xref: /openbmc/linux/drivers/md/dm-thin.c (revision 236bdb88)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011-2012 Red Hat UK.
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include "dm-thin-metadata.h"
9 #include "dm-bio-prison-v1.h"
10 #include "dm.h"
11 
12 #include <linux/device-mapper.h>
13 #include <linux/dm-io.h>
14 #include <linux/dm-kcopyd.h>
15 #include <linux/jiffies.h>
16 #include <linux/log2.h>
17 #include <linux/list.h>
18 #include <linux/rculist.h>
19 #include <linux/init.h>
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sort.h>
24 #include <linux/rbtree.h>
25 
26 #define	DM_MSG_PREFIX	"thin"
27 
28 /*
29  * Tunable constants
30  */
31 #define ENDIO_HOOK_POOL_SIZE 1024
32 #define MAPPING_POOL_SIZE 1024
33 #define COMMIT_PERIOD HZ
34 #define NO_SPACE_TIMEOUT_SECS 60
35 
36 static unsigned int no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
37 
38 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
39 		"A percentage of time allocated for copy on write");
40 
41 /*
42  * The block size of the device holding pool data must be
43  * between 64KB and 1GB.
44  */
45 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
46 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
47 
48 /*
49  * Device id is restricted to 24 bits.
50  */
51 #define MAX_DEV_ID ((1 << 24) - 1)
52 
53 /*
54  * How do we handle breaking sharing of data blocks?
55  * =================================================
56  *
57  * We use a standard copy-on-write btree to store the mappings for the
58  * devices (note I'm talking about copy-on-write of the metadata here, not
59  * the data).  When you take an internal snapshot you clone the root node
60  * of the origin btree.  After this there is no concept of an origin or a
61  * snapshot.  They are just two device trees that happen to point to the
62  * same data blocks.
63  *
64  * When we get a write in we decide if it's to a shared data block using
65  * some timestamp magic.  If it is, we have to break sharing.
66  *
67  * Let's say we write to a shared block in what was the origin.  The
68  * steps are:
69  *
70  * i) plug io further to this physical block. (see bio_prison code).
71  *
72  * ii) quiesce any read io to that shared data block.  Obviously
73  * including all devices that share this block.  (see dm_deferred_set code)
74  *
75  * iii) copy the data block to a newly allocate block.  This step can be
76  * missed out if the io covers the block. (schedule_copy).
77  *
78  * iv) insert the new mapping into the origin's btree
79  * (process_prepared_mapping).  This act of inserting breaks some
80  * sharing of btree nodes between the two devices.  Breaking sharing only
81  * effects the btree of that specific device.  Btrees for the other
82  * devices that share the block never change.  The btree for the origin
83  * device as it was after the last commit is untouched, ie. we're using
84  * persistent data structures in the functional programming sense.
85  *
86  * v) unplug io to this physical block, including the io that triggered
87  * the breaking of sharing.
88  *
89  * Steps (ii) and (iii) occur in parallel.
90  *
91  * The metadata _doesn't_ need to be committed before the io continues.  We
92  * get away with this because the io is always written to a _new_ block.
93  * If there's a crash, then:
94  *
95  * - The origin mapping will point to the old origin block (the shared
96  * one).  This will contain the data as it was before the io that triggered
97  * the breaking of sharing came in.
98  *
99  * - The snap mapping still points to the old block.  As it would after
100  * the commit.
101  *
102  * The downside of this scheme is the timestamp magic isn't perfect, and
103  * will continue to think that data block in the snapshot device is shared
104  * even after the write to the origin has broken sharing.  I suspect data
105  * blocks will typically be shared by many different devices, so we're
106  * breaking sharing n + 1 times, rather than n, where n is the number of
107  * devices that reference this data block.  At the moment I think the
108  * benefits far, far outweigh the disadvantages.
109  */
110 
111 /*----------------------------------------------------------------*/
112 
113 /*
114  * Key building.
115  */
116 enum lock_space {
117 	VIRTUAL,
118 	PHYSICAL
119 };
120 
121 static void build_key(struct dm_thin_device *td, enum lock_space ls,
122 		      dm_block_t b, dm_block_t e, struct dm_cell_key *key)
123 {
124 	key->virtual = (ls == VIRTUAL);
125 	key->dev = dm_thin_dev_id(td);
126 	key->block_begin = b;
127 	key->block_end = e;
128 }
129 
130 static void build_data_key(struct dm_thin_device *td, dm_block_t b,
131 			   struct dm_cell_key *key)
132 {
133 	build_key(td, PHYSICAL, b, b + 1llu, key);
134 }
135 
136 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
137 			      struct dm_cell_key *key)
138 {
139 	build_key(td, VIRTUAL, b, b + 1llu, key);
140 }
141 
142 /*----------------------------------------------------------------*/
143 
144 #define THROTTLE_THRESHOLD (1 * HZ)
145 
146 struct throttle {
147 	struct rw_semaphore lock;
148 	unsigned long threshold;
149 	bool throttle_applied;
150 };
151 
152 static void throttle_init(struct throttle *t)
153 {
154 	init_rwsem(&t->lock);
155 	t->throttle_applied = false;
156 }
157 
158 static void throttle_work_start(struct throttle *t)
159 {
160 	t->threshold = jiffies + THROTTLE_THRESHOLD;
161 }
162 
163 static void throttle_work_update(struct throttle *t)
164 {
165 	if (!t->throttle_applied && time_is_before_jiffies(t->threshold)) {
166 		down_write(&t->lock);
167 		t->throttle_applied = true;
168 	}
169 }
170 
171 static void throttle_work_complete(struct throttle *t)
172 {
173 	if (t->throttle_applied) {
174 		t->throttle_applied = false;
175 		up_write(&t->lock);
176 	}
177 }
178 
179 static void throttle_lock(struct throttle *t)
180 {
181 	down_read(&t->lock);
182 }
183 
184 static void throttle_unlock(struct throttle *t)
185 {
186 	up_read(&t->lock);
187 }
188 
189 /*----------------------------------------------------------------*/
190 
191 /*
192  * A pool device ties together a metadata device and a data device.  It
193  * also provides the interface for creating and destroying internal
194  * devices.
195  */
196 struct dm_thin_new_mapping;
197 
198 /*
199  * The pool runs in various modes.  Ordered in degraded order for comparisons.
200  */
201 enum pool_mode {
202 	PM_WRITE,		/* metadata may be changed */
203 	PM_OUT_OF_DATA_SPACE,	/* metadata may be changed, though data may not be allocated */
204 
205 	/*
206 	 * Like READ_ONLY, except may switch back to WRITE on metadata resize. Reported as READ_ONLY.
207 	 */
208 	PM_OUT_OF_METADATA_SPACE,
209 	PM_READ_ONLY,		/* metadata may not be changed */
210 
211 	PM_FAIL,		/* all I/O fails */
212 };
213 
214 struct pool_features {
215 	enum pool_mode mode;
216 
217 	bool zero_new_blocks:1;
218 	bool discard_enabled:1;
219 	bool discard_passdown:1;
220 	bool error_if_no_space:1;
221 };
222 
223 struct thin_c;
224 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
225 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
226 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
227 
228 #define CELL_SORT_ARRAY_SIZE 8192
229 
230 struct pool {
231 	struct list_head list;
232 	struct dm_target *ti;	/* Only set if a pool target is bound */
233 
234 	struct mapped_device *pool_md;
235 	struct block_device *data_dev;
236 	struct block_device *md_dev;
237 	struct dm_pool_metadata *pmd;
238 
239 	dm_block_t low_water_blocks;
240 	uint32_t sectors_per_block;
241 	int sectors_per_block_shift;
242 
243 	struct pool_features pf;
244 	bool low_water_triggered:1;	/* A dm event has been sent */
245 	bool suspended:1;
246 	bool out_of_data_space:1;
247 
248 	struct dm_bio_prison *prison;
249 	struct dm_kcopyd_client *copier;
250 
251 	struct work_struct worker;
252 	struct workqueue_struct *wq;
253 	struct throttle throttle;
254 	struct delayed_work waker;
255 	struct delayed_work no_space_timeout;
256 
257 	unsigned long last_commit_jiffies;
258 	unsigned int ref_count;
259 
260 	spinlock_t lock;
261 	struct bio_list deferred_flush_bios;
262 	struct bio_list deferred_flush_completions;
263 	struct list_head prepared_mappings;
264 	struct list_head prepared_discards;
265 	struct list_head prepared_discards_pt2;
266 	struct list_head active_thins;
267 
268 	struct dm_deferred_set *shared_read_ds;
269 	struct dm_deferred_set *all_io_ds;
270 
271 	struct dm_thin_new_mapping *next_mapping;
272 
273 	process_bio_fn process_bio;
274 	process_bio_fn process_discard;
275 
276 	process_cell_fn process_cell;
277 	process_cell_fn process_discard_cell;
278 
279 	process_mapping_fn process_prepared_mapping;
280 	process_mapping_fn process_prepared_discard;
281 	process_mapping_fn process_prepared_discard_pt2;
282 
283 	struct dm_bio_prison_cell **cell_sort_array;
284 
285 	mempool_t mapping_pool;
286 };
287 
288 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
289 
290 static enum pool_mode get_pool_mode(struct pool *pool)
291 {
292 	return pool->pf.mode;
293 }
294 
295 static void notify_of_pool_mode_change(struct pool *pool)
296 {
297 	static const char *descs[] = {
298 		"write",
299 		"out-of-data-space",
300 		"read-only",
301 		"read-only",
302 		"fail"
303 	};
304 	const char *extra_desc = NULL;
305 	enum pool_mode mode = get_pool_mode(pool);
306 
307 	if (mode == PM_OUT_OF_DATA_SPACE) {
308 		if (!pool->pf.error_if_no_space)
309 			extra_desc = " (queue IO)";
310 		else
311 			extra_desc = " (error IO)";
312 	}
313 
314 	dm_table_event(pool->ti->table);
315 	DMINFO("%s: switching pool to %s%s mode",
316 	       dm_device_name(pool->pool_md),
317 	       descs[(int)mode], extra_desc ? : "");
318 }
319 
320 /*
321  * Target context for a pool.
322  */
323 struct pool_c {
324 	struct dm_target *ti;
325 	struct pool *pool;
326 	struct dm_dev *data_dev;
327 	struct dm_dev *metadata_dev;
328 
329 	dm_block_t low_water_blocks;
330 	struct pool_features requested_pf; /* Features requested during table load */
331 	struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
332 };
333 
334 /*
335  * Target context for a thin.
336  */
337 struct thin_c {
338 	struct list_head list;
339 	struct dm_dev *pool_dev;
340 	struct dm_dev *origin_dev;
341 	sector_t origin_size;
342 	dm_thin_id dev_id;
343 
344 	struct pool *pool;
345 	struct dm_thin_device *td;
346 	struct mapped_device *thin_md;
347 
348 	bool requeue_mode:1;
349 	spinlock_t lock;
350 	struct list_head deferred_cells;
351 	struct bio_list deferred_bio_list;
352 	struct bio_list retry_on_resume_list;
353 	struct rb_root sort_bio_list; /* sorted list of deferred bios */
354 
355 	/*
356 	 * Ensures the thin is not destroyed until the worker has finished
357 	 * iterating the active_thins list.
358 	 */
359 	refcount_t refcount;
360 	struct completion can_destroy;
361 };
362 
363 /*----------------------------------------------------------------*/
364 
365 static bool block_size_is_power_of_two(struct pool *pool)
366 {
367 	return pool->sectors_per_block_shift >= 0;
368 }
369 
370 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
371 {
372 	return block_size_is_power_of_two(pool) ?
373 		(b << pool->sectors_per_block_shift) :
374 		(b * pool->sectors_per_block);
375 }
376 
377 /*----------------------------------------------------------------*/
378 
379 struct discard_op {
380 	struct thin_c *tc;
381 	struct blk_plug plug;
382 	struct bio *parent_bio;
383 	struct bio *bio;
384 };
385 
386 static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
387 {
388 	BUG_ON(!parent);
389 
390 	op->tc = tc;
391 	blk_start_plug(&op->plug);
392 	op->parent_bio = parent;
393 	op->bio = NULL;
394 }
395 
396 static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
397 {
398 	struct thin_c *tc = op->tc;
399 	sector_t s = block_to_sectors(tc->pool, data_b);
400 	sector_t len = block_to_sectors(tc->pool, data_e - data_b);
401 
402 	return __blkdev_issue_discard(tc->pool_dev->bdev, s, len, GFP_NOWAIT,
403 				      &op->bio);
404 }
405 
406 static void end_discard(struct discard_op *op, int r)
407 {
408 	if (op->bio) {
409 		/*
410 		 * Even if one of the calls to issue_discard failed, we
411 		 * need to wait for the chain to complete.
412 		 */
413 		bio_chain(op->bio, op->parent_bio);
414 		op->bio->bi_opf = REQ_OP_DISCARD;
415 		submit_bio(op->bio);
416 	}
417 
418 	blk_finish_plug(&op->plug);
419 
420 	/*
421 	 * Even if r is set, there could be sub discards in flight that we
422 	 * need to wait for.
423 	 */
424 	if (r && !op->parent_bio->bi_status)
425 		op->parent_bio->bi_status = errno_to_blk_status(r);
426 	bio_endio(op->parent_bio);
427 }
428 
429 /*----------------------------------------------------------------*/
430 
431 /*
432  * wake_worker() is used when new work is queued and when pool_resume is
433  * ready to continue deferred IO processing.
434  */
435 static void wake_worker(struct pool *pool)
436 {
437 	queue_work(pool->wq, &pool->worker);
438 }
439 
440 /*----------------------------------------------------------------*/
441 
442 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
443 		      struct dm_bio_prison_cell **cell_result)
444 {
445 	int r;
446 	struct dm_bio_prison_cell *cell_prealloc;
447 
448 	/*
449 	 * Allocate a cell from the prison's mempool.
450 	 * This might block but it can't fail.
451 	 */
452 	cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
453 
454 	r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
455 	if (r)
456 		/*
457 		 * We reused an old cell; we can get rid of
458 		 * the new one.
459 		 */
460 		dm_bio_prison_free_cell(pool->prison, cell_prealloc);
461 
462 	return r;
463 }
464 
465 static void cell_release(struct pool *pool,
466 			 struct dm_bio_prison_cell *cell,
467 			 struct bio_list *bios)
468 {
469 	dm_cell_release(pool->prison, cell, bios);
470 	dm_bio_prison_free_cell(pool->prison, cell);
471 }
472 
473 static void cell_visit_release(struct pool *pool,
474 			       void (*fn)(void *, struct dm_bio_prison_cell *),
475 			       void *context,
476 			       struct dm_bio_prison_cell *cell)
477 {
478 	dm_cell_visit_release(pool->prison, fn, context, cell);
479 	dm_bio_prison_free_cell(pool->prison, cell);
480 }
481 
482 static void cell_release_no_holder(struct pool *pool,
483 				   struct dm_bio_prison_cell *cell,
484 				   struct bio_list *bios)
485 {
486 	dm_cell_release_no_holder(pool->prison, cell, bios);
487 	dm_bio_prison_free_cell(pool->prison, cell);
488 }
489 
490 static void cell_error_with_code(struct pool *pool,
491 		struct dm_bio_prison_cell *cell, blk_status_t error_code)
492 {
493 	dm_cell_error(pool->prison, cell, error_code);
494 	dm_bio_prison_free_cell(pool->prison, cell);
495 }
496 
497 static blk_status_t get_pool_io_error_code(struct pool *pool)
498 {
499 	return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
500 }
501 
502 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
503 {
504 	cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
505 }
506 
507 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
508 {
509 	cell_error_with_code(pool, cell, 0);
510 }
511 
512 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
513 {
514 	cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
515 }
516 
517 /*----------------------------------------------------------------*/
518 
519 /*
520  * A global list of pools that uses a struct mapped_device as a key.
521  */
522 static struct dm_thin_pool_table {
523 	struct mutex mutex;
524 	struct list_head pools;
525 } dm_thin_pool_table;
526 
527 static void pool_table_init(void)
528 {
529 	mutex_init(&dm_thin_pool_table.mutex);
530 	INIT_LIST_HEAD(&dm_thin_pool_table.pools);
531 }
532 
533 static void pool_table_exit(void)
534 {
535 	mutex_destroy(&dm_thin_pool_table.mutex);
536 }
537 
538 static void __pool_table_insert(struct pool *pool)
539 {
540 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
541 	list_add(&pool->list, &dm_thin_pool_table.pools);
542 }
543 
544 static void __pool_table_remove(struct pool *pool)
545 {
546 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
547 	list_del(&pool->list);
548 }
549 
550 static struct pool *__pool_table_lookup(struct mapped_device *md)
551 {
552 	struct pool *pool = NULL, *tmp;
553 
554 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
555 
556 	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
557 		if (tmp->pool_md == md) {
558 			pool = tmp;
559 			break;
560 		}
561 	}
562 
563 	return pool;
564 }
565 
566 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
567 {
568 	struct pool *pool = NULL, *tmp;
569 
570 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
571 
572 	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
573 		if (tmp->md_dev == md_dev) {
574 			pool = tmp;
575 			break;
576 		}
577 	}
578 
579 	return pool;
580 }
581 
582 /*----------------------------------------------------------------*/
583 
584 struct dm_thin_endio_hook {
585 	struct thin_c *tc;
586 	struct dm_deferred_entry *shared_read_entry;
587 	struct dm_deferred_entry *all_io_entry;
588 	struct dm_thin_new_mapping *overwrite_mapping;
589 	struct rb_node rb_node;
590 	struct dm_bio_prison_cell *cell;
591 };
592 
593 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
594 {
595 	bio_list_merge(bios, master);
596 	bio_list_init(master);
597 }
598 
599 static void error_bio_list(struct bio_list *bios, blk_status_t error)
600 {
601 	struct bio *bio;
602 
603 	while ((bio = bio_list_pop(bios))) {
604 		bio->bi_status = error;
605 		bio_endio(bio);
606 	}
607 }
608 
609 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
610 		blk_status_t error)
611 {
612 	struct bio_list bios;
613 
614 	bio_list_init(&bios);
615 
616 	spin_lock_irq(&tc->lock);
617 	__merge_bio_list(&bios, master);
618 	spin_unlock_irq(&tc->lock);
619 
620 	error_bio_list(&bios, error);
621 }
622 
623 static void requeue_deferred_cells(struct thin_c *tc)
624 {
625 	struct pool *pool = tc->pool;
626 	struct list_head cells;
627 	struct dm_bio_prison_cell *cell, *tmp;
628 
629 	INIT_LIST_HEAD(&cells);
630 
631 	spin_lock_irq(&tc->lock);
632 	list_splice_init(&tc->deferred_cells, &cells);
633 	spin_unlock_irq(&tc->lock);
634 
635 	list_for_each_entry_safe(cell, tmp, &cells, user_list)
636 		cell_requeue(pool, cell);
637 }
638 
639 static void requeue_io(struct thin_c *tc)
640 {
641 	struct bio_list bios;
642 
643 	bio_list_init(&bios);
644 
645 	spin_lock_irq(&tc->lock);
646 	__merge_bio_list(&bios, &tc->deferred_bio_list);
647 	__merge_bio_list(&bios, &tc->retry_on_resume_list);
648 	spin_unlock_irq(&tc->lock);
649 
650 	error_bio_list(&bios, BLK_STS_DM_REQUEUE);
651 	requeue_deferred_cells(tc);
652 }
653 
654 static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
655 {
656 	struct thin_c *tc;
657 
658 	rcu_read_lock();
659 	list_for_each_entry_rcu(tc, &pool->active_thins, list)
660 		error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
661 	rcu_read_unlock();
662 }
663 
664 static void error_retry_list(struct pool *pool)
665 {
666 	error_retry_list_with_code(pool, get_pool_io_error_code(pool));
667 }
668 
669 /*
670  * This section of code contains the logic for processing a thin device's IO.
671  * Much of the code depends on pool object resources (lists, workqueues, etc)
672  * but most is exclusively called from the thin target rather than the thin-pool
673  * target.
674  */
675 
676 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
677 {
678 	struct pool *pool = tc->pool;
679 	sector_t block_nr = bio->bi_iter.bi_sector;
680 
681 	if (block_size_is_power_of_two(pool))
682 		block_nr >>= pool->sectors_per_block_shift;
683 	else
684 		(void) sector_div(block_nr, pool->sectors_per_block);
685 
686 	return block_nr;
687 }
688 
689 /*
690  * Returns the _complete_ blocks that this bio covers.
691  */
692 static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
693 				dm_block_t *begin, dm_block_t *end)
694 {
695 	struct pool *pool = tc->pool;
696 	sector_t b = bio->bi_iter.bi_sector;
697 	sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
698 
699 	b += pool->sectors_per_block - 1ull; /* so we round up */
700 
701 	if (block_size_is_power_of_two(pool)) {
702 		b >>= pool->sectors_per_block_shift;
703 		e >>= pool->sectors_per_block_shift;
704 	} else {
705 		(void) sector_div(b, pool->sectors_per_block);
706 		(void) sector_div(e, pool->sectors_per_block);
707 	}
708 
709 	if (e < b)
710 		/* Can happen if the bio is within a single block. */
711 		e = b;
712 
713 	*begin = b;
714 	*end = e;
715 }
716 
717 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
718 {
719 	struct pool *pool = tc->pool;
720 	sector_t bi_sector = bio->bi_iter.bi_sector;
721 
722 	bio_set_dev(bio, tc->pool_dev->bdev);
723 	if (block_size_is_power_of_two(pool))
724 		bio->bi_iter.bi_sector =
725 			(block << pool->sectors_per_block_shift) |
726 			(bi_sector & (pool->sectors_per_block - 1));
727 	else
728 		bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
729 				 sector_div(bi_sector, pool->sectors_per_block);
730 }
731 
732 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
733 {
734 	bio_set_dev(bio, tc->origin_dev->bdev);
735 }
736 
737 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
738 {
739 	return op_is_flush(bio->bi_opf) &&
740 		dm_thin_changed_this_transaction(tc->td);
741 }
742 
743 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
744 {
745 	struct dm_thin_endio_hook *h;
746 
747 	if (bio_op(bio) == REQ_OP_DISCARD)
748 		return;
749 
750 	h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
751 	h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
752 }
753 
754 static void issue(struct thin_c *tc, struct bio *bio)
755 {
756 	struct pool *pool = tc->pool;
757 
758 	if (!bio_triggers_commit(tc, bio)) {
759 		dm_submit_bio_remap(bio, NULL);
760 		return;
761 	}
762 
763 	/*
764 	 * Complete bio with an error if earlier I/O caused changes to
765 	 * the metadata that can't be committed e.g, due to I/O errors
766 	 * on the metadata device.
767 	 */
768 	if (dm_thin_aborted_changes(tc->td)) {
769 		bio_io_error(bio);
770 		return;
771 	}
772 
773 	/*
774 	 * Batch together any bios that trigger commits and then issue a
775 	 * single commit for them in process_deferred_bios().
776 	 */
777 	spin_lock_irq(&pool->lock);
778 	bio_list_add(&pool->deferred_flush_bios, bio);
779 	spin_unlock_irq(&pool->lock);
780 }
781 
782 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
783 {
784 	remap_to_origin(tc, bio);
785 	issue(tc, bio);
786 }
787 
788 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
789 			    dm_block_t block)
790 {
791 	remap(tc, bio, block);
792 	issue(tc, bio);
793 }
794 
795 /*----------------------------------------------------------------*/
796 
797 /*
798  * Bio endio functions.
799  */
800 struct dm_thin_new_mapping {
801 	struct list_head list;
802 
803 	bool pass_discard:1;
804 	bool maybe_shared:1;
805 
806 	/*
807 	 * Track quiescing, copying and zeroing preparation actions.  When this
808 	 * counter hits zero the block is prepared and can be inserted into the
809 	 * btree.
810 	 */
811 	atomic_t prepare_actions;
812 
813 	blk_status_t status;
814 	struct thin_c *tc;
815 	dm_block_t virt_begin, virt_end;
816 	dm_block_t data_block;
817 	struct dm_bio_prison_cell *cell;
818 
819 	/*
820 	 * If the bio covers the whole area of a block then we can avoid
821 	 * zeroing or copying.  Instead this bio is hooked.  The bio will
822 	 * still be in the cell, so care has to be taken to avoid issuing
823 	 * the bio twice.
824 	 */
825 	struct bio *bio;
826 	bio_end_io_t *saved_bi_end_io;
827 };
828 
829 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
830 {
831 	struct pool *pool = m->tc->pool;
832 
833 	if (atomic_dec_and_test(&m->prepare_actions)) {
834 		list_add_tail(&m->list, &pool->prepared_mappings);
835 		wake_worker(pool);
836 	}
837 }
838 
839 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
840 {
841 	unsigned long flags;
842 	struct pool *pool = m->tc->pool;
843 
844 	spin_lock_irqsave(&pool->lock, flags);
845 	__complete_mapping_preparation(m);
846 	spin_unlock_irqrestore(&pool->lock, flags);
847 }
848 
849 static void copy_complete(int read_err, unsigned long write_err, void *context)
850 {
851 	struct dm_thin_new_mapping *m = context;
852 
853 	m->status = read_err || write_err ? BLK_STS_IOERR : 0;
854 	complete_mapping_preparation(m);
855 }
856 
857 static void overwrite_endio(struct bio *bio)
858 {
859 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
860 	struct dm_thin_new_mapping *m = h->overwrite_mapping;
861 
862 	bio->bi_end_io = m->saved_bi_end_io;
863 
864 	m->status = bio->bi_status;
865 	complete_mapping_preparation(m);
866 }
867 
868 /*----------------------------------------------------------------*/
869 
870 /*
871  * Workqueue.
872  */
873 
874 /*
875  * Prepared mapping jobs.
876  */
877 
878 /*
879  * This sends the bios in the cell, except the original holder, back
880  * to the deferred_bios list.
881  */
882 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
883 {
884 	struct pool *pool = tc->pool;
885 	unsigned long flags;
886 	int has_work;
887 
888 	spin_lock_irqsave(&tc->lock, flags);
889 	cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
890 	has_work = !bio_list_empty(&tc->deferred_bio_list);
891 	spin_unlock_irqrestore(&tc->lock, flags);
892 
893 	if (has_work)
894 		wake_worker(pool);
895 }
896 
897 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
898 
899 struct remap_info {
900 	struct thin_c *tc;
901 	struct bio_list defer_bios;
902 	struct bio_list issue_bios;
903 };
904 
905 static void __inc_remap_and_issue_cell(void *context,
906 				       struct dm_bio_prison_cell *cell)
907 {
908 	struct remap_info *info = context;
909 	struct bio *bio;
910 
911 	while ((bio = bio_list_pop(&cell->bios))) {
912 		if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
913 			bio_list_add(&info->defer_bios, bio);
914 		else {
915 			inc_all_io_entry(info->tc->pool, bio);
916 
917 			/*
918 			 * We can't issue the bios with the bio prison lock
919 			 * held, so we add them to a list to issue on
920 			 * return from this function.
921 			 */
922 			bio_list_add(&info->issue_bios, bio);
923 		}
924 	}
925 }
926 
927 static void inc_remap_and_issue_cell(struct thin_c *tc,
928 				     struct dm_bio_prison_cell *cell,
929 				     dm_block_t block)
930 {
931 	struct bio *bio;
932 	struct remap_info info;
933 
934 	info.tc = tc;
935 	bio_list_init(&info.defer_bios);
936 	bio_list_init(&info.issue_bios);
937 
938 	/*
939 	 * We have to be careful to inc any bios we're about to issue
940 	 * before the cell is released, and avoid a race with new bios
941 	 * being added to the cell.
942 	 */
943 	cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
944 			   &info, cell);
945 
946 	while ((bio = bio_list_pop(&info.defer_bios)))
947 		thin_defer_bio(tc, bio);
948 
949 	while ((bio = bio_list_pop(&info.issue_bios)))
950 		remap_and_issue(info.tc, bio, block);
951 }
952 
953 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
954 {
955 	cell_error(m->tc->pool, m->cell);
956 	list_del(&m->list);
957 	mempool_free(m, &m->tc->pool->mapping_pool);
958 }
959 
960 static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
961 {
962 	struct pool *pool = tc->pool;
963 
964 	/*
965 	 * If the bio has the REQ_FUA flag set we must commit the metadata
966 	 * before signaling its completion.
967 	 */
968 	if (!bio_triggers_commit(tc, bio)) {
969 		bio_endio(bio);
970 		return;
971 	}
972 
973 	/*
974 	 * Complete bio with an error if earlier I/O caused changes to the
975 	 * metadata that can't be committed, e.g, due to I/O errors on the
976 	 * metadata device.
977 	 */
978 	if (dm_thin_aborted_changes(tc->td)) {
979 		bio_io_error(bio);
980 		return;
981 	}
982 
983 	/*
984 	 * Batch together any bios that trigger commits and then issue a
985 	 * single commit for them in process_deferred_bios().
986 	 */
987 	spin_lock_irq(&pool->lock);
988 	bio_list_add(&pool->deferred_flush_completions, bio);
989 	spin_unlock_irq(&pool->lock);
990 }
991 
992 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
993 {
994 	struct thin_c *tc = m->tc;
995 	struct pool *pool = tc->pool;
996 	struct bio *bio = m->bio;
997 	int r;
998 
999 	if (m->status) {
1000 		cell_error(pool, m->cell);
1001 		goto out;
1002 	}
1003 
1004 	/*
1005 	 * Commit the prepared block into the mapping btree.
1006 	 * Any I/O for this block arriving after this point will get
1007 	 * remapped to it directly.
1008 	 */
1009 	r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1010 	if (r) {
1011 		metadata_operation_failed(pool, "dm_thin_insert_block", r);
1012 		cell_error(pool, m->cell);
1013 		goto out;
1014 	}
1015 
1016 	/*
1017 	 * Release any bios held while the block was being provisioned.
1018 	 * If we are processing a write bio that completely covers the block,
1019 	 * we already processed it so can ignore it now when processing
1020 	 * the bios in the cell.
1021 	 */
1022 	if (bio) {
1023 		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1024 		complete_overwrite_bio(tc, bio);
1025 	} else {
1026 		inc_all_io_entry(tc->pool, m->cell->holder);
1027 		remap_and_issue(tc, m->cell->holder, m->data_block);
1028 		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1029 	}
1030 
1031 out:
1032 	list_del(&m->list);
1033 	mempool_free(m, &pool->mapping_pool);
1034 }
1035 
1036 /*----------------------------------------------------------------*/
1037 
1038 static void free_discard_mapping(struct dm_thin_new_mapping *m)
1039 {
1040 	struct thin_c *tc = m->tc;
1041 
1042 	if (m->cell)
1043 		cell_defer_no_holder(tc, m->cell);
1044 	mempool_free(m, &tc->pool->mapping_pool);
1045 }
1046 
1047 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1048 {
1049 	bio_io_error(m->bio);
1050 	free_discard_mapping(m);
1051 }
1052 
1053 static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1054 {
1055 	bio_endio(m->bio);
1056 	free_discard_mapping(m);
1057 }
1058 
1059 static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1060 {
1061 	int r;
1062 	struct thin_c *tc = m->tc;
1063 
1064 	r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1065 	if (r) {
1066 		metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1067 		bio_io_error(m->bio);
1068 	} else
1069 		bio_endio(m->bio);
1070 
1071 	cell_defer_no_holder(tc, m->cell);
1072 	mempool_free(m, &tc->pool->mapping_pool);
1073 }
1074 
1075 /*----------------------------------------------------------------*/
1076 
1077 static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1078 						   struct bio *discard_parent)
1079 {
1080 	/*
1081 	 * We've already unmapped this range of blocks, but before we
1082 	 * passdown we have to check that these blocks are now unused.
1083 	 */
1084 	int r = 0;
1085 	bool shared = true;
1086 	struct thin_c *tc = m->tc;
1087 	struct pool *pool = tc->pool;
1088 	dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1089 	struct discard_op op;
1090 
1091 	begin_discard(&op, tc, discard_parent);
1092 	while (b != end) {
1093 		/* find start of unmapped run */
1094 		for (; b < end; b++) {
1095 			r = dm_pool_block_is_shared(pool->pmd, b, &shared);
1096 			if (r)
1097 				goto out;
1098 
1099 			if (!shared)
1100 				break;
1101 		}
1102 
1103 		if (b == end)
1104 			break;
1105 
1106 		/* find end of run */
1107 		for (e = b + 1; e != end; e++) {
1108 			r = dm_pool_block_is_shared(pool->pmd, e, &shared);
1109 			if (r)
1110 				goto out;
1111 
1112 			if (shared)
1113 				break;
1114 		}
1115 
1116 		r = issue_discard(&op, b, e);
1117 		if (r)
1118 			goto out;
1119 
1120 		b = e;
1121 	}
1122 out:
1123 	end_discard(&op, r);
1124 }
1125 
1126 static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1127 {
1128 	unsigned long flags;
1129 	struct pool *pool = m->tc->pool;
1130 
1131 	spin_lock_irqsave(&pool->lock, flags);
1132 	list_add_tail(&m->list, &pool->prepared_discards_pt2);
1133 	spin_unlock_irqrestore(&pool->lock, flags);
1134 	wake_worker(pool);
1135 }
1136 
1137 static void passdown_endio(struct bio *bio)
1138 {
1139 	/*
1140 	 * It doesn't matter if the passdown discard failed, we still want
1141 	 * to unmap (we ignore err).
1142 	 */
1143 	queue_passdown_pt2(bio->bi_private);
1144 	bio_put(bio);
1145 }
1146 
1147 static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1148 {
1149 	int r;
1150 	struct thin_c *tc = m->tc;
1151 	struct pool *pool = tc->pool;
1152 	struct bio *discard_parent;
1153 	dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1154 
1155 	/*
1156 	 * Only this thread allocates blocks, so we can be sure that the
1157 	 * newly unmapped blocks will not be allocated before the end of
1158 	 * the function.
1159 	 */
1160 	r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1161 	if (r) {
1162 		metadata_operation_failed(pool, "dm_thin_remove_range", r);
1163 		bio_io_error(m->bio);
1164 		cell_defer_no_holder(tc, m->cell);
1165 		mempool_free(m, &pool->mapping_pool);
1166 		return;
1167 	}
1168 
1169 	/*
1170 	 * Increment the unmapped blocks.  This prevents a race between the
1171 	 * passdown io and reallocation of freed blocks.
1172 	 */
1173 	r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1174 	if (r) {
1175 		metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1176 		bio_io_error(m->bio);
1177 		cell_defer_no_holder(tc, m->cell);
1178 		mempool_free(m, &pool->mapping_pool);
1179 		return;
1180 	}
1181 
1182 	discard_parent = bio_alloc(NULL, 1, 0, GFP_NOIO);
1183 	discard_parent->bi_end_io = passdown_endio;
1184 	discard_parent->bi_private = m;
1185 	if (m->maybe_shared)
1186 		passdown_double_checking_shared_status(m, discard_parent);
1187 	else {
1188 		struct discard_op op;
1189 
1190 		begin_discard(&op, tc, discard_parent);
1191 		r = issue_discard(&op, m->data_block, data_end);
1192 		end_discard(&op, r);
1193 	}
1194 }
1195 
1196 static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1197 {
1198 	int r;
1199 	struct thin_c *tc = m->tc;
1200 	struct pool *pool = tc->pool;
1201 
1202 	/*
1203 	 * The passdown has completed, so now we can decrement all those
1204 	 * unmapped blocks.
1205 	 */
1206 	r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1207 				   m->data_block + (m->virt_end - m->virt_begin));
1208 	if (r) {
1209 		metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1210 		bio_io_error(m->bio);
1211 	} else
1212 		bio_endio(m->bio);
1213 
1214 	cell_defer_no_holder(tc, m->cell);
1215 	mempool_free(m, &pool->mapping_pool);
1216 }
1217 
1218 static void process_prepared(struct pool *pool, struct list_head *head,
1219 			     process_mapping_fn *fn)
1220 {
1221 	struct list_head maps;
1222 	struct dm_thin_new_mapping *m, *tmp;
1223 
1224 	INIT_LIST_HEAD(&maps);
1225 	spin_lock_irq(&pool->lock);
1226 	list_splice_init(head, &maps);
1227 	spin_unlock_irq(&pool->lock);
1228 
1229 	list_for_each_entry_safe(m, tmp, &maps, list)
1230 		(*fn)(m);
1231 }
1232 
1233 /*
1234  * Deferred bio jobs.
1235  */
1236 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1237 {
1238 	return bio->bi_iter.bi_size ==
1239 		(pool->sectors_per_block << SECTOR_SHIFT);
1240 }
1241 
1242 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1243 {
1244 	return (bio_data_dir(bio) == WRITE) &&
1245 		io_overlaps_block(pool, bio);
1246 }
1247 
1248 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1249 			       bio_end_io_t *fn)
1250 {
1251 	*save = bio->bi_end_io;
1252 	bio->bi_end_io = fn;
1253 }
1254 
1255 static int ensure_next_mapping(struct pool *pool)
1256 {
1257 	if (pool->next_mapping)
1258 		return 0;
1259 
1260 	pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1261 
1262 	return pool->next_mapping ? 0 : -ENOMEM;
1263 }
1264 
1265 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1266 {
1267 	struct dm_thin_new_mapping *m = pool->next_mapping;
1268 
1269 	BUG_ON(!pool->next_mapping);
1270 
1271 	memset(m, 0, sizeof(struct dm_thin_new_mapping));
1272 	INIT_LIST_HEAD(&m->list);
1273 	m->bio = NULL;
1274 
1275 	pool->next_mapping = NULL;
1276 
1277 	return m;
1278 }
1279 
1280 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1281 		    sector_t begin, sector_t end)
1282 {
1283 	struct dm_io_region to;
1284 
1285 	to.bdev = tc->pool_dev->bdev;
1286 	to.sector = begin;
1287 	to.count = end - begin;
1288 
1289 	dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1290 }
1291 
1292 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1293 				      dm_block_t data_begin,
1294 				      struct dm_thin_new_mapping *m)
1295 {
1296 	struct pool *pool = tc->pool;
1297 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1298 
1299 	h->overwrite_mapping = m;
1300 	m->bio = bio;
1301 	save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1302 	inc_all_io_entry(pool, bio);
1303 	remap_and_issue(tc, bio, data_begin);
1304 }
1305 
1306 /*
1307  * A partial copy also needs to zero the uncopied region.
1308  */
1309 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1310 			  struct dm_dev *origin, dm_block_t data_origin,
1311 			  dm_block_t data_dest,
1312 			  struct dm_bio_prison_cell *cell, struct bio *bio,
1313 			  sector_t len)
1314 {
1315 	struct pool *pool = tc->pool;
1316 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1317 
1318 	m->tc = tc;
1319 	m->virt_begin = virt_block;
1320 	m->virt_end = virt_block + 1u;
1321 	m->data_block = data_dest;
1322 	m->cell = cell;
1323 
1324 	/*
1325 	 * quiesce action + copy action + an extra reference held for the
1326 	 * duration of this function (we may need to inc later for a
1327 	 * partial zero).
1328 	 */
1329 	atomic_set(&m->prepare_actions, 3);
1330 
1331 	if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1332 		complete_mapping_preparation(m); /* already quiesced */
1333 
1334 	/*
1335 	 * IO to pool_dev remaps to the pool target's data_dev.
1336 	 *
1337 	 * If the whole block of data is being overwritten, we can issue the
1338 	 * bio immediately. Otherwise we use kcopyd to clone the data first.
1339 	 */
1340 	if (io_overwrites_block(pool, bio))
1341 		remap_and_issue_overwrite(tc, bio, data_dest, m);
1342 	else {
1343 		struct dm_io_region from, to;
1344 
1345 		from.bdev = origin->bdev;
1346 		from.sector = data_origin * pool->sectors_per_block;
1347 		from.count = len;
1348 
1349 		to.bdev = tc->pool_dev->bdev;
1350 		to.sector = data_dest * pool->sectors_per_block;
1351 		to.count = len;
1352 
1353 		dm_kcopyd_copy(pool->copier, &from, 1, &to,
1354 			       0, copy_complete, m);
1355 
1356 		/*
1357 		 * Do we need to zero a tail region?
1358 		 */
1359 		if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1360 			atomic_inc(&m->prepare_actions);
1361 			ll_zero(tc, m,
1362 				data_dest * pool->sectors_per_block + len,
1363 				(data_dest + 1) * pool->sectors_per_block);
1364 		}
1365 	}
1366 
1367 	complete_mapping_preparation(m); /* drop our ref */
1368 }
1369 
1370 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1371 				   dm_block_t data_origin, dm_block_t data_dest,
1372 				   struct dm_bio_prison_cell *cell, struct bio *bio)
1373 {
1374 	schedule_copy(tc, virt_block, tc->pool_dev,
1375 		      data_origin, data_dest, cell, bio,
1376 		      tc->pool->sectors_per_block);
1377 }
1378 
1379 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1380 			  dm_block_t data_block, struct dm_bio_prison_cell *cell,
1381 			  struct bio *bio)
1382 {
1383 	struct pool *pool = tc->pool;
1384 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1385 
1386 	atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1387 	m->tc = tc;
1388 	m->virt_begin = virt_block;
1389 	m->virt_end = virt_block + 1u;
1390 	m->data_block = data_block;
1391 	m->cell = cell;
1392 
1393 	/*
1394 	 * If the whole block of data is being overwritten or we are not
1395 	 * zeroing pre-existing data, we can issue the bio immediately.
1396 	 * Otherwise we use kcopyd to zero the data first.
1397 	 */
1398 	if (pool->pf.zero_new_blocks) {
1399 		if (io_overwrites_block(pool, bio))
1400 			remap_and_issue_overwrite(tc, bio, data_block, m);
1401 		else
1402 			ll_zero(tc, m, data_block * pool->sectors_per_block,
1403 				(data_block + 1) * pool->sectors_per_block);
1404 	} else
1405 		process_prepared_mapping(m);
1406 }
1407 
1408 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1409 				   dm_block_t data_dest,
1410 				   struct dm_bio_prison_cell *cell, struct bio *bio)
1411 {
1412 	struct pool *pool = tc->pool;
1413 	sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1414 	sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1415 
1416 	if (virt_block_end <= tc->origin_size)
1417 		schedule_copy(tc, virt_block, tc->origin_dev,
1418 			      virt_block, data_dest, cell, bio,
1419 			      pool->sectors_per_block);
1420 
1421 	else if (virt_block_begin < tc->origin_size)
1422 		schedule_copy(tc, virt_block, tc->origin_dev,
1423 			      virt_block, data_dest, cell, bio,
1424 			      tc->origin_size - virt_block_begin);
1425 
1426 	else
1427 		schedule_zero(tc, virt_block, data_dest, cell, bio);
1428 }
1429 
1430 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1431 
1432 static void requeue_bios(struct pool *pool);
1433 
1434 static bool is_read_only_pool_mode(enum pool_mode mode)
1435 {
1436 	return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1437 }
1438 
1439 static bool is_read_only(struct pool *pool)
1440 {
1441 	return is_read_only_pool_mode(get_pool_mode(pool));
1442 }
1443 
1444 static void check_for_metadata_space(struct pool *pool)
1445 {
1446 	int r;
1447 	const char *ooms_reason = NULL;
1448 	dm_block_t nr_free;
1449 
1450 	r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1451 	if (r)
1452 		ooms_reason = "Could not get free metadata blocks";
1453 	else if (!nr_free)
1454 		ooms_reason = "No free metadata blocks";
1455 
1456 	if (ooms_reason && !is_read_only(pool)) {
1457 		DMERR("%s", ooms_reason);
1458 		set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1459 	}
1460 }
1461 
1462 static void check_for_data_space(struct pool *pool)
1463 {
1464 	int r;
1465 	dm_block_t nr_free;
1466 
1467 	if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1468 		return;
1469 
1470 	r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1471 	if (r)
1472 		return;
1473 
1474 	if (nr_free) {
1475 		set_pool_mode(pool, PM_WRITE);
1476 		requeue_bios(pool);
1477 	}
1478 }
1479 
1480 /*
1481  * A non-zero return indicates read_only or fail_io mode.
1482  * Many callers don't care about the return value.
1483  */
1484 static int commit(struct pool *pool)
1485 {
1486 	int r;
1487 
1488 	if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1489 		return -EINVAL;
1490 
1491 	r = dm_pool_commit_metadata(pool->pmd);
1492 	if (r)
1493 		metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1494 	else {
1495 		check_for_metadata_space(pool);
1496 		check_for_data_space(pool);
1497 	}
1498 
1499 	return r;
1500 }
1501 
1502 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1503 {
1504 	if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1505 		DMWARN("%s: reached low water mark for data device: sending event.",
1506 		       dm_device_name(pool->pool_md));
1507 		spin_lock_irq(&pool->lock);
1508 		pool->low_water_triggered = true;
1509 		spin_unlock_irq(&pool->lock);
1510 		dm_table_event(pool->ti->table);
1511 	}
1512 }
1513 
1514 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1515 {
1516 	int r;
1517 	dm_block_t free_blocks;
1518 	struct pool *pool = tc->pool;
1519 
1520 	if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1521 		return -EINVAL;
1522 
1523 	r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1524 	if (r) {
1525 		metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1526 		return r;
1527 	}
1528 
1529 	check_low_water_mark(pool, free_blocks);
1530 
1531 	if (!free_blocks) {
1532 		/*
1533 		 * Try to commit to see if that will free up some
1534 		 * more space.
1535 		 */
1536 		r = commit(pool);
1537 		if (r)
1538 			return r;
1539 
1540 		r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1541 		if (r) {
1542 			metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1543 			return r;
1544 		}
1545 
1546 		if (!free_blocks) {
1547 			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1548 			return -ENOSPC;
1549 		}
1550 	}
1551 
1552 	r = dm_pool_alloc_data_block(pool->pmd, result);
1553 	if (r) {
1554 		if (r == -ENOSPC)
1555 			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1556 		else
1557 			metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1558 		return r;
1559 	}
1560 
1561 	r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1562 	if (r) {
1563 		metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1564 		return r;
1565 	}
1566 
1567 	if (!free_blocks) {
1568 		/* Let's commit before we use up the metadata reserve. */
1569 		r = commit(pool);
1570 		if (r)
1571 			return r;
1572 	}
1573 
1574 	return 0;
1575 }
1576 
1577 /*
1578  * If we have run out of space, queue bios until the device is
1579  * resumed, presumably after having been reloaded with more space.
1580  */
1581 static void retry_on_resume(struct bio *bio)
1582 {
1583 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1584 	struct thin_c *tc = h->tc;
1585 
1586 	spin_lock_irq(&tc->lock);
1587 	bio_list_add(&tc->retry_on_resume_list, bio);
1588 	spin_unlock_irq(&tc->lock);
1589 }
1590 
1591 static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1592 {
1593 	enum pool_mode m = get_pool_mode(pool);
1594 
1595 	switch (m) {
1596 	case PM_WRITE:
1597 		/* Shouldn't get here */
1598 		DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1599 		return BLK_STS_IOERR;
1600 
1601 	case PM_OUT_OF_DATA_SPACE:
1602 		return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1603 
1604 	case PM_OUT_OF_METADATA_SPACE:
1605 	case PM_READ_ONLY:
1606 	case PM_FAIL:
1607 		return BLK_STS_IOERR;
1608 	default:
1609 		/* Shouldn't get here */
1610 		DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1611 		return BLK_STS_IOERR;
1612 	}
1613 }
1614 
1615 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1616 {
1617 	blk_status_t error = should_error_unserviceable_bio(pool);
1618 
1619 	if (error) {
1620 		bio->bi_status = error;
1621 		bio_endio(bio);
1622 	} else
1623 		retry_on_resume(bio);
1624 }
1625 
1626 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1627 {
1628 	struct bio *bio;
1629 	struct bio_list bios;
1630 	blk_status_t error;
1631 
1632 	error = should_error_unserviceable_bio(pool);
1633 	if (error) {
1634 		cell_error_with_code(pool, cell, error);
1635 		return;
1636 	}
1637 
1638 	bio_list_init(&bios);
1639 	cell_release(pool, cell, &bios);
1640 
1641 	while ((bio = bio_list_pop(&bios)))
1642 		retry_on_resume(bio);
1643 }
1644 
1645 static void process_discard_cell_no_passdown(struct thin_c *tc,
1646 					     struct dm_bio_prison_cell *virt_cell)
1647 {
1648 	struct pool *pool = tc->pool;
1649 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1650 
1651 	/*
1652 	 * We don't need to lock the data blocks, since there's no
1653 	 * passdown.  We only lock data blocks for allocation and breaking sharing.
1654 	 */
1655 	m->tc = tc;
1656 	m->virt_begin = virt_cell->key.block_begin;
1657 	m->virt_end = virt_cell->key.block_end;
1658 	m->cell = virt_cell;
1659 	m->bio = virt_cell->holder;
1660 
1661 	if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1662 		pool->process_prepared_discard(m);
1663 }
1664 
1665 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1666 				 struct bio *bio)
1667 {
1668 	struct pool *pool = tc->pool;
1669 
1670 	int r;
1671 	bool maybe_shared;
1672 	struct dm_cell_key data_key;
1673 	struct dm_bio_prison_cell *data_cell;
1674 	struct dm_thin_new_mapping *m;
1675 	dm_block_t virt_begin, virt_end, data_begin;
1676 
1677 	while (begin != end) {
1678 		r = ensure_next_mapping(pool);
1679 		if (r)
1680 			/* we did our best */
1681 			return;
1682 
1683 		r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1684 					      &data_begin, &maybe_shared);
1685 		if (r)
1686 			/*
1687 			 * Silently fail, letting any mappings we've
1688 			 * created complete.
1689 			 */
1690 			break;
1691 
1692 		build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1693 		if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1694 			/* contention, we'll give up with this range */
1695 			begin = virt_end;
1696 			continue;
1697 		}
1698 
1699 		/*
1700 		 * IO may still be going to the destination block.  We must
1701 		 * quiesce before we can do the removal.
1702 		 */
1703 		m = get_next_mapping(pool);
1704 		m->tc = tc;
1705 		m->maybe_shared = maybe_shared;
1706 		m->virt_begin = virt_begin;
1707 		m->virt_end = virt_end;
1708 		m->data_block = data_begin;
1709 		m->cell = data_cell;
1710 		m->bio = bio;
1711 
1712 		/*
1713 		 * The parent bio must not complete before sub discard bios are
1714 		 * chained to it (see end_discard's bio_chain)!
1715 		 *
1716 		 * This per-mapping bi_remaining increment is paired with
1717 		 * the implicit decrement that occurs via bio_endio() in
1718 		 * end_discard().
1719 		 */
1720 		bio_inc_remaining(bio);
1721 		if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1722 			pool->process_prepared_discard(m);
1723 
1724 		begin = virt_end;
1725 	}
1726 }
1727 
1728 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1729 {
1730 	struct bio *bio = virt_cell->holder;
1731 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1732 
1733 	/*
1734 	 * The virt_cell will only get freed once the origin bio completes.
1735 	 * This means it will remain locked while all the individual
1736 	 * passdown bios are in flight.
1737 	 */
1738 	h->cell = virt_cell;
1739 	break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1740 
1741 	/*
1742 	 * We complete the bio now, knowing that the bi_remaining field
1743 	 * will prevent completion until the sub range discards have
1744 	 * completed.
1745 	 */
1746 	bio_endio(bio);
1747 }
1748 
1749 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1750 {
1751 	dm_block_t begin, end;
1752 	struct dm_cell_key virt_key;
1753 	struct dm_bio_prison_cell *virt_cell;
1754 
1755 	get_bio_block_range(tc, bio, &begin, &end);
1756 	if (begin == end) {
1757 		/*
1758 		 * The discard covers less than a block.
1759 		 */
1760 		bio_endio(bio);
1761 		return;
1762 	}
1763 
1764 	build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1765 	if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1766 		/*
1767 		 * Potential starvation issue: We're relying on the
1768 		 * fs/application being well behaved, and not trying to
1769 		 * send IO to a region at the same time as discarding it.
1770 		 * If they do this persistently then it's possible this
1771 		 * cell will never be granted.
1772 		 */
1773 		return;
1774 
1775 	tc->pool->process_discard_cell(tc, virt_cell);
1776 }
1777 
1778 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1779 			  struct dm_cell_key *key,
1780 			  struct dm_thin_lookup_result *lookup_result,
1781 			  struct dm_bio_prison_cell *cell)
1782 {
1783 	int r;
1784 	dm_block_t data_block;
1785 	struct pool *pool = tc->pool;
1786 
1787 	r = alloc_data_block(tc, &data_block);
1788 	switch (r) {
1789 	case 0:
1790 		schedule_internal_copy(tc, block, lookup_result->block,
1791 				       data_block, cell, bio);
1792 		break;
1793 
1794 	case -ENOSPC:
1795 		retry_bios_on_resume(pool, cell);
1796 		break;
1797 
1798 	default:
1799 		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1800 			    __func__, r);
1801 		cell_error(pool, cell);
1802 		break;
1803 	}
1804 }
1805 
1806 static void __remap_and_issue_shared_cell(void *context,
1807 					  struct dm_bio_prison_cell *cell)
1808 {
1809 	struct remap_info *info = context;
1810 	struct bio *bio;
1811 
1812 	while ((bio = bio_list_pop(&cell->bios))) {
1813 		if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1814 		    bio_op(bio) == REQ_OP_DISCARD)
1815 			bio_list_add(&info->defer_bios, bio);
1816 		else {
1817 			struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1818 
1819 			h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1820 			inc_all_io_entry(info->tc->pool, bio);
1821 			bio_list_add(&info->issue_bios, bio);
1822 		}
1823 	}
1824 }
1825 
1826 static void remap_and_issue_shared_cell(struct thin_c *tc,
1827 					struct dm_bio_prison_cell *cell,
1828 					dm_block_t block)
1829 {
1830 	struct bio *bio;
1831 	struct remap_info info;
1832 
1833 	info.tc = tc;
1834 	bio_list_init(&info.defer_bios);
1835 	bio_list_init(&info.issue_bios);
1836 
1837 	cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1838 			   &info, cell);
1839 
1840 	while ((bio = bio_list_pop(&info.defer_bios)))
1841 		thin_defer_bio(tc, bio);
1842 
1843 	while ((bio = bio_list_pop(&info.issue_bios)))
1844 		remap_and_issue(tc, bio, block);
1845 }
1846 
1847 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1848 			       dm_block_t block,
1849 			       struct dm_thin_lookup_result *lookup_result,
1850 			       struct dm_bio_prison_cell *virt_cell)
1851 {
1852 	struct dm_bio_prison_cell *data_cell;
1853 	struct pool *pool = tc->pool;
1854 	struct dm_cell_key key;
1855 
1856 	/*
1857 	 * If cell is already occupied, then sharing is already in the process
1858 	 * of being broken so we have nothing further to do here.
1859 	 */
1860 	build_data_key(tc->td, lookup_result->block, &key);
1861 	if (bio_detain(pool, &key, bio, &data_cell)) {
1862 		cell_defer_no_holder(tc, virt_cell);
1863 		return;
1864 	}
1865 
1866 	if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1867 		break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1868 		cell_defer_no_holder(tc, virt_cell);
1869 	} else {
1870 		struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1871 
1872 		h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1873 		inc_all_io_entry(pool, bio);
1874 		remap_and_issue(tc, bio, lookup_result->block);
1875 
1876 		remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1877 		remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1878 	}
1879 }
1880 
1881 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1882 			    struct dm_bio_prison_cell *cell)
1883 {
1884 	int r;
1885 	dm_block_t data_block;
1886 	struct pool *pool = tc->pool;
1887 
1888 	/*
1889 	 * Remap empty bios (flushes) immediately, without provisioning.
1890 	 */
1891 	if (!bio->bi_iter.bi_size) {
1892 		inc_all_io_entry(pool, bio);
1893 		cell_defer_no_holder(tc, cell);
1894 
1895 		remap_and_issue(tc, bio, 0);
1896 		return;
1897 	}
1898 
1899 	/*
1900 	 * Fill read bios with zeroes and complete them immediately.
1901 	 */
1902 	if (bio_data_dir(bio) == READ) {
1903 		zero_fill_bio(bio);
1904 		cell_defer_no_holder(tc, cell);
1905 		bio_endio(bio);
1906 		return;
1907 	}
1908 
1909 	r = alloc_data_block(tc, &data_block);
1910 	switch (r) {
1911 	case 0:
1912 		if (tc->origin_dev)
1913 			schedule_external_copy(tc, block, data_block, cell, bio);
1914 		else
1915 			schedule_zero(tc, block, data_block, cell, bio);
1916 		break;
1917 
1918 	case -ENOSPC:
1919 		retry_bios_on_resume(pool, cell);
1920 		break;
1921 
1922 	default:
1923 		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1924 			    __func__, r);
1925 		cell_error(pool, cell);
1926 		break;
1927 	}
1928 }
1929 
1930 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1931 {
1932 	int r;
1933 	struct pool *pool = tc->pool;
1934 	struct bio *bio = cell->holder;
1935 	dm_block_t block = get_bio_block(tc, bio);
1936 	struct dm_thin_lookup_result lookup_result;
1937 
1938 	if (tc->requeue_mode) {
1939 		cell_requeue(pool, cell);
1940 		return;
1941 	}
1942 
1943 	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1944 	switch (r) {
1945 	case 0:
1946 		if (lookup_result.shared)
1947 			process_shared_bio(tc, bio, block, &lookup_result, cell);
1948 		else {
1949 			inc_all_io_entry(pool, bio);
1950 			remap_and_issue(tc, bio, lookup_result.block);
1951 			inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1952 		}
1953 		break;
1954 
1955 	case -ENODATA:
1956 		if (bio_data_dir(bio) == READ && tc->origin_dev) {
1957 			inc_all_io_entry(pool, bio);
1958 			cell_defer_no_holder(tc, cell);
1959 
1960 			if (bio_end_sector(bio) <= tc->origin_size)
1961 				remap_to_origin_and_issue(tc, bio);
1962 
1963 			else if (bio->bi_iter.bi_sector < tc->origin_size) {
1964 				zero_fill_bio(bio);
1965 				bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1966 				remap_to_origin_and_issue(tc, bio);
1967 
1968 			} else {
1969 				zero_fill_bio(bio);
1970 				bio_endio(bio);
1971 			}
1972 		} else
1973 			provision_block(tc, bio, block, cell);
1974 		break;
1975 
1976 	default:
1977 		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1978 			    __func__, r);
1979 		cell_defer_no_holder(tc, cell);
1980 		bio_io_error(bio);
1981 		break;
1982 	}
1983 }
1984 
1985 static void process_bio(struct thin_c *tc, struct bio *bio)
1986 {
1987 	struct pool *pool = tc->pool;
1988 	dm_block_t block = get_bio_block(tc, bio);
1989 	struct dm_bio_prison_cell *cell;
1990 	struct dm_cell_key key;
1991 
1992 	/*
1993 	 * If cell is already occupied, then the block is already
1994 	 * being provisioned so we have nothing further to do here.
1995 	 */
1996 	build_virtual_key(tc->td, block, &key);
1997 	if (bio_detain(pool, &key, bio, &cell))
1998 		return;
1999 
2000 	process_cell(tc, cell);
2001 }
2002 
2003 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2004 				    struct dm_bio_prison_cell *cell)
2005 {
2006 	int r;
2007 	int rw = bio_data_dir(bio);
2008 	dm_block_t block = get_bio_block(tc, bio);
2009 	struct dm_thin_lookup_result lookup_result;
2010 
2011 	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2012 	switch (r) {
2013 	case 0:
2014 		if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2015 			handle_unserviceable_bio(tc->pool, bio);
2016 			if (cell)
2017 				cell_defer_no_holder(tc, cell);
2018 		} else {
2019 			inc_all_io_entry(tc->pool, bio);
2020 			remap_and_issue(tc, bio, lookup_result.block);
2021 			if (cell)
2022 				inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2023 		}
2024 		break;
2025 
2026 	case -ENODATA:
2027 		if (cell)
2028 			cell_defer_no_holder(tc, cell);
2029 		if (rw != READ) {
2030 			handle_unserviceable_bio(tc->pool, bio);
2031 			break;
2032 		}
2033 
2034 		if (tc->origin_dev) {
2035 			inc_all_io_entry(tc->pool, bio);
2036 			remap_to_origin_and_issue(tc, bio);
2037 			break;
2038 		}
2039 
2040 		zero_fill_bio(bio);
2041 		bio_endio(bio);
2042 		break;
2043 
2044 	default:
2045 		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2046 			    __func__, r);
2047 		if (cell)
2048 			cell_defer_no_holder(tc, cell);
2049 		bio_io_error(bio);
2050 		break;
2051 	}
2052 }
2053 
2054 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2055 {
2056 	__process_bio_read_only(tc, bio, NULL);
2057 }
2058 
2059 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2060 {
2061 	__process_bio_read_only(tc, cell->holder, cell);
2062 }
2063 
2064 static void process_bio_success(struct thin_c *tc, struct bio *bio)
2065 {
2066 	bio_endio(bio);
2067 }
2068 
2069 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2070 {
2071 	bio_io_error(bio);
2072 }
2073 
2074 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2075 {
2076 	cell_success(tc->pool, cell);
2077 }
2078 
2079 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2080 {
2081 	cell_error(tc->pool, cell);
2082 }
2083 
2084 /*
2085  * FIXME: should we also commit due to size of transaction, measured in
2086  * metadata blocks?
2087  */
2088 static int need_commit_due_to_time(struct pool *pool)
2089 {
2090 	return !time_in_range(jiffies, pool->last_commit_jiffies,
2091 			      pool->last_commit_jiffies + COMMIT_PERIOD);
2092 }
2093 
2094 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2095 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2096 
2097 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2098 {
2099 	struct rb_node **rbp, *parent;
2100 	struct dm_thin_endio_hook *pbd;
2101 	sector_t bi_sector = bio->bi_iter.bi_sector;
2102 
2103 	rbp = &tc->sort_bio_list.rb_node;
2104 	parent = NULL;
2105 	while (*rbp) {
2106 		parent = *rbp;
2107 		pbd = thin_pbd(parent);
2108 
2109 		if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2110 			rbp = &(*rbp)->rb_left;
2111 		else
2112 			rbp = &(*rbp)->rb_right;
2113 	}
2114 
2115 	pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2116 	rb_link_node(&pbd->rb_node, parent, rbp);
2117 	rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2118 }
2119 
2120 static void __extract_sorted_bios(struct thin_c *tc)
2121 {
2122 	struct rb_node *node;
2123 	struct dm_thin_endio_hook *pbd;
2124 	struct bio *bio;
2125 
2126 	for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2127 		pbd = thin_pbd(node);
2128 		bio = thin_bio(pbd);
2129 
2130 		bio_list_add(&tc->deferred_bio_list, bio);
2131 		rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2132 	}
2133 
2134 	WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2135 }
2136 
2137 static void __sort_thin_deferred_bios(struct thin_c *tc)
2138 {
2139 	struct bio *bio;
2140 	struct bio_list bios;
2141 
2142 	bio_list_init(&bios);
2143 	bio_list_merge(&bios, &tc->deferred_bio_list);
2144 	bio_list_init(&tc->deferred_bio_list);
2145 
2146 	/* Sort deferred_bio_list using rb-tree */
2147 	while ((bio = bio_list_pop(&bios)))
2148 		__thin_bio_rb_add(tc, bio);
2149 
2150 	/*
2151 	 * Transfer the sorted bios in sort_bio_list back to
2152 	 * deferred_bio_list to allow lockless submission of
2153 	 * all bios.
2154 	 */
2155 	__extract_sorted_bios(tc);
2156 }
2157 
2158 static void process_thin_deferred_bios(struct thin_c *tc)
2159 {
2160 	struct pool *pool = tc->pool;
2161 	struct bio *bio;
2162 	struct bio_list bios;
2163 	struct blk_plug plug;
2164 	unsigned int count = 0;
2165 
2166 	if (tc->requeue_mode) {
2167 		error_thin_bio_list(tc, &tc->deferred_bio_list,
2168 				BLK_STS_DM_REQUEUE);
2169 		return;
2170 	}
2171 
2172 	bio_list_init(&bios);
2173 
2174 	spin_lock_irq(&tc->lock);
2175 
2176 	if (bio_list_empty(&tc->deferred_bio_list)) {
2177 		spin_unlock_irq(&tc->lock);
2178 		return;
2179 	}
2180 
2181 	__sort_thin_deferred_bios(tc);
2182 
2183 	bio_list_merge(&bios, &tc->deferred_bio_list);
2184 	bio_list_init(&tc->deferred_bio_list);
2185 
2186 	spin_unlock_irq(&tc->lock);
2187 
2188 	blk_start_plug(&plug);
2189 	while ((bio = bio_list_pop(&bios))) {
2190 		/*
2191 		 * If we've got no free new_mapping structs, and processing
2192 		 * this bio might require one, we pause until there are some
2193 		 * prepared mappings to process.
2194 		 */
2195 		if (ensure_next_mapping(pool)) {
2196 			spin_lock_irq(&tc->lock);
2197 			bio_list_add(&tc->deferred_bio_list, bio);
2198 			bio_list_merge(&tc->deferred_bio_list, &bios);
2199 			spin_unlock_irq(&tc->lock);
2200 			break;
2201 		}
2202 
2203 		if (bio_op(bio) == REQ_OP_DISCARD)
2204 			pool->process_discard(tc, bio);
2205 		else
2206 			pool->process_bio(tc, bio);
2207 
2208 		if ((count++ & 127) == 0) {
2209 			throttle_work_update(&pool->throttle);
2210 			dm_pool_issue_prefetches(pool->pmd);
2211 		}
2212 		cond_resched();
2213 	}
2214 	blk_finish_plug(&plug);
2215 }
2216 
2217 static int cmp_cells(const void *lhs, const void *rhs)
2218 {
2219 	struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2220 	struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2221 
2222 	BUG_ON(!lhs_cell->holder);
2223 	BUG_ON(!rhs_cell->holder);
2224 
2225 	if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2226 		return -1;
2227 
2228 	if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2229 		return 1;
2230 
2231 	return 0;
2232 }
2233 
2234 static unsigned int sort_cells(struct pool *pool, struct list_head *cells)
2235 {
2236 	unsigned int count = 0;
2237 	struct dm_bio_prison_cell *cell, *tmp;
2238 
2239 	list_for_each_entry_safe(cell, tmp, cells, user_list) {
2240 		if (count >= CELL_SORT_ARRAY_SIZE)
2241 			break;
2242 
2243 		pool->cell_sort_array[count++] = cell;
2244 		list_del(&cell->user_list);
2245 	}
2246 
2247 	sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2248 
2249 	return count;
2250 }
2251 
2252 static void process_thin_deferred_cells(struct thin_c *tc)
2253 {
2254 	struct pool *pool = tc->pool;
2255 	struct list_head cells;
2256 	struct dm_bio_prison_cell *cell;
2257 	unsigned int i, j, count;
2258 
2259 	INIT_LIST_HEAD(&cells);
2260 
2261 	spin_lock_irq(&tc->lock);
2262 	list_splice_init(&tc->deferred_cells, &cells);
2263 	spin_unlock_irq(&tc->lock);
2264 
2265 	if (list_empty(&cells))
2266 		return;
2267 
2268 	do {
2269 		count = sort_cells(tc->pool, &cells);
2270 
2271 		for (i = 0; i < count; i++) {
2272 			cell = pool->cell_sort_array[i];
2273 			BUG_ON(!cell->holder);
2274 
2275 			/*
2276 			 * If we've got no free new_mapping structs, and processing
2277 			 * this bio might require one, we pause until there are some
2278 			 * prepared mappings to process.
2279 			 */
2280 			if (ensure_next_mapping(pool)) {
2281 				for (j = i; j < count; j++)
2282 					list_add(&pool->cell_sort_array[j]->user_list, &cells);
2283 
2284 				spin_lock_irq(&tc->lock);
2285 				list_splice(&cells, &tc->deferred_cells);
2286 				spin_unlock_irq(&tc->lock);
2287 				return;
2288 			}
2289 
2290 			if (bio_op(cell->holder) == REQ_OP_DISCARD)
2291 				pool->process_discard_cell(tc, cell);
2292 			else
2293 				pool->process_cell(tc, cell);
2294 		}
2295 		cond_resched();
2296 	} while (!list_empty(&cells));
2297 }
2298 
2299 static void thin_get(struct thin_c *tc);
2300 static void thin_put(struct thin_c *tc);
2301 
2302 /*
2303  * We can't hold rcu_read_lock() around code that can block.  So we
2304  * find a thin with the rcu lock held; bump a refcount; then drop
2305  * the lock.
2306  */
2307 static struct thin_c *get_first_thin(struct pool *pool)
2308 {
2309 	struct thin_c *tc = NULL;
2310 
2311 	rcu_read_lock();
2312 	if (!list_empty(&pool->active_thins)) {
2313 		tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2314 		thin_get(tc);
2315 	}
2316 	rcu_read_unlock();
2317 
2318 	return tc;
2319 }
2320 
2321 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2322 {
2323 	struct thin_c *old_tc = tc;
2324 
2325 	rcu_read_lock();
2326 	list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2327 		thin_get(tc);
2328 		thin_put(old_tc);
2329 		rcu_read_unlock();
2330 		return tc;
2331 	}
2332 	thin_put(old_tc);
2333 	rcu_read_unlock();
2334 
2335 	return NULL;
2336 }
2337 
2338 static void process_deferred_bios(struct pool *pool)
2339 {
2340 	struct bio *bio;
2341 	struct bio_list bios, bio_completions;
2342 	struct thin_c *tc;
2343 
2344 	tc = get_first_thin(pool);
2345 	while (tc) {
2346 		process_thin_deferred_cells(tc);
2347 		process_thin_deferred_bios(tc);
2348 		tc = get_next_thin(pool, tc);
2349 	}
2350 
2351 	/*
2352 	 * If there are any deferred flush bios, we must commit the metadata
2353 	 * before issuing them or signaling their completion.
2354 	 */
2355 	bio_list_init(&bios);
2356 	bio_list_init(&bio_completions);
2357 
2358 	spin_lock_irq(&pool->lock);
2359 	bio_list_merge(&bios, &pool->deferred_flush_bios);
2360 	bio_list_init(&pool->deferred_flush_bios);
2361 
2362 	bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2363 	bio_list_init(&pool->deferred_flush_completions);
2364 	spin_unlock_irq(&pool->lock);
2365 
2366 	if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2367 	    !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2368 		return;
2369 
2370 	if (commit(pool)) {
2371 		bio_list_merge(&bios, &bio_completions);
2372 
2373 		while ((bio = bio_list_pop(&bios)))
2374 			bio_io_error(bio);
2375 		return;
2376 	}
2377 	pool->last_commit_jiffies = jiffies;
2378 
2379 	while ((bio = bio_list_pop(&bio_completions)))
2380 		bio_endio(bio);
2381 
2382 	while ((bio = bio_list_pop(&bios))) {
2383 		/*
2384 		 * The data device was flushed as part of metadata commit,
2385 		 * so complete redundant flushes immediately.
2386 		 */
2387 		if (bio->bi_opf & REQ_PREFLUSH)
2388 			bio_endio(bio);
2389 		else
2390 			dm_submit_bio_remap(bio, NULL);
2391 	}
2392 }
2393 
2394 static void do_worker(struct work_struct *ws)
2395 {
2396 	struct pool *pool = container_of(ws, struct pool, worker);
2397 
2398 	throttle_work_start(&pool->throttle);
2399 	dm_pool_issue_prefetches(pool->pmd);
2400 	throttle_work_update(&pool->throttle);
2401 	process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2402 	throttle_work_update(&pool->throttle);
2403 	process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2404 	throttle_work_update(&pool->throttle);
2405 	process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2406 	throttle_work_update(&pool->throttle);
2407 	process_deferred_bios(pool);
2408 	throttle_work_complete(&pool->throttle);
2409 }
2410 
2411 /*
2412  * We want to commit periodically so that not too much
2413  * unwritten data builds up.
2414  */
2415 static void do_waker(struct work_struct *ws)
2416 {
2417 	struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2418 
2419 	wake_worker(pool);
2420 	queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2421 }
2422 
2423 /*
2424  * We're holding onto IO to allow userland time to react.  After the
2425  * timeout either the pool will have been resized (and thus back in
2426  * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2427  */
2428 static void do_no_space_timeout(struct work_struct *ws)
2429 {
2430 	struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2431 					 no_space_timeout);
2432 
2433 	if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2434 		pool->pf.error_if_no_space = true;
2435 		notify_of_pool_mode_change(pool);
2436 		error_retry_list_with_code(pool, BLK_STS_NOSPC);
2437 	}
2438 }
2439 
2440 /*----------------------------------------------------------------*/
2441 
2442 struct pool_work {
2443 	struct work_struct worker;
2444 	struct completion complete;
2445 };
2446 
2447 static struct pool_work *to_pool_work(struct work_struct *ws)
2448 {
2449 	return container_of(ws, struct pool_work, worker);
2450 }
2451 
2452 static void pool_work_complete(struct pool_work *pw)
2453 {
2454 	complete(&pw->complete);
2455 }
2456 
2457 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2458 			   void (*fn)(struct work_struct *))
2459 {
2460 	INIT_WORK_ONSTACK(&pw->worker, fn);
2461 	init_completion(&pw->complete);
2462 	queue_work(pool->wq, &pw->worker);
2463 	wait_for_completion(&pw->complete);
2464 }
2465 
2466 /*----------------------------------------------------------------*/
2467 
2468 struct noflush_work {
2469 	struct pool_work pw;
2470 	struct thin_c *tc;
2471 };
2472 
2473 static struct noflush_work *to_noflush(struct work_struct *ws)
2474 {
2475 	return container_of(to_pool_work(ws), struct noflush_work, pw);
2476 }
2477 
2478 static void do_noflush_start(struct work_struct *ws)
2479 {
2480 	struct noflush_work *w = to_noflush(ws);
2481 
2482 	w->tc->requeue_mode = true;
2483 	requeue_io(w->tc);
2484 	pool_work_complete(&w->pw);
2485 }
2486 
2487 static void do_noflush_stop(struct work_struct *ws)
2488 {
2489 	struct noflush_work *w = to_noflush(ws);
2490 
2491 	w->tc->requeue_mode = false;
2492 	pool_work_complete(&w->pw);
2493 }
2494 
2495 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2496 {
2497 	struct noflush_work w;
2498 
2499 	w.tc = tc;
2500 	pool_work_wait(&w.pw, tc->pool, fn);
2501 }
2502 
2503 /*----------------------------------------------------------------*/
2504 
2505 static bool passdown_enabled(struct pool_c *pt)
2506 {
2507 	return pt->adjusted_pf.discard_passdown;
2508 }
2509 
2510 static void set_discard_callbacks(struct pool *pool)
2511 {
2512 	struct pool_c *pt = pool->ti->private;
2513 
2514 	if (passdown_enabled(pt)) {
2515 		pool->process_discard_cell = process_discard_cell_passdown;
2516 		pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2517 		pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2518 	} else {
2519 		pool->process_discard_cell = process_discard_cell_no_passdown;
2520 		pool->process_prepared_discard = process_prepared_discard_no_passdown;
2521 	}
2522 }
2523 
2524 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2525 {
2526 	struct pool_c *pt = pool->ti->private;
2527 	bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2528 	enum pool_mode old_mode = get_pool_mode(pool);
2529 	unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2530 
2531 	/*
2532 	 * Never allow the pool to transition to PM_WRITE mode if user
2533 	 * intervention is required to verify metadata and data consistency.
2534 	 */
2535 	if (new_mode == PM_WRITE && needs_check) {
2536 		DMERR("%s: unable to switch pool to write mode until repaired.",
2537 		      dm_device_name(pool->pool_md));
2538 		if (old_mode != new_mode)
2539 			new_mode = old_mode;
2540 		else
2541 			new_mode = PM_READ_ONLY;
2542 	}
2543 	/*
2544 	 * If we were in PM_FAIL mode, rollback of metadata failed.  We're
2545 	 * not going to recover without a thin_repair.	So we never let the
2546 	 * pool move out of the old mode.
2547 	 */
2548 	if (old_mode == PM_FAIL)
2549 		new_mode = old_mode;
2550 
2551 	switch (new_mode) {
2552 	case PM_FAIL:
2553 		dm_pool_metadata_read_only(pool->pmd);
2554 		pool->process_bio = process_bio_fail;
2555 		pool->process_discard = process_bio_fail;
2556 		pool->process_cell = process_cell_fail;
2557 		pool->process_discard_cell = process_cell_fail;
2558 		pool->process_prepared_mapping = process_prepared_mapping_fail;
2559 		pool->process_prepared_discard = process_prepared_discard_fail;
2560 
2561 		error_retry_list(pool);
2562 		break;
2563 
2564 	case PM_OUT_OF_METADATA_SPACE:
2565 	case PM_READ_ONLY:
2566 		dm_pool_metadata_read_only(pool->pmd);
2567 		pool->process_bio = process_bio_read_only;
2568 		pool->process_discard = process_bio_success;
2569 		pool->process_cell = process_cell_read_only;
2570 		pool->process_discard_cell = process_cell_success;
2571 		pool->process_prepared_mapping = process_prepared_mapping_fail;
2572 		pool->process_prepared_discard = process_prepared_discard_success;
2573 
2574 		error_retry_list(pool);
2575 		break;
2576 
2577 	case PM_OUT_OF_DATA_SPACE:
2578 		/*
2579 		 * Ideally we'd never hit this state; the low water mark
2580 		 * would trigger userland to extend the pool before we
2581 		 * completely run out of data space.  However, many small
2582 		 * IOs to unprovisioned space can consume data space at an
2583 		 * alarming rate.  Adjust your low water mark if you're
2584 		 * frequently seeing this mode.
2585 		 */
2586 		pool->out_of_data_space = true;
2587 		pool->process_bio = process_bio_read_only;
2588 		pool->process_discard = process_discard_bio;
2589 		pool->process_cell = process_cell_read_only;
2590 		pool->process_prepared_mapping = process_prepared_mapping;
2591 		set_discard_callbacks(pool);
2592 
2593 		if (!pool->pf.error_if_no_space && no_space_timeout)
2594 			queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2595 		break;
2596 
2597 	case PM_WRITE:
2598 		if (old_mode == PM_OUT_OF_DATA_SPACE)
2599 			cancel_delayed_work_sync(&pool->no_space_timeout);
2600 		pool->out_of_data_space = false;
2601 		pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2602 		dm_pool_metadata_read_write(pool->pmd);
2603 		pool->process_bio = process_bio;
2604 		pool->process_discard = process_discard_bio;
2605 		pool->process_cell = process_cell;
2606 		pool->process_prepared_mapping = process_prepared_mapping;
2607 		set_discard_callbacks(pool);
2608 		break;
2609 	}
2610 
2611 	pool->pf.mode = new_mode;
2612 	/*
2613 	 * The pool mode may have changed, sync it so bind_control_target()
2614 	 * doesn't cause an unexpected mode transition on resume.
2615 	 */
2616 	pt->adjusted_pf.mode = new_mode;
2617 
2618 	if (old_mode != new_mode)
2619 		notify_of_pool_mode_change(pool);
2620 }
2621 
2622 static void abort_transaction(struct pool *pool)
2623 {
2624 	const char *dev_name = dm_device_name(pool->pool_md);
2625 
2626 	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2627 	if (dm_pool_abort_metadata(pool->pmd)) {
2628 		DMERR("%s: failed to abort metadata transaction", dev_name);
2629 		set_pool_mode(pool, PM_FAIL);
2630 	}
2631 
2632 	if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2633 		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2634 		set_pool_mode(pool, PM_FAIL);
2635 	}
2636 }
2637 
2638 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2639 {
2640 	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2641 		    dm_device_name(pool->pool_md), op, r);
2642 
2643 	abort_transaction(pool);
2644 	set_pool_mode(pool, PM_READ_ONLY);
2645 }
2646 
2647 /*----------------------------------------------------------------*/
2648 
2649 /*
2650  * Mapping functions.
2651  */
2652 
2653 /*
2654  * Called only while mapping a thin bio to hand it over to the workqueue.
2655  */
2656 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2657 {
2658 	struct pool *pool = tc->pool;
2659 
2660 	spin_lock_irq(&tc->lock);
2661 	bio_list_add(&tc->deferred_bio_list, bio);
2662 	spin_unlock_irq(&tc->lock);
2663 
2664 	wake_worker(pool);
2665 }
2666 
2667 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2668 {
2669 	struct pool *pool = tc->pool;
2670 
2671 	throttle_lock(&pool->throttle);
2672 	thin_defer_bio(tc, bio);
2673 	throttle_unlock(&pool->throttle);
2674 }
2675 
2676 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2677 {
2678 	struct pool *pool = tc->pool;
2679 
2680 	throttle_lock(&pool->throttle);
2681 	spin_lock_irq(&tc->lock);
2682 	list_add_tail(&cell->user_list, &tc->deferred_cells);
2683 	spin_unlock_irq(&tc->lock);
2684 	throttle_unlock(&pool->throttle);
2685 
2686 	wake_worker(pool);
2687 }
2688 
2689 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2690 {
2691 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2692 
2693 	h->tc = tc;
2694 	h->shared_read_entry = NULL;
2695 	h->all_io_entry = NULL;
2696 	h->overwrite_mapping = NULL;
2697 	h->cell = NULL;
2698 }
2699 
2700 /*
2701  * Non-blocking function called from the thin target's map function.
2702  */
2703 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2704 {
2705 	int r;
2706 	struct thin_c *tc = ti->private;
2707 	dm_block_t block = get_bio_block(tc, bio);
2708 	struct dm_thin_device *td = tc->td;
2709 	struct dm_thin_lookup_result result;
2710 	struct dm_bio_prison_cell *virt_cell, *data_cell;
2711 	struct dm_cell_key key;
2712 
2713 	thin_hook_bio(tc, bio);
2714 
2715 	if (tc->requeue_mode) {
2716 		bio->bi_status = BLK_STS_DM_REQUEUE;
2717 		bio_endio(bio);
2718 		return DM_MAPIO_SUBMITTED;
2719 	}
2720 
2721 	if (get_pool_mode(tc->pool) == PM_FAIL) {
2722 		bio_io_error(bio);
2723 		return DM_MAPIO_SUBMITTED;
2724 	}
2725 
2726 	if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2727 		thin_defer_bio_with_throttle(tc, bio);
2728 		return DM_MAPIO_SUBMITTED;
2729 	}
2730 
2731 	/*
2732 	 * We must hold the virtual cell before doing the lookup, otherwise
2733 	 * there's a race with discard.
2734 	 */
2735 	build_virtual_key(tc->td, block, &key);
2736 	if (bio_detain(tc->pool, &key, bio, &virt_cell))
2737 		return DM_MAPIO_SUBMITTED;
2738 
2739 	r = dm_thin_find_block(td, block, 0, &result);
2740 
2741 	/*
2742 	 * Note that we defer readahead too.
2743 	 */
2744 	switch (r) {
2745 	case 0:
2746 		if (unlikely(result.shared)) {
2747 			/*
2748 			 * We have a race condition here between the
2749 			 * result.shared value returned by the lookup and
2750 			 * snapshot creation, which may cause new
2751 			 * sharing.
2752 			 *
2753 			 * To avoid this always quiesce the origin before
2754 			 * taking the snap.  You want to do this anyway to
2755 			 * ensure a consistent application view
2756 			 * (i.e. lockfs).
2757 			 *
2758 			 * More distant ancestors are irrelevant. The
2759 			 * shared flag will be set in their case.
2760 			 */
2761 			thin_defer_cell(tc, virt_cell);
2762 			return DM_MAPIO_SUBMITTED;
2763 		}
2764 
2765 		build_data_key(tc->td, result.block, &key);
2766 		if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2767 			cell_defer_no_holder(tc, virt_cell);
2768 			return DM_MAPIO_SUBMITTED;
2769 		}
2770 
2771 		inc_all_io_entry(tc->pool, bio);
2772 		cell_defer_no_holder(tc, data_cell);
2773 		cell_defer_no_holder(tc, virt_cell);
2774 
2775 		remap(tc, bio, result.block);
2776 		return DM_MAPIO_REMAPPED;
2777 
2778 	case -ENODATA:
2779 	case -EWOULDBLOCK:
2780 		thin_defer_cell(tc, virt_cell);
2781 		return DM_MAPIO_SUBMITTED;
2782 
2783 	default:
2784 		/*
2785 		 * Must always call bio_io_error on failure.
2786 		 * dm_thin_find_block can fail with -EINVAL if the
2787 		 * pool is switched to fail-io mode.
2788 		 */
2789 		bio_io_error(bio);
2790 		cell_defer_no_holder(tc, virt_cell);
2791 		return DM_MAPIO_SUBMITTED;
2792 	}
2793 }
2794 
2795 static void requeue_bios(struct pool *pool)
2796 {
2797 	struct thin_c *tc;
2798 
2799 	rcu_read_lock();
2800 	list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2801 		spin_lock_irq(&tc->lock);
2802 		bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2803 		bio_list_init(&tc->retry_on_resume_list);
2804 		spin_unlock_irq(&tc->lock);
2805 	}
2806 	rcu_read_unlock();
2807 }
2808 
2809 /*
2810  *--------------------------------------------------------------
2811  * Binding of control targets to a pool object
2812  *--------------------------------------------------------------
2813  */
2814 static bool is_factor(sector_t block_size, uint32_t n)
2815 {
2816 	return !sector_div(block_size, n);
2817 }
2818 
2819 /*
2820  * If discard_passdown was enabled verify that the data device
2821  * supports discards.  Disable discard_passdown if not.
2822  */
2823 static void disable_passdown_if_not_supported(struct pool_c *pt)
2824 {
2825 	struct pool *pool = pt->pool;
2826 	struct block_device *data_bdev = pt->data_dev->bdev;
2827 	struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2828 	const char *reason = NULL;
2829 
2830 	if (!pt->adjusted_pf.discard_passdown)
2831 		return;
2832 
2833 	if (!bdev_max_discard_sectors(pt->data_dev->bdev))
2834 		reason = "discard unsupported";
2835 
2836 	else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2837 		reason = "max discard sectors smaller than a block";
2838 
2839 	if (reason) {
2840 		DMWARN("Data device (%pg) %s: Disabling discard passdown.", data_bdev, reason);
2841 		pt->adjusted_pf.discard_passdown = false;
2842 	}
2843 }
2844 
2845 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2846 {
2847 	struct pool_c *pt = ti->private;
2848 
2849 	/*
2850 	 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2851 	 */
2852 	enum pool_mode old_mode = get_pool_mode(pool);
2853 	enum pool_mode new_mode = pt->adjusted_pf.mode;
2854 
2855 	/*
2856 	 * Don't change the pool's mode until set_pool_mode() below.
2857 	 * Otherwise the pool's process_* function pointers may
2858 	 * not match the desired pool mode.
2859 	 */
2860 	pt->adjusted_pf.mode = old_mode;
2861 
2862 	pool->ti = ti;
2863 	pool->pf = pt->adjusted_pf;
2864 	pool->low_water_blocks = pt->low_water_blocks;
2865 
2866 	set_pool_mode(pool, new_mode);
2867 
2868 	return 0;
2869 }
2870 
2871 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2872 {
2873 	if (pool->ti == ti)
2874 		pool->ti = NULL;
2875 }
2876 
2877 /*
2878  *--------------------------------------------------------------
2879  * Pool creation
2880  *--------------------------------------------------------------
2881  */
2882 /* Initialize pool features. */
2883 static void pool_features_init(struct pool_features *pf)
2884 {
2885 	pf->mode = PM_WRITE;
2886 	pf->zero_new_blocks = true;
2887 	pf->discard_enabled = true;
2888 	pf->discard_passdown = true;
2889 	pf->error_if_no_space = false;
2890 }
2891 
2892 static void __pool_destroy(struct pool *pool)
2893 {
2894 	__pool_table_remove(pool);
2895 
2896 	vfree(pool->cell_sort_array);
2897 	if (dm_pool_metadata_close(pool->pmd) < 0)
2898 		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2899 
2900 	dm_bio_prison_destroy(pool->prison);
2901 	dm_kcopyd_client_destroy(pool->copier);
2902 
2903 	cancel_delayed_work_sync(&pool->waker);
2904 	cancel_delayed_work_sync(&pool->no_space_timeout);
2905 	if (pool->wq)
2906 		destroy_workqueue(pool->wq);
2907 
2908 	if (pool->next_mapping)
2909 		mempool_free(pool->next_mapping, &pool->mapping_pool);
2910 	mempool_exit(&pool->mapping_pool);
2911 	dm_deferred_set_destroy(pool->shared_read_ds);
2912 	dm_deferred_set_destroy(pool->all_io_ds);
2913 	kfree(pool);
2914 }
2915 
2916 static struct kmem_cache *_new_mapping_cache;
2917 
2918 static struct pool *pool_create(struct mapped_device *pool_md,
2919 				struct block_device *metadata_dev,
2920 				struct block_device *data_dev,
2921 				unsigned long block_size,
2922 				int read_only, char **error)
2923 {
2924 	int r;
2925 	void *err_p;
2926 	struct pool *pool;
2927 	struct dm_pool_metadata *pmd;
2928 	bool format_device = read_only ? false : true;
2929 
2930 	pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2931 	if (IS_ERR(pmd)) {
2932 		*error = "Error creating metadata object";
2933 		return (struct pool *)pmd;
2934 	}
2935 
2936 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2937 	if (!pool) {
2938 		*error = "Error allocating memory for pool";
2939 		err_p = ERR_PTR(-ENOMEM);
2940 		goto bad_pool;
2941 	}
2942 
2943 	pool->pmd = pmd;
2944 	pool->sectors_per_block = block_size;
2945 	if (block_size & (block_size - 1))
2946 		pool->sectors_per_block_shift = -1;
2947 	else
2948 		pool->sectors_per_block_shift = __ffs(block_size);
2949 	pool->low_water_blocks = 0;
2950 	pool_features_init(&pool->pf);
2951 	pool->prison = dm_bio_prison_create();
2952 	if (!pool->prison) {
2953 		*error = "Error creating pool's bio prison";
2954 		err_p = ERR_PTR(-ENOMEM);
2955 		goto bad_prison;
2956 	}
2957 
2958 	pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2959 	if (IS_ERR(pool->copier)) {
2960 		r = PTR_ERR(pool->copier);
2961 		*error = "Error creating pool's kcopyd client";
2962 		err_p = ERR_PTR(r);
2963 		goto bad_kcopyd_client;
2964 	}
2965 
2966 	/*
2967 	 * Create singlethreaded workqueue that will service all devices
2968 	 * that use this metadata.
2969 	 */
2970 	pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2971 	if (!pool->wq) {
2972 		*error = "Error creating pool's workqueue";
2973 		err_p = ERR_PTR(-ENOMEM);
2974 		goto bad_wq;
2975 	}
2976 
2977 	throttle_init(&pool->throttle);
2978 	INIT_WORK(&pool->worker, do_worker);
2979 	INIT_DELAYED_WORK(&pool->waker, do_waker);
2980 	INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2981 	spin_lock_init(&pool->lock);
2982 	bio_list_init(&pool->deferred_flush_bios);
2983 	bio_list_init(&pool->deferred_flush_completions);
2984 	INIT_LIST_HEAD(&pool->prepared_mappings);
2985 	INIT_LIST_HEAD(&pool->prepared_discards);
2986 	INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2987 	INIT_LIST_HEAD(&pool->active_thins);
2988 	pool->low_water_triggered = false;
2989 	pool->suspended = true;
2990 	pool->out_of_data_space = false;
2991 
2992 	pool->shared_read_ds = dm_deferred_set_create();
2993 	if (!pool->shared_read_ds) {
2994 		*error = "Error creating pool's shared read deferred set";
2995 		err_p = ERR_PTR(-ENOMEM);
2996 		goto bad_shared_read_ds;
2997 	}
2998 
2999 	pool->all_io_ds = dm_deferred_set_create();
3000 	if (!pool->all_io_ds) {
3001 		*error = "Error creating pool's all io deferred set";
3002 		err_p = ERR_PTR(-ENOMEM);
3003 		goto bad_all_io_ds;
3004 	}
3005 
3006 	pool->next_mapping = NULL;
3007 	r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
3008 				   _new_mapping_cache);
3009 	if (r) {
3010 		*error = "Error creating pool's mapping mempool";
3011 		err_p = ERR_PTR(r);
3012 		goto bad_mapping_pool;
3013 	}
3014 
3015 	pool->cell_sort_array =
3016 		vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3017 				   sizeof(*pool->cell_sort_array)));
3018 	if (!pool->cell_sort_array) {
3019 		*error = "Error allocating cell sort array";
3020 		err_p = ERR_PTR(-ENOMEM);
3021 		goto bad_sort_array;
3022 	}
3023 
3024 	pool->ref_count = 1;
3025 	pool->last_commit_jiffies = jiffies;
3026 	pool->pool_md = pool_md;
3027 	pool->md_dev = metadata_dev;
3028 	pool->data_dev = data_dev;
3029 	__pool_table_insert(pool);
3030 
3031 	return pool;
3032 
3033 bad_sort_array:
3034 	mempool_exit(&pool->mapping_pool);
3035 bad_mapping_pool:
3036 	dm_deferred_set_destroy(pool->all_io_ds);
3037 bad_all_io_ds:
3038 	dm_deferred_set_destroy(pool->shared_read_ds);
3039 bad_shared_read_ds:
3040 	destroy_workqueue(pool->wq);
3041 bad_wq:
3042 	dm_kcopyd_client_destroy(pool->copier);
3043 bad_kcopyd_client:
3044 	dm_bio_prison_destroy(pool->prison);
3045 bad_prison:
3046 	kfree(pool);
3047 bad_pool:
3048 	if (dm_pool_metadata_close(pmd))
3049 		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3050 
3051 	return err_p;
3052 }
3053 
3054 static void __pool_inc(struct pool *pool)
3055 {
3056 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3057 	pool->ref_count++;
3058 }
3059 
3060 static void __pool_dec(struct pool *pool)
3061 {
3062 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3063 	BUG_ON(!pool->ref_count);
3064 	if (!--pool->ref_count)
3065 		__pool_destroy(pool);
3066 }
3067 
3068 static struct pool *__pool_find(struct mapped_device *pool_md,
3069 				struct block_device *metadata_dev,
3070 				struct block_device *data_dev,
3071 				unsigned long block_size, int read_only,
3072 				char **error, int *created)
3073 {
3074 	struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3075 
3076 	if (pool) {
3077 		if (pool->pool_md != pool_md) {
3078 			*error = "metadata device already in use by a pool";
3079 			return ERR_PTR(-EBUSY);
3080 		}
3081 		if (pool->data_dev != data_dev) {
3082 			*error = "data device already in use by a pool";
3083 			return ERR_PTR(-EBUSY);
3084 		}
3085 		__pool_inc(pool);
3086 
3087 	} else {
3088 		pool = __pool_table_lookup(pool_md);
3089 		if (pool) {
3090 			if (pool->md_dev != metadata_dev || pool->data_dev != data_dev) {
3091 				*error = "different pool cannot replace a pool";
3092 				return ERR_PTR(-EINVAL);
3093 			}
3094 			__pool_inc(pool);
3095 
3096 		} else {
3097 			pool = pool_create(pool_md, metadata_dev, data_dev, block_size, read_only, error);
3098 			*created = 1;
3099 		}
3100 	}
3101 
3102 	return pool;
3103 }
3104 
3105 /*
3106  *--------------------------------------------------------------
3107  * Pool target methods
3108  *--------------------------------------------------------------
3109  */
3110 static void pool_dtr(struct dm_target *ti)
3111 {
3112 	struct pool_c *pt = ti->private;
3113 
3114 	mutex_lock(&dm_thin_pool_table.mutex);
3115 
3116 	unbind_control_target(pt->pool, ti);
3117 	__pool_dec(pt->pool);
3118 	dm_put_device(ti, pt->metadata_dev);
3119 	dm_put_device(ti, pt->data_dev);
3120 	kfree(pt);
3121 
3122 	mutex_unlock(&dm_thin_pool_table.mutex);
3123 }
3124 
3125 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3126 			       struct dm_target *ti)
3127 {
3128 	int r;
3129 	unsigned int argc;
3130 	const char *arg_name;
3131 
3132 	static const struct dm_arg _args[] = {
3133 		{0, 4, "Invalid number of pool feature arguments"},
3134 	};
3135 
3136 	/*
3137 	 * No feature arguments supplied.
3138 	 */
3139 	if (!as->argc)
3140 		return 0;
3141 
3142 	r = dm_read_arg_group(_args, as, &argc, &ti->error);
3143 	if (r)
3144 		return -EINVAL;
3145 
3146 	while (argc && !r) {
3147 		arg_name = dm_shift_arg(as);
3148 		argc--;
3149 
3150 		if (!strcasecmp(arg_name, "skip_block_zeroing"))
3151 			pf->zero_new_blocks = false;
3152 
3153 		else if (!strcasecmp(arg_name, "ignore_discard"))
3154 			pf->discard_enabled = false;
3155 
3156 		else if (!strcasecmp(arg_name, "no_discard_passdown"))
3157 			pf->discard_passdown = false;
3158 
3159 		else if (!strcasecmp(arg_name, "read_only"))
3160 			pf->mode = PM_READ_ONLY;
3161 
3162 		else if (!strcasecmp(arg_name, "error_if_no_space"))
3163 			pf->error_if_no_space = true;
3164 
3165 		else {
3166 			ti->error = "Unrecognised pool feature requested";
3167 			r = -EINVAL;
3168 			break;
3169 		}
3170 	}
3171 
3172 	return r;
3173 }
3174 
3175 static void metadata_low_callback(void *context)
3176 {
3177 	struct pool *pool = context;
3178 
3179 	DMWARN("%s: reached low water mark for metadata device: sending event.",
3180 	       dm_device_name(pool->pool_md));
3181 
3182 	dm_table_event(pool->ti->table);
3183 }
3184 
3185 /*
3186  * We need to flush the data device **before** committing the metadata.
3187  *
3188  * This ensures that the data blocks of any newly inserted mappings are
3189  * properly written to non-volatile storage and won't be lost in case of a
3190  * crash.
3191  *
3192  * Failure to do so can result in data corruption in the case of internal or
3193  * external snapshots and in the case of newly provisioned blocks, when block
3194  * zeroing is enabled.
3195  */
3196 static int metadata_pre_commit_callback(void *context)
3197 {
3198 	struct pool *pool = context;
3199 
3200 	return blkdev_issue_flush(pool->data_dev);
3201 }
3202 
3203 static sector_t get_dev_size(struct block_device *bdev)
3204 {
3205 	return bdev_nr_sectors(bdev);
3206 }
3207 
3208 static void warn_if_metadata_device_too_big(struct block_device *bdev)
3209 {
3210 	sector_t metadata_dev_size = get_dev_size(bdev);
3211 
3212 	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3213 		DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
3214 		       bdev, THIN_METADATA_MAX_SECTORS);
3215 }
3216 
3217 static sector_t get_metadata_dev_size(struct block_device *bdev)
3218 {
3219 	sector_t metadata_dev_size = get_dev_size(bdev);
3220 
3221 	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3222 		metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3223 
3224 	return metadata_dev_size;
3225 }
3226 
3227 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3228 {
3229 	sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3230 
3231 	sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3232 
3233 	return metadata_dev_size;
3234 }
3235 
3236 /*
3237  * When a metadata threshold is crossed a dm event is triggered, and
3238  * userland should respond by growing the metadata device.  We could let
3239  * userland set the threshold, like we do with the data threshold, but I'm
3240  * not sure they know enough to do this well.
3241  */
3242 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3243 {
3244 	/*
3245 	 * 4M is ample for all ops with the possible exception of thin
3246 	 * device deletion which is harmless if it fails (just retry the
3247 	 * delete after you've grown the device).
3248 	 */
3249 	dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3250 
3251 	return min((dm_block_t)1024ULL /* 4M */, quarter);
3252 }
3253 
3254 /*
3255  * thin-pool <metadata dev> <data dev>
3256  *	     <data block size (sectors)>
3257  *	     <low water mark (blocks)>
3258  *	     [<#feature args> [<arg>]*]
3259  *
3260  * Optional feature arguments are:
3261  *	     skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3262  *	     ignore_discard: disable discard
3263  *	     no_discard_passdown: don't pass discards down to the data device
3264  *	     read_only: Don't allow any changes to be made to the pool metadata.
3265  *	     error_if_no_space: error IOs, instead of queueing, if no space.
3266  */
3267 static int pool_ctr(struct dm_target *ti, unsigned int argc, char **argv)
3268 {
3269 	int r, pool_created = 0;
3270 	struct pool_c *pt;
3271 	struct pool *pool;
3272 	struct pool_features pf;
3273 	struct dm_arg_set as;
3274 	struct dm_dev *data_dev;
3275 	unsigned long block_size;
3276 	dm_block_t low_water_blocks;
3277 	struct dm_dev *metadata_dev;
3278 	fmode_t metadata_mode;
3279 
3280 	/*
3281 	 * FIXME Remove validation from scope of lock.
3282 	 */
3283 	mutex_lock(&dm_thin_pool_table.mutex);
3284 
3285 	if (argc < 4) {
3286 		ti->error = "Invalid argument count";
3287 		r = -EINVAL;
3288 		goto out_unlock;
3289 	}
3290 
3291 	as.argc = argc;
3292 	as.argv = argv;
3293 
3294 	/* make sure metadata and data are different devices */
3295 	if (!strcmp(argv[0], argv[1])) {
3296 		ti->error = "Error setting metadata or data device";
3297 		r = -EINVAL;
3298 		goto out_unlock;
3299 	}
3300 
3301 	/*
3302 	 * Set default pool features.
3303 	 */
3304 	pool_features_init(&pf);
3305 
3306 	dm_consume_args(&as, 4);
3307 	r = parse_pool_features(&as, &pf, ti);
3308 	if (r)
3309 		goto out_unlock;
3310 
3311 	metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3312 	r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3313 	if (r) {
3314 		ti->error = "Error opening metadata block device";
3315 		goto out_unlock;
3316 	}
3317 	warn_if_metadata_device_too_big(metadata_dev->bdev);
3318 
3319 	r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3320 	if (r) {
3321 		ti->error = "Error getting data device";
3322 		goto out_metadata;
3323 	}
3324 
3325 	if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3326 	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3327 	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3328 	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3329 		ti->error = "Invalid block size";
3330 		r = -EINVAL;
3331 		goto out;
3332 	}
3333 
3334 	if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3335 		ti->error = "Invalid low water mark";
3336 		r = -EINVAL;
3337 		goto out;
3338 	}
3339 
3340 	pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3341 	if (!pt) {
3342 		r = -ENOMEM;
3343 		goto out;
3344 	}
3345 
3346 	pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev, data_dev->bdev,
3347 			   block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3348 	if (IS_ERR(pool)) {
3349 		r = PTR_ERR(pool);
3350 		goto out_free_pt;
3351 	}
3352 
3353 	/*
3354 	 * 'pool_created' reflects whether this is the first table load.
3355 	 * Top level discard support is not allowed to be changed after
3356 	 * initial load.  This would require a pool reload to trigger thin
3357 	 * device changes.
3358 	 */
3359 	if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3360 		ti->error = "Discard support cannot be disabled once enabled";
3361 		r = -EINVAL;
3362 		goto out_flags_changed;
3363 	}
3364 
3365 	pt->pool = pool;
3366 	pt->ti = ti;
3367 	pt->metadata_dev = metadata_dev;
3368 	pt->data_dev = data_dev;
3369 	pt->low_water_blocks = low_water_blocks;
3370 	pt->adjusted_pf = pt->requested_pf = pf;
3371 	ti->num_flush_bios = 1;
3372 
3373 	/*
3374 	 * Only need to enable discards if the pool should pass
3375 	 * them down to the data device.  The thin device's discard
3376 	 * processing will cause mappings to be removed from the btree.
3377 	 */
3378 	if (pf.discard_enabled && pf.discard_passdown) {
3379 		ti->num_discard_bios = 1;
3380 
3381 		/*
3382 		 * Setting 'discards_supported' circumvents the normal
3383 		 * stacking of discard limits (this keeps the pool and
3384 		 * thin devices' discard limits consistent).
3385 		 */
3386 		ti->discards_supported = true;
3387 	}
3388 	ti->private = pt;
3389 
3390 	r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3391 						calc_metadata_threshold(pt),
3392 						metadata_low_callback,
3393 						pool);
3394 	if (r) {
3395 		ti->error = "Error registering metadata threshold";
3396 		goto out_flags_changed;
3397 	}
3398 
3399 	dm_pool_register_pre_commit_callback(pool->pmd,
3400 					     metadata_pre_commit_callback, pool);
3401 
3402 	mutex_unlock(&dm_thin_pool_table.mutex);
3403 
3404 	return 0;
3405 
3406 out_flags_changed:
3407 	__pool_dec(pool);
3408 out_free_pt:
3409 	kfree(pt);
3410 out:
3411 	dm_put_device(ti, data_dev);
3412 out_metadata:
3413 	dm_put_device(ti, metadata_dev);
3414 out_unlock:
3415 	mutex_unlock(&dm_thin_pool_table.mutex);
3416 
3417 	return r;
3418 }
3419 
3420 static int pool_map(struct dm_target *ti, struct bio *bio)
3421 {
3422 	int r;
3423 	struct pool_c *pt = ti->private;
3424 	struct pool *pool = pt->pool;
3425 
3426 	/*
3427 	 * As this is a singleton target, ti->begin is always zero.
3428 	 */
3429 	spin_lock_irq(&pool->lock);
3430 	bio_set_dev(bio, pt->data_dev->bdev);
3431 	r = DM_MAPIO_REMAPPED;
3432 	spin_unlock_irq(&pool->lock);
3433 
3434 	return r;
3435 }
3436 
3437 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3438 {
3439 	int r;
3440 	struct pool_c *pt = ti->private;
3441 	struct pool *pool = pt->pool;
3442 	sector_t data_size = ti->len;
3443 	dm_block_t sb_data_size;
3444 
3445 	*need_commit = false;
3446 
3447 	(void) sector_div(data_size, pool->sectors_per_block);
3448 
3449 	r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3450 	if (r) {
3451 		DMERR("%s: failed to retrieve data device size",
3452 		      dm_device_name(pool->pool_md));
3453 		return r;
3454 	}
3455 
3456 	if (data_size < sb_data_size) {
3457 		DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3458 		      dm_device_name(pool->pool_md),
3459 		      (unsigned long long)data_size, sb_data_size);
3460 		return -EINVAL;
3461 
3462 	} else if (data_size > sb_data_size) {
3463 		if (dm_pool_metadata_needs_check(pool->pmd)) {
3464 			DMERR("%s: unable to grow the data device until repaired.",
3465 			      dm_device_name(pool->pool_md));
3466 			return 0;
3467 		}
3468 
3469 		if (sb_data_size)
3470 			DMINFO("%s: growing the data device from %llu to %llu blocks",
3471 			       dm_device_name(pool->pool_md),
3472 			       sb_data_size, (unsigned long long)data_size);
3473 		r = dm_pool_resize_data_dev(pool->pmd, data_size);
3474 		if (r) {
3475 			metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3476 			return r;
3477 		}
3478 
3479 		*need_commit = true;
3480 	}
3481 
3482 	return 0;
3483 }
3484 
3485 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3486 {
3487 	int r;
3488 	struct pool_c *pt = ti->private;
3489 	struct pool *pool = pt->pool;
3490 	dm_block_t metadata_dev_size, sb_metadata_dev_size;
3491 
3492 	*need_commit = false;
3493 
3494 	metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3495 
3496 	r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3497 	if (r) {
3498 		DMERR("%s: failed to retrieve metadata device size",
3499 		      dm_device_name(pool->pool_md));
3500 		return r;
3501 	}
3502 
3503 	if (metadata_dev_size < sb_metadata_dev_size) {
3504 		DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3505 		      dm_device_name(pool->pool_md),
3506 		      metadata_dev_size, sb_metadata_dev_size);
3507 		return -EINVAL;
3508 
3509 	} else if (metadata_dev_size > sb_metadata_dev_size) {
3510 		if (dm_pool_metadata_needs_check(pool->pmd)) {
3511 			DMERR("%s: unable to grow the metadata device until repaired.",
3512 			      dm_device_name(pool->pool_md));
3513 			return 0;
3514 		}
3515 
3516 		warn_if_metadata_device_too_big(pool->md_dev);
3517 		DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3518 		       dm_device_name(pool->pool_md),
3519 		       sb_metadata_dev_size, metadata_dev_size);
3520 
3521 		if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3522 			set_pool_mode(pool, PM_WRITE);
3523 
3524 		r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3525 		if (r) {
3526 			metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3527 			return r;
3528 		}
3529 
3530 		*need_commit = true;
3531 	}
3532 
3533 	return 0;
3534 }
3535 
3536 /*
3537  * Retrieves the number of blocks of the data device from
3538  * the superblock and compares it to the actual device size,
3539  * thus resizing the data device in case it has grown.
3540  *
3541  * This both copes with opening preallocated data devices in the ctr
3542  * being followed by a resume
3543  * -and-
3544  * calling the resume method individually after userspace has
3545  * grown the data device in reaction to a table event.
3546  */
3547 static int pool_preresume(struct dm_target *ti)
3548 {
3549 	int r;
3550 	bool need_commit1, need_commit2;
3551 	struct pool_c *pt = ti->private;
3552 	struct pool *pool = pt->pool;
3553 
3554 	/*
3555 	 * Take control of the pool object.
3556 	 */
3557 	r = bind_control_target(pool, ti);
3558 	if (r)
3559 		goto out;
3560 
3561 	r = maybe_resize_data_dev(ti, &need_commit1);
3562 	if (r)
3563 		goto out;
3564 
3565 	r = maybe_resize_metadata_dev(ti, &need_commit2);
3566 	if (r)
3567 		goto out;
3568 
3569 	if (need_commit1 || need_commit2)
3570 		(void) commit(pool);
3571 out:
3572 	/*
3573 	 * When a thin-pool is PM_FAIL, it cannot be rebuilt if
3574 	 * bio is in deferred list. Therefore need to return 0
3575 	 * to allow pool_resume() to flush IO.
3576 	 */
3577 	if (r && get_pool_mode(pool) == PM_FAIL)
3578 		r = 0;
3579 
3580 	return r;
3581 }
3582 
3583 static void pool_suspend_active_thins(struct pool *pool)
3584 {
3585 	struct thin_c *tc;
3586 
3587 	/* Suspend all active thin devices */
3588 	tc = get_first_thin(pool);
3589 	while (tc) {
3590 		dm_internal_suspend_noflush(tc->thin_md);
3591 		tc = get_next_thin(pool, tc);
3592 	}
3593 }
3594 
3595 static void pool_resume_active_thins(struct pool *pool)
3596 {
3597 	struct thin_c *tc;
3598 
3599 	/* Resume all active thin devices */
3600 	tc = get_first_thin(pool);
3601 	while (tc) {
3602 		dm_internal_resume(tc->thin_md);
3603 		tc = get_next_thin(pool, tc);
3604 	}
3605 }
3606 
3607 static void pool_resume(struct dm_target *ti)
3608 {
3609 	struct pool_c *pt = ti->private;
3610 	struct pool *pool = pt->pool;
3611 
3612 	/*
3613 	 * Must requeue active_thins' bios and then resume
3614 	 * active_thins _before_ clearing 'suspend' flag.
3615 	 */
3616 	requeue_bios(pool);
3617 	pool_resume_active_thins(pool);
3618 
3619 	spin_lock_irq(&pool->lock);
3620 	pool->low_water_triggered = false;
3621 	pool->suspended = false;
3622 	spin_unlock_irq(&pool->lock);
3623 
3624 	do_waker(&pool->waker.work);
3625 }
3626 
3627 static void pool_presuspend(struct dm_target *ti)
3628 {
3629 	struct pool_c *pt = ti->private;
3630 	struct pool *pool = pt->pool;
3631 
3632 	spin_lock_irq(&pool->lock);
3633 	pool->suspended = true;
3634 	spin_unlock_irq(&pool->lock);
3635 
3636 	pool_suspend_active_thins(pool);
3637 }
3638 
3639 static void pool_presuspend_undo(struct dm_target *ti)
3640 {
3641 	struct pool_c *pt = ti->private;
3642 	struct pool *pool = pt->pool;
3643 
3644 	pool_resume_active_thins(pool);
3645 
3646 	spin_lock_irq(&pool->lock);
3647 	pool->suspended = false;
3648 	spin_unlock_irq(&pool->lock);
3649 }
3650 
3651 static void pool_postsuspend(struct dm_target *ti)
3652 {
3653 	struct pool_c *pt = ti->private;
3654 	struct pool *pool = pt->pool;
3655 
3656 	cancel_delayed_work_sync(&pool->waker);
3657 	cancel_delayed_work_sync(&pool->no_space_timeout);
3658 	flush_workqueue(pool->wq);
3659 	(void) commit(pool);
3660 }
3661 
3662 static int check_arg_count(unsigned int argc, unsigned int args_required)
3663 {
3664 	if (argc != args_required) {
3665 		DMWARN("Message received with %u arguments instead of %u.",
3666 		       argc, args_required);
3667 		return -EINVAL;
3668 	}
3669 
3670 	return 0;
3671 }
3672 
3673 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3674 {
3675 	if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3676 	    *dev_id <= MAX_DEV_ID)
3677 		return 0;
3678 
3679 	if (warning)
3680 		DMWARN("Message received with invalid device id: %s", arg);
3681 
3682 	return -EINVAL;
3683 }
3684 
3685 static int process_create_thin_mesg(unsigned int argc, char **argv, struct pool *pool)
3686 {
3687 	dm_thin_id dev_id;
3688 	int r;
3689 
3690 	r = check_arg_count(argc, 2);
3691 	if (r)
3692 		return r;
3693 
3694 	r = read_dev_id(argv[1], &dev_id, 1);
3695 	if (r)
3696 		return r;
3697 
3698 	r = dm_pool_create_thin(pool->pmd, dev_id);
3699 	if (r) {
3700 		DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3701 		       argv[1]);
3702 		return r;
3703 	}
3704 
3705 	return 0;
3706 }
3707 
3708 static int process_create_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3709 {
3710 	dm_thin_id dev_id;
3711 	dm_thin_id origin_dev_id;
3712 	int r;
3713 
3714 	r = check_arg_count(argc, 3);
3715 	if (r)
3716 		return r;
3717 
3718 	r = read_dev_id(argv[1], &dev_id, 1);
3719 	if (r)
3720 		return r;
3721 
3722 	r = read_dev_id(argv[2], &origin_dev_id, 1);
3723 	if (r)
3724 		return r;
3725 
3726 	r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3727 	if (r) {
3728 		DMWARN("Creation of new snapshot %s of device %s failed.",
3729 		       argv[1], argv[2]);
3730 		return r;
3731 	}
3732 
3733 	return 0;
3734 }
3735 
3736 static int process_delete_mesg(unsigned int argc, char **argv, struct pool *pool)
3737 {
3738 	dm_thin_id dev_id;
3739 	int r;
3740 
3741 	r = check_arg_count(argc, 2);
3742 	if (r)
3743 		return r;
3744 
3745 	r = read_dev_id(argv[1], &dev_id, 1);
3746 	if (r)
3747 		return r;
3748 
3749 	r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3750 	if (r)
3751 		DMWARN("Deletion of thin device %s failed.", argv[1]);
3752 
3753 	return r;
3754 }
3755 
3756 static int process_set_transaction_id_mesg(unsigned int argc, char **argv, struct pool *pool)
3757 {
3758 	dm_thin_id old_id, new_id;
3759 	int r;
3760 
3761 	r = check_arg_count(argc, 3);
3762 	if (r)
3763 		return r;
3764 
3765 	if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3766 		DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3767 		return -EINVAL;
3768 	}
3769 
3770 	if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3771 		DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3772 		return -EINVAL;
3773 	}
3774 
3775 	r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3776 	if (r) {
3777 		DMWARN("Failed to change transaction id from %s to %s.",
3778 		       argv[1], argv[2]);
3779 		return r;
3780 	}
3781 
3782 	return 0;
3783 }
3784 
3785 static int process_reserve_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3786 {
3787 	int r;
3788 
3789 	r = check_arg_count(argc, 1);
3790 	if (r)
3791 		return r;
3792 
3793 	(void) commit(pool);
3794 
3795 	r = dm_pool_reserve_metadata_snap(pool->pmd);
3796 	if (r)
3797 		DMWARN("reserve_metadata_snap message failed.");
3798 
3799 	return r;
3800 }
3801 
3802 static int process_release_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3803 {
3804 	int r;
3805 
3806 	r = check_arg_count(argc, 1);
3807 	if (r)
3808 		return r;
3809 
3810 	r = dm_pool_release_metadata_snap(pool->pmd);
3811 	if (r)
3812 		DMWARN("release_metadata_snap message failed.");
3813 
3814 	return r;
3815 }
3816 
3817 /*
3818  * Messages supported:
3819  *   create_thin	<dev_id>
3820  *   create_snap	<dev_id> <origin_id>
3821  *   delete		<dev_id>
3822  *   set_transaction_id <current_trans_id> <new_trans_id>
3823  *   reserve_metadata_snap
3824  *   release_metadata_snap
3825  */
3826 static int pool_message(struct dm_target *ti, unsigned int argc, char **argv,
3827 			char *result, unsigned int maxlen)
3828 {
3829 	int r = -EINVAL;
3830 	struct pool_c *pt = ti->private;
3831 	struct pool *pool = pt->pool;
3832 
3833 	if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE) {
3834 		DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3835 		      dm_device_name(pool->pool_md));
3836 		return -EOPNOTSUPP;
3837 	}
3838 
3839 	if (!strcasecmp(argv[0], "create_thin"))
3840 		r = process_create_thin_mesg(argc, argv, pool);
3841 
3842 	else if (!strcasecmp(argv[0], "create_snap"))
3843 		r = process_create_snap_mesg(argc, argv, pool);
3844 
3845 	else if (!strcasecmp(argv[0], "delete"))
3846 		r = process_delete_mesg(argc, argv, pool);
3847 
3848 	else if (!strcasecmp(argv[0], "set_transaction_id"))
3849 		r = process_set_transaction_id_mesg(argc, argv, pool);
3850 
3851 	else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3852 		r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3853 
3854 	else if (!strcasecmp(argv[0], "release_metadata_snap"))
3855 		r = process_release_metadata_snap_mesg(argc, argv, pool);
3856 
3857 	else
3858 		DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3859 
3860 	if (!r)
3861 		(void) commit(pool);
3862 
3863 	return r;
3864 }
3865 
3866 static void emit_flags(struct pool_features *pf, char *result,
3867 		       unsigned int sz, unsigned int maxlen)
3868 {
3869 	unsigned int count = !pf->zero_new_blocks + !pf->discard_enabled +
3870 		!pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3871 		pf->error_if_no_space;
3872 	DMEMIT("%u ", count);
3873 
3874 	if (!pf->zero_new_blocks)
3875 		DMEMIT("skip_block_zeroing ");
3876 
3877 	if (!pf->discard_enabled)
3878 		DMEMIT("ignore_discard ");
3879 
3880 	if (!pf->discard_passdown)
3881 		DMEMIT("no_discard_passdown ");
3882 
3883 	if (pf->mode == PM_READ_ONLY)
3884 		DMEMIT("read_only ");
3885 
3886 	if (pf->error_if_no_space)
3887 		DMEMIT("error_if_no_space ");
3888 }
3889 
3890 /*
3891  * Status line is:
3892  *    <transaction id> <used metadata sectors>/<total metadata sectors>
3893  *    <used data sectors>/<total data sectors> <held metadata root>
3894  *    <pool mode> <discard config> <no space config> <needs_check>
3895  */
3896 static void pool_status(struct dm_target *ti, status_type_t type,
3897 			unsigned int status_flags, char *result, unsigned int maxlen)
3898 {
3899 	int r;
3900 	unsigned int sz = 0;
3901 	uint64_t transaction_id;
3902 	dm_block_t nr_free_blocks_data;
3903 	dm_block_t nr_free_blocks_metadata;
3904 	dm_block_t nr_blocks_data;
3905 	dm_block_t nr_blocks_metadata;
3906 	dm_block_t held_root;
3907 	enum pool_mode mode;
3908 	char buf[BDEVNAME_SIZE];
3909 	char buf2[BDEVNAME_SIZE];
3910 	struct pool_c *pt = ti->private;
3911 	struct pool *pool = pt->pool;
3912 
3913 	switch (type) {
3914 	case STATUSTYPE_INFO:
3915 		if (get_pool_mode(pool) == PM_FAIL) {
3916 			DMEMIT("Fail");
3917 			break;
3918 		}
3919 
3920 		/* Commit to ensure statistics aren't out-of-date */
3921 		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3922 			(void) commit(pool);
3923 
3924 		r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3925 		if (r) {
3926 			DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3927 			      dm_device_name(pool->pool_md), r);
3928 			goto err;
3929 		}
3930 
3931 		r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3932 		if (r) {
3933 			DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3934 			      dm_device_name(pool->pool_md), r);
3935 			goto err;
3936 		}
3937 
3938 		r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3939 		if (r) {
3940 			DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3941 			      dm_device_name(pool->pool_md), r);
3942 			goto err;
3943 		}
3944 
3945 		r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3946 		if (r) {
3947 			DMERR("%s: dm_pool_get_free_block_count returned %d",
3948 			      dm_device_name(pool->pool_md), r);
3949 			goto err;
3950 		}
3951 
3952 		r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3953 		if (r) {
3954 			DMERR("%s: dm_pool_get_data_dev_size returned %d",
3955 			      dm_device_name(pool->pool_md), r);
3956 			goto err;
3957 		}
3958 
3959 		r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3960 		if (r) {
3961 			DMERR("%s: dm_pool_get_metadata_snap returned %d",
3962 			      dm_device_name(pool->pool_md), r);
3963 			goto err;
3964 		}
3965 
3966 		DMEMIT("%llu %llu/%llu %llu/%llu ",
3967 		       (unsigned long long)transaction_id,
3968 		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3969 		       (unsigned long long)nr_blocks_metadata,
3970 		       (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3971 		       (unsigned long long)nr_blocks_data);
3972 
3973 		if (held_root)
3974 			DMEMIT("%llu ", held_root);
3975 		else
3976 			DMEMIT("- ");
3977 
3978 		mode = get_pool_mode(pool);
3979 		if (mode == PM_OUT_OF_DATA_SPACE)
3980 			DMEMIT("out_of_data_space ");
3981 		else if (is_read_only_pool_mode(mode))
3982 			DMEMIT("ro ");
3983 		else
3984 			DMEMIT("rw ");
3985 
3986 		if (!pool->pf.discard_enabled)
3987 			DMEMIT("ignore_discard ");
3988 		else if (pool->pf.discard_passdown)
3989 			DMEMIT("discard_passdown ");
3990 		else
3991 			DMEMIT("no_discard_passdown ");
3992 
3993 		if (pool->pf.error_if_no_space)
3994 			DMEMIT("error_if_no_space ");
3995 		else
3996 			DMEMIT("queue_if_no_space ");
3997 
3998 		if (dm_pool_metadata_needs_check(pool->pmd))
3999 			DMEMIT("needs_check ");
4000 		else
4001 			DMEMIT("- ");
4002 
4003 		DMEMIT("%llu ", (unsigned long long)calc_metadata_threshold(pt));
4004 
4005 		break;
4006 
4007 	case STATUSTYPE_TABLE:
4008 		DMEMIT("%s %s %lu %llu ",
4009 		       format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
4010 		       format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
4011 		       (unsigned long)pool->sectors_per_block,
4012 		       (unsigned long long)pt->low_water_blocks);
4013 		emit_flags(&pt->requested_pf, result, sz, maxlen);
4014 		break;
4015 
4016 	case STATUSTYPE_IMA:
4017 		*result = '\0';
4018 		break;
4019 	}
4020 	return;
4021 
4022 err:
4023 	DMEMIT("Error");
4024 }
4025 
4026 static int pool_iterate_devices(struct dm_target *ti,
4027 				iterate_devices_callout_fn fn, void *data)
4028 {
4029 	struct pool_c *pt = ti->private;
4030 
4031 	return fn(ti, pt->data_dev, 0, ti->len, data);
4032 }
4033 
4034 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
4035 {
4036 	struct pool_c *pt = ti->private;
4037 	struct pool *pool = pt->pool;
4038 	sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
4039 
4040 	/*
4041 	 * If max_sectors is smaller than pool->sectors_per_block adjust it
4042 	 * to the highest possible power-of-2 factor of pool->sectors_per_block.
4043 	 * This is especially beneficial when the pool's data device is a RAID
4044 	 * device that has a full stripe width that matches pool->sectors_per_block
4045 	 * -- because even though partial RAID stripe-sized IOs will be issued to a
4046 	 *    single RAID stripe; when aggregated they will end on a full RAID stripe
4047 	 *    boundary.. which avoids additional partial RAID stripe writes cascading
4048 	 */
4049 	if (limits->max_sectors < pool->sectors_per_block) {
4050 		while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
4051 			if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
4052 				limits->max_sectors--;
4053 			limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
4054 		}
4055 	}
4056 
4057 	/*
4058 	 * If the system-determined stacked limits are compatible with the
4059 	 * pool's blocksize (io_opt is a factor) do not override them.
4060 	 */
4061 	if (io_opt_sectors < pool->sectors_per_block ||
4062 	    !is_factor(io_opt_sectors, pool->sectors_per_block)) {
4063 		if (is_factor(pool->sectors_per_block, limits->max_sectors))
4064 			blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
4065 		else
4066 			blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
4067 		blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
4068 	}
4069 
4070 	/*
4071 	 * pt->adjusted_pf is a staging area for the actual features to use.
4072 	 * They get transferred to the live pool in bind_control_target()
4073 	 * called from pool_preresume().
4074 	 */
4075 	if (!pt->adjusted_pf.discard_enabled) {
4076 		/*
4077 		 * Must explicitly disallow stacking discard limits otherwise the
4078 		 * block layer will stack them if pool's data device has support.
4079 		 */
4080 		limits->discard_granularity = 0;
4081 		return;
4082 	}
4083 
4084 	disable_passdown_if_not_supported(pt);
4085 
4086 	/*
4087 	 * The pool uses the same discard limits as the underlying data
4088 	 * device.  DM core has already set this up.
4089 	 */
4090 }
4091 
4092 static struct target_type pool_target = {
4093 	.name = "thin-pool",
4094 	.features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4095 		    DM_TARGET_IMMUTABLE,
4096 	.version = {1, 22, 0},
4097 	.module = THIS_MODULE,
4098 	.ctr = pool_ctr,
4099 	.dtr = pool_dtr,
4100 	.map = pool_map,
4101 	.presuspend = pool_presuspend,
4102 	.presuspend_undo = pool_presuspend_undo,
4103 	.postsuspend = pool_postsuspend,
4104 	.preresume = pool_preresume,
4105 	.resume = pool_resume,
4106 	.message = pool_message,
4107 	.status = pool_status,
4108 	.iterate_devices = pool_iterate_devices,
4109 	.io_hints = pool_io_hints,
4110 };
4111 
4112 /*
4113  *--------------------------------------------------------------
4114  * Thin target methods
4115  *--------------------------------------------------------------
4116  */
4117 static void thin_get(struct thin_c *tc)
4118 {
4119 	refcount_inc(&tc->refcount);
4120 }
4121 
4122 static void thin_put(struct thin_c *tc)
4123 {
4124 	if (refcount_dec_and_test(&tc->refcount))
4125 		complete(&tc->can_destroy);
4126 }
4127 
4128 static void thin_dtr(struct dm_target *ti)
4129 {
4130 	struct thin_c *tc = ti->private;
4131 
4132 	spin_lock_irq(&tc->pool->lock);
4133 	list_del_rcu(&tc->list);
4134 	spin_unlock_irq(&tc->pool->lock);
4135 	synchronize_rcu();
4136 
4137 	thin_put(tc);
4138 	wait_for_completion(&tc->can_destroy);
4139 
4140 	mutex_lock(&dm_thin_pool_table.mutex);
4141 
4142 	__pool_dec(tc->pool);
4143 	dm_pool_close_thin_device(tc->td);
4144 	dm_put_device(ti, tc->pool_dev);
4145 	if (tc->origin_dev)
4146 		dm_put_device(ti, tc->origin_dev);
4147 	kfree(tc);
4148 
4149 	mutex_unlock(&dm_thin_pool_table.mutex);
4150 }
4151 
4152 /*
4153  * Thin target parameters:
4154  *
4155  * <pool_dev> <dev_id> [origin_dev]
4156  *
4157  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4158  * dev_id: the internal device identifier
4159  * origin_dev: a device external to the pool that should act as the origin
4160  *
4161  * If the pool device has discards disabled, they get disabled for the thin
4162  * device as well.
4163  */
4164 static int thin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
4165 {
4166 	int r;
4167 	struct thin_c *tc;
4168 	struct dm_dev *pool_dev, *origin_dev;
4169 	struct mapped_device *pool_md;
4170 
4171 	mutex_lock(&dm_thin_pool_table.mutex);
4172 
4173 	if (argc != 2 && argc != 3) {
4174 		ti->error = "Invalid argument count";
4175 		r = -EINVAL;
4176 		goto out_unlock;
4177 	}
4178 
4179 	tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4180 	if (!tc) {
4181 		ti->error = "Out of memory";
4182 		r = -ENOMEM;
4183 		goto out_unlock;
4184 	}
4185 	tc->thin_md = dm_table_get_md(ti->table);
4186 	spin_lock_init(&tc->lock);
4187 	INIT_LIST_HEAD(&tc->deferred_cells);
4188 	bio_list_init(&tc->deferred_bio_list);
4189 	bio_list_init(&tc->retry_on_resume_list);
4190 	tc->sort_bio_list = RB_ROOT;
4191 
4192 	if (argc == 3) {
4193 		if (!strcmp(argv[0], argv[2])) {
4194 			ti->error = "Error setting origin device";
4195 			r = -EINVAL;
4196 			goto bad_origin_dev;
4197 		}
4198 
4199 		r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4200 		if (r) {
4201 			ti->error = "Error opening origin device";
4202 			goto bad_origin_dev;
4203 		}
4204 		tc->origin_dev = origin_dev;
4205 	}
4206 
4207 	r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4208 	if (r) {
4209 		ti->error = "Error opening pool device";
4210 		goto bad_pool_dev;
4211 	}
4212 	tc->pool_dev = pool_dev;
4213 
4214 	if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4215 		ti->error = "Invalid device id";
4216 		r = -EINVAL;
4217 		goto bad_common;
4218 	}
4219 
4220 	pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4221 	if (!pool_md) {
4222 		ti->error = "Couldn't get pool mapped device";
4223 		r = -EINVAL;
4224 		goto bad_common;
4225 	}
4226 
4227 	tc->pool = __pool_table_lookup(pool_md);
4228 	if (!tc->pool) {
4229 		ti->error = "Couldn't find pool object";
4230 		r = -EINVAL;
4231 		goto bad_pool_lookup;
4232 	}
4233 	__pool_inc(tc->pool);
4234 
4235 	if (get_pool_mode(tc->pool) == PM_FAIL) {
4236 		ti->error = "Couldn't open thin device, Pool is in fail mode";
4237 		r = -EINVAL;
4238 		goto bad_pool;
4239 	}
4240 
4241 	r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4242 	if (r) {
4243 		ti->error = "Couldn't open thin internal device";
4244 		goto bad_pool;
4245 	}
4246 
4247 	r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4248 	if (r)
4249 		goto bad;
4250 
4251 	ti->num_flush_bios = 1;
4252 	ti->flush_supported = true;
4253 	ti->accounts_remapped_io = true;
4254 	ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4255 
4256 	/* In case the pool supports discards, pass them on. */
4257 	if (tc->pool->pf.discard_enabled) {
4258 		ti->discards_supported = true;
4259 		ti->num_discard_bios = 1;
4260 	}
4261 
4262 	mutex_unlock(&dm_thin_pool_table.mutex);
4263 
4264 	spin_lock_irq(&tc->pool->lock);
4265 	if (tc->pool->suspended) {
4266 		spin_unlock_irq(&tc->pool->lock);
4267 		mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4268 		ti->error = "Unable to activate thin device while pool is suspended";
4269 		r = -EINVAL;
4270 		goto bad;
4271 	}
4272 	refcount_set(&tc->refcount, 1);
4273 	init_completion(&tc->can_destroy);
4274 	list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4275 	spin_unlock_irq(&tc->pool->lock);
4276 	/*
4277 	 * This synchronize_rcu() call is needed here otherwise we risk a
4278 	 * wake_worker() call finding no bios to process (because the newly
4279 	 * added tc isn't yet visible).  So this reduces latency since we
4280 	 * aren't then dependent on the periodic commit to wake_worker().
4281 	 */
4282 	synchronize_rcu();
4283 
4284 	dm_put(pool_md);
4285 
4286 	return 0;
4287 
4288 bad:
4289 	dm_pool_close_thin_device(tc->td);
4290 bad_pool:
4291 	__pool_dec(tc->pool);
4292 bad_pool_lookup:
4293 	dm_put(pool_md);
4294 bad_common:
4295 	dm_put_device(ti, tc->pool_dev);
4296 bad_pool_dev:
4297 	if (tc->origin_dev)
4298 		dm_put_device(ti, tc->origin_dev);
4299 bad_origin_dev:
4300 	kfree(tc);
4301 out_unlock:
4302 	mutex_unlock(&dm_thin_pool_table.mutex);
4303 
4304 	return r;
4305 }
4306 
4307 static int thin_map(struct dm_target *ti, struct bio *bio)
4308 {
4309 	bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4310 
4311 	return thin_bio_map(ti, bio);
4312 }
4313 
4314 static int thin_endio(struct dm_target *ti, struct bio *bio,
4315 		blk_status_t *err)
4316 {
4317 	unsigned long flags;
4318 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4319 	struct list_head work;
4320 	struct dm_thin_new_mapping *m, *tmp;
4321 	struct pool *pool = h->tc->pool;
4322 
4323 	if (h->shared_read_entry) {
4324 		INIT_LIST_HEAD(&work);
4325 		dm_deferred_entry_dec(h->shared_read_entry, &work);
4326 
4327 		spin_lock_irqsave(&pool->lock, flags);
4328 		list_for_each_entry_safe(m, tmp, &work, list) {
4329 			list_del(&m->list);
4330 			__complete_mapping_preparation(m);
4331 		}
4332 		spin_unlock_irqrestore(&pool->lock, flags);
4333 	}
4334 
4335 	if (h->all_io_entry) {
4336 		INIT_LIST_HEAD(&work);
4337 		dm_deferred_entry_dec(h->all_io_entry, &work);
4338 		if (!list_empty(&work)) {
4339 			spin_lock_irqsave(&pool->lock, flags);
4340 			list_for_each_entry_safe(m, tmp, &work, list)
4341 				list_add_tail(&m->list, &pool->prepared_discards);
4342 			spin_unlock_irqrestore(&pool->lock, flags);
4343 			wake_worker(pool);
4344 		}
4345 	}
4346 
4347 	if (h->cell)
4348 		cell_defer_no_holder(h->tc, h->cell);
4349 
4350 	return DM_ENDIO_DONE;
4351 }
4352 
4353 static void thin_presuspend(struct dm_target *ti)
4354 {
4355 	struct thin_c *tc = ti->private;
4356 
4357 	if (dm_noflush_suspending(ti))
4358 		noflush_work(tc, do_noflush_start);
4359 }
4360 
4361 static void thin_postsuspend(struct dm_target *ti)
4362 {
4363 	struct thin_c *tc = ti->private;
4364 
4365 	/*
4366 	 * The dm_noflush_suspending flag has been cleared by now, so
4367 	 * unfortunately we must always run this.
4368 	 */
4369 	noflush_work(tc, do_noflush_stop);
4370 }
4371 
4372 static int thin_preresume(struct dm_target *ti)
4373 {
4374 	struct thin_c *tc = ti->private;
4375 
4376 	if (tc->origin_dev)
4377 		tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4378 
4379 	return 0;
4380 }
4381 
4382 /*
4383  * <nr mapped sectors> <highest mapped sector>
4384  */
4385 static void thin_status(struct dm_target *ti, status_type_t type,
4386 			unsigned int status_flags, char *result, unsigned int maxlen)
4387 {
4388 	int r;
4389 	ssize_t sz = 0;
4390 	dm_block_t mapped, highest;
4391 	char buf[BDEVNAME_SIZE];
4392 	struct thin_c *tc = ti->private;
4393 
4394 	if (get_pool_mode(tc->pool) == PM_FAIL) {
4395 		DMEMIT("Fail");
4396 		return;
4397 	}
4398 
4399 	if (!tc->td)
4400 		DMEMIT("-");
4401 	else {
4402 		switch (type) {
4403 		case STATUSTYPE_INFO:
4404 			r = dm_thin_get_mapped_count(tc->td, &mapped);
4405 			if (r) {
4406 				DMERR("dm_thin_get_mapped_count returned %d", r);
4407 				goto err;
4408 			}
4409 
4410 			r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4411 			if (r < 0) {
4412 				DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4413 				goto err;
4414 			}
4415 
4416 			DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4417 			if (r)
4418 				DMEMIT("%llu", ((highest + 1) *
4419 						tc->pool->sectors_per_block) - 1);
4420 			else
4421 				DMEMIT("-");
4422 			break;
4423 
4424 		case STATUSTYPE_TABLE:
4425 			DMEMIT("%s %lu",
4426 			       format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4427 			       (unsigned long) tc->dev_id);
4428 			if (tc->origin_dev)
4429 				DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4430 			break;
4431 
4432 		case STATUSTYPE_IMA:
4433 			*result = '\0';
4434 			break;
4435 		}
4436 	}
4437 
4438 	return;
4439 
4440 err:
4441 	DMEMIT("Error");
4442 }
4443 
4444 static int thin_iterate_devices(struct dm_target *ti,
4445 				iterate_devices_callout_fn fn, void *data)
4446 {
4447 	sector_t blocks;
4448 	struct thin_c *tc = ti->private;
4449 	struct pool *pool = tc->pool;
4450 
4451 	/*
4452 	 * We can't call dm_pool_get_data_dev_size() since that blocks.  So
4453 	 * we follow a more convoluted path through to the pool's target.
4454 	 */
4455 	if (!pool->ti)
4456 		return 0;	/* nothing is bound */
4457 
4458 	blocks = pool->ti->len;
4459 	(void) sector_div(blocks, pool->sectors_per_block);
4460 	if (blocks)
4461 		return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4462 
4463 	return 0;
4464 }
4465 
4466 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4467 {
4468 	struct thin_c *tc = ti->private;
4469 	struct pool *pool = tc->pool;
4470 
4471 	if (!pool->pf.discard_enabled)
4472 		return;
4473 
4474 	limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4475 	limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4476 }
4477 
4478 static struct target_type thin_target = {
4479 	.name = "thin",
4480 	.version = {1, 22, 0},
4481 	.module	= THIS_MODULE,
4482 	.ctr = thin_ctr,
4483 	.dtr = thin_dtr,
4484 	.map = thin_map,
4485 	.end_io = thin_endio,
4486 	.preresume = thin_preresume,
4487 	.presuspend = thin_presuspend,
4488 	.postsuspend = thin_postsuspend,
4489 	.status = thin_status,
4490 	.iterate_devices = thin_iterate_devices,
4491 	.io_hints = thin_io_hints,
4492 };
4493 
4494 /*----------------------------------------------------------------*/
4495 
4496 static int __init dm_thin_init(void)
4497 {
4498 	int r = -ENOMEM;
4499 
4500 	pool_table_init();
4501 
4502 	_new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4503 	if (!_new_mapping_cache)
4504 		return r;
4505 
4506 	r = dm_register_target(&thin_target);
4507 	if (r)
4508 		goto bad_new_mapping_cache;
4509 
4510 	r = dm_register_target(&pool_target);
4511 	if (r)
4512 		goto bad_thin_target;
4513 
4514 	return 0;
4515 
4516 bad_thin_target:
4517 	dm_unregister_target(&thin_target);
4518 bad_new_mapping_cache:
4519 	kmem_cache_destroy(_new_mapping_cache);
4520 
4521 	return r;
4522 }
4523 
4524 static void dm_thin_exit(void)
4525 {
4526 	dm_unregister_target(&thin_target);
4527 	dm_unregister_target(&pool_target);
4528 
4529 	kmem_cache_destroy(_new_mapping_cache);
4530 
4531 	pool_table_exit();
4532 }
4533 
4534 module_init(dm_thin_init);
4535 module_exit(dm_thin_exit);
4536 
4537 module_param_named(no_space_timeout, no_space_timeout_secs, uint, 0644);
4538 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4539 
4540 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4541 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4542 MODULE_LICENSE("GPL");
4543