
872 
873 	blk_rq_unprep_clone(clone);
874 	free_rq_tio(tio);
875 }
876 
877 /*
878  * Complete the clone and the original request.
879  * Must be called without queue lock.
880  */
881 static void dm_end_request(struct request *clone, int error)
882 {
883 	int rw = rq_data_dir(clone);
884 	struct dm_rq_target_io *tio = clone->end_io_data;
885 	struct mapped_device *md = tio->md;
886 	struct request *rq = tio->orig;
887 
888 	if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
889 		rq->errors = clone->errors;
890 		rq->resid_len = clone->resid_len;
891 
892 		if (rq->sense)
893 			/*
894 			 * We are using the sense buffer of the original
895 			 * request.
896 			 * So setting the length of the sense data is enough.
897 			 */
898 			rq->sense_len = clone->sense_len;
899 	}
900 
901 	free_rq_clone(clone);
902 	blk_end_request_all(rq, error);
903 	rq_completed(md, rw, true);
904 }
905 
906 static void dm_unprep_request(struct request *rq)
907 {
908 	struct request *clone = rq->special;
909 
910 	rq->special = NULL;
911 	rq->cmd_flags &= ~REQ_DONTPREP;
912 
913 	free_rq_clone(clone);
914 }
915 
916 /*
917  * Requeue the original request of a clone.
918  */
919 void dm_requeue_unmapped_request(struct request *clone)
920 {
921 	int rw = rq_data_dir(clone);
922 	struct dm_rq_target_io *tio = clone->end_io_data;
923 	struct mapped_device *md = tio->md;
924 	struct request *rq = tio->orig;
925 	struct request_queue *q = rq->q;
926 	unsigned long flags;
927 
928 	dm_unprep_request(rq);
929 
930 	spin_lock_irqsave(q->queue_lock, flags);
931 	blk_requeue_request(q, rq);
932 	spin_unlock_irqrestore(q->queue_lock, flags);
933 
934 	rq_completed(md, rw, 0);
935 }
936 EXPORT_SYMBOL_GPL(dm_requeue_unmapped_request);
937 
938 static void __stop_queue(struct request_queue *q)
939 {
940 	blk_stop_queue(q);
941 }
942 
943 static void stop_queue(struct request_queue *q)
944 {
945 	unsigned long flags;
946 
947 	spin_lock_irqsave(q->queue_lock, flags);
948 	__stop_queue(q);
949 	spin_unlock_irqrestore(q->queue_lock, flags);
950 }
951 
952 static void __start_queue(struct request_queue *q)
953 {
954 	if (blk_queue_stopped(q))
955 		blk_start_queue(q);
956 }
957 
958 static void start_queue(struct request_queue *q)
959 {
960 	unsigned long flags;
961 
962 	spin_lock_irqsave(q->queue_lock, flags);
963 	__start_queue(q);
964 	spin_unlock_irqrestore(q->queue_lock, flags);
965 }
966 
967 static void dm_done(struct request *clone, int error, bool mapped)
968 {
969 	int r = error;
970 	struct dm_rq_target_io *tio = clone->end_io_data;
971 	dm_request_endio_fn rq_end_io = NULL;
972 
973 	if (tio->ti) {
974 		rq_end_io = tio->ti->type->rq_end_io;
975 
976 		if (mapped && rq_end_io)
977 			r = rq_end_io(tio->ti, clone, error, &tio->info);
978 	}
979 
980 	if (r <= 0)
981 		/* The target wants to complete the I/O */
982 		dm_end_request(clone, r);
983 	else if (r == DM_ENDIO_INCOMPLETE)
984 		/* The target will handle the I/O */
985 		return;
986 	else if (r == DM_ENDIO_REQUEUE)
987 		/* The target wants to requeue the I/O */
988 		dm_requeue_unmapped_request(clone);
989 	else {
990 		DMWARN("unimplemented target endio return value: %d", r);
991 		BUG();
992 	}
993 }
994 
995 /*
996  * Request completion handler for request-based dm
997  */
998 static void dm_softirq_done(struct request *rq)
999 {
1000 	bool mapped = true;
1001 	struct request *clone = rq->completion_data;
1002 	struct dm_rq_target_io *tio = clone->end_io_data;
1003 
1004 	if (rq->cmd_flags & REQ_FAILED)
1005 		mapped = false;
1006 
1007 	dm_done(clone, tio->error, mapped);
1008 }
1009 
1010 /*
1011  * Complete the clone and the original request with the error status
1012  * through softirq context.
1013  */
1014 static void dm_complete_request(struct request *clone, int error)
1015 {
1016 	struct dm_rq_target_io *tio = clone->end_io_data;
1017 	struct request *rq = tio->orig;
1018 
1019 	tio->error = error;
1020 	rq->completion_data = clone;
1021 	blk_complete_request(rq);
1022 }
1023 
1024 /*
1025  * Complete the not-mapped clone and the original request with the error status
1026  * through softirq context.
1027  * Target's rq_end_io() function isn't called.
1028  * This may be used when the target's map_rq() function fails.
1029  */
1030 void dm_kill_unmapped_request(struct request *clone, int error)
1031 {
1032 	struct dm_rq_target_io *tio = clone->end_io_data;
1033 	struct request *rq = tio->orig;
1034 
1035 	rq->cmd_flags |= REQ_FAILED;
1036 	dm_complete_request(clone, error);
1037 }
1038 EXPORT_SYMBOL_GPL(dm_kill_unmapped_request);
1039 
1040 /*
1041  * Called with the queue lock held
1042  */
1043 static void end_clone_request(struct request *clone, int error)
1044 {
1045 	/*
1046 	 * For just cleaning up the information of the queue in which
1047 	 * the clone was dispatched.
1048 	 * The clone is *NOT* freed actually here because it is alloced from
1049 	 * dm own mempool and REQ_ALLOCED isn't set in clone->cmd_flags.
1050 	 */
1051 	__blk_put_request(clone->q, clone);
1052 
1053 	/*
1054 	 * Actual request completion is done in a softirq context which doesn't
1055 	 * hold the queue lock.  Otherwise, deadlock could occur because:
1056 	 *     - another request may be submitted by the upper level driver
1057 	 *       of the stacking during the completion
1058 	 *     - the submission which requires queue lock may be done
1059 	 *       against this queue
1060 	 */
1061 	dm_complete_request(clone, error);
1062 }
1063 
1064 /*
1065  * Return maximum size of I/O possible at the supplied sector up to the current
1066  * target boundary.
1067  */
1068 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1069 {
1070 	sector_t target_offset = dm_target_offset(ti, sector);
1071 
1072 	return ti->len - target_offset;
1073 }
1074 
1075 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1076 {
1077 	sector_t len = max_io_len_target_boundary(sector, ti);
1078 	sector_t offset, max_len;
1079 
1080 	/*
1081 	 * Does the target need to split even further?
1082 	 */
1083 	if (ti->max_io_len) {
1084 		offset = dm_target_offset(ti, sector);
1085 		if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1086 			max_len = sector_div(offset, ti->max_io_len);
1087 		else
1088 			max_len = offset & (ti->max_io_len - 1);
1089 		max_len = ti->max_io_len - max_len;
1090 
1091 		if (len > max_len)
1092 			len = max_len;
1093 	}
1094 
1095 	return len;
1096 }
1097 
1098 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1099 {
1100 	if (len > UINT_MAX) {
1101 		DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1102 		      (unsigned long long)len, UINT_MAX);
1103 		ti->error = "Maximum size of target IO is too large";
1104 		return -EINVAL;
1105 	}
1106 
1107 	ti->max_io_len = (uint32_t) len;
1108 
1109 	return 0;
1110 }
1111 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1112 
1113 static void __map_bio(struct dm_target_io *tio)
1114 {
1115 	int r;
1116 	sector_t sector;
1117 	struct mapped_device *md;
1118 	struct bio *clone = &tio->clone;
1119 	struct dm_target *ti = tio->ti;
1120 
1121 	clone->bi_end_io = clone_endio;
1122 
1123 	/*
1124 	 * Map the clone.  If r == 0 we don't need to do
1125 	 * anything, the target has assumed ownership of
1126 	 * this io.
1127 	 */
1128 	atomic_inc(&tio->io->io_count);
1129 	sector = clone->bi_iter.bi_sector;
1130 	r = ti->type->map(ti, clone);
1131 	if (r == DM_MAPIO_REMAPPED) {
1132 		/* the bio has been remapped so dispatch it */
1133 
1134 		trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1135 				      tio->io->bio->bi_bdev->bd_dev, sector);
1136 
1137 		generic_make_request(clone);
1138 	} else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1139 		/* error the io and bail out, or requeue it if needed */
1140 		md = tio->io->md;
1141 		dec_pending(tio->io, r);
1142 		free_tio(md, tio);
1143 	} else if (r) {
1144 		DMWARN("unimplemented target map return value: %d", r);
1145 		BUG();
1146 	}
1147 }
1148 
1149 struct clone_info {
1150 	struct mapped_device *md;
1151 	struct dm_table *map;
1152 	struct bio *bio;
1153 	struct dm_io *io;
1154 	sector_t sector;
1155 	sector_t sector_count;
1156 };
1157 
1158 static void bio_setup_sector(struct bio *bio, sector_t sector, sector_t len)
1159 {
1160 	bio->bi_iter.bi_sector = sector;
1161 	bio->bi_iter.bi_size = to_bytes(len);
1162 }
1163 
1164 /*
1165  * Creates a bio that consists of range of complete bvecs.
1166  */
1167 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1168 		      sector_t sector, unsigned len)
1169 {
1170 	struct bio *clone = &tio->clone;
1171 
1172 	__bio_clone_fast(clone, bio);
1173 
1174 	if (bio_integrity(bio))
1175 		bio_integrity_clone(clone, bio, GFP_NOIO);
1176 
1177 	bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1178 	clone->bi_iter.bi_size = to_bytes(len);
1179 
1180 	if (bio_integrity(bio))
1181 		bio_integrity_trim(clone, 0, len);
1182 }
1183 
1184 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1185 				      struct dm_target *ti, int nr_iovecs,
1186 				      unsigned target_bio_nr)
1187 {
1188 	struct dm_target_io *tio;
1189 	struct bio *clone;
1190 
1191 	clone = bio_alloc_bioset(GFP_NOIO, nr_iovecs, ci->md->bs);
1192 	tio = container_of(clone, struct dm_target_io, clone);
1193 
1194 	tio->io = ci->io;
1195 	tio->ti = ti;
1196 	tio->target_bio_nr = target_bio_nr;
1197 
1198 	return tio;
1199 }
1200 
1201 static void __clone_and_map_simple_bio(struct clone_info *ci,
1202 				       struct dm_target *ti,
1203 				       unsigned target_bio_nr, sector_t len)
1204 {
1205 	struct dm_target_io *tio = alloc_tio(ci, ti, ci->bio->bi_max_vecs, target_bio_nr);
1206 	struct bio *clone = &tio->clone;
1207 
1208 	/*
1209 	 * Discard requests require the bio's inline iovecs be initialized.
1210 	 * ci->bio->bi_max_vecs is BIO_INLINE_VECS anyway, for both flush
1211 	 * and discard, so no need for concern about wasted bvec allocations.
1212 	 */
1213 	 __bio_clone_fast(clone, ci->bio);
1214 	if (len)
1215 		bio_setup_sector(clone, ci->sector, len);
1216 
1217 	__map_bio(tio);
1218 }
1219 
1220 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1221 				  unsigned num_bios, sector_t len)
1222 {
1223 	unsigned target_bio_nr;
1224 
1225 	for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1226 		__clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1227 }
1228 
1229 static int __send_empty_flush(struct clone_info *ci)
1230 {
1231 	unsigned target_nr = 0;
1232 	struct dm_target *ti;
1233 
1234 	BUG_ON(bio_has_data(ci->bio));
1235 	while ((ti = dm_table_get_target(ci->map, target_nr++)))
1236 		__send_duplicate_bios(ci, ti, ti->num_flush_bios, 0);
1237 
1238 	return 0;
1239 }
1240 
1241 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1242 				     sector_t sector, unsigned len)
1243 {
1244 	struct bio *bio = ci->bio;
1245 	struct dm_target_io *tio;
1246 	unsigned target_bio_nr;
1247 	unsigned num_target_bios = 1;
1248 
1249 	/*
1250 	 * Does the target want to receive duplicate copies of the bio?
1251 	 */
1252 	if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1253 		num_target_bios = ti->num_write_bios(ti, bio);
1254 
1255 	for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1256 		tio = alloc_tio(ci, ti, 0, target_bio_nr);
1257 		clone_bio(tio, bio, sector, len);
1258 		__map_bio(tio);
1259 	}
1260 }
1261 
1262 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1263 
1264 static unsigned get_num_discard_bios(struct dm_target *ti)
1265 {
1266 	return ti->num_discard_bios;
1267 }
1268 
1269 static unsigned get_num_write_same_bios(struct dm_target *ti)
1270 {
1271 	return ti->num_write_same_bios;
1272 }
1273 
1274 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1275 
1276 static bool is_split_required_for_discard(struct dm_target *ti)
1277 {
1278 	return ti->split_discard_bios;
1279 }
1280 
1281 static int __send_changing_extent_only(struct clone_info *ci,
1282 				       get_num_bios_fn get_num_bios,
1283 				       is_split_required_fn is_split_required)
1284 {
1285 	struct dm_target *ti;
1286 	sector_t len;
1287 	unsigned num_bios;
1288 
1289 	do {
1290 		ti = dm_table_find_target(ci->map, ci->sector);
1291 		if (!dm_target_is_valid(ti))
1292 			return -EIO;
1293 
1294 		/*
1295 		 * Even though the device advertised support for this type of
1296 		 * request, that does not mean every target supports it, and
1297 		 * reconfiguration might also have changed that since the
1298 		 * check was performed.
1299 		 */
1300 		num_bios = get_num_bios ? get_num_bios(ti) : 0;
1301 		if (!num_bios)
1302 			return -EOPNOTSUPP;
1303 
1304 		if (is_split_required && !is_split_required(ti))
1305 			len = min(ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1306 		else
1307 			len = min(ci->sector_count, max_io_len(ci->sector, ti));
1308 
1309 		__send_duplicate_bios(ci, ti, num_bios, len);
1310 
1311 		ci->sector += len;
1312 	} while (ci->sector_count -= len);
1313 
1314 	return 0;
1315 }
1316 
1317 static int __send_discard(struct clone_info *ci)
1318 {
1319 	return __send_changing_extent_only(ci, get_num_discard_bios,
1320 					   is_split_required_for_discard);
1321 }
1322 
1323 static int __send_write_same(struct clone_info *ci)
1324 {
1325 	return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1326 }
1327 
1328 /*
1329  * Select the correct strategy for processing a non-flush bio.
1330  */
1331 static int __split_and_process_non_flush(struct clone_info *ci)
1332 {
1333 	struct bio *bio = ci->bio;
1334 	struct dm_target *ti;
1335 	unsigned len;
1336 
1337 	if (unlikely(bio->bi_rw & REQ_DISCARD))
1338 		return __send_discard(ci);
1339 	else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1340 		return __send_write_same(ci);
1341 
1342 	ti = dm_table_find_target(ci->map, ci->sector);
1343 	if (!dm_target_is_valid(ti))
1344 		return -EIO;
1345 
1346 	len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1347 
1348 	__clone_and_map_data_bio(ci, ti, ci->sector, len);
1349 
1350 	ci->sector += len;
1351 	ci->sector_count -= len;
1352 
1353 	return 0;
1354 }
1355 
1356 /*
1357  * Entry point to split a bio into clones and submit them to the targets.
1358  */
1359 static void __split_and_process_bio(struct mapped_device *md,
1360 				    struct dm_table *map, struct bio *bio)
1361 {
1362 	struct clone_info ci;
1363 	int error = 0;
1364 
1365 	if (unlikely(!map)) {
1366 		bio_io_error(bio);
1367 		return;
1368 	}
1369 
1370 	ci.map = map;
1371 	ci.md = md;
1372 	ci.io = alloc_io(md);
1373 	ci.io->error = 0;
1374 	atomic_set(&ci.io->io_count, 1);
1375 	ci.io->bio = bio;
1376 	ci.io->md = md;
1377 	spin_lock_init(&ci.io->endio_lock);
1378 	ci.sector = bio->bi_iter.bi_sector;
1379 
1380 	start_io_acct(ci.io);
1381 
1382 	if (bio->bi_rw & REQ_FLUSH) {
1383 		ci.bio = &ci.md->flush_bio;
1384 		ci.sector_count = 0;
1385 		error = __send_empty_flush(&ci);
1386 		/* dec_pending submits any data associated with flush */
1387 	} else {
1388 		ci.bio = bio;
1389 		ci.sector_count = bio_sectors(bio);
1390 		while (ci.sector_count && !error)
1391 			error = __split_and_process_non_flush(&ci);
1392 	}
1393 
1394 	/* drop the extra reference count */
1395 	dec_pending(ci.io, error);
1396 }
1397 /*-----------------------------------------------------------------
1398  * CRUD END
1399  *---------------------------------------------------------------*/
1400 
1401 static int dm_merge_bvec(struct request_queue *q,
1402 			 struct bvec_merge_data *bvm,
1403 			 struct bio_vec *biovec)
1404 {
1405 	struct mapped_device *md = q->queuedata;
1406 	struct dm_table *map = dm_get_live_table_fast(md);
1407 	struct dm_target *ti;
1408 	sector_t max_sectors;
1409 	int max_size = 0;
1410 
1411 	if (unlikely(!map))
1412 		goto out;
1413 
1414 	ti = dm_table_find_target(map, bvm->bi_sector);
1415 	if (!dm_target_is_valid(ti))
1416 		goto out;
1417 
1418 	/*
1419 	 * Find maximum amount of I/O that won't need splitting
1420 	 */
1421 	max_sectors = min(max_io_len(bvm->bi_sector, ti),
1422 			  (sector_t) BIO_MAX_SECTORS);
1423 	max_size = (max_sectors << SECTOR_SHIFT) - bvm->bi_size;
1424 	if (max_size < 0)
1425 		max_size = 0;
1426 
1427 	/*
1428 	 * merge_bvec_fn() returns number of bytes
1429 	 * it can accept at this offset
1430 	 * max is precomputed maximal io size
1431 	 */
1432 	if (max_size && ti->type->merge)
1433 		max_size = ti->type->merge(ti, bvm, biovec, max_size);
1434 	/*
1435 	 * If the target doesn't support merge method and some of the devices
1436 	 * provided their merge_bvec method (we know this by looking at
1437 	 * queue_max_hw_sectors), then we can't allow bios with multiple vector
1438 	 * entries.  So always set max_size to 0, and the code below allows
1439 	 * just one page.
1440 	 */
1441 	else if (queue_max_hw_sectors(q) <= PAGE_SIZE >> 9)
1442 
1443 		max_size = 0;
1444 
1445 out:
1446 	dm_put_live_table_fast(md);
1447 	/*
1448 	 * Always allow an entire first page
1449 	 */
1450 	if (max_size <= biovec->bv_len && !(bvm->bi_size >> SECTOR_SHIFT))
1451 		max_size = biovec->bv_len;
1452 
1453 	return max_size;
1454 }
1455 
1456 /*
1457  * The request function that just remaps the bio built up by
1458  * dm_merge_bvec.
1459  */
1460 static void _dm_request(struct request_queue *q, struct bio *bio)
1461 {
1462 	int rw = bio_data_dir(bio);
1463 	struct mapped_device *md = q->queuedata;
1464 	int cpu;
1465 	int srcu_idx;
1466 	struct dm_table *map;
1467 
1468 	map = dm_get_live_table(md, &srcu_idx);
1469 
1470 	cpu = part_stat_lock();
1471 	part_stat_inc(cpu, &dm_disk(md)->part0, ios[rw]);
1472 	part_stat_add(cpu, &dm_disk(md)->part0, sectors[rw], bio_sectors(bio));
1473 	part_stat_unlock();
1474 
1475 	/* if we're suspended, we have to queue this io for later */
1476 	if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1477 		dm_put_live_table(md, srcu_idx);
1478 
1479 		if (bio_rw(bio) != READA)
1480 			queue_io(md, bio);
1481 		else
1482 			bio_io_error(bio);
1483 		return;
1484 	}
1485 
1486 	__split_and_process_bio(md, map, bio);
1487 	dm_put_live_table(md, srcu_idx);
1488 	return;
1489 }
1490 
1491 int dm_request_based(struct mapped_device *md)
1492 {
1493 	return blk_queue_stackable(md->queue);
1494 }
1495 
1496 static void dm_request(struct request_queue *q, struct bio *bio)
1497 {
1498 	struct mapped_device *md = q->queuedata;
1499 
1500 	if (dm_request_based(md))
1501 		blk_queue_bio(q, bio);
1502 	else
1503 		_dm_request(q, bio);
1504 }
1505 
1506 void dm_dispatch_request(struct request *rq)
1507 {
1508 	int r;
1509 
1510 	if (blk_queue_io_stat(rq->q))
1511 		rq->cmd_flags |= REQ_IO_STAT;
1512 
1513 	rq->start_time = jiffies;
1514 	r = blk_insert_cloned_request(rq->q, rq);
1515 	if (r)
1516 		dm_complete_request(rq, r);
1517 }
1518 EXPORT_SYMBOL_GPL(dm_dispatch_request);
1519 
1520 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1521 				 void *data)
1522 {
1523 	struct dm_rq_target_io *tio = data;
1524 	struct dm_rq_clone_bio_info *info =
1525 		container_of(bio, struct dm_rq_clone_bio_info, clone);
1526 
1527 	info->orig = bio_orig;
1528 	info->tio = tio;
1529 	bio->bi_end_io = end_clone_bio;
1530 
1531 	return 0;
1532 }
1533 
1534 static int setup_clone(struct request *clone, struct request *rq,
1535 		       struct dm_rq_target_io *tio)
1536 {
1537 	int r;
1538 
1539 	r = blk_rq_prep_clone(clone, rq, tio->md->bs, GFP_ATOMIC,
1540 			      dm_rq_bio_constructor, tio);
1541 	if (r)
1542 		return r;
1543 
1544 	clone->cmd = rq->cmd;
1545 	clone->cmd_len = rq->cmd_len;
1546 	clone->sense = rq->sense;
1547 	clone->buffer = rq->buffer;
1548 	clone->end_io = end_clone_request;
1549 	clone->end_io_data = tio;
1550 
1551 	return 0;
1552 }
1553 
1554 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1555 				gfp_t gfp_mask)
1556 {
1557 	struct request *clone;
1558 	struct dm_rq_target_io *tio;
1559 
1560 	tio = alloc_rq_tio(md, gfp_mask);
1561 	if (!tio)
1562 		return NULL;
1563 
1564 	tio->md = md;
1565 	tio->ti = NULL;
1566 	tio->orig = rq;
1567 	tio->error = 0;
1568 	memset(&tio->info, 0, sizeof(tio->info));
1569 
1570 	clone = &tio->clone;
1571 	if (setup_clone(clone, rq, tio)) {
1572 		/* -ENOMEM */
1573 		free_rq_tio(tio);
1574 		return NULL;
1575 	}
1576 
1577 	return clone;
1578 }
1579 
1580 /*
1581  * Called with the queue lock held.
1582  */
1583 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1584 {
1585 	struct mapped_device *md = q->queuedata;
1586 	struct request *clone;
1587 
1588 	if (unlikely(rq->special)) {
1589 		DMWARN("Already has something in rq->special.");
1590 		return BLKPREP_KILL;
1591 	}
1592 
1593 	clone = clone_rq(rq, md, GFP_ATOMIC);
1594 	if (!clone)
1595 		return BLKPREP_DEFER;
1596 
1597 	rq->special = clone;
1598 	rq->cmd_flags |= REQ_DONTPREP;
1599 
1600 	return BLKPREP_OK;
1601 }
1602 
1603 /*
1604  * Returns:
1605  * 0  : the request has been processed (not requeued)
1606  * !0 : the request has been requeued
1607  */
1608 static int map_request(struct dm_target *ti, struct request *clone,
1609 		       struct mapped_device *md)
1610 {
1611 	int r, requeued = 0;
1612 	struct dm_rq_target_io *tio = clone->end_io_data;
1613 
1614 	tio->ti = ti;
1615 	r = ti->type->map_rq(ti, clone, &tio->info);
1616 	switch (r) {
1617 	case DM_MAPIO_SUBMITTED:
1618 		/* The target has taken the I/O to submit by itself later */
1619 		break;
1620 	case DM_MAPIO_REMAPPED:
1621 		/* The target has remapped the I/O so dispatch it */
1622 		trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1623 				     blk_rq_pos(tio->orig));
1624 		dm_dispatch_request(clone);
1625 		break;
1626 	case DM_MAPIO_REQUEUE:
1627 		/* The target wants to requeue the I/O */
1628 		dm_requeue_unmapped_request(clone);
1629 		requeued = 1;
1630 		break;
1631 	default:
1632 		if (r > 0) {
1633 			DMWARN("unimplemented target map return value: %d", r);
1634 			BUG();
1635 		}
1636 
1637 		/* The target wants to complete the I/O */
1638 		dm_kill_unmapped_request(clone, r);
1639 		break;
1640 	}
1641 
1642 	return requeued;
1643 }
1644 
1645 static struct request *dm_start_request(struct mapped_device *md, struct request *orig)
1646 {
1647 	struct request *clone;
1648 
1649 	blk_start_request(orig);
1650 	clone = orig->special;
1651 	atomic_inc(&md->pending[rq_data_dir(clone)]);
1652 
1653 	/*
1654 	 * Hold the md reference here for the in-flight I/O.
1655 	 * We can't rely on the reference count by device opener,
1656 	 * because the device may be closed during the request completion
1657 	 * when all bios are completed.
1658 	 * See the comment in rq_completed() too.
1659 	 */
1660 	dm_get(md);
1661 
1662 	return clone;
1663 }
1664 
1665 /*
1666  * q->request_fn for request-based dm.
1667  * Called with the queue lock held.
1668  */
1669 static void dm_request_fn(struct request_queue *q)
1670 {
1671 	struct mapped_device *md = q->queuedata;
1672 	int srcu_idx;
1673 	struct dm_table *map = dm_get_live_table(md, &srcu_idx);
1674 	struct dm_target *ti;
1675 	struct request *rq, *clone;
1676 	sector_t pos;
1677 
1678 	/*
1679 	 * For suspend, check blk_queue_stopped() and increment
1680 	 * ->pending within a single queue_lock not to increment the
1681 	 * number of in-flight I/Os after the queue is stopped in
1682 	 * dm_suspend().
1683 	 */
1684 	while (!blk_queue_stopped(q)) {
1685 		rq = blk_peek_request(q);
1686 		if (!rq)
1687 			goto delay_and_out;
1688 
1689 		/* always use block 0 to find the target for flushes for now */
1690 		pos = 0;
1691 		if (!(rq->cmd_flags & REQ_FLUSH))
1692 			pos = blk_rq_pos(rq);
1693 
1694 		ti = dm_table_find_target(map, pos);
1695 		if (!dm_target_is_valid(ti)) {
1696 			/*
1697 			 * Must perform setup, that dm_done() requires,
1698 			 * before calling dm_kill_unmapped_request
1699 			 */
1700 			DMERR_LIMIT("request attempted access beyond the end of device");
1701 			clone = dm_start_request(md, rq);
1702 			dm_kill_unmapped_request(clone, -EIO);
1703 			continue;
1704 		}
1705 
1706 		if (ti->type->busy && ti->type->busy(ti))
1707 			goto delay_and_out;
1708 
1709 		clone = dm_start_request(md, rq);
1710 
1711 		spin_unlock(q->queue_lock);
1712 		if (map_request(ti, clone, md))
1713 			goto requeued;
1714 
1715 		BUG_ON(!irqs_disabled());
1716 		spin_lock(q->queue_lock);
1717 	}
1718 
1719 	goto out;
1720 
1721 requeued:
1722 	BUG_ON(!irqs_disabled());
1723 	spin_lock(q->queue_lock);
1724 
1725 delay_and_out:
1726 	blk_delay_queue(q, HZ / 10);
1727 out:
1728 	dm_put_live_table(md, srcu_idx);
1729 }
1730 
1731 int dm_underlying_device_busy(struct request_queue *q)
1732 {
1733 	return blk_lld_busy(q);
1734 }
1735 EXPORT_SYMBOL_GPL(dm_underlying_device_busy);
1736 
1737 static int dm_lld_busy(struct request_queue *q)
1738 {
1739 	int r;
1740 	struct mapped_device *md = q->queuedata;
1741 	struct dm_table *map = dm_get_live_table_fast(md);
1742 
1743 	if (!map || test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))
1744 		r = 1;
1745 	else
1746 		r = dm_table_any_busy_target(map);
1747 
1748 	dm_put_live_table_fast(md);
1749 
1750 	return r;
1751 }
1752 
1753 static int dm_any_congested(void *congested_data, int bdi_bits)
1754 {
1755 	int r = bdi_bits;
1756 	struct mapped_device *md = congested_data;
1757 	struct dm_table *map;
1758 
1759 	if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
1760 		map = dm_get_live_table_fast(md);
1761 		if (map) {
1762 			/*
1763 			 * Request-based dm cares about only own queue for
1764 			 * the query about congestion status of request_queue
1765 			 */
1766 			if (dm_request_based(md))
1767 				r = md->queue->backing_dev_info.state &
1768 				    bdi_bits;
1769 			else
1770 				r = dm_table_any_congested(map, bdi_bits);
1771 		}
1772 		dm_put_live_table_fast(md);
1773 	}
1774 
1775 	return r;
1776 }
1777 
1778 /*-----------------------------------------------------------------
1779  * An IDR is used to keep track of allocated minor numbers.
1780  *---------------------------------------------------------------*/
1781 static void free_minor(int minor)
1782 {
1783 	spin_lock(&_minor_lock);
1784 	idr_remove(&_minor_idr, minor);
1785 	spin_unlock(&_minor_lock);
1786 }
1787 
1788 /*
1789  * See if the device with a specific minor # is free.
1790  */
1791 static int specific_minor(int minor)
1792 {
1793 	int r;
1794 
1795 	if (minor >= (1 << MINORBITS))
1796 		return -EINVAL;
1797 
1798 	idr_preload(GFP_KERNEL);
1799 	spin_lock(&_minor_lock);
1800 
1801 	r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1802 
1803 	spin_unlock(&_minor_lock);
1804 	idr_preload_end();
1805 	if (r < 0)
1806 		return r == -ENOSPC ? -EBUSY : r;
1807 	return 0;
1808 }
1809 
1810 static int next_free_minor(int *minor)
1811 {
1812 	int r;
1813 
1814 	idr_preload(GFP_KERNEL);
1815 	spin_lock(&_minor_lock);
1816 
1817 	r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1818 
1819 	spin_unlock(&_minor_lock);
1820 	idr_preload_end();
1821 	if (r < 0)
1822 		return r;
1823 	*minor = r;
1824 	return 0;
1825 }
1826 
1827 static const struct block_device_operations dm_blk_dops;
1828 
1829 static void dm_wq_work(struct work_struct *work);
1830 
1831 static void dm_init_md_queue(struct mapped_device *md)
1832 {
1833 	/*
1834 	 * Request-based dm devices cannot be stacked on top of bio-based dm
1835 	 * devices.  The type of this dm device has not been decided yet.
1836 	 * The type is decided at the first table loading time.
1837 	 * To prevent problematic device stacking, clear the queue flag
1838 	 * for request stacking support until then.
1839 	 *
1840 	 * This queue is new, so no concurrency on the queue_flags.
1841 	 */
1842 	queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
1843 
1844 	md->queue->queuedata = md;
1845 	md->queue->backing_dev_info.congested_fn = dm_any_congested;
1846 	md->queue->backing_dev_info.congested_data = md;
1847 	blk_queue_make_request(md->queue, dm_request);
1848 	blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
1849 	blk_queue_merge_bvec(md->queue, dm_merge_bvec);
1850 }
1851 
1852 /*
1853  * Allocate and initialise a blank device with a given minor.
1854  */
1855 static struct mapped_device *alloc_dev(int minor)
1856 {
1857 	int r;
1858 	struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
1859 	void *old_md;
1860 
1861 	if (!md) {
1862 		DMWARN("unable to allocate device, out of memory.");
1863 		return NULL;
1864 	}
1865 
1866 	if (!try_module_get(THIS_MODULE))
1867 		goto bad_module_get;
1868 
1869 	/* get a minor number for the dev */
1870 	if (minor == DM_ANY_MINOR)
1871 		r = next_free_minor(&minor);
1872 	else
1873 		r = specific_minor(minor);
1874 	if (r < 0)
1875 		goto bad_minor;
1876 
1877 	r = init_srcu_struct(&md->io_barrier);
1878 	if (r < 0)
1879 		goto bad_io_barrier;
1880 
1881 	md->type = DM_TYPE_NONE;
1882 	mutex_init(&md->suspend_lock);
1883 	mutex_init(&md->type_lock);
1884 	spin_lock_init(&md->deferred_lock);
1885 	atomic_set(&md->holders, 1);
1886 	atomic_set(&md->open_count, 0);
1887 	atomic_set(&md->event_nr, 0);
1888 	atomic_set(&md->uevent_seq, 0);
1889 	INIT_LIST_HEAD(&md->uevent_list);
1890 	spin_lock_init(&md->uevent_lock);
1891 
1892 	md->queue = blk_alloc_queue(GFP_KERNEL);
1893 	if (!md->queue)
1894 		goto bad_queue;
1895 
1896 	dm_init_md_queue(md);
1897 
1898 	md->disk = alloc_disk(1);
1899 	if (!md->disk)
1900 		goto bad_disk;
1901 
1902 	atomic_set(&md->pending[0], 0);
1903 	atomic_set(&md->pending[1], 0);
1904 	init_waitqueue_head(&md->wait);
1905 	INIT_WORK(&md->work, dm_wq_work);
1906 	init_waitqueue_head(&md->eventq);
1907 	init_completion(&md->kobj_holder.completion);
1908 
1909 	md->disk->major = _major;
1910 	md->disk->first_minor = minor;
1911 	md->disk->fops = &dm_blk_dops;
1912 	md->disk->queue = md->queue;
1913 	md->disk->private_data = md;
1914 	sprintf(md->disk->disk_name, "dm-%d", minor);
1915 	add_disk(md->disk);
1916 	format_dev_t(md->name, MKDEV(_major, minor));
1917 
1918 	md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
1919 	if (!md->wq)
1920 		goto bad_thread;
1921 
1922 	md->bdev = bdget_disk(md->disk, 0);
1923 	if (!md->bdev)
1924 		goto bad_bdev;
1925 
1926 	bio_init(&md->flush_bio);
1927 	md->flush_bio.bi_bdev = md->bdev;
1928 	md->flush_bio.bi_rw = WRITE_FLUSH;
1929 
1930 	dm_stats_init(&md->stats);
1931 
1932 	/* Populate the mapping, nobody knows we exist yet */
1933 	spin_lock(&_minor_lock);
1934 	old_md = idr_replace(&_minor_idr, md, minor);
1935 	spin_unlock(&_minor_lock);
1936 
1937 	BUG_ON(old_md != MINOR_ALLOCED);
1938 
1939 	return md;
1940 
1941 bad_bdev:
1942 	destroy_workqueue(md->wq);
1943 bad_thread:
1944 	del_gendisk(md->disk);
1945 	put_disk(md->disk);
1946 bad_disk:
1947 	blk_cleanup_queue(md->queue);
1948 bad_queue:
1949 	cleanup_srcu_struct(&md->io_barrier);
1950 bad_io_barrier:
1951 	free_minor(minor);
1952 bad_minor:
1953 	module_put(THIS_MODULE);
1954 bad_module_get:
1955 	kfree(md);
1956 	return NULL;
1957 }
1958 
1959 static void unlock_fs(struct mapped_device *md);
1960 
1961 static void free_dev(struct mapped_device *md)
1962 {
1963 	int minor = MINOR(disk_devt(md->disk));
1964 
1965 	unlock_fs(md);
1966 	bdput(md->bdev);
1967 	destroy_workqueue(md->wq);
1968 	if (md->io_pool)
1969 		mempool_destroy(md->io_pool);
1970 	if (md->bs)
1971 		bioset_free(md->bs);
1972 	blk_integrity_unregister(md->disk);
1973 	del_gendisk(md->disk);
1974 	cleanup_srcu_struct(&md->io_barrier);
1975 	free_minor(minor);
1976 
1977 	spin_lock(&_minor_lock);
1978 	md->disk->private_data = NULL;
1979 	spin_unlock(&_minor_lock);
1980 
1981 	put_disk(md->disk);
1982 	blk_cleanup_queue(md->queue);
1983 	dm_stats_cleanup(&md->stats);
1984 	module_put(THIS_MODULE);
1985 	kfree(md);
1986 }
1987 
1988 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
1989 {
1990 	struct dm_md_mempools *p = dm_table_get_md_mempools(t);
1991 
1992 	if (md->io_pool && md->bs) {
1993 		/* The md already has necessary mempools. */
1994 		if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
1995 			/*
1996 			 * Reload bioset because front_pad may have changed
1997 			 * because a different table was loaded.
1998 			 */
1999 			bioset_free(md->bs);
2000 			md->bs = p->bs;
2001 			p->bs = NULL;
2002 		} else if (dm_table_get_type(t) == DM_TYPE_REQUEST_BASED) {
2003 			/*
2004 			 * There's no need to reload with request-based dm
2005 			 * because the size of front_pad doesn't change.
2006 			 * Note for future: If you are to reload bioset,
2007 			 * prep-ed requests in the queue may refer
2008 			 * to bio from the old bioset, so you must walk
2009 			 * through the queue to unprep.
2010 			 */
2011 		}
2012 		goto out;
2013 	}
2014 
2015 	BUG_ON(!p || md->io_pool || md->bs);
2016 
2017 	md->io_pool = p->io_pool;
2018 	p->io_pool = NULL;
2019 	md->bs = p->bs;
2020 	p->bs = NULL;
2021 
2022 out:
2023 	/* mempool bind completed, now no need any mempools in the table */
2024 	dm_table_free_md_mempools(t);
2025 }
2026 
2027 /*
2028  * Bind a table to the device.
2029  */
2030 static void event_callback(void *context)
2031 {
2032 	unsigned long flags;
2033 	LIST_HEAD(uevents);
2034 	struct mapped_device *md = (struct mapped_device *) context;
2035 
2036 	spin_lock_irqsave(&md->uevent_lock, flags);
2037 	list_splice_init(&md->uevent_list, &uevents);
2038 	spin_unlock_irqrestore(&md->uevent_lock, flags);
2039 
2040 	dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2041 
2042 	atomic_inc(&md->event_nr);
2043 	wake_up(&md->eventq);
2044 }
2045 
2046 /*
2047  * Protected by md->suspend_lock obtained by dm_swap_table().
2048  */
2049 static void __set_size(struct mapped_device *md, sector_t size)
2050 {
2051 	set_capacity(md->disk, size);
2052 
2053 	i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2054 }
2055 
2056 /*
2057  * Return 1 if the queue has a compulsory merge_bvec_fn function.
2058  *
2059  * If this function returns 0, then the device is either a non-dm
2060  * device without a merge_bvec_fn, or it is a dm device that is
2061  * able to split any bios it receives that are too big.
2062  */
2063 int dm_queue_merge_is_compulsory(struct request_queue *q)
2064 {
2065 	struct mapped_device *dev_md;
2066 
2067 	if (!q->merge_bvec_fn)
2068 		return 0;
2069 
2070 	if (q->make_request_fn == dm_request) {
2071 		dev_md = q->queuedata;
2072 		if (test_bit(DMF_MERGE_IS_OPTIONAL, &dev_md->flags))
2073 			return 0;
2074 	}
2075 
2076 	return 1;
2077 }
2078 
2079 static int dm_device_merge_is_compulsory(struct dm_target *ti,
2080 					 struct dm_dev *dev, sector_t start,
2081 					 sector_t len, void *data)
2082 {
2083 	struct block_device *bdev = dev->bdev;
2084 	struct request_queue *q = bdev_get_queue(bdev);
2085 
2086 	return dm_queue_merge_is_compulsory(q);
2087 }
2088 
2089 /*
2090  * Return 1 if it is acceptable to ignore merge_bvec_fn based
2091  * on the properties of the underlying devices.
2092  */
2093 static int dm_table_merge_is_optional(struct dm_table *table)
2094 {
2095 	unsigned i = 0;
2096 	struct dm_target *ti;
2097 
2098 	while (i < dm_table_get_num_targets(table)) {
2099 		ti = dm_table_get_target(table, i++);
2100 
2101 		if (ti->type->iterate_devices &&
2102 		    ti->type->iterate_devices(ti, dm_device_merge_is_compulsory, NULL))
2103 			return 0;
2104 	}
2105 
2106 	return 1;
2107 }
2108 
2109 /*
2110  * Returns old map, which caller must destroy.
2111  */
2112 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2113 			       struct queue_limits *limits)
2114 {
2115 	struct dm_table *old_map;
2116 	struct request_queue *q = md->queue;
2117 	sector_t size;
2118 	int merge_is_optional;
2119 
2120 	size = dm_table_get_size(t);
2121 
2122 	/*
2123 	 * Wipe any geometry if the size of the table changed.
2124 	 */
2125 	if (size != dm_get_size(md))
2126 		memset(&md->geometry, 0, sizeof(md->geometry));
2127 
2128 	__set_size(md, size);
2129 
2130 	dm_table_event_callback(t, event_callback, md);
2131 
2132 	/*
2133 	 * The queue hasn't been stopped yet, if the old table type wasn't
2134 	 * for request-based during suspension.  So stop it to prevent
2135 	 * I/O mapping before resume.
2136 	 * This must be done before setting the queue restrictions,
2137 	 * because request-based dm may be run just after the setting.
2138 	 */
2139 	if (dm_table_request_based(t) && !blk_queue_stopped(q))
2140 		stop_queue(q);
2141 
2142 	__bind_mempools(md, t);
2143 
2144 	merge_is_optional = dm_table_merge_is_optional(t);
2145 
2146 	old_map = md->map;
2147 	rcu_assign_pointer(md->map, t);
2148 	md->immutable_target_type = dm_table_get_immutable_target_type(t);
2149 
2150 	dm_table_set_restrictions(t, q, limits);
2151 	if (merge_is_optional)
2152 		set_bit(DMF_MERGE_IS_OPTIONAL, &md->flags);
2153 	else
2154 		clear_bit(DMF_MERGE_IS_OPTIONAL, &md->flags);
2155 	dm_sync_table(md);
2156 
2157 	return old_map;
2158 }
2159 
2160 /*
2161  * Returns unbound table for the caller to free.
2162  */
2163 static struct dm_table *__unbind(struct mapped_device *md)
2164 {
2165 	struct dm_table *map = md->map;
2166 
2167 	if (!map)
2168 		return NULL;
2169 
2170 	dm_table_event_callback(map, NULL, NULL);
2171 	RCU_INIT_POINTER(md->map, NULL);
2172 	dm_sync_table(md);
2173 
2174 	return map;
2175 }
2176 
2177 /*
2178  * Constructor for a new device.
2179  */
2180 int dm_create(int minor, struct mapped_device **result)
2181 {
2182 	struct mapped_device *md;
2183 
2184 	md = alloc_dev(minor);
2185 	if (!md)
2186 		return -ENXIO;
2187 
2188 	dm_sysfs_init(md);
2189 
2190 	*result = md;
2191 	return 0;
2192 }
2193 
2194 /*
2195  * Functions to manage md->type.
2196  * All are required to hold md->type_lock.
2197  */
2198 void dm_lock_md_type(struct mapped_device *md)
2199 {
2200 	mutex_lock(&md->type_lock);
2201 }
2202 
2203 void dm_unlock_md_type(struct mapped_device *md)
2204 {
2205 	mutex_unlock(&md->type_lock);
2206 }
2207 
2208 void dm_set_md_type(struct mapped_device *md, unsigned type)
2209 {
2210 	BUG_ON(!mutex_is_locked(&md->type_lock));
2211 	md->type = type;
2212 }
2213 
2214 unsigned dm_get_md_type(struct mapped_device *md)
2215 {
2216 	BUG_ON(!mutex_is_locked(&md->type_lock));
2217 	return md->type;
2218 }
2219 
2220 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2221 {
2222 	return md->immutable_target_type;
2223 }
2224 
2225 /*
2226  * The queue_limits are only valid as long as you have a reference
2227  * count on 'md'.
2228  */
2229 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2230 {
2231 	BUG_ON(!atomic_read(&md->holders));
2232 	return &md->queue->limits;
2233 }
2234 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2235 
2236 /*
2237  * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2238  */
2239 static int dm_init_request_based_queue(struct mapped_device *md)
2240 {
2241 	struct request_queue *q = NULL;
2242 
2243 	if (md->queue->elevator)
2244 		return 1;
2245 
2246 	/* Fully initialize the queue */
2247 	q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2248 	if (!q)
2249 		return 0;
2250 
2251 	md->queue = q;
2252 	dm_init_md_queue(md);
2253 	blk_queue_softirq_done(md->queue, dm_softirq_done);
2254 	blk_queue_prep_rq(md->queue, dm_prep_fn);
2255 	blk_queue_lld_busy(md->queue, dm_lld_busy);
2256 
2257 	elv_register_queue(md->queue);
2258 
2259 	return 1;
2260 }
2261 
2262 /*
2263  * Setup the DM device's queue based on md's type
2264  */
2265 int dm_setup_md_queue(struct mapped_device *md)
2266 {
2267 	if ((dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) &&
2268 	    !dm_init_request_based_queue(md)) {
2269 		DMWARN("Cannot initialize queue for request-based mapped device");
2270 		return -EINVAL;
2271 	}
2272 
2273 	return 0;
2274 }
2275 
2276 static struct mapped_device *dm_find_md(dev_t dev)
2277 {
2278 	struct mapped_device *md;
2279 	unsigned minor = MINOR(dev);
2280 
2281 	if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2282 		return NULL;
2283 
2284 	spin_lock(&_minor_lock);
2285 
2286 	md = idr_find(&_minor_idr, minor);
2287 	if (md && (md == MINOR_ALLOCED ||
2288 		   (MINOR(disk_devt(dm_disk(md))) != minor) ||
2289 		   dm_deleting_md(md) ||
2290 		   test_bit(DMF_FREEING, &md->flags))) {
2291 		md = NULL;
2292 		goto out;
2293 	}
2294 
2295 out:
2296 	spin_unlock(&_minor_lock);
2297 
2298 	return md;
2299 }
2300 
2301 struct mapped_device *dm_get_md(dev_t dev)
2302 {
2303 	struct mapped_device *md = dm_find_md(dev);
2304 
2305 	if (md)
2306 		dm_get(md);
2307 
2308 	return md;
2309 }
2310 EXPORT_SYMBOL_GPL(dm_get_md);
2311 
2312 void *dm_get_mdptr(struct mapped_device *md)
2313 {
2314 	return md->interface_ptr;
2315 }
2316 
2317 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2318 {
2319 	md->interface_ptr = ptr;
2320 }
2321 
2322 void dm_get(struct mapped_device *md)
2323 {
2324 	atomic_inc(&md->holders);
2325 	BUG_ON(test_bit(DMF_FREEING, &md->flags));
2326 }
2327 
2328 const char *dm_device_name(struct mapped_device *md)
2329 {
2330 	return md->name;
2331 }
2332 EXPORT_SYMBOL_GPL(dm_device_name);
2333 
2334 static void __dm_destroy(struct mapped_device *md, bool wait)
2335 {
2336 	struct dm_table *map;
2337 	int srcu_idx;
2338 
2339 	might_sleep();
2340 
2341 	spin_lock(&_minor_lock);
2342 	map = dm_get_live_table(md, &srcu_idx);
2343 	idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2344 	set_bit(DMF_FREEING, &md->flags);
2345 	spin_unlock(&_minor_lock);
2346 
2347 	if (!dm_suspended_md(md)) {
2348 		dm_table_presuspend_targets(map);
2349 		dm_table_postsuspend_targets(map);
2350 	}
2351 
2352 	/* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2353 	dm_put_live_table(md, srcu_idx);
2354 
2355 	/*
2356 	 * Rare, but there may be I/O requests still going to complete,
2357 	 * for example.  Wait for all references to disappear.
2358 	 * No one should increment the reference count of the mapped_device,
2359 	 * after the mapped_device state becomes DMF_FREEING.
2360 	 */
2361 	if (wait)
2362 		while (atomic_read(&md->holders))
2363 			msleep(1);
2364 	else if (atomic_read(&md->holders))
2365 		DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2366 		       dm_device_name(md), atomic_read(&md->holders));
2367 
2368 	dm_sysfs_exit(md);
2369 	dm_table_destroy(__unbind(md));
2370 	free_dev(md);
2371 }
2372 
2373 void dm_destroy(struct mapped_device *md)
2374 {
2375 	__dm_destroy(md, true);
2376 }
2377 
2378 void dm_destroy_immediate(struct mapped_device *md)
2379 {
2380 	__dm_destroy(md, false);
2381 }
2382 
2383 void dm_put(struct mapped_device *md)
2384 {
2385 	atomic_dec(&md->holders);
2386 }
2387 EXPORT_SYMBOL_GPL(dm_put);
2388 
2389 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2390 {
2391 	int r = 0;
2392 	DECLARE_WAITQUEUE(wait, current);
2393 
2394 	add_wait_queue(&md->wait, &wait);
2395 
2396 	while (1) {
2397 		set_current_state(interruptible);
2398 
2399 		if (!md_in_flight(md))
2400 			break;
2401 
2402 		if (interruptible == TASK_INTERRUPTIBLE &&
2403 		    signal_pending(current)) {
2404 			r = -EINTR;
2405 			break;
2406 		}
2407 
2408 		io_schedule();
2409 	}
2410 	set_current_state(TASK_RUNNING);
2411 
2412 	remove_wait_queue(&md->wait, &wait);
2413 
2414 	return r;
2415 }
2416 
2417 /*
2418  * Process the deferred bios
2419  */
2420 static void dm_wq_work(struct work_struct *work)
2421 {
2422 	struct mapped_device *md = container_of(work, struct mapped_device,
2423 						work);
2424 	struct bio *c;
2425 	int srcu_idx;
2426 	struct dm_table *map;
2427 
2428 	map = dm_get_live_table(md, &srcu_idx);
2429 
2430 	while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2431 		spin_lock_irq(&md->deferred_lock);
2432 		c = bio_list_pop(&md->deferred);
2433 		spin_unlock_irq(&md->deferred_lock);
2434 
2435 		if (!c)
2436 			break;
2437 
2438 		if (dm_request_based(md))
2439 			generic_make_request(c);
2440 		else
2441 			__split_and_process_bio(md, map, c);
2442 	}
2443 
2444 	dm_put_live_table(md, srcu_idx);
2445 }
2446 
2447 static void dm_queue_flush(struct mapped_device *md)
2448 {
2449 	clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2450 	smp_mb__after_clear_bit();
2451 	queue_work(md->wq, &md->work);
2452 }
2453 
2454 /*
2455  * Swap in a new table, returning the old one for the caller to destroy.
2456  */
2457 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2458 {
2459 	struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2460 	struct queue_limits limits;
2461 	int r;
2462 
2463 	mutex_lock(&md->suspend_lock);
2464 
2465 	/* device must be suspended */
2466 	if (!dm_suspended_md(md))
2467 		goto out;
2468 
2469 	/*
2470 	 * If the new table has no data devices, retain the existing limits.
2471 	 * This helps multipath with queue_if_no_path if all paths disappear,
2472 	 * then new I/O is queued based on these limits, and then some paths
2473 	 * reappear.
2474 	 */
2475 	if (dm_table_has_no_data_devices(table)) {
2476 		live_map = dm_get_live_table_fast(md);
2477 		if (live_map)
2478 			limits = md->queue->limits;
2479 		dm_put_live_table_fast(md);
2480 	}
2481 
2482 	if (!live_map) {
2483 		r = dm_calculate_queue_limits(table, &limits);
2484 		if (r) {
2485 			map = ERR_PTR(r);
2486 			goto out;
2487 		}
2488 	}
2489 
2490 	map = __bind(md, table, &limits);
2491 
2492 out:
2493 	mutex_unlock(&md->suspend_lock);
2494 	return map;
2495 }
2496 
2497 /*
2498  * Functions to lock and unlock any filesystem running on the
2499  * device.
2500  */
2501 static int lock_fs(struct mapped_device *md)
2502 {
2503 	int r;
2504 
2505 	WARN_ON(md->frozen_sb);
2506 
2507 	md->frozen_sb = freeze_bdev(md->bdev);
2508 	if (IS_ERR(md->frozen_sb)) {
2509 		r = PTR_ERR(md->frozen_sb);
2510 		md->frozen_sb = NULL;
2511 		return r;
2512 	}
2513 
2514 	set_bit(DMF_FROZEN, &md->flags);
2515 
2516 	return 0;
2517 }
2518 
2519 static void unlock_fs(struct mapped_device *md)
2520 {
2521 	if (!test_bit(DMF_FROZEN, &md->flags))
2522 		return;
2523 
2524 	thaw_bdev(md->bdev, md->frozen_sb);
2525 	md->frozen_sb = NULL;
2526 	clear_bit(DMF_FROZEN, &md->flags);
2527 }
2528 
2529 /*
2530  * We need to be able to change a mapping table under a mounted
2531  * filesystem.  For example we might want to move some data in
2532  * the background.  Before the table can be swapped with
2533  * dm_bind_table, dm_suspend must be called to flush any in
2534  * flight bios and ensure that any further io gets deferred.
2535  */
2536 /*
2537  * Suspend mechanism in request-based dm.
2538  *
2539  * 1. Flush all I/Os by lock_fs() if needed.
2540  * 2. Stop dispatching any I/O by stopping the request_queue.
2541  * 3. Wait for all in-flight I/Os to be completed or requeued.
2542  *
2543  * To abort suspend, start the request_queue.
2544  */
2545 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
2546 {
2547 	struct dm_table *map = NULL;
2548 	int r = 0;
2549 	int do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG ? 1 : 0;
2550 	int noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG ? 1 : 0;
2551 
2552 	mutex_lock(&md->suspend_lock);
2553 
2554 	if (dm_suspended_md(md)) {
2555 		r = -EINVAL;
2556 		goto out_unlock;
2557 	}
2558 
2559 	map = md->map;
2560 
2561 	/*
2562 	 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2563 	 * This flag is cleared before dm_suspend returns.
2564 	 */
2565 	if (noflush)
2566 		set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2567 
2568 	/* This does not get reverted if there's an error later. */
2569 	dm_table_presuspend_targets(map);
2570 
2571 	/*
2572 	 * Flush I/O to the device.
2573 	 * Any I/O submitted after lock_fs() may not be flushed.
2574 	 * noflush takes precedence over do_lockfs.
2575 	 * (lock_fs() flushes I/Os and waits for them to complete.)
2576 	 */
2577 	if (!noflush && do_lockfs) {
2578 		r = lock_fs(md);
2579 		if (r)
2580 			goto out_unlock;
2581 	}
2582 
2583 	/*
2584 	 * Here we must make sure that no processes are submitting requests
2585 	 * to target drivers i.e. no one may be executing
2586 	 * __split_and_process_bio. This is called from dm_request and
2587 	 * dm_wq_work.
2588 	 *
2589 	 * To get all processes out of __split_and_process_bio in dm_request,
2590 	 * we take the write lock. To prevent any process from reentering
2591 	 * __split_and_process_bio from dm_request and quiesce the thread
2592 	 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
2593 	 * flush_workqueue(md->wq).
2594 	 */
2595 	set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2596 	synchronize_srcu(&md->io_barrier);
2597 
2598 	/*
2599 	 * Stop md->queue before flushing md->wq in case request-based
2600 	 * dm defers requests to md->wq from md->queue.
2601 	 */
2602 	if (dm_request_based(md))
2603 		stop_queue(md->queue);
2604 
2605 	flush_workqueue(md->wq);
2606 
2607 	/*
2608 	 * At this point no more requests are entering target request routines.
2609 	 * We call dm_wait_for_completion to wait for all existing requests
2610 	 * to finish.
2611 	 */
2612 	r = dm_wait_for_completion(md, TASK_INTERRUPTIBLE);
2613 
2614 	if (noflush)
2615 		clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2616 	synchronize_srcu(&md->io_barrier);
2617 
2618 	/* were we interrupted ? */
2619 	if (r < 0) {
2620 		dm_queue_flush(md);
2621 
2622 		if (dm_request_based(md))
2623 			start_queue(md->queue);
2624 
2625 		unlock_fs(md);
2626 		goto out_unlock; /* pushback list is already flushed, so skip flush */
2627 	}
2628 
2629 	/*
2630 	 * If dm_wait_for_completion returned 0, the device is completely
2631 	 * quiescent now. There is no request-processing activity. All new
2632 	 * requests are being added to md->deferred list.
2633 	 */
2634 
2635 	set_bit(DMF_SUSPENDED, &md->flags);
2636 
2637 	dm_table_postsuspend_targets(map);
2638 
2639 out_unlock:
2640 	mutex_unlock(&md->suspend_lock);
2641 	return r;
2642 }
2643 
2644 int dm_resume(struct mapped_device *md)
2645 {
2646 	int r = -EINVAL;
2647 	struct dm_table *map = NULL;
2648 
2649 	mutex_lock(&md->suspend_lock);
2650 	if (!dm_suspended_md(md))
2651 		goto out;
2652 
2653 	map = md->map;
2654 	if (!map || !dm_table_get_size(map))
2655 		goto out;
2656 
2657 	r = dm_table_resume_targets(map);
2658 	if (r)
2659 		goto out;
2660 
2661 	dm_queue_flush(md);
2662 
2663 	/*
2664 	 * Flushing deferred I/Os must be done after targets are resumed
2665 	 * so that mapping of targets can work correctly.
2666 	 * Request-based dm is queueing the deferred I/Os in its request_queue.
2667 	 */
2668 	if (dm_request_based(md))
2669 		start_queue(md->queue);
2670 
2671 	unlock_fs(md);
2672 
2673 	clear_bit(DMF_SUSPENDED, &md->flags);
2674 
2675 	r = 0;
2676 out:
2677 	mutex_unlock(&md->suspend_lock);
2678 
2679 	return r;
2680 }
2681 
2682 /*
2683  * Internal suspend/resume works like userspace-driven suspend. It waits
2684  * until all bios finish and prevents issuing new bios to the target drivers.
2685  * It may be used only from the kernel.
2686  *
2687  * Internal suspend holds md->suspend_lock, which prevents interaction with
2688  * userspace-driven suspend.
2689  */
2690 
2691 void dm_internal_suspend(struct mapped_device *md)
2692 {
2693 	mutex_lock(&md->suspend_lock);
2694 	if (dm_suspended_md(md))
2695 		return;
2696 
2697 	set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2698 	synchronize_srcu(&md->io_barrier);
2699 	flush_workqueue(md->wq);
2700 	dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
2701 }
2702 
2703 void dm_internal_resume(struct mapped_device *md)
2704 {
2705 	if (dm_suspended_md(md))
2706 		goto done;
2707 
2708 	dm_queue_flush(md);
2709 
2710 done:
2711 	mutex_unlock(&md->suspend_lock);
2712 }
2713 
2714 /*-----------------------------------------------------------------
2715  * Event notification.
2716  *---------------------------------------------------------------*/
2717 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
2718 		       unsigned cookie)
2719 {
2720 	char udev_cookie[DM_COOKIE_LENGTH];
2721 	char *envp[] = { udev_cookie, NULL };
2722 
2723 	if (!cookie)
2724 		return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
2725 	else {
2726 		snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
2727 			 DM_COOKIE_ENV_VAR_NAME, cookie);
2728 		return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
2729 					  action, envp);
2730 	}
2731 }
2732 
2733 uint32_t dm_next_uevent_seq(struct mapped_device *md)
2734 {
2735 	return atomic_add_return(1, &md->uevent_seq);
2736 }
2737 
2738 uint32_t dm_get_event_nr(struct mapped_device *md)
2739 {
2740 	return atomic_read(&md->event_nr);
2741 }
2742 
2743 int dm_wait_event(struct mapped_device *md, int event_nr)
2744 {
2745 	return wait_event_interruptible(md->eventq,
2746 			(event_nr != atomic_read(&md->event_nr)));
2747 }
2748 
2749 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
2750 {
2751 	unsigned long flags;
2752 
2753 	spin_lock_irqsave(&md->uevent_lock, flags);
2754 	list_add(elist, &md->uevent_list);
2755 	spin_unlock_irqrestore(&md->uevent_lock, flags);
2756 }
2757 
2758 /*
2759  * The gendisk is only valid as long as you have a reference
2760  * count on 'md'.
2761  */
2762 struct gendisk *dm_disk(struct mapped_device *md)
2763 {
2764 	return md->disk;
2765 }
2766 
2767 struct kobject *dm_kobject(struct mapped_device *md)
2768 {
2769 	return &md->kobj_holder.kobj;
2770 }
2771 
2772 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
2773 {
2774 	struct mapped_device *md;
2775 
2776 	md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
2777 
2778 	if (test_bit(DMF_FREEING, &md->flags) ||
2779 	    dm_deleting_md(md))
2780 		return NULL;
2781 
2782 	dm_get(md);
2783 	return md;
2784 }
2785 
2786 int dm_suspended_md(struct mapped_device *md)
2787 {
2788 	return test_bit(DMF_SUSPENDED, &md->flags);
2789 }
2790 
2791 int dm_test_deferred_remove_flag(struct mapped_device *md)
2792 {
2793 	return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
2794 }
2795 
2796 int dm_suspended(struct dm_target *ti)
2797 {
2798 	return dm_suspended_md(dm_table_get_md(ti->table));
2799 }
2800 EXPORT_SYMBOL_GPL(dm_suspended);
2801 
2802 int dm_noflush_suspending(struct dm_target *ti)
2803 {
2804 	return __noflush_suspending(dm_table_get_md(ti->table));
2805 }
2806 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
2807 
2808 struct dm_md_mempools *dm_alloc_md_mempools(unsigned type, unsigned integrity, unsigned per_bio_data_size)
2809 {
2810 	struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
2811 	struct kmem_cache *cachep;
2812 	unsigned int pool_size;
2813 	unsigned int front_pad;
2814 
2815 	if (!pools)
2816 		return NULL;
2817 
2818 	if (type == DM_TYPE_BIO_BASED) {
2819 		cachep = _io_cache;
2820 		pool_size = dm_get_reserved_bio_based_ios();
2821 		front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
2822 	} else if (type == DM_TYPE_REQUEST_BASED) {
2823 		cachep = _rq_tio_cache;
2824 		pool_size = dm_get_reserved_rq_based_ios();
2825 		front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
2826 		/* per_bio_data_size is not used. See __bind_mempools(). */
2827 		WARN_ON(per_bio_data_size != 0);
2828 	} else
2829 		goto out;
2830 
2831 	pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
2832 	if (!pools->io_pool)
2833 		goto out;
2834 
2835 	pools->bs = bioset_create(pool_size, front_pad);
2836 	if (!pools->bs)
2837 		goto out;
2838 
2839 	if (integrity && bioset_integrity_create(pools->bs, pool_size))
2840 		goto out;
2841 
2842 	return pools;
2843 
2844 out:
2845 	dm_free_md_mempools(pools);
2846 
2847 	return NULL;
2848 }
2849 
2850 void dm_free_md_mempools(struct dm_md_mempools *pools)
2851 {
2852 	if (!pools)
2853 		return;
2854 
2855 	if (pools->io_pool)
2856 		mempool_destroy(pools->io_pool);
2857 
2858 	if (pools->bs)
2859 		bioset_free(pools->bs);
2860 
2861 	kfree(pools);
2862 }
2863 
2864 static const struct block_device_operations dm_blk_dops = {
2865 	.open = dm_blk_open,
2866 	.release = dm_blk_close,
2867 	.ioctl = dm_blk_ioctl,
2868 	.getgeo = dm_blk_getgeo,
2869 	.owner = THIS_MODULE
2870 };
2871 
2872 /*
2873  * module hooks
2874  */
2875 module_init(dm_init);
2876 module_exit(dm_exit);
2877 
2878 module_param(major, uint, 0);
2879 MODULE_PARM_DESC(major, "The major number of the device mapper");
2880 
2881 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
2882 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
2883 
2884 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
2885 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
2886 
2887 MODULE_DESCRIPTION(DM_NAME " driver");
2888 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2889 MODULE_LICENSE("GPL");
2890