xref: /openbmc/linux/drivers/md/dm-thin.c (revision 25b892b5)
1 /*
2  * Copyright (C) 2011-2012 Red Hat UK.
3  *
4  * This file is released under the GPL.
5  */
6 
7 #include "dm-thin-metadata.h"
8 #include "dm-bio-prison-v1.h"
9 #include "dm.h"
10 
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/jiffies.h>
15 #include <linux/log2.h>
16 #include <linux/list.h>
17 #include <linux/rculist.h>
18 #include <linux/init.h>
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
22 #include <linux/sort.h>
23 #include <linux/rbtree.h>
24 
25 #define	DM_MSG_PREFIX	"thin"
26 
27 /*
28  * Tunable constants
29  */
30 #define ENDIO_HOOK_POOL_SIZE 1024
31 #define MAPPING_POOL_SIZE 1024
32 #define COMMIT_PERIOD HZ
33 #define NO_SPACE_TIMEOUT_SECS 60
34 
35 static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
36 
37 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
38 		"A percentage of time allocated for copy on write");
39 
40 /*
41  * The block size of the device holding pool data must be
42  * between 64KB and 1GB.
43  */
44 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
45 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
46 
47 /*
48  * Device id is restricted to 24 bits.
49  */
50 #define MAX_DEV_ID ((1 << 24) - 1)
51 
52 /*
53  * How do we handle breaking sharing of data blocks?
54  * =================================================
55  *
56  * We use a standard copy-on-write btree to store the mappings for the
57  * devices (note I'm talking about copy-on-write of the metadata here, not
58  * the data).  When you take an internal snapshot you clone the root node
59  * of the origin btree.  After this there is no concept of an origin or a
60  * snapshot.  They are just two device trees that happen to point to the
61  * same data blocks.
62  *
63  * When we get a write in we decide if it's to a shared data block using
64  * some timestamp magic.  If it is, we have to break sharing.
65  *
66  * Let's say we write to a shared block in what was the origin.  The
67  * steps are:
68  *
69  * i) plug io further to this physical block. (see bio_prison code).
70  *
71  * ii) quiesce any read io to that shared data block.  Obviously
72  * including all devices that share this block.  (see dm_deferred_set code)
73  *
74  * iii) copy the data block to a newly allocate block.  This step can be
75  * missed out if the io covers the block. (schedule_copy).
76  *
77  * iv) insert the new mapping into the origin's btree
78  * (process_prepared_mapping).  This act of inserting breaks some
79  * sharing of btree nodes between the two devices.  Breaking sharing only
80  * effects the btree of that specific device.  Btrees for the other
81  * devices that share the block never change.  The btree for the origin
82  * device as it was after the last commit is untouched, ie. we're using
83  * persistent data structures in the functional programming sense.
84  *
85  * v) unplug io to this physical block, including the io that triggered
86  * the breaking of sharing.
87  *
88  * Steps (ii) and (iii) occur in parallel.
89  *
90  * The metadata _doesn't_ need to be committed before the io continues.  We
91  * get away with this because the io is always written to a _new_ block.
92  * If there's a crash, then:
93  *
94  * - The origin mapping will point to the old origin block (the shared
95  * one).  This will contain the data as it was before the io that triggered
96  * the breaking of sharing came in.
97  *
98  * - The snap mapping still points to the old block.  As it would after
99  * the commit.
100  *
101  * The downside of this scheme is the timestamp magic isn't perfect, and
102  * will continue to think that data block in the snapshot device is shared
103  * even after the write to the origin has broken sharing.  I suspect data
104  * blocks will typically be shared by many different devices, so we're
105  * breaking sharing n + 1 times, rather than n, where n is the number of
106  * devices that reference this data block.  At the moment I think the
107  * benefits far, far outweigh the disadvantages.
108  */
109 
110 /*----------------------------------------------------------------*/
111 
112 /*
113  * Key building.
114  */
115 enum lock_space {
116 	VIRTUAL,
117 	PHYSICAL
118 };
119 
120 static void build_key(struct dm_thin_device *td, enum lock_space ls,
121 		      dm_block_t b, dm_block_t e, struct dm_cell_key *key)
122 {
123 	key->virtual = (ls == VIRTUAL);
124 	key->dev = dm_thin_dev_id(td);
125 	key->block_begin = b;
126 	key->block_end = e;
127 }
128 
129 static void build_data_key(struct dm_thin_device *td, dm_block_t b,
130 			   struct dm_cell_key *key)
131 {
132 	build_key(td, PHYSICAL, b, b + 1llu, key);
133 }
134 
135 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
136 			      struct dm_cell_key *key)
137 {
138 	build_key(td, VIRTUAL, b, b + 1llu, key);
139 }
140 
141 /*----------------------------------------------------------------*/
142 
143 #define THROTTLE_THRESHOLD (1 * HZ)
144 
145 struct throttle {
146 	struct rw_semaphore lock;
147 	unsigned long threshold;
148 	bool throttle_applied;
149 };
150 
151 static void throttle_init(struct throttle *t)
152 {
153 	init_rwsem(&t->lock);
154 	t->throttle_applied = false;
155 }
156 
157 static void throttle_work_start(struct throttle *t)
158 {
159 	t->threshold = jiffies + THROTTLE_THRESHOLD;
160 }
161 
162 static void throttle_work_update(struct throttle *t)
163 {
164 	if (!t->throttle_applied && jiffies > t->threshold) {
165 		down_write(&t->lock);
166 		t->throttle_applied = true;
167 	}
168 }
169 
170 static void throttle_work_complete(struct throttle *t)
171 {
172 	if (t->throttle_applied) {
173 		t->throttle_applied = false;
174 		up_write(&t->lock);
175 	}
176 }
177 
178 static void throttle_lock(struct throttle *t)
179 {
180 	down_read(&t->lock);
181 }
182 
183 static void throttle_unlock(struct throttle *t)
184 {
185 	up_read(&t->lock);
186 }
187 
188 /*----------------------------------------------------------------*/
189 
190 /*
191  * A pool device ties together a metadata device and a data device.  It
192  * also provides the interface for creating and destroying internal
193  * devices.
194  */
195 struct dm_thin_new_mapping;
196 
197 /*
198  * The pool runs in various modes.  Ordered in degraded order for comparisons.
199  */
200 enum pool_mode {
201 	PM_WRITE,		/* metadata may be changed */
202 	PM_OUT_OF_DATA_SPACE,	/* metadata may be changed, though data may not be allocated */
203 
204 	/*
205 	 * Like READ_ONLY, except may switch back to WRITE on metadata resize. Reported as READ_ONLY.
206 	 */
207 	PM_OUT_OF_METADATA_SPACE,
208 	PM_READ_ONLY,		/* metadata may not be changed */
209 
210 	PM_FAIL,		/* all I/O fails */
211 };
212 
213 struct pool_features {
214 	enum pool_mode mode;
215 
216 	bool zero_new_blocks:1;
217 	bool discard_enabled:1;
218 	bool discard_passdown:1;
219 	bool error_if_no_space:1;
220 };
221 
222 struct thin_c;
223 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
224 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
225 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
226 
227 #define CELL_SORT_ARRAY_SIZE 8192
228 
229 struct pool {
230 	struct list_head list;
231 	struct dm_target *ti;	/* Only set if a pool target is bound */
232 
233 	struct mapped_device *pool_md;
234 	struct block_device *data_dev;
235 	struct block_device *md_dev;
236 	struct dm_pool_metadata *pmd;
237 
238 	dm_block_t low_water_blocks;
239 	uint32_t sectors_per_block;
240 	int sectors_per_block_shift;
241 
242 	struct pool_features pf;
243 	bool low_water_triggered:1;	/* A dm event has been sent */
244 	bool suspended:1;
245 	bool out_of_data_space:1;
246 
247 	struct dm_bio_prison *prison;
248 	struct dm_kcopyd_client *copier;
249 
250 	struct work_struct worker;
251 	struct workqueue_struct *wq;
252 	struct throttle throttle;
253 	struct delayed_work waker;
254 	struct delayed_work no_space_timeout;
255 
256 	unsigned long last_commit_jiffies;
257 	unsigned ref_count;
258 
259 	spinlock_t lock;
260 	struct bio_list deferred_flush_bios;
261 	struct bio_list deferred_flush_completions;
262 	struct list_head prepared_mappings;
263 	struct list_head prepared_discards;
264 	struct list_head prepared_discards_pt2;
265 	struct list_head active_thins;
266 
267 	struct dm_deferred_set *shared_read_ds;
268 	struct dm_deferred_set *all_io_ds;
269 
270 	struct dm_thin_new_mapping *next_mapping;
271 
272 	process_bio_fn process_bio;
273 	process_bio_fn process_discard;
274 
275 	process_cell_fn process_cell;
276 	process_cell_fn process_discard_cell;
277 
278 	process_mapping_fn process_prepared_mapping;
279 	process_mapping_fn process_prepared_discard;
280 	process_mapping_fn process_prepared_discard_pt2;
281 
282 	struct dm_bio_prison_cell **cell_sort_array;
283 
284 	mempool_t mapping_pool;
285 
286 	struct bio flush_bio;
287 };
288 
289 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
290 
291 static enum pool_mode get_pool_mode(struct pool *pool)
292 {
293 	return pool->pf.mode;
294 }
295 
296 static void notify_of_pool_mode_change(struct pool *pool)
297 {
298 	const char *descs[] = {
299 		"write",
300 		"out-of-data-space",
301 		"read-only",
302 		"read-only",
303 		"fail"
304 	};
305 	const char *extra_desc = NULL;
306 	enum pool_mode mode = get_pool_mode(pool);
307 
308 	if (mode == PM_OUT_OF_DATA_SPACE) {
309 		if (!pool->pf.error_if_no_space)
310 			extra_desc = " (queue IO)";
311 		else
312 			extra_desc = " (error IO)";
313 	}
314 
315 	dm_table_event(pool->ti->table);
316 	DMINFO("%s: switching pool to %s%s mode",
317 	       dm_device_name(pool->pool_md),
318 	       descs[(int)mode], extra_desc ? : "");
319 }
320 
321 /*
322  * Target context for a pool.
323  */
324 struct pool_c {
325 	struct dm_target *ti;
326 	struct pool *pool;
327 	struct dm_dev *data_dev;
328 	struct dm_dev *metadata_dev;
329 
330 	dm_block_t low_water_blocks;
331 	struct pool_features requested_pf; /* Features requested during table load */
332 	struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
333 };
334 
335 /*
336  * Target context for a thin.
337  */
338 struct thin_c {
339 	struct list_head list;
340 	struct dm_dev *pool_dev;
341 	struct dm_dev *origin_dev;
342 	sector_t origin_size;
343 	dm_thin_id dev_id;
344 
345 	struct pool *pool;
346 	struct dm_thin_device *td;
347 	struct mapped_device *thin_md;
348 
349 	bool requeue_mode:1;
350 	spinlock_t lock;
351 	struct list_head deferred_cells;
352 	struct bio_list deferred_bio_list;
353 	struct bio_list retry_on_resume_list;
354 	struct rb_root sort_bio_list; /* sorted list of deferred bios */
355 
356 	/*
357 	 * Ensures the thin is not destroyed until the worker has finished
358 	 * iterating the active_thins list.
359 	 */
360 	refcount_t refcount;
361 	struct completion can_destroy;
362 };
363 
364 /*----------------------------------------------------------------*/
365 
366 static bool block_size_is_power_of_two(struct pool *pool)
367 {
368 	return pool->sectors_per_block_shift >= 0;
369 }
370 
371 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
372 {
373 	return block_size_is_power_of_two(pool) ?
374 		(b << pool->sectors_per_block_shift) :
375 		(b * pool->sectors_per_block);
376 }
377 
378 /*----------------------------------------------------------------*/
379 
380 struct discard_op {
381 	struct thin_c *tc;
382 	struct blk_plug plug;
383 	struct bio *parent_bio;
384 	struct bio *bio;
385 };
386 
387 static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
388 {
389 	BUG_ON(!parent);
390 
391 	op->tc = tc;
392 	blk_start_plug(&op->plug);
393 	op->parent_bio = parent;
394 	op->bio = NULL;
395 }
396 
397 static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
398 {
399 	struct thin_c *tc = op->tc;
400 	sector_t s = block_to_sectors(tc->pool, data_b);
401 	sector_t len = block_to_sectors(tc->pool, data_e - data_b);
402 
403 	return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
404 				      GFP_NOWAIT, 0, &op->bio);
405 }
406 
407 static void end_discard(struct discard_op *op, int r)
408 {
409 	if (op->bio) {
410 		/*
411 		 * Even if one of the calls to issue_discard failed, we
412 		 * need to wait for the chain to complete.
413 		 */
414 		bio_chain(op->bio, op->parent_bio);
415 		bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
416 		submit_bio(op->bio);
417 	}
418 
419 	blk_finish_plug(&op->plug);
420 
421 	/*
422 	 * Even if r is set, there could be sub discards in flight that we
423 	 * need to wait for.
424 	 */
425 	if (r && !op->parent_bio->bi_status)
426 		op->parent_bio->bi_status = errno_to_blk_status(r);
427 	bio_endio(op->parent_bio);
428 }
429 
430 /*----------------------------------------------------------------*/
431 
432 /*
433  * wake_worker() is used when new work is queued and when pool_resume is
434  * ready to continue deferred IO processing.
435  */
436 static void wake_worker(struct pool *pool)
437 {
438 	queue_work(pool->wq, &pool->worker);
439 }
440 
441 /*----------------------------------------------------------------*/
442 
443 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
444 		      struct dm_bio_prison_cell **cell_result)
445 {
446 	int r;
447 	struct dm_bio_prison_cell *cell_prealloc;
448 
449 	/*
450 	 * Allocate a cell from the prison's mempool.
451 	 * This might block but it can't fail.
452 	 */
453 	cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
454 
455 	r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
456 	if (r)
457 		/*
458 		 * We reused an old cell; we can get rid of
459 		 * the new one.
460 		 */
461 		dm_bio_prison_free_cell(pool->prison, cell_prealloc);
462 
463 	return r;
464 }
465 
466 static void cell_release(struct pool *pool,
467 			 struct dm_bio_prison_cell *cell,
468 			 struct bio_list *bios)
469 {
470 	dm_cell_release(pool->prison, cell, bios);
471 	dm_bio_prison_free_cell(pool->prison, cell);
472 }
473 
474 static void cell_visit_release(struct pool *pool,
475 			       void (*fn)(void *, struct dm_bio_prison_cell *),
476 			       void *context,
477 			       struct dm_bio_prison_cell *cell)
478 {
479 	dm_cell_visit_release(pool->prison, fn, context, cell);
480 	dm_bio_prison_free_cell(pool->prison, cell);
481 }
482 
483 static void cell_release_no_holder(struct pool *pool,
484 				   struct dm_bio_prison_cell *cell,
485 				   struct bio_list *bios)
486 {
487 	dm_cell_release_no_holder(pool->prison, cell, bios);
488 	dm_bio_prison_free_cell(pool->prison, cell);
489 }
490 
491 static void cell_error_with_code(struct pool *pool,
492 		struct dm_bio_prison_cell *cell, blk_status_t error_code)
493 {
494 	dm_cell_error(pool->prison, cell, error_code);
495 	dm_bio_prison_free_cell(pool->prison, cell);
496 }
497 
498 static blk_status_t get_pool_io_error_code(struct pool *pool)
499 {
500 	return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
501 }
502 
503 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
504 {
505 	cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
506 }
507 
508 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
509 {
510 	cell_error_with_code(pool, cell, 0);
511 }
512 
513 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
514 {
515 	cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
516 }
517 
518 /*----------------------------------------------------------------*/
519 
520 /*
521  * A global list of pools that uses a struct mapped_device as a key.
522  */
523 static struct dm_thin_pool_table {
524 	struct mutex mutex;
525 	struct list_head pools;
526 } dm_thin_pool_table;
527 
528 static void pool_table_init(void)
529 {
530 	mutex_init(&dm_thin_pool_table.mutex);
531 	INIT_LIST_HEAD(&dm_thin_pool_table.pools);
532 }
533 
534 static void pool_table_exit(void)
535 {
536 	mutex_destroy(&dm_thin_pool_table.mutex);
537 }
538 
539 static void __pool_table_insert(struct pool *pool)
540 {
541 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
542 	list_add(&pool->list, &dm_thin_pool_table.pools);
543 }
544 
545 static void __pool_table_remove(struct pool *pool)
546 {
547 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
548 	list_del(&pool->list);
549 }
550 
551 static struct pool *__pool_table_lookup(struct mapped_device *md)
552 {
553 	struct pool *pool = NULL, *tmp;
554 
555 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
556 
557 	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
558 		if (tmp->pool_md == md) {
559 			pool = tmp;
560 			break;
561 		}
562 	}
563 
564 	return pool;
565 }
566 
567 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
568 {
569 	struct pool *pool = NULL, *tmp;
570 
571 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
572 
573 	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
574 		if (tmp->md_dev == md_dev) {
575 			pool = tmp;
576 			break;
577 		}
578 	}
579 
580 	return pool;
581 }
582 
583 /*----------------------------------------------------------------*/
584 
585 struct dm_thin_endio_hook {
586 	struct thin_c *tc;
587 	struct dm_deferred_entry *shared_read_entry;
588 	struct dm_deferred_entry *all_io_entry;
589 	struct dm_thin_new_mapping *overwrite_mapping;
590 	struct rb_node rb_node;
591 	struct dm_bio_prison_cell *cell;
592 };
593 
594 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
595 {
596 	bio_list_merge(bios, master);
597 	bio_list_init(master);
598 }
599 
600 static void error_bio_list(struct bio_list *bios, blk_status_t error)
601 {
602 	struct bio *bio;
603 
604 	while ((bio = bio_list_pop(bios))) {
605 		bio->bi_status = error;
606 		bio_endio(bio);
607 	}
608 }
609 
610 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
611 		blk_status_t error)
612 {
613 	struct bio_list bios;
614 
615 	bio_list_init(&bios);
616 
617 	spin_lock_irq(&tc->lock);
618 	__merge_bio_list(&bios, master);
619 	spin_unlock_irq(&tc->lock);
620 
621 	error_bio_list(&bios, error);
622 }
623 
624 static void requeue_deferred_cells(struct thin_c *tc)
625 {
626 	struct pool *pool = tc->pool;
627 	struct list_head cells;
628 	struct dm_bio_prison_cell *cell, *tmp;
629 
630 	INIT_LIST_HEAD(&cells);
631 
632 	spin_lock_irq(&tc->lock);
633 	list_splice_init(&tc->deferred_cells, &cells);
634 	spin_unlock_irq(&tc->lock);
635 
636 	list_for_each_entry_safe(cell, tmp, &cells, user_list)
637 		cell_requeue(pool, cell);
638 }
639 
640 static void requeue_io(struct thin_c *tc)
641 {
642 	struct bio_list bios;
643 
644 	bio_list_init(&bios);
645 
646 	spin_lock_irq(&tc->lock);
647 	__merge_bio_list(&bios, &tc->deferred_bio_list);
648 	__merge_bio_list(&bios, &tc->retry_on_resume_list);
649 	spin_unlock_irq(&tc->lock);
650 
651 	error_bio_list(&bios, BLK_STS_DM_REQUEUE);
652 	requeue_deferred_cells(tc);
653 }
654 
655 static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
656 {
657 	struct thin_c *tc;
658 
659 	rcu_read_lock();
660 	list_for_each_entry_rcu(tc, &pool->active_thins, list)
661 		error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
662 	rcu_read_unlock();
663 }
664 
665 static void error_retry_list(struct pool *pool)
666 {
667 	error_retry_list_with_code(pool, get_pool_io_error_code(pool));
668 }
669 
670 /*
671  * This section of code contains the logic for processing a thin device's IO.
672  * Much of the code depends on pool object resources (lists, workqueues, etc)
673  * but most is exclusively called from the thin target rather than the thin-pool
674  * target.
675  */
676 
677 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
678 {
679 	struct pool *pool = tc->pool;
680 	sector_t block_nr = bio->bi_iter.bi_sector;
681 
682 	if (block_size_is_power_of_two(pool))
683 		block_nr >>= pool->sectors_per_block_shift;
684 	else
685 		(void) sector_div(block_nr, pool->sectors_per_block);
686 
687 	return block_nr;
688 }
689 
690 /*
691  * Returns the _complete_ blocks that this bio covers.
692  */
693 static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
694 				dm_block_t *begin, dm_block_t *end)
695 {
696 	struct pool *pool = tc->pool;
697 	sector_t b = bio->bi_iter.bi_sector;
698 	sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
699 
700 	b += pool->sectors_per_block - 1ull; /* so we round up */
701 
702 	if (block_size_is_power_of_two(pool)) {
703 		b >>= pool->sectors_per_block_shift;
704 		e >>= pool->sectors_per_block_shift;
705 	} else {
706 		(void) sector_div(b, pool->sectors_per_block);
707 		(void) sector_div(e, pool->sectors_per_block);
708 	}
709 
710 	if (e < b)
711 		/* Can happen if the bio is within a single block. */
712 		e = b;
713 
714 	*begin = b;
715 	*end = e;
716 }
717 
718 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
719 {
720 	struct pool *pool = tc->pool;
721 	sector_t bi_sector = bio->bi_iter.bi_sector;
722 
723 	bio_set_dev(bio, tc->pool_dev->bdev);
724 	if (block_size_is_power_of_two(pool))
725 		bio->bi_iter.bi_sector =
726 			(block << pool->sectors_per_block_shift) |
727 			(bi_sector & (pool->sectors_per_block - 1));
728 	else
729 		bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
730 				 sector_div(bi_sector, pool->sectors_per_block);
731 }
732 
733 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
734 {
735 	bio_set_dev(bio, tc->origin_dev->bdev);
736 }
737 
738 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
739 {
740 	return op_is_flush(bio->bi_opf) &&
741 		dm_thin_changed_this_transaction(tc->td);
742 }
743 
744 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
745 {
746 	struct dm_thin_endio_hook *h;
747 
748 	if (bio_op(bio) == REQ_OP_DISCARD)
749 		return;
750 
751 	h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
752 	h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
753 }
754 
755 static void issue(struct thin_c *tc, struct bio *bio)
756 {
757 	struct pool *pool = tc->pool;
758 
759 	if (!bio_triggers_commit(tc, bio)) {
760 		submit_bio_noacct(bio);
761 		return;
762 	}
763 
764 	/*
765 	 * Complete bio with an error if earlier I/O caused changes to
766 	 * the metadata that can't be committed e.g, due to I/O errors
767 	 * on the metadata device.
768 	 */
769 	if (dm_thin_aborted_changes(tc->td)) {
770 		bio_io_error(bio);
771 		return;
772 	}
773 
774 	/*
775 	 * Batch together any bios that trigger commits and then issue a
776 	 * single commit for them in process_deferred_bios().
777 	 */
778 	spin_lock_irq(&pool->lock);
779 	bio_list_add(&pool->deferred_flush_bios, bio);
780 	spin_unlock_irq(&pool->lock);
781 }
782 
783 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
784 {
785 	remap_to_origin(tc, bio);
786 	issue(tc, bio);
787 }
788 
789 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
790 			    dm_block_t block)
791 {
792 	remap(tc, bio, block);
793 	issue(tc, bio);
794 }
795 
796 /*----------------------------------------------------------------*/
797 
798 /*
799  * Bio endio functions.
800  */
801 struct dm_thin_new_mapping {
802 	struct list_head list;
803 
804 	bool pass_discard:1;
805 	bool maybe_shared:1;
806 
807 	/*
808 	 * Track quiescing, copying and zeroing preparation actions.  When this
809 	 * counter hits zero the block is prepared and can be inserted into the
810 	 * btree.
811 	 */
812 	atomic_t prepare_actions;
813 
814 	blk_status_t status;
815 	struct thin_c *tc;
816 	dm_block_t virt_begin, virt_end;
817 	dm_block_t data_block;
818 	struct dm_bio_prison_cell *cell;
819 
820 	/*
821 	 * If the bio covers the whole area of a block then we can avoid
822 	 * zeroing or copying.  Instead this bio is hooked.  The bio will
823 	 * still be in the cell, so care has to be taken to avoid issuing
824 	 * the bio twice.
825 	 */
826 	struct bio *bio;
827 	bio_end_io_t *saved_bi_end_io;
828 };
829 
830 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
831 {
832 	struct pool *pool = m->tc->pool;
833 
834 	if (atomic_dec_and_test(&m->prepare_actions)) {
835 		list_add_tail(&m->list, &pool->prepared_mappings);
836 		wake_worker(pool);
837 	}
838 }
839 
840 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
841 {
842 	unsigned long flags;
843 	struct pool *pool = m->tc->pool;
844 
845 	spin_lock_irqsave(&pool->lock, flags);
846 	__complete_mapping_preparation(m);
847 	spin_unlock_irqrestore(&pool->lock, flags);
848 }
849 
850 static void copy_complete(int read_err, unsigned long write_err, void *context)
851 {
852 	struct dm_thin_new_mapping *m = context;
853 
854 	m->status = read_err || write_err ? BLK_STS_IOERR : 0;
855 	complete_mapping_preparation(m);
856 }
857 
858 static void overwrite_endio(struct bio *bio)
859 {
860 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
861 	struct dm_thin_new_mapping *m = h->overwrite_mapping;
862 
863 	bio->bi_end_io = m->saved_bi_end_io;
864 
865 	m->status = bio->bi_status;
866 	complete_mapping_preparation(m);
867 }
868 
869 /*----------------------------------------------------------------*/
870 
871 /*
872  * Workqueue.
873  */
874 
875 /*
876  * Prepared mapping jobs.
877  */
878 
879 /*
880  * This sends the bios in the cell, except the original holder, back
881  * to the deferred_bios list.
882  */
883 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
884 {
885 	struct pool *pool = tc->pool;
886 	unsigned long flags;
887 	int has_work;
888 
889 	spin_lock_irqsave(&tc->lock, flags);
890 	cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
891 	has_work = !bio_list_empty(&tc->deferred_bio_list);
892 	spin_unlock_irqrestore(&tc->lock, flags);
893 
894 	if (has_work)
895 		wake_worker(pool);
896 }
897 
898 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
899 
900 struct remap_info {
901 	struct thin_c *tc;
902 	struct bio_list defer_bios;
903 	struct bio_list issue_bios;
904 };
905 
906 static void __inc_remap_and_issue_cell(void *context,
907 				       struct dm_bio_prison_cell *cell)
908 {
909 	struct remap_info *info = context;
910 	struct bio *bio;
911 
912 	while ((bio = bio_list_pop(&cell->bios))) {
913 		if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
914 			bio_list_add(&info->defer_bios, bio);
915 		else {
916 			inc_all_io_entry(info->tc->pool, bio);
917 
918 			/*
919 			 * We can't issue the bios with the bio prison lock
920 			 * held, so we add them to a list to issue on
921 			 * return from this function.
922 			 */
923 			bio_list_add(&info->issue_bios, bio);
924 		}
925 	}
926 }
927 
928 static void inc_remap_and_issue_cell(struct thin_c *tc,
929 				     struct dm_bio_prison_cell *cell,
930 				     dm_block_t block)
931 {
932 	struct bio *bio;
933 	struct remap_info info;
934 
935 	info.tc = tc;
936 	bio_list_init(&info.defer_bios);
937 	bio_list_init(&info.issue_bios);
938 
939 	/*
940 	 * We have to be careful to inc any bios we're about to issue
941 	 * before the cell is released, and avoid a race with new bios
942 	 * being added to the cell.
943 	 */
944 	cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
945 			   &info, cell);
946 
947 	while ((bio = bio_list_pop(&info.defer_bios)))
948 		thin_defer_bio(tc, bio);
949 
950 	while ((bio = bio_list_pop(&info.issue_bios)))
951 		remap_and_issue(info.tc, bio, block);
952 }
953 
954 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
955 {
956 	cell_error(m->tc->pool, m->cell);
957 	list_del(&m->list);
958 	mempool_free(m, &m->tc->pool->mapping_pool);
959 }
960 
961 static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
962 {
963 	struct pool *pool = tc->pool;
964 
965 	/*
966 	 * If the bio has the REQ_FUA flag set we must commit the metadata
967 	 * before signaling its completion.
968 	 */
969 	if (!bio_triggers_commit(tc, bio)) {
970 		bio_endio(bio);
971 		return;
972 	}
973 
974 	/*
975 	 * Complete bio with an error if earlier I/O caused changes to the
976 	 * metadata that can't be committed, e.g, due to I/O errors on the
977 	 * metadata device.
978 	 */
979 	if (dm_thin_aborted_changes(tc->td)) {
980 		bio_io_error(bio);
981 		return;
982 	}
983 
984 	/*
985 	 * Batch together any bios that trigger commits and then issue a
986 	 * single commit for them in process_deferred_bios().
987 	 */
988 	spin_lock_irq(&pool->lock);
989 	bio_list_add(&pool->deferred_flush_completions, bio);
990 	spin_unlock_irq(&pool->lock);
991 }
992 
993 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
994 {
995 	struct thin_c *tc = m->tc;
996 	struct pool *pool = tc->pool;
997 	struct bio *bio = m->bio;
998 	int r;
999 
1000 	if (m->status) {
1001 		cell_error(pool, m->cell);
1002 		goto out;
1003 	}
1004 
1005 	/*
1006 	 * Commit the prepared block into the mapping btree.
1007 	 * Any I/O for this block arriving after this point will get
1008 	 * remapped to it directly.
1009 	 */
1010 	r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1011 	if (r) {
1012 		metadata_operation_failed(pool, "dm_thin_insert_block", r);
1013 		cell_error(pool, m->cell);
1014 		goto out;
1015 	}
1016 
1017 	/*
1018 	 * Release any bios held while the block was being provisioned.
1019 	 * If we are processing a write bio that completely covers the block,
1020 	 * we already processed it so can ignore it now when processing
1021 	 * the bios in the cell.
1022 	 */
1023 	if (bio) {
1024 		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1025 		complete_overwrite_bio(tc, bio);
1026 	} else {
1027 		inc_all_io_entry(tc->pool, m->cell->holder);
1028 		remap_and_issue(tc, m->cell->holder, m->data_block);
1029 		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1030 	}
1031 
1032 out:
1033 	list_del(&m->list);
1034 	mempool_free(m, &pool->mapping_pool);
1035 }
1036 
1037 /*----------------------------------------------------------------*/
1038 
1039 static void free_discard_mapping(struct dm_thin_new_mapping *m)
1040 {
1041 	struct thin_c *tc = m->tc;
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(GFP_NOIO, 1);
1183 	if (!discard_parent) {
1184 		DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1185 		       dm_device_name(tc->pool->pool_md));
1186 		queue_passdown_pt2(m);
1187 
1188 	} else {
1189 		discard_parent->bi_end_io = passdown_endio;
1190 		discard_parent->bi_private = m;
1191 
1192 		if (m->maybe_shared)
1193 			passdown_double_checking_shared_status(m, discard_parent);
1194 		else {
1195 			struct discard_op op;
1196 
1197 			begin_discard(&op, tc, discard_parent);
1198 			r = issue_discard(&op, m->data_block, data_end);
1199 			end_discard(&op, r);
1200 		}
1201 	}
1202 }
1203 
1204 static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1205 {
1206 	int r;
1207 	struct thin_c *tc = m->tc;
1208 	struct pool *pool = tc->pool;
1209 
1210 	/*
1211 	 * The passdown has completed, so now we can decrement all those
1212 	 * unmapped blocks.
1213 	 */
1214 	r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1215 				   m->data_block + (m->virt_end - m->virt_begin));
1216 	if (r) {
1217 		metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1218 		bio_io_error(m->bio);
1219 	} else
1220 		bio_endio(m->bio);
1221 
1222 	cell_defer_no_holder(tc, m->cell);
1223 	mempool_free(m, &pool->mapping_pool);
1224 }
1225 
1226 static void process_prepared(struct pool *pool, struct list_head *head,
1227 			     process_mapping_fn *fn)
1228 {
1229 	struct list_head maps;
1230 	struct dm_thin_new_mapping *m, *tmp;
1231 
1232 	INIT_LIST_HEAD(&maps);
1233 	spin_lock_irq(&pool->lock);
1234 	list_splice_init(head, &maps);
1235 	spin_unlock_irq(&pool->lock);
1236 
1237 	list_for_each_entry_safe(m, tmp, &maps, list)
1238 		(*fn)(m);
1239 }
1240 
1241 /*
1242  * Deferred bio jobs.
1243  */
1244 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1245 {
1246 	return bio->bi_iter.bi_size ==
1247 		(pool->sectors_per_block << SECTOR_SHIFT);
1248 }
1249 
1250 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1251 {
1252 	return (bio_data_dir(bio) == WRITE) &&
1253 		io_overlaps_block(pool, bio);
1254 }
1255 
1256 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1257 			       bio_end_io_t *fn)
1258 {
1259 	*save = bio->bi_end_io;
1260 	bio->bi_end_io = fn;
1261 }
1262 
1263 static int ensure_next_mapping(struct pool *pool)
1264 {
1265 	if (pool->next_mapping)
1266 		return 0;
1267 
1268 	pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1269 
1270 	return pool->next_mapping ? 0 : -ENOMEM;
1271 }
1272 
1273 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1274 {
1275 	struct dm_thin_new_mapping *m = pool->next_mapping;
1276 
1277 	BUG_ON(!pool->next_mapping);
1278 
1279 	memset(m, 0, sizeof(struct dm_thin_new_mapping));
1280 	INIT_LIST_HEAD(&m->list);
1281 	m->bio = NULL;
1282 
1283 	pool->next_mapping = NULL;
1284 
1285 	return m;
1286 }
1287 
1288 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1289 		    sector_t begin, sector_t end)
1290 {
1291 	struct dm_io_region to;
1292 
1293 	to.bdev = tc->pool_dev->bdev;
1294 	to.sector = begin;
1295 	to.count = end - begin;
1296 
1297 	dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1298 }
1299 
1300 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1301 				      dm_block_t data_begin,
1302 				      struct dm_thin_new_mapping *m)
1303 {
1304 	struct pool *pool = tc->pool;
1305 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1306 
1307 	h->overwrite_mapping = m;
1308 	m->bio = bio;
1309 	save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1310 	inc_all_io_entry(pool, bio);
1311 	remap_and_issue(tc, bio, data_begin);
1312 }
1313 
1314 /*
1315  * A partial copy also needs to zero the uncopied region.
1316  */
1317 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1318 			  struct dm_dev *origin, dm_block_t data_origin,
1319 			  dm_block_t data_dest,
1320 			  struct dm_bio_prison_cell *cell, struct bio *bio,
1321 			  sector_t len)
1322 {
1323 	struct pool *pool = tc->pool;
1324 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1325 
1326 	m->tc = tc;
1327 	m->virt_begin = virt_block;
1328 	m->virt_end = virt_block + 1u;
1329 	m->data_block = data_dest;
1330 	m->cell = cell;
1331 
1332 	/*
1333 	 * quiesce action + copy action + an extra reference held for the
1334 	 * duration of this function (we may need to inc later for a
1335 	 * partial zero).
1336 	 */
1337 	atomic_set(&m->prepare_actions, 3);
1338 
1339 	if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1340 		complete_mapping_preparation(m); /* already quiesced */
1341 
1342 	/*
1343 	 * IO to pool_dev remaps to the pool target's data_dev.
1344 	 *
1345 	 * If the whole block of data is being overwritten, we can issue the
1346 	 * bio immediately. Otherwise we use kcopyd to clone the data first.
1347 	 */
1348 	if (io_overwrites_block(pool, bio))
1349 		remap_and_issue_overwrite(tc, bio, data_dest, m);
1350 	else {
1351 		struct dm_io_region from, to;
1352 
1353 		from.bdev = origin->bdev;
1354 		from.sector = data_origin * pool->sectors_per_block;
1355 		from.count = len;
1356 
1357 		to.bdev = tc->pool_dev->bdev;
1358 		to.sector = data_dest * pool->sectors_per_block;
1359 		to.count = len;
1360 
1361 		dm_kcopyd_copy(pool->copier, &from, 1, &to,
1362 			       0, copy_complete, m);
1363 
1364 		/*
1365 		 * Do we need to zero a tail region?
1366 		 */
1367 		if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1368 			atomic_inc(&m->prepare_actions);
1369 			ll_zero(tc, m,
1370 				data_dest * pool->sectors_per_block + len,
1371 				(data_dest + 1) * pool->sectors_per_block);
1372 		}
1373 	}
1374 
1375 	complete_mapping_preparation(m); /* drop our ref */
1376 }
1377 
1378 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1379 				   dm_block_t data_origin, dm_block_t data_dest,
1380 				   struct dm_bio_prison_cell *cell, struct bio *bio)
1381 {
1382 	schedule_copy(tc, virt_block, tc->pool_dev,
1383 		      data_origin, data_dest, cell, bio,
1384 		      tc->pool->sectors_per_block);
1385 }
1386 
1387 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1388 			  dm_block_t data_block, struct dm_bio_prison_cell *cell,
1389 			  struct bio *bio)
1390 {
1391 	struct pool *pool = tc->pool;
1392 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1393 
1394 	atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1395 	m->tc = tc;
1396 	m->virt_begin = virt_block;
1397 	m->virt_end = virt_block + 1u;
1398 	m->data_block = data_block;
1399 	m->cell = cell;
1400 
1401 	/*
1402 	 * If the whole block of data is being overwritten or we are not
1403 	 * zeroing pre-existing data, we can issue the bio immediately.
1404 	 * Otherwise we use kcopyd to zero the data first.
1405 	 */
1406 	if (pool->pf.zero_new_blocks) {
1407 		if (io_overwrites_block(pool, bio))
1408 			remap_and_issue_overwrite(tc, bio, data_block, m);
1409 		else
1410 			ll_zero(tc, m, data_block * pool->sectors_per_block,
1411 				(data_block + 1) * pool->sectors_per_block);
1412 	} else
1413 		process_prepared_mapping(m);
1414 }
1415 
1416 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1417 				   dm_block_t data_dest,
1418 				   struct dm_bio_prison_cell *cell, struct bio *bio)
1419 {
1420 	struct pool *pool = tc->pool;
1421 	sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1422 	sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1423 
1424 	if (virt_block_end <= tc->origin_size)
1425 		schedule_copy(tc, virt_block, tc->origin_dev,
1426 			      virt_block, data_dest, cell, bio,
1427 			      pool->sectors_per_block);
1428 
1429 	else if (virt_block_begin < tc->origin_size)
1430 		schedule_copy(tc, virt_block, tc->origin_dev,
1431 			      virt_block, data_dest, cell, bio,
1432 			      tc->origin_size - virt_block_begin);
1433 
1434 	else
1435 		schedule_zero(tc, virt_block, data_dest, cell, bio);
1436 }
1437 
1438 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1439 
1440 static void requeue_bios(struct pool *pool);
1441 
1442 static bool is_read_only_pool_mode(enum pool_mode mode)
1443 {
1444 	return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1445 }
1446 
1447 static bool is_read_only(struct pool *pool)
1448 {
1449 	return is_read_only_pool_mode(get_pool_mode(pool));
1450 }
1451 
1452 static void check_for_metadata_space(struct pool *pool)
1453 {
1454 	int r;
1455 	const char *ooms_reason = NULL;
1456 	dm_block_t nr_free;
1457 
1458 	r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1459 	if (r)
1460 		ooms_reason = "Could not get free metadata blocks";
1461 	else if (!nr_free)
1462 		ooms_reason = "No free metadata blocks";
1463 
1464 	if (ooms_reason && !is_read_only(pool)) {
1465 		DMERR("%s", ooms_reason);
1466 		set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1467 	}
1468 }
1469 
1470 static void check_for_data_space(struct pool *pool)
1471 {
1472 	int r;
1473 	dm_block_t nr_free;
1474 
1475 	if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1476 		return;
1477 
1478 	r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1479 	if (r)
1480 		return;
1481 
1482 	if (nr_free) {
1483 		set_pool_mode(pool, PM_WRITE);
1484 		requeue_bios(pool);
1485 	}
1486 }
1487 
1488 /*
1489  * A non-zero return indicates read_only or fail_io mode.
1490  * Many callers don't care about the return value.
1491  */
1492 static int commit(struct pool *pool)
1493 {
1494 	int r;
1495 
1496 	if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1497 		return -EINVAL;
1498 
1499 	r = dm_pool_commit_metadata(pool->pmd);
1500 	if (r)
1501 		metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1502 	else {
1503 		check_for_metadata_space(pool);
1504 		check_for_data_space(pool);
1505 	}
1506 
1507 	return r;
1508 }
1509 
1510 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1511 {
1512 	if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1513 		DMWARN("%s: reached low water mark for data device: sending event.",
1514 		       dm_device_name(pool->pool_md));
1515 		spin_lock_irq(&pool->lock);
1516 		pool->low_water_triggered = true;
1517 		spin_unlock_irq(&pool->lock);
1518 		dm_table_event(pool->ti->table);
1519 	}
1520 }
1521 
1522 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1523 {
1524 	int r;
1525 	dm_block_t free_blocks;
1526 	struct pool *pool = tc->pool;
1527 
1528 	if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1529 		return -EINVAL;
1530 
1531 	r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1532 	if (r) {
1533 		metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1534 		return r;
1535 	}
1536 
1537 	check_low_water_mark(pool, free_blocks);
1538 
1539 	if (!free_blocks) {
1540 		/*
1541 		 * Try to commit to see if that will free up some
1542 		 * more space.
1543 		 */
1544 		r = commit(pool);
1545 		if (r)
1546 			return r;
1547 
1548 		r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1549 		if (r) {
1550 			metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1551 			return r;
1552 		}
1553 
1554 		if (!free_blocks) {
1555 			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1556 			return -ENOSPC;
1557 		}
1558 	}
1559 
1560 	r = dm_pool_alloc_data_block(pool->pmd, result);
1561 	if (r) {
1562 		if (r == -ENOSPC)
1563 			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1564 		else
1565 			metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1566 		return r;
1567 	}
1568 
1569 	r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1570 	if (r) {
1571 		metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1572 		return r;
1573 	}
1574 
1575 	if (!free_blocks) {
1576 		/* Let's commit before we use up the metadata reserve. */
1577 		r = commit(pool);
1578 		if (r)
1579 			return r;
1580 	}
1581 
1582 	return 0;
1583 }
1584 
1585 /*
1586  * If we have run out of space, queue bios until the device is
1587  * resumed, presumably after having been reloaded with more space.
1588  */
1589 static void retry_on_resume(struct bio *bio)
1590 {
1591 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1592 	struct thin_c *tc = h->tc;
1593 
1594 	spin_lock_irq(&tc->lock);
1595 	bio_list_add(&tc->retry_on_resume_list, bio);
1596 	spin_unlock_irq(&tc->lock);
1597 }
1598 
1599 static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1600 {
1601 	enum pool_mode m = get_pool_mode(pool);
1602 
1603 	switch (m) {
1604 	case PM_WRITE:
1605 		/* Shouldn't get here */
1606 		DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1607 		return BLK_STS_IOERR;
1608 
1609 	case PM_OUT_OF_DATA_SPACE:
1610 		return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1611 
1612 	case PM_OUT_OF_METADATA_SPACE:
1613 	case PM_READ_ONLY:
1614 	case PM_FAIL:
1615 		return BLK_STS_IOERR;
1616 	default:
1617 		/* Shouldn't get here */
1618 		DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1619 		return BLK_STS_IOERR;
1620 	}
1621 }
1622 
1623 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1624 {
1625 	blk_status_t error = should_error_unserviceable_bio(pool);
1626 
1627 	if (error) {
1628 		bio->bi_status = error;
1629 		bio_endio(bio);
1630 	} else
1631 		retry_on_resume(bio);
1632 }
1633 
1634 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1635 {
1636 	struct bio *bio;
1637 	struct bio_list bios;
1638 	blk_status_t error;
1639 
1640 	error = should_error_unserviceable_bio(pool);
1641 	if (error) {
1642 		cell_error_with_code(pool, cell, error);
1643 		return;
1644 	}
1645 
1646 	bio_list_init(&bios);
1647 	cell_release(pool, cell, &bios);
1648 
1649 	while ((bio = bio_list_pop(&bios)))
1650 		retry_on_resume(bio);
1651 }
1652 
1653 static void process_discard_cell_no_passdown(struct thin_c *tc,
1654 					     struct dm_bio_prison_cell *virt_cell)
1655 {
1656 	struct pool *pool = tc->pool;
1657 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1658 
1659 	/*
1660 	 * We don't need to lock the data blocks, since there's no
1661 	 * passdown.  We only lock data blocks for allocation and breaking sharing.
1662 	 */
1663 	m->tc = tc;
1664 	m->virt_begin = virt_cell->key.block_begin;
1665 	m->virt_end = virt_cell->key.block_end;
1666 	m->cell = virt_cell;
1667 	m->bio = virt_cell->holder;
1668 
1669 	if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1670 		pool->process_prepared_discard(m);
1671 }
1672 
1673 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1674 				 struct bio *bio)
1675 {
1676 	struct pool *pool = tc->pool;
1677 
1678 	int r;
1679 	bool maybe_shared;
1680 	struct dm_cell_key data_key;
1681 	struct dm_bio_prison_cell *data_cell;
1682 	struct dm_thin_new_mapping *m;
1683 	dm_block_t virt_begin, virt_end, data_begin;
1684 
1685 	while (begin != end) {
1686 		r = ensure_next_mapping(pool);
1687 		if (r)
1688 			/* we did our best */
1689 			return;
1690 
1691 		r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1692 					      &data_begin, &maybe_shared);
1693 		if (r)
1694 			/*
1695 			 * Silently fail, letting any mappings we've
1696 			 * created complete.
1697 			 */
1698 			break;
1699 
1700 		build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1701 		if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1702 			/* contention, we'll give up with this range */
1703 			begin = virt_end;
1704 			continue;
1705 		}
1706 
1707 		/*
1708 		 * IO may still be going to the destination block.  We must
1709 		 * quiesce before we can do the removal.
1710 		 */
1711 		m = get_next_mapping(pool);
1712 		m->tc = tc;
1713 		m->maybe_shared = maybe_shared;
1714 		m->virt_begin = virt_begin;
1715 		m->virt_end = virt_end;
1716 		m->data_block = data_begin;
1717 		m->cell = data_cell;
1718 		m->bio = bio;
1719 
1720 		/*
1721 		 * The parent bio must not complete before sub discard bios are
1722 		 * chained to it (see end_discard's bio_chain)!
1723 		 *
1724 		 * This per-mapping bi_remaining increment is paired with
1725 		 * the implicit decrement that occurs via bio_endio() in
1726 		 * end_discard().
1727 		 */
1728 		bio_inc_remaining(bio);
1729 		if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1730 			pool->process_prepared_discard(m);
1731 
1732 		begin = virt_end;
1733 	}
1734 }
1735 
1736 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1737 {
1738 	struct bio *bio = virt_cell->holder;
1739 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1740 
1741 	/*
1742 	 * The virt_cell will only get freed once the origin bio completes.
1743 	 * This means it will remain locked while all the individual
1744 	 * passdown bios are in flight.
1745 	 */
1746 	h->cell = virt_cell;
1747 	break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1748 
1749 	/*
1750 	 * We complete the bio now, knowing that the bi_remaining field
1751 	 * will prevent completion until the sub range discards have
1752 	 * completed.
1753 	 */
1754 	bio_endio(bio);
1755 }
1756 
1757 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1758 {
1759 	dm_block_t begin, end;
1760 	struct dm_cell_key virt_key;
1761 	struct dm_bio_prison_cell *virt_cell;
1762 
1763 	get_bio_block_range(tc, bio, &begin, &end);
1764 	if (begin == end) {
1765 		/*
1766 		 * The discard covers less than a block.
1767 		 */
1768 		bio_endio(bio);
1769 		return;
1770 	}
1771 
1772 	build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1773 	if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1774 		/*
1775 		 * Potential starvation issue: We're relying on the
1776 		 * fs/application being well behaved, and not trying to
1777 		 * send IO to a region at the same time as discarding it.
1778 		 * If they do this persistently then it's possible this
1779 		 * cell will never be granted.
1780 		 */
1781 		return;
1782 
1783 	tc->pool->process_discard_cell(tc, virt_cell);
1784 }
1785 
1786 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1787 			  struct dm_cell_key *key,
1788 			  struct dm_thin_lookup_result *lookup_result,
1789 			  struct dm_bio_prison_cell *cell)
1790 {
1791 	int r;
1792 	dm_block_t data_block;
1793 	struct pool *pool = tc->pool;
1794 
1795 	r = alloc_data_block(tc, &data_block);
1796 	switch (r) {
1797 	case 0:
1798 		schedule_internal_copy(tc, block, lookup_result->block,
1799 				       data_block, cell, bio);
1800 		break;
1801 
1802 	case -ENOSPC:
1803 		retry_bios_on_resume(pool, cell);
1804 		break;
1805 
1806 	default:
1807 		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1808 			    __func__, r);
1809 		cell_error(pool, cell);
1810 		break;
1811 	}
1812 }
1813 
1814 static void __remap_and_issue_shared_cell(void *context,
1815 					  struct dm_bio_prison_cell *cell)
1816 {
1817 	struct remap_info *info = context;
1818 	struct bio *bio;
1819 
1820 	while ((bio = bio_list_pop(&cell->bios))) {
1821 		if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1822 		    bio_op(bio) == REQ_OP_DISCARD)
1823 			bio_list_add(&info->defer_bios, bio);
1824 		else {
1825 			struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1826 
1827 			h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1828 			inc_all_io_entry(info->tc->pool, bio);
1829 			bio_list_add(&info->issue_bios, bio);
1830 		}
1831 	}
1832 }
1833 
1834 static void remap_and_issue_shared_cell(struct thin_c *tc,
1835 					struct dm_bio_prison_cell *cell,
1836 					dm_block_t block)
1837 {
1838 	struct bio *bio;
1839 	struct remap_info info;
1840 
1841 	info.tc = tc;
1842 	bio_list_init(&info.defer_bios);
1843 	bio_list_init(&info.issue_bios);
1844 
1845 	cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1846 			   &info, cell);
1847 
1848 	while ((bio = bio_list_pop(&info.defer_bios)))
1849 		thin_defer_bio(tc, bio);
1850 
1851 	while ((bio = bio_list_pop(&info.issue_bios)))
1852 		remap_and_issue(tc, bio, block);
1853 }
1854 
1855 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1856 			       dm_block_t block,
1857 			       struct dm_thin_lookup_result *lookup_result,
1858 			       struct dm_bio_prison_cell *virt_cell)
1859 {
1860 	struct dm_bio_prison_cell *data_cell;
1861 	struct pool *pool = tc->pool;
1862 	struct dm_cell_key key;
1863 
1864 	/*
1865 	 * If cell is already occupied, then sharing is already in the process
1866 	 * of being broken so we have nothing further to do here.
1867 	 */
1868 	build_data_key(tc->td, lookup_result->block, &key);
1869 	if (bio_detain(pool, &key, bio, &data_cell)) {
1870 		cell_defer_no_holder(tc, virt_cell);
1871 		return;
1872 	}
1873 
1874 	if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1875 		break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1876 		cell_defer_no_holder(tc, virt_cell);
1877 	} else {
1878 		struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1879 
1880 		h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1881 		inc_all_io_entry(pool, bio);
1882 		remap_and_issue(tc, bio, lookup_result->block);
1883 
1884 		remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1885 		remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1886 	}
1887 }
1888 
1889 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1890 			    struct dm_bio_prison_cell *cell)
1891 {
1892 	int r;
1893 	dm_block_t data_block;
1894 	struct pool *pool = tc->pool;
1895 
1896 	/*
1897 	 * Remap empty bios (flushes) immediately, without provisioning.
1898 	 */
1899 	if (!bio->bi_iter.bi_size) {
1900 		inc_all_io_entry(pool, bio);
1901 		cell_defer_no_holder(tc, cell);
1902 
1903 		remap_and_issue(tc, bio, 0);
1904 		return;
1905 	}
1906 
1907 	/*
1908 	 * Fill read bios with zeroes and complete them immediately.
1909 	 */
1910 	if (bio_data_dir(bio) == READ) {
1911 		zero_fill_bio(bio);
1912 		cell_defer_no_holder(tc, cell);
1913 		bio_endio(bio);
1914 		return;
1915 	}
1916 
1917 	r = alloc_data_block(tc, &data_block);
1918 	switch (r) {
1919 	case 0:
1920 		if (tc->origin_dev)
1921 			schedule_external_copy(tc, block, data_block, cell, bio);
1922 		else
1923 			schedule_zero(tc, block, data_block, cell, bio);
1924 		break;
1925 
1926 	case -ENOSPC:
1927 		retry_bios_on_resume(pool, cell);
1928 		break;
1929 
1930 	default:
1931 		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1932 			    __func__, r);
1933 		cell_error(pool, cell);
1934 		break;
1935 	}
1936 }
1937 
1938 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1939 {
1940 	int r;
1941 	struct pool *pool = tc->pool;
1942 	struct bio *bio = cell->holder;
1943 	dm_block_t block = get_bio_block(tc, bio);
1944 	struct dm_thin_lookup_result lookup_result;
1945 
1946 	if (tc->requeue_mode) {
1947 		cell_requeue(pool, cell);
1948 		return;
1949 	}
1950 
1951 	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1952 	switch (r) {
1953 	case 0:
1954 		if (lookup_result.shared)
1955 			process_shared_bio(tc, bio, block, &lookup_result, cell);
1956 		else {
1957 			inc_all_io_entry(pool, bio);
1958 			remap_and_issue(tc, bio, lookup_result.block);
1959 			inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1960 		}
1961 		break;
1962 
1963 	case -ENODATA:
1964 		if (bio_data_dir(bio) == READ && tc->origin_dev) {
1965 			inc_all_io_entry(pool, bio);
1966 			cell_defer_no_holder(tc, cell);
1967 
1968 			if (bio_end_sector(bio) <= tc->origin_size)
1969 				remap_to_origin_and_issue(tc, bio);
1970 
1971 			else if (bio->bi_iter.bi_sector < tc->origin_size) {
1972 				zero_fill_bio(bio);
1973 				bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1974 				remap_to_origin_and_issue(tc, bio);
1975 
1976 			} else {
1977 				zero_fill_bio(bio);
1978 				bio_endio(bio);
1979 			}
1980 		} else
1981 			provision_block(tc, bio, block, cell);
1982 		break;
1983 
1984 	default:
1985 		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1986 			    __func__, r);
1987 		cell_defer_no_holder(tc, cell);
1988 		bio_io_error(bio);
1989 		break;
1990 	}
1991 }
1992 
1993 static void process_bio(struct thin_c *tc, struct bio *bio)
1994 {
1995 	struct pool *pool = tc->pool;
1996 	dm_block_t block = get_bio_block(tc, bio);
1997 	struct dm_bio_prison_cell *cell;
1998 	struct dm_cell_key key;
1999 
2000 	/*
2001 	 * If cell is already occupied, then the block is already
2002 	 * being provisioned so we have nothing further to do here.
2003 	 */
2004 	build_virtual_key(tc->td, block, &key);
2005 	if (bio_detain(pool, &key, bio, &cell))
2006 		return;
2007 
2008 	process_cell(tc, cell);
2009 }
2010 
2011 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2012 				    struct dm_bio_prison_cell *cell)
2013 {
2014 	int r;
2015 	int rw = bio_data_dir(bio);
2016 	dm_block_t block = get_bio_block(tc, bio);
2017 	struct dm_thin_lookup_result lookup_result;
2018 
2019 	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2020 	switch (r) {
2021 	case 0:
2022 		if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2023 			handle_unserviceable_bio(tc->pool, bio);
2024 			if (cell)
2025 				cell_defer_no_holder(tc, cell);
2026 		} else {
2027 			inc_all_io_entry(tc->pool, bio);
2028 			remap_and_issue(tc, bio, lookup_result.block);
2029 			if (cell)
2030 				inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2031 		}
2032 		break;
2033 
2034 	case -ENODATA:
2035 		if (cell)
2036 			cell_defer_no_holder(tc, cell);
2037 		if (rw != READ) {
2038 			handle_unserviceable_bio(tc->pool, bio);
2039 			break;
2040 		}
2041 
2042 		if (tc->origin_dev) {
2043 			inc_all_io_entry(tc->pool, bio);
2044 			remap_to_origin_and_issue(tc, bio);
2045 			break;
2046 		}
2047 
2048 		zero_fill_bio(bio);
2049 		bio_endio(bio);
2050 		break;
2051 
2052 	default:
2053 		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2054 			    __func__, r);
2055 		if (cell)
2056 			cell_defer_no_holder(tc, cell);
2057 		bio_io_error(bio);
2058 		break;
2059 	}
2060 }
2061 
2062 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2063 {
2064 	__process_bio_read_only(tc, bio, NULL);
2065 }
2066 
2067 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2068 {
2069 	__process_bio_read_only(tc, cell->holder, cell);
2070 }
2071 
2072 static void process_bio_success(struct thin_c *tc, struct bio *bio)
2073 {
2074 	bio_endio(bio);
2075 }
2076 
2077 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2078 {
2079 	bio_io_error(bio);
2080 }
2081 
2082 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2083 {
2084 	cell_success(tc->pool, cell);
2085 }
2086 
2087 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2088 {
2089 	cell_error(tc->pool, cell);
2090 }
2091 
2092 /*
2093  * FIXME: should we also commit due to size of transaction, measured in
2094  * metadata blocks?
2095  */
2096 static int need_commit_due_to_time(struct pool *pool)
2097 {
2098 	return !time_in_range(jiffies, pool->last_commit_jiffies,
2099 			      pool->last_commit_jiffies + COMMIT_PERIOD);
2100 }
2101 
2102 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2103 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2104 
2105 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2106 {
2107 	struct rb_node **rbp, *parent;
2108 	struct dm_thin_endio_hook *pbd;
2109 	sector_t bi_sector = bio->bi_iter.bi_sector;
2110 
2111 	rbp = &tc->sort_bio_list.rb_node;
2112 	parent = NULL;
2113 	while (*rbp) {
2114 		parent = *rbp;
2115 		pbd = thin_pbd(parent);
2116 
2117 		if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2118 			rbp = &(*rbp)->rb_left;
2119 		else
2120 			rbp = &(*rbp)->rb_right;
2121 	}
2122 
2123 	pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2124 	rb_link_node(&pbd->rb_node, parent, rbp);
2125 	rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2126 }
2127 
2128 static void __extract_sorted_bios(struct thin_c *tc)
2129 {
2130 	struct rb_node *node;
2131 	struct dm_thin_endio_hook *pbd;
2132 	struct bio *bio;
2133 
2134 	for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2135 		pbd = thin_pbd(node);
2136 		bio = thin_bio(pbd);
2137 
2138 		bio_list_add(&tc->deferred_bio_list, bio);
2139 		rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2140 	}
2141 
2142 	WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2143 }
2144 
2145 static void __sort_thin_deferred_bios(struct thin_c *tc)
2146 {
2147 	struct bio *bio;
2148 	struct bio_list bios;
2149 
2150 	bio_list_init(&bios);
2151 	bio_list_merge(&bios, &tc->deferred_bio_list);
2152 	bio_list_init(&tc->deferred_bio_list);
2153 
2154 	/* Sort deferred_bio_list using rb-tree */
2155 	while ((bio = bio_list_pop(&bios)))
2156 		__thin_bio_rb_add(tc, bio);
2157 
2158 	/*
2159 	 * Transfer the sorted bios in sort_bio_list back to
2160 	 * deferred_bio_list to allow lockless submission of
2161 	 * all bios.
2162 	 */
2163 	__extract_sorted_bios(tc);
2164 }
2165 
2166 static void process_thin_deferred_bios(struct thin_c *tc)
2167 {
2168 	struct pool *pool = tc->pool;
2169 	struct bio *bio;
2170 	struct bio_list bios;
2171 	struct blk_plug plug;
2172 	unsigned count = 0;
2173 
2174 	if (tc->requeue_mode) {
2175 		error_thin_bio_list(tc, &tc->deferred_bio_list,
2176 				BLK_STS_DM_REQUEUE);
2177 		return;
2178 	}
2179 
2180 	bio_list_init(&bios);
2181 
2182 	spin_lock_irq(&tc->lock);
2183 
2184 	if (bio_list_empty(&tc->deferred_bio_list)) {
2185 		spin_unlock_irq(&tc->lock);
2186 		return;
2187 	}
2188 
2189 	__sort_thin_deferred_bios(tc);
2190 
2191 	bio_list_merge(&bios, &tc->deferred_bio_list);
2192 	bio_list_init(&tc->deferred_bio_list);
2193 
2194 	spin_unlock_irq(&tc->lock);
2195 
2196 	blk_start_plug(&plug);
2197 	while ((bio = bio_list_pop(&bios))) {
2198 		/*
2199 		 * If we've got no free new_mapping structs, and processing
2200 		 * this bio might require one, we pause until there are some
2201 		 * prepared mappings to process.
2202 		 */
2203 		if (ensure_next_mapping(pool)) {
2204 			spin_lock_irq(&tc->lock);
2205 			bio_list_add(&tc->deferred_bio_list, bio);
2206 			bio_list_merge(&tc->deferred_bio_list, &bios);
2207 			spin_unlock_irq(&tc->lock);
2208 			break;
2209 		}
2210 
2211 		if (bio_op(bio) == REQ_OP_DISCARD)
2212 			pool->process_discard(tc, bio);
2213 		else
2214 			pool->process_bio(tc, bio);
2215 
2216 		if ((count++ & 127) == 0) {
2217 			throttle_work_update(&pool->throttle);
2218 			dm_pool_issue_prefetches(pool->pmd);
2219 		}
2220 	}
2221 	blk_finish_plug(&plug);
2222 }
2223 
2224 static int cmp_cells(const void *lhs, const void *rhs)
2225 {
2226 	struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2227 	struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2228 
2229 	BUG_ON(!lhs_cell->holder);
2230 	BUG_ON(!rhs_cell->holder);
2231 
2232 	if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2233 		return -1;
2234 
2235 	if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2236 		return 1;
2237 
2238 	return 0;
2239 }
2240 
2241 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2242 {
2243 	unsigned count = 0;
2244 	struct dm_bio_prison_cell *cell, *tmp;
2245 
2246 	list_for_each_entry_safe(cell, tmp, cells, user_list) {
2247 		if (count >= CELL_SORT_ARRAY_SIZE)
2248 			break;
2249 
2250 		pool->cell_sort_array[count++] = cell;
2251 		list_del(&cell->user_list);
2252 	}
2253 
2254 	sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2255 
2256 	return count;
2257 }
2258 
2259 static void process_thin_deferred_cells(struct thin_c *tc)
2260 {
2261 	struct pool *pool = tc->pool;
2262 	struct list_head cells;
2263 	struct dm_bio_prison_cell *cell;
2264 	unsigned i, j, count;
2265 
2266 	INIT_LIST_HEAD(&cells);
2267 
2268 	spin_lock_irq(&tc->lock);
2269 	list_splice_init(&tc->deferred_cells, &cells);
2270 	spin_unlock_irq(&tc->lock);
2271 
2272 	if (list_empty(&cells))
2273 		return;
2274 
2275 	do {
2276 		count = sort_cells(tc->pool, &cells);
2277 
2278 		for (i = 0; i < count; i++) {
2279 			cell = pool->cell_sort_array[i];
2280 			BUG_ON(!cell->holder);
2281 
2282 			/*
2283 			 * If we've got no free new_mapping structs, and processing
2284 			 * this bio might require one, we pause until there are some
2285 			 * prepared mappings to process.
2286 			 */
2287 			if (ensure_next_mapping(pool)) {
2288 				for (j = i; j < count; j++)
2289 					list_add(&pool->cell_sort_array[j]->user_list, &cells);
2290 
2291 				spin_lock_irq(&tc->lock);
2292 				list_splice(&cells, &tc->deferred_cells);
2293 				spin_unlock_irq(&tc->lock);
2294 				return;
2295 			}
2296 
2297 			if (bio_op(cell->holder) == REQ_OP_DISCARD)
2298 				pool->process_discard_cell(tc, cell);
2299 			else
2300 				pool->process_cell(tc, cell);
2301 		}
2302 	} while (!list_empty(&cells));
2303 }
2304 
2305 static void thin_get(struct thin_c *tc);
2306 static void thin_put(struct thin_c *tc);
2307 
2308 /*
2309  * We can't hold rcu_read_lock() around code that can block.  So we
2310  * find a thin with the rcu lock held; bump a refcount; then drop
2311  * the lock.
2312  */
2313 static struct thin_c *get_first_thin(struct pool *pool)
2314 {
2315 	struct thin_c *tc = NULL;
2316 
2317 	rcu_read_lock();
2318 	if (!list_empty(&pool->active_thins)) {
2319 		tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2320 		thin_get(tc);
2321 	}
2322 	rcu_read_unlock();
2323 
2324 	return tc;
2325 }
2326 
2327 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2328 {
2329 	struct thin_c *old_tc = tc;
2330 
2331 	rcu_read_lock();
2332 	list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2333 		thin_get(tc);
2334 		thin_put(old_tc);
2335 		rcu_read_unlock();
2336 		return tc;
2337 	}
2338 	thin_put(old_tc);
2339 	rcu_read_unlock();
2340 
2341 	return NULL;
2342 }
2343 
2344 static void process_deferred_bios(struct pool *pool)
2345 {
2346 	struct bio *bio;
2347 	struct bio_list bios, bio_completions;
2348 	struct thin_c *tc;
2349 
2350 	tc = get_first_thin(pool);
2351 	while (tc) {
2352 		process_thin_deferred_cells(tc);
2353 		process_thin_deferred_bios(tc);
2354 		tc = get_next_thin(pool, tc);
2355 	}
2356 
2357 	/*
2358 	 * If there are any deferred flush bios, we must commit the metadata
2359 	 * before issuing them or signaling their completion.
2360 	 */
2361 	bio_list_init(&bios);
2362 	bio_list_init(&bio_completions);
2363 
2364 	spin_lock_irq(&pool->lock);
2365 	bio_list_merge(&bios, &pool->deferred_flush_bios);
2366 	bio_list_init(&pool->deferred_flush_bios);
2367 
2368 	bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2369 	bio_list_init(&pool->deferred_flush_completions);
2370 	spin_unlock_irq(&pool->lock);
2371 
2372 	if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2373 	    !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2374 		return;
2375 
2376 	if (commit(pool)) {
2377 		bio_list_merge(&bios, &bio_completions);
2378 
2379 		while ((bio = bio_list_pop(&bios)))
2380 			bio_io_error(bio);
2381 		return;
2382 	}
2383 	pool->last_commit_jiffies = jiffies;
2384 
2385 	while ((bio = bio_list_pop(&bio_completions)))
2386 		bio_endio(bio);
2387 
2388 	while ((bio = bio_list_pop(&bios))) {
2389 		/*
2390 		 * The data device was flushed as part of metadata commit,
2391 		 * so complete redundant flushes immediately.
2392 		 */
2393 		if (bio->bi_opf & REQ_PREFLUSH)
2394 			bio_endio(bio);
2395 		else
2396 			submit_bio_noacct(bio);
2397 	}
2398 }
2399 
2400 static void do_worker(struct work_struct *ws)
2401 {
2402 	struct pool *pool = container_of(ws, struct pool, worker);
2403 
2404 	throttle_work_start(&pool->throttle);
2405 	dm_pool_issue_prefetches(pool->pmd);
2406 	throttle_work_update(&pool->throttle);
2407 	process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2408 	throttle_work_update(&pool->throttle);
2409 	process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2410 	throttle_work_update(&pool->throttle);
2411 	process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2412 	throttle_work_update(&pool->throttle);
2413 	process_deferred_bios(pool);
2414 	throttle_work_complete(&pool->throttle);
2415 }
2416 
2417 /*
2418  * We want to commit periodically so that not too much
2419  * unwritten data builds up.
2420  */
2421 static void do_waker(struct work_struct *ws)
2422 {
2423 	struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2424 	wake_worker(pool);
2425 	queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2426 }
2427 
2428 /*
2429  * We're holding onto IO to allow userland time to react.  After the
2430  * timeout either the pool will have been resized (and thus back in
2431  * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2432  */
2433 static void do_no_space_timeout(struct work_struct *ws)
2434 {
2435 	struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2436 					 no_space_timeout);
2437 
2438 	if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2439 		pool->pf.error_if_no_space = true;
2440 		notify_of_pool_mode_change(pool);
2441 		error_retry_list_with_code(pool, BLK_STS_NOSPC);
2442 	}
2443 }
2444 
2445 /*----------------------------------------------------------------*/
2446 
2447 struct pool_work {
2448 	struct work_struct worker;
2449 	struct completion complete;
2450 };
2451 
2452 static struct pool_work *to_pool_work(struct work_struct *ws)
2453 {
2454 	return container_of(ws, struct pool_work, worker);
2455 }
2456 
2457 static void pool_work_complete(struct pool_work *pw)
2458 {
2459 	complete(&pw->complete);
2460 }
2461 
2462 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2463 			   void (*fn)(struct work_struct *))
2464 {
2465 	INIT_WORK_ONSTACK(&pw->worker, fn);
2466 	init_completion(&pw->complete);
2467 	queue_work(pool->wq, &pw->worker);
2468 	wait_for_completion(&pw->complete);
2469 }
2470 
2471 /*----------------------------------------------------------------*/
2472 
2473 struct noflush_work {
2474 	struct pool_work pw;
2475 	struct thin_c *tc;
2476 };
2477 
2478 static struct noflush_work *to_noflush(struct work_struct *ws)
2479 {
2480 	return container_of(to_pool_work(ws), struct noflush_work, pw);
2481 }
2482 
2483 static void do_noflush_start(struct work_struct *ws)
2484 {
2485 	struct noflush_work *w = to_noflush(ws);
2486 	w->tc->requeue_mode = true;
2487 	requeue_io(w->tc);
2488 	pool_work_complete(&w->pw);
2489 }
2490 
2491 static void do_noflush_stop(struct work_struct *ws)
2492 {
2493 	struct noflush_work *w = to_noflush(ws);
2494 	w->tc->requeue_mode = false;
2495 	pool_work_complete(&w->pw);
2496 }
2497 
2498 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2499 {
2500 	struct noflush_work w;
2501 
2502 	w.tc = tc;
2503 	pool_work_wait(&w.pw, tc->pool, fn);
2504 }
2505 
2506 /*----------------------------------------------------------------*/
2507 
2508 static bool passdown_enabled(struct pool_c *pt)
2509 {
2510 	return pt->adjusted_pf.discard_passdown;
2511 }
2512 
2513 static void set_discard_callbacks(struct pool *pool)
2514 {
2515 	struct pool_c *pt = pool->ti->private;
2516 
2517 	if (passdown_enabled(pt)) {
2518 		pool->process_discard_cell = process_discard_cell_passdown;
2519 		pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2520 		pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2521 	} else {
2522 		pool->process_discard_cell = process_discard_cell_no_passdown;
2523 		pool->process_prepared_discard = process_prepared_discard_no_passdown;
2524 	}
2525 }
2526 
2527 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2528 {
2529 	struct pool_c *pt = pool->ti->private;
2530 	bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2531 	enum pool_mode old_mode = get_pool_mode(pool);
2532 	unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2533 
2534 	/*
2535 	 * Never allow the pool to transition to PM_WRITE mode if user
2536 	 * intervention is required to verify metadata and data consistency.
2537 	 */
2538 	if (new_mode == PM_WRITE && needs_check) {
2539 		DMERR("%s: unable to switch pool to write mode until repaired.",
2540 		      dm_device_name(pool->pool_md));
2541 		if (old_mode != new_mode)
2542 			new_mode = old_mode;
2543 		else
2544 			new_mode = PM_READ_ONLY;
2545 	}
2546 	/*
2547 	 * If we were in PM_FAIL mode, rollback of metadata failed.  We're
2548 	 * not going to recover without a thin_repair.	So we never let the
2549 	 * pool move out of the old mode.
2550 	 */
2551 	if (old_mode == PM_FAIL)
2552 		new_mode = old_mode;
2553 
2554 	switch (new_mode) {
2555 	case PM_FAIL:
2556 		dm_pool_metadata_read_only(pool->pmd);
2557 		pool->process_bio = process_bio_fail;
2558 		pool->process_discard = process_bio_fail;
2559 		pool->process_cell = process_cell_fail;
2560 		pool->process_discard_cell = process_cell_fail;
2561 		pool->process_prepared_mapping = process_prepared_mapping_fail;
2562 		pool->process_prepared_discard = process_prepared_discard_fail;
2563 
2564 		error_retry_list(pool);
2565 		break;
2566 
2567 	case PM_OUT_OF_METADATA_SPACE:
2568 	case PM_READ_ONLY:
2569 		dm_pool_metadata_read_only(pool->pmd);
2570 		pool->process_bio = process_bio_read_only;
2571 		pool->process_discard = process_bio_success;
2572 		pool->process_cell = process_cell_read_only;
2573 		pool->process_discard_cell = process_cell_success;
2574 		pool->process_prepared_mapping = process_prepared_mapping_fail;
2575 		pool->process_prepared_discard = process_prepared_discard_success;
2576 
2577 		error_retry_list(pool);
2578 		break;
2579 
2580 	case PM_OUT_OF_DATA_SPACE:
2581 		/*
2582 		 * Ideally we'd never hit this state; the low water mark
2583 		 * would trigger userland to extend the pool before we
2584 		 * completely run out of data space.  However, many small
2585 		 * IOs to unprovisioned space can consume data space at an
2586 		 * alarming rate.  Adjust your low water mark if you're
2587 		 * frequently seeing this mode.
2588 		 */
2589 		pool->out_of_data_space = true;
2590 		pool->process_bio = process_bio_read_only;
2591 		pool->process_discard = process_discard_bio;
2592 		pool->process_cell = process_cell_read_only;
2593 		pool->process_prepared_mapping = process_prepared_mapping;
2594 		set_discard_callbacks(pool);
2595 
2596 		if (!pool->pf.error_if_no_space && no_space_timeout)
2597 			queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2598 		break;
2599 
2600 	case PM_WRITE:
2601 		if (old_mode == PM_OUT_OF_DATA_SPACE)
2602 			cancel_delayed_work_sync(&pool->no_space_timeout);
2603 		pool->out_of_data_space = false;
2604 		pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2605 		dm_pool_metadata_read_write(pool->pmd);
2606 		pool->process_bio = process_bio;
2607 		pool->process_discard = process_discard_bio;
2608 		pool->process_cell = process_cell;
2609 		pool->process_prepared_mapping = process_prepared_mapping;
2610 		set_discard_callbacks(pool);
2611 		break;
2612 	}
2613 
2614 	pool->pf.mode = new_mode;
2615 	/*
2616 	 * The pool mode may have changed, sync it so bind_control_target()
2617 	 * doesn't cause an unexpected mode transition on resume.
2618 	 */
2619 	pt->adjusted_pf.mode = new_mode;
2620 
2621 	if (old_mode != new_mode)
2622 		notify_of_pool_mode_change(pool);
2623 }
2624 
2625 static void abort_transaction(struct pool *pool)
2626 {
2627 	const char *dev_name = dm_device_name(pool->pool_md);
2628 
2629 	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2630 	if (dm_pool_abort_metadata(pool->pmd)) {
2631 		DMERR("%s: failed to abort metadata transaction", dev_name);
2632 		set_pool_mode(pool, PM_FAIL);
2633 	}
2634 
2635 	if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2636 		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2637 		set_pool_mode(pool, PM_FAIL);
2638 	}
2639 }
2640 
2641 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2642 {
2643 	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2644 		    dm_device_name(pool->pool_md), op, r);
2645 
2646 	abort_transaction(pool);
2647 	set_pool_mode(pool, PM_READ_ONLY);
2648 }
2649 
2650 /*----------------------------------------------------------------*/
2651 
2652 /*
2653  * Mapping functions.
2654  */
2655 
2656 /*
2657  * Called only while mapping a thin bio to hand it over to the workqueue.
2658  */
2659 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2660 {
2661 	struct pool *pool = tc->pool;
2662 
2663 	spin_lock_irq(&tc->lock);
2664 	bio_list_add(&tc->deferred_bio_list, bio);
2665 	spin_unlock_irq(&tc->lock);
2666 
2667 	wake_worker(pool);
2668 }
2669 
2670 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2671 {
2672 	struct pool *pool = tc->pool;
2673 
2674 	throttle_lock(&pool->throttle);
2675 	thin_defer_bio(tc, bio);
2676 	throttle_unlock(&pool->throttle);
2677 }
2678 
2679 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2680 {
2681 	struct pool *pool = tc->pool;
2682 
2683 	throttle_lock(&pool->throttle);
2684 	spin_lock_irq(&tc->lock);
2685 	list_add_tail(&cell->user_list, &tc->deferred_cells);
2686 	spin_unlock_irq(&tc->lock);
2687 	throttle_unlock(&pool->throttle);
2688 
2689 	wake_worker(pool);
2690 }
2691 
2692 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2693 {
2694 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2695 
2696 	h->tc = tc;
2697 	h->shared_read_entry = NULL;
2698 	h->all_io_entry = NULL;
2699 	h->overwrite_mapping = NULL;
2700 	h->cell = NULL;
2701 }
2702 
2703 /*
2704  * Non-blocking function called from the thin target's map function.
2705  */
2706 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2707 {
2708 	int r;
2709 	struct thin_c *tc = ti->private;
2710 	dm_block_t block = get_bio_block(tc, bio);
2711 	struct dm_thin_device *td = tc->td;
2712 	struct dm_thin_lookup_result result;
2713 	struct dm_bio_prison_cell *virt_cell, *data_cell;
2714 	struct dm_cell_key key;
2715 
2716 	thin_hook_bio(tc, bio);
2717 
2718 	if (tc->requeue_mode) {
2719 		bio->bi_status = BLK_STS_DM_REQUEUE;
2720 		bio_endio(bio);
2721 		return DM_MAPIO_SUBMITTED;
2722 	}
2723 
2724 	if (get_pool_mode(tc->pool) == PM_FAIL) {
2725 		bio_io_error(bio);
2726 		return DM_MAPIO_SUBMITTED;
2727 	}
2728 
2729 	if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2730 		thin_defer_bio_with_throttle(tc, bio);
2731 		return DM_MAPIO_SUBMITTED;
2732 	}
2733 
2734 	/*
2735 	 * We must hold the virtual cell before doing the lookup, otherwise
2736 	 * there's a race with discard.
2737 	 */
2738 	build_virtual_key(tc->td, block, &key);
2739 	if (bio_detain(tc->pool, &key, bio, &virt_cell))
2740 		return DM_MAPIO_SUBMITTED;
2741 
2742 	r = dm_thin_find_block(td, block, 0, &result);
2743 
2744 	/*
2745 	 * Note that we defer readahead too.
2746 	 */
2747 	switch (r) {
2748 	case 0:
2749 		if (unlikely(result.shared)) {
2750 			/*
2751 			 * We have a race condition here between the
2752 			 * result.shared value returned by the lookup and
2753 			 * snapshot creation, which may cause new
2754 			 * sharing.
2755 			 *
2756 			 * To avoid this always quiesce the origin before
2757 			 * taking the snap.  You want to do this anyway to
2758 			 * ensure a consistent application view
2759 			 * (i.e. lockfs).
2760 			 *
2761 			 * More distant ancestors are irrelevant. The
2762 			 * shared flag will be set in their case.
2763 			 */
2764 			thin_defer_cell(tc, virt_cell);
2765 			return DM_MAPIO_SUBMITTED;
2766 		}
2767 
2768 		build_data_key(tc->td, result.block, &key);
2769 		if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2770 			cell_defer_no_holder(tc, virt_cell);
2771 			return DM_MAPIO_SUBMITTED;
2772 		}
2773 
2774 		inc_all_io_entry(tc->pool, bio);
2775 		cell_defer_no_holder(tc, data_cell);
2776 		cell_defer_no_holder(tc, virt_cell);
2777 
2778 		remap(tc, bio, result.block);
2779 		return DM_MAPIO_REMAPPED;
2780 
2781 	case -ENODATA:
2782 	case -EWOULDBLOCK:
2783 		thin_defer_cell(tc, virt_cell);
2784 		return DM_MAPIO_SUBMITTED;
2785 
2786 	default:
2787 		/*
2788 		 * Must always call bio_io_error on failure.
2789 		 * dm_thin_find_block can fail with -EINVAL if the
2790 		 * pool is switched to fail-io mode.
2791 		 */
2792 		bio_io_error(bio);
2793 		cell_defer_no_holder(tc, virt_cell);
2794 		return DM_MAPIO_SUBMITTED;
2795 	}
2796 }
2797 
2798 static void requeue_bios(struct pool *pool)
2799 {
2800 	struct thin_c *tc;
2801 
2802 	rcu_read_lock();
2803 	list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2804 		spin_lock_irq(&tc->lock);
2805 		bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2806 		bio_list_init(&tc->retry_on_resume_list);
2807 		spin_unlock_irq(&tc->lock);
2808 	}
2809 	rcu_read_unlock();
2810 }
2811 
2812 /*----------------------------------------------------------------
2813  * Binding of control targets to a pool object
2814  *--------------------------------------------------------------*/
2815 static bool data_dev_supports_discard(struct pool_c *pt)
2816 {
2817 	struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2818 
2819 	return blk_queue_discard(q);
2820 }
2821 
2822 static bool is_factor(sector_t block_size, uint32_t n)
2823 {
2824 	return !sector_div(block_size, n);
2825 }
2826 
2827 /*
2828  * If discard_passdown was enabled verify that the data device
2829  * supports discards.  Disable discard_passdown if not.
2830  */
2831 static void disable_passdown_if_not_supported(struct pool_c *pt)
2832 {
2833 	struct pool *pool = pt->pool;
2834 	struct block_device *data_bdev = pt->data_dev->bdev;
2835 	struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2836 	const char *reason = NULL;
2837 	char buf[BDEVNAME_SIZE];
2838 
2839 	if (!pt->adjusted_pf.discard_passdown)
2840 		return;
2841 
2842 	if (!data_dev_supports_discard(pt))
2843 		reason = "discard unsupported";
2844 
2845 	else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2846 		reason = "max discard sectors smaller than a block";
2847 
2848 	if (reason) {
2849 		DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2850 		pt->adjusted_pf.discard_passdown = false;
2851 	}
2852 }
2853 
2854 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2855 {
2856 	struct pool_c *pt = ti->private;
2857 
2858 	/*
2859 	 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2860 	 */
2861 	enum pool_mode old_mode = get_pool_mode(pool);
2862 	enum pool_mode new_mode = pt->adjusted_pf.mode;
2863 
2864 	/*
2865 	 * Don't change the pool's mode until set_pool_mode() below.
2866 	 * Otherwise the pool's process_* function pointers may
2867 	 * not match the desired pool mode.
2868 	 */
2869 	pt->adjusted_pf.mode = old_mode;
2870 
2871 	pool->ti = ti;
2872 	pool->pf = pt->adjusted_pf;
2873 	pool->low_water_blocks = pt->low_water_blocks;
2874 
2875 	set_pool_mode(pool, new_mode);
2876 
2877 	return 0;
2878 }
2879 
2880 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2881 {
2882 	if (pool->ti == ti)
2883 		pool->ti = NULL;
2884 }
2885 
2886 /*----------------------------------------------------------------
2887  * Pool creation
2888  *--------------------------------------------------------------*/
2889 /* Initialize pool features. */
2890 static void pool_features_init(struct pool_features *pf)
2891 {
2892 	pf->mode = PM_WRITE;
2893 	pf->zero_new_blocks = true;
2894 	pf->discard_enabled = true;
2895 	pf->discard_passdown = true;
2896 	pf->error_if_no_space = false;
2897 }
2898 
2899 static void __pool_destroy(struct pool *pool)
2900 {
2901 	__pool_table_remove(pool);
2902 
2903 	vfree(pool->cell_sort_array);
2904 	if (dm_pool_metadata_close(pool->pmd) < 0)
2905 		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2906 
2907 	dm_bio_prison_destroy(pool->prison);
2908 	dm_kcopyd_client_destroy(pool->copier);
2909 
2910 	if (pool->wq)
2911 		destroy_workqueue(pool->wq);
2912 
2913 	if (pool->next_mapping)
2914 		mempool_free(pool->next_mapping, &pool->mapping_pool);
2915 	mempool_exit(&pool->mapping_pool);
2916 	bio_uninit(&pool->flush_bio);
2917 	dm_deferred_set_destroy(pool->shared_read_ds);
2918 	dm_deferred_set_destroy(pool->all_io_ds);
2919 	kfree(pool);
2920 }
2921 
2922 static struct kmem_cache *_new_mapping_cache;
2923 
2924 static struct pool *pool_create(struct mapped_device *pool_md,
2925 				struct block_device *metadata_dev,
2926 				struct block_device *data_dev,
2927 				unsigned long block_size,
2928 				int read_only, char **error)
2929 {
2930 	int r;
2931 	void *err_p;
2932 	struct pool *pool;
2933 	struct dm_pool_metadata *pmd;
2934 	bool format_device = read_only ? false : true;
2935 
2936 	pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2937 	if (IS_ERR(pmd)) {
2938 		*error = "Error creating metadata object";
2939 		return (struct pool *)pmd;
2940 	}
2941 
2942 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2943 	if (!pool) {
2944 		*error = "Error allocating memory for pool";
2945 		err_p = ERR_PTR(-ENOMEM);
2946 		goto bad_pool;
2947 	}
2948 
2949 	pool->pmd = pmd;
2950 	pool->sectors_per_block = block_size;
2951 	if (block_size & (block_size - 1))
2952 		pool->sectors_per_block_shift = -1;
2953 	else
2954 		pool->sectors_per_block_shift = __ffs(block_size);
2955 	pool->low_water_blocks = 0;
2956 	pool_features_init(&pool->pf);
2957 	pool->prison = dm_bio_prison_create();
2958 	if (!pool->prison) {
2959 		*error = "Error creating pool's bio prison";
2960 		err_p = ERR_PTR(-ENOMEM);
2961 		goto bad_prison;
2962 	}
2963 
2964 	pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2965 	if (IS_ERR(pool->copier)) {
2966 		r = PTR_ERR(pool->copier);
2967 		*error = "Error creating pool's kcopyd client";
2968 		err_p = ERR_PTR(r);
2969 		goto bad_kcopyd_client;
2970 	}
2971 
2972 	/*
2973 	 * Create singlethreaded workqueue that will service all devices
2974 	 * that use this metadata.
2975 	 */
2976 	pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2977 	if (!pool->wq) {
2978 		*error = "Error creating pool's workqueue";
2979 		err_p = ERR_PTR(-ENOMEM);
2980 		goto bad_wq;
2981 	}
2982 
2983 	throttle_init(&pool->throttle);
2984 	INIT_WORK(&pool->worker, do_worker);
2985 	INIT_DELAYED_WORK(&pool->waker, do_waker);
2986 	INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2987 	spin_lock_init(&pool->lock);
2988 	bio_list_init(&pool->deferred_flush_bios);
2989 	bio_list_init(&pool->deferred_flush_completions);
2990 	INIT_LIST_HEAD(&pool->prepared_mappings);
2991 	INIT_LIST_HEAD(&pool->prepared_discards);
2992 	INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2993 	INIT_LIST_HEAD(&pool->active_thins);
2994 	pool->low_water_triggered = false;
2995 	pool->suspended = true;
2996 	pool->out_of_data_space = false;
2997 	bio_init(&pool->flush_bio, NULL, 0);
2998 
2999 	pool->shared_read_ds = dm_deferred_set_create();
3000 	if (!pool->shared_read_ds) {
3001 		*error = "Error creating pool's shared read deferred set";
3002 		err_p = ERR_PTR(-ENOMEM);
3003 		goto bad_shared_read_ds;
3004 	}
3005 
3006 	pool->all_io_ds = dm_deferred_set_create();
3007 	if (!pool->all_io_ds) {
3008 		*error = "Error creating pool's all io deferred set";
3009 		err_p = ERR_PTR(-ENOMEM);
3010 		goto bad_all_io_ds;
3011 	}
3012 
3013 	pool->next_mapping = NULL;
3014 	r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
3015 				   _new_mapping_cache);
3016 	if (r) {
3017 		*error = "Error creating pool's mapping mempool";
3018 		err_p = ERR_PTR(r);
3019 		goto bad_mapping_pool;
3020 	}
3021 
3022 	pool->cell_sort_array =
3023 		vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3024 				   sizeof(*pool->cell_sort_array)));
3025 	if (!pool->cell_sort_array) {
3026 		*error = "Error allocating cell sort array";
3027 		err_p = ERR_PTR(-ENOMEM);
3028 		goto bad_sort_array;
3029 	}
3030 
3031 	pool->ref_count = 1;
3032 	pool->last_commit_jiffies = jiffies;
3033 	pool->pool_md = pool_md;
3034 	pool->md_dev = metadata_dev;
3035 	pool->data_dev = data_dev;
3036 	__pool_table_insert(pool);
3037 
3038 	return pool;
3039 
3040 bad_sort_array:
3041 	mempool_exit(&pool->mapping_pool);
3042 bad_mapping_pool:
3043 	dm_deferred_set_destroy(pool->all_io_ds);
3044 bad_all_io_ds:
3045 	dm_deferred_set_destroy(pool->shared_read_ds);
3046 bad_shared_read_ds:
3047 	destroy_workqueue(pool->wq);
3048 bad_wq:
3049 	dm_kcopyd_client_destroy(pool->copier);
3050 bad_kcopyd_client:
3051 	dm_bio_prison_destroy(pool->prison);
3052 bad_prison:
3053 	kfree(pool);
3054 bad_pool:
3055 	if (dm_pool_metadata_close(pmd))
3056 		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3057 
3058 	return err_p;
3059 }
3060 
3061 static void __pool_inc(struct pool *pool)
3062 {
3063 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3064 	pool->ref_count++;
3065 }
3066 
3067 static void __pool_dec(struct pool *pool)
3068 {
3069 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3070 	BUG_ON(!pool->ref_count);
3071 	if (!--pool->ref_count)
3072 		__pool_destroy(pool);
3073 }
3074 
3075 static struct pool *__pool_find(struct mapped_device *pool_md,
3076 				struct block_device *metadata_dev,
3077 				struct block_device *data_dev,
3078 				unsigned long block_size, int read_only,
3079 				char **error, int *created)
3080 {
3081 	struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3082 
3083 	if (pool) {
3084 		if (pool->pool_md != pool_md) {
3085 			*error = "metadata device already in use by a pool";
3086 			return ERR_PTR(-EBUSY);
3087 		}
3088 		if (pool->data_dev != data_dev) {
3089 			*error = "data device already in use by a pool";
3090 			return ERR_PTR(-EBUSY);
3091 		}
3092 		__pool_inc(pool);
3093 
3094 	} else {
3095 		pool = __pool_table_lookup(pool_md);
3096 		if (pool) {
3097 			if (pool->md_dev != metadata_dev || pool->data_dev != data_dev) {
3098 				*error = "different pool cannot replace a pool";
3099 				return ERR_PTR(-EINVAL);
3100 			}
3101 			__pool_inc(pool);
3102 
3103 		} else {
3104 			pool = pool_create(pool_md, metadata_dev, data_dev, block_size, read_only, error);
3105 			*created = 1;
3106 		}
3107 	}
3108 
3109 	return pool;
3110 }
3111 
3112 /*----------------------------------------------------------------
3113  * Pool target methods
3114  *--------------------------------------------------------------*/
3115 static void pool_dtr(struct dm_target *ti)
3116 {
3117 	struct pool_c *pt = ti->private;
3118 
3119 	mutex_lock(&dm_thin_pool_table.mutex);
3120 
3121 	unbind_control_target(pt->pool, ti);
3122 	__pool_dec(pt->pool);
3123 	dm_put_device(ti, pt->metadata_dev);
3124 	dm_put_device(ti, pt->data_dev);
3125 	kfree(pt);
3126 
3127 	mutex_unlock(&dm_thin_pool_table.mutex);
3128 }
3129 
3130 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3131 			       struct dm_target *ti)
3132 {
3133 	int r;
3134 	unsigned argc;
3135 	const char *arg_name;
3136 
3137 	static const struct dm_arg _args[] = {
3138 		{0, 4, "Invalid number of pool feature arguments"},
3139 	};
3140 
3141 	/*
3142 	 * No feature arguments supplied.
3143 	 */
3144 	if (!as->argc)
3145 		return 0;
3146 
3147 	r = dm_read_arg_group(_args, as, &argc, &ti->error);
3148 	if (r)
3149 		return -EINVAL;
3150 
3151 	while (argc && !r) {
3152 		arg_name = dm_shift_arg(as);
3153 		argc--;
3154 
3155 		if (!strcasecmp(arg_name, "skip_block_zeroing"))
3156 			pf->zero_new_blocks = false;
3157 
3158 		else if (!strcasecmp(arg_name, "ignore_discard"))
3159 			pf->discard_enabled = false;
3160 
3161 		else if (!strcasecmp(arg_name, "no_discard_passdown"))
3162 			pf->discard_passdown = false;
3163 
3164 		else if (!strcasecmp(arg_name, "read_only"))
3165 			pf->mode = PM_READ_ONLY;
3166 
3167 		else if (!strcasecmp(arg_name, "error_if_no_space"))
3168 			pf->error_if_no_space = true;
3169 
3170 		else {
3171 			ti->error = "Unrecognised pool feature requested";
3172 			r = -EINVAL;
3173 			break;
3174 		}
3175 	}
3176 
3177 	return r;
3178 }
3179 
3180 static void metadata_low_callback(void *context)
3181 {
3182 	struct pool *pool = context;
3183 
3184 	DMWARN("%s: reached low water mark for metadata device: sending event.",
3185 	       dm_device_name(pool->pool_md));
3186 
3187 	dm_table_event(pool->ti->table);
3188 }
3189 
3190 /*
3191  * We need to flush the data device **before** committing the metadata.
3192  *
3193  * This ensures that the data blocks of any newly inserted mappings are
3194  * properly written to non-volatile storage and won't be lost in case of a
3195  * crash.
3196  *
3197  * Failure to do so can result in data corruption in the case of internal or
3198  * external snapshots and in the case of newly provisioned blocks, when block
3199  * zeroing is enabled.
3200  */
3201 static int metadata_pre_commit_callback(void *context)
3202 {
3203 	struct pool *pool = context;
3204 	struct bio *flush_bio = &pool->flush_bio;
3205 
3206 	bio_reset(flush_bio);
3207 	bio_set_dev(flush_bio, pool->data_dev);
3208 	flush_bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
3209 
3210 	return submit_bio_wait(flush_bio);
3211 }
3212 
3213 static sector_t get_dev_size(struct block_device *bdev)
3214 {
3215 	return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3216 }
3217 
3218 static void warn_if_metadata_device_too_big(struct block_device *bdev)
3219 {
3220 	sector_t metadata_dev_size = get_dev_size(bdev);
3221 	char buffer[BDEVNAME_SIZE];
3222 
3223 	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3224 		DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3225 		       bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3226 }
3227 
3228 static sector_t get_metadata_dev_size(struct block_device *bdev)
3229 {
3230 	sector_t metadata_dev_size = get_dev_size(bdev);
3231 
3232 	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3233 		metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3234 
3235 	return metadata_dev_size;
3236 }
3237 
3238 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3239 {
3240 	sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3241 
3242 	sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3243 
3244 	return metadata_dev_size;
3245 }
3246 
3247 /*
3248  * When a metadata threshold is crossed a dm event is triggered, and
3249  * userland should respond by growing the metadata device.  We could let
3250  * userland set the threshold, like we do with the data threshold, but I'm
3251  * not sure they know enough to do this well.
3252  */
3253 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3254 {
3255 	/*
3256 	 * 4M is ample for all ops with the possible exception of thin
3257 	 * device deletion which is harmless if it fails (just retry the
3258 	 * delete after you've grown the device).
3259 	 */
3260 	dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3261 	return min((dm_block_t)1024ULL /* 4M */, quarter);
3262 }
3263 
3264 /*
3265  * thin-pool <metadata dev> <data dev>
3266  *	     <data block size (sectors)>
3267  *	     <low water mark (blocks)>
3268  *	     [<#feature args> [<arg>]*]
3269  *
3270  * Optional feature arguments are:
3271  *	     skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3272  *	     ignore_discard: disable discard
3273  *	     no_discard_passdown: don't pass discards down to the data device
3274  *	     read_only: Don't allow any changes to be made to the pool metadata.
3275  *	     error_if_no_space: error IOs, instead of queueing, if no space.
3276  */
3277 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3278 {
3279 	int r, pool_created = 0;
3280 	struct pool_c *pt;
3281 	struct pool *pool;
3282 	struct pool_features pf;
3283 	struct dm_arg_set as;
3284 	struct dm_dev *data_dev;
3285 	unsigned long block_size;
3286 	dm_block_t low_water_blocks;
3287 	struct dm_dev *metadata_dev;
3288 	fmode_t metadata_mode;
3289 
3290 	/*
3291 	 * FIXME Remove validation from scope of lock.
3292 	 */
3293 	mutex_lock(&dm_thin_pool_table.mutex);
3294 
3295 	if (argc < 4) {
3296 		ti->error = "Invalid argument count";
3297 		r = -EINVAL;
3298 		goto out_unlock;
3299 	}
3300 
3301 	as.argc = argc;
3302 	as.argv = argv;
3303 
3304 	/* make sure metadata and data are different devices */
3305 	if (!strcmp(argv[0], argv[1])) {
3306 		ti->error = "Error setting metadata or data device";
3307 		r = -EINVAL;
3308 		goto out_unlock;
3309 	}
3310 
3311 	/*
3312 	 * Set default pool features.
3313 	 */
3314 	pool_features_init(&pf);
3315 
3316 	dm_consume_args(&as, 4);
3317 	r = parse_pool_features(&as, &pf, ti);
3318 	if (r)
3319 		goto out_unlock;
3320 
3321 	metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3322 	r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3323 	if (r) {
3324 		ti->error = "Error opening metadata block device";
3325 		goto out_unlock;
3326 	}
3327 	warn_if_metadata_device_too_big(metadata_dev->bdev);
3328 
3329 	r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3330 	if (r) {
3331 		ti->error = "Error getting data device";
3332 		goto out_metadata;
3333 	}
3334 
3335 	if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3336 	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3337 	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3338 	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3339 		ti->error = "Invalid block size";
3340 		r = -EINVAL;
3341 		goto out;
3342 	}
3343 
3344 	if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3345 		ti->error = "Invalid low water mark";
3346 		r = -EINVAL;
3347 		goto out;
3348 	}
3349 
3350 	pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3351 	if (!pt) {
3352 		r = -ENOMEM;
3353 		goto out;
3354 	}
3355 
3356 	pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev, data_dev->bdev,
3357 			   block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3358 	if (IS_ERR(pool)) {
3359 		r = PTR_ERR(pool);
3360 		goto out_free_pt;
3361 	}
3362 
3363 	/*
3364 	 * 'pool_created' reflects whether this is the first table load.
3365 	 * Top level discard support is not allowed to be changed after
3366 	 * initial load.  This would require a pool reload to trigger thin
3367 	 * device changes.
3368 	 */
3369 	if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3370 		ti->error = "Discard support cannot be disabled once enabled";
3371 		r = -EINVAL;
3372 		goto out_flags_changed;
3373 	}
3374 
3375 	pt->pool = pool;
3376 	pt->ti = ti;
3377 	pt->metadata_dev = metadata_dev;
3378 	pt->data_dev = data_dev;
3379 	pt->low_water_blocks = low_water_blocks;
3380 	pt->adjusted_pf = pt->requested_pf = pf;
3381 	ti->num_flush_bios = 1;
3382 
3383 	/*
3384 	 * Only need to enable discards if the pool should pass
3385 	 * them down to the data device.  The thin device's discard
3386 	 * processing will cause mappings to be removed from the btree.
3387 	 */
3388 	if (pf.discard_enabled && pf.discard_passdown) {
3389 		ti->num_discard_bios = 1;
3390 
3391 		/*
3392 		 * Setting 'discards_supported' circumvents the normal
3393 		 * stacking of discard limits (this keeps the pool and
3394 		 * thin devices' discard limits consistent).
3395 		 */
3396 		ti->discards_supported = true;
3397 	}
3398 	ti->private = pt;
3399 
3400 	r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3401 						calc_metadata_threshold(pt),
3402 						metadata_low_callback,
3403 						pool);
3404 	if (r)
3405 		goto out_flags_changed;
3406 
3407 	dm_pool_register_pre_commit_callback(pool->pmd,
3408 					     metadata_pre_commit_callback, pool);
3409 
3410 	mutex_unlock(&dm_thin_pool_table.mutex);
3411 
3412 	return 0;
3413 
3414 out_flags_changed:
3415 	__pool_dec(pool);
3416 out_free_pt:
3417 	kfree(pt);
3418 out:
3419 	dm_put_device(ti, data_dev);
3420 out_metadata:
3421 	dm_put_device(ti, metadata_dev);
3422 out_unlock:
3423 	mutex_unlock(&dm_thin_pool_table.mutex);
3424 
3425 	return r;
3426 }
3427 
3428 static int pool_map(struct dm_target *ti, struct bio *bio)
3429 {
3430 	int r;
3431 	struct pool_c *pt = ti->private;
3432 	struct pool *pool = pt->pool;
3433 
3434 	/*
3435 	 * As this is a singleton target, ti->begin is always zero.
3436 	 */
3437 	spin_lock_irq(&pool->lock);
3438 	bio_set_dev(bio, pt->data_dev->bdev);
3439 	r = DM_MAPIO_REMAPPED;
3440 	spin_unlock_irq(&pool->lock);
3441 
3442 	return r;
3443 }
3444 
3445 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3446 {
3447 	int r;
3448 	struct pool_c *pt = ti->private;
3449 	struct pool *pool = pt->pool;
3450 	sector_t data_size = ti->len;
3451 	dm_block_t sb_data_size;
3452 
3453 	*need_commit = false;
3454 
3455 	(void) sector_div(data_size, pool->sectors_per_block);
3456 
3457 	r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3458 	if (r) {
3459 		DMERR("%s: failed to retrieve data device size",
3460 		      dm_device_name(pool->pool_md));
3461 		return r;
3462 	}
3463 
3464 	if (data_size < sb_data_size) {
3465 		DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3466 		      dm_device_name(pool->pool_md),
3467 		      (unsigned long long)data_size, sb_data_size);
3468 		return -EINVAL;
3469 
3470 	} else if (data_size > sb_data_size) {
3471 		if (dm_pool_metadata_needs_check(pool->pmd)) {
3472 			DMERR("%s: unable to grow the data device until repaired.",
3473 			      dm_device_name(pool->pool_md));
3474 			return 0;
3475 		}
3476 
3477 		if (sb_data_size)
3478 			DMINFO("%s: growing the data device from %llu to %llu blocks",
3479 			       dm_device_name(pool->pool_md),
3480 			       sb_data_size, (unsigned long long)data_size);
3481 		r = dm_pool_resize_data_dev(pool->pmd, data_size);
3482 		if (r) {
3483 			metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3484 			return r;
3485 		}
3486 
3487 		*need_commit = true;
3488 	}
3489 
3490 	return 0;
3491 }
3492 
3493 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3494 {
3495 	int r;
3496 	struct pool_c *pt = ti->private;
3497 	struct pool *pool = pt->pool;
3498 	dm_block_t metadata_dev_size, sb_metadata_dev_size;
3499 
3500 	*need_commit = false;
3501 
3502 	metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3503 
3504 	r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3505 	if (r) {
3506 		DMERR("%s: failed to retrieve metadata device size",
3507 		      dm_device_name(pool->pool_md));
3508 		return r;
3509 	}
3510 
3511 	if (metadata_dev_size < sb_metadata_dev_size) {
3512 		DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3513 		      dm_device_name(pool->pool_md),
3514 		      metadata_dev_size, sb_metadata_dev_size);
3515 		return -EINVAL;
3516 
3517 	} else if (metadata_dev_size > sb_metadata_dev_size) {
3518 		if (dm_pool_metadata_needs_check(pool->pmd)) {
3519 			DMERR("%s: unable to grow the metadata device until repaired.",
3520 			      dm_device_name(pool->pool_md));
3521 			return 0;
3522 		}
3523 
3524 		warn_if_metadata_device_too_big(pool->md_dev);
3525 		DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3526 		       dm_device_name(pool->pool_md),
3527 		       sb_metadata_dev_size, metadata_dev_size);
3528 
3529 		if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3530 			set_pool_mode(pool, PM_WRITE);
3531 
3532 		r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3533 		if (r) {
3534 			metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3535 			return r;
3536 		}
3537 
3538 		*need_commit = true;
3539 	}
3540 
3541 	return 0;
3542 }
3543 
3544 /*
3545  * Retrieves the number of blocks of the data device from
3546  * the superblock and compares it to the actual device size,
3547  * thus resizing the data device in case it has grown.
3548  *
3549  * This both copes with opening preallocated data devices in the ctr
3550  * being followed by a resume
3551  * -and-
3552  * calling the resume method individually after userspace has
3553  * grown the data device in reaction to a table event.
3554  */
3555 static int pool_preresume(struct dm_target *ti)
3556 {
3557 	int r;
3558 	bool need_commit1, need_commit2;
3559 	struct pool_c *pt = ti->private;
3560 	struct pool *pool = pt->pool;
3561 
3562 	/*
3563 	 * Take control of the pool object.
3564 	 */
3565 	r = bind_control_target(pool, ti);
3566 	if (r)
3567 		return r;
3568 
3569 	r = maybe_resize_data_dev(ti, &need_commit1);
3570 	if (r)
3571 		return r;
3572 
3573 	r = maybe_resize_metadata_dev(ti, &need_commit2);
3574 	if (r)
3575 		return r;
3576 
3577 	if (need_commit1 || need_commit2)
3578 		(void) commit(pool);
3579 
3580 	return 0;
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 argc, unsigned 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 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 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 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 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 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 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 argc, char **argv,
3827 			char *result, unsigned 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 sz, unsigned maxlen)
3868 {
3869 	unsigned 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 status_flags, char *result, unsigned maxlen)
3898 {
3899 	int r;
3900 	unsigned 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 		 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
4080 		 * user to see that, so make sure to set all discard limits to 0.
4081 		 */
4082 		limits->discard_granularity = 0;
4083 		return;
4084 	}
4085 
4086 	disable_passdown_if_not_supported(pt);
4087 
4088 	/*
4089 	 * The pool uses the same discard limits as the underlying data
4090 	 * device.  DM core has already set this up.
4091 	 */
4092 }
4093 
4094 static struct target_type pool_target = {
4095 	.name = "thin-pool",
4096 	.features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4097 		    DM_TARGET_IMMUTABLE,
4098 	.version = {1, 22, 0},
4099 	.module = THIS_MODULE,
4100 	.ctr = pool_ctr,
4101 	.dtr = pool_dtr,
4102 	.map = pool_map,
4103 	.presuspend = pool_presuspend,
4104 	.presuspend_undo = pool_presuspend_undo,
4105 	.postsuspend = pool_postsuspend,
4106 	.preresume = pool_preresume,
4107 	.resume = pool_resume,
4108 	.message = pool_message,
4109 	.status = pool_status,
4110 	.iterate_devices = pool_iterate_devices,
4111 	.io_hints = pool_io_hints,
4112 };
4113 
4114 /*----------------------------------------------------------------
4115  * Thin target methods
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 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->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4254 
4255 	/* In case the pool supports discards, pass them on. */
4256 	if (tc->pool->pf.discard_enabled) {
4257 		ti->discards_supported = true;
4258 		ti->num_discard_bios = 1;
4259 	}
4260 
4261 	mutex_unlock(&dm_thin_pool_table.mutex);
4262 
4263 	spin_lock_irq(&tc->pool->lock);
4264 	if (tc->pool->suspended) {
4265 		spin_unlock_irq(&tc->pool->lock);
4266 		mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4267 		ti->error = "Unable to activate thin device while pool is suspended";
4268 		r = -EINVAL;
4269 		goto bad;
4270 	}
4271 	refcount_set(&tc->refcount, 1);
4272 	init_completion(&tc->can_destroy);
4273 	list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4274 	spin_unlock_irq(&tc->pool->lock);
4275 	/*
4276 	 * This synchronize_rcu() call is needed here otherwise we risk a
4277 	 * wake_worker() call finding no bios to process (because the newly
4278 	 * added tc isn't yet visible).  So this reduces latency since we
4279 	 * aren't then dependent on the periodic commit to wake_worker().
4280 	 */
4281 	synchronize_rcu();
4282 
4283 	dm_put(pool_md);
4284 
4285 	return 0;
4286 
4287 bad:
4288 	dm_pool_close_thin_device(tc->td);
4289 bad_pool:
4290 	__pool_dec(tc->pool);
4291 bad_pool_lookup:
4292 	dm_put(pool_md);
4293 bad_common:
4294 	dm_put_device(ti, tc->pool_dev);
4295 bad_pool_dev:
4296 	if (tc->origin_dev)
4297 		dm_put_device(ti, tc->origin_dev);
4298 bad_origin_dev:
4299 	kfree(tc);
4300 out_unlock:
4301 	mutex_unlock(&dm_thin_pool_table.mutex);
4302 
4303 	return r;
4304 }
4305 
4306 static int thin_map(struct dm_target *ti, struct bio *bio)
4307 {
4308 	bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4309 
4310 	return thin_bio_map(ti, bio);
4311 }
4312 
4313 static int thin_endio(struct dm_target *ti, struct bio *bio,
4314 		blk_status_t *err)
4315 {
4316 	unsigned long flags;
4317 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4318 	struct list_head work;
4319 	struct dm_thin_new_mapping *m, *tmp;
4320 	struct pool *pool = h->tc->pool;
4321 
4322 	if (h->shared_read_entry) {
4323 		INIT_LIST_HEAD(&work);
4324 		dm_deferred_entry_dec(h->shared_read_entry, &work);
4325 
4326 		spin_lock_irqsave(&pool->lock, flags);
4327 		list_for_each_entry_safe(m, tmp, &work, list) {
4328 			list_del(&m->list);
4329 			__complete_mapping_preparation(m);
4330 		}
4331 		spin_unlock_irqrestore(&pool->lock, flags);
4332 	}
4333 
4334 	if (h->all_io_entry) {
4335 		INIT_LIST_HEAD(&work);
4336 		dm_deferred_entry_dec(h->all_io_entry, &work);
4337 		if (!list_empty(&work)) {
4338 			spin_lock_irqsave(&pool->lock, flags);
4339 			list_for_each_entry_safe(m, tmp, &work, list)
4340 				list_add_tail(&m->list, &pool->prepared_discards);
4341 			spin_unlock_irqrestore(&pool->lock, flags);
4342 			wake_worker(pool);
4343 		}
4344 	}
4345 
4346 	if (h->cell)
4347 		cell_defer_no_holder(h->tc, h->cell);
4348 
4349 	return DM_ENDIO_DONE;
4350 }
4351 
4352 static void thin_presuspend(struct dm_target *ti)
4353 {
4354 	struct thin_c *tc = ti->private;
4355 
4356 	if (dm_noflush_suspending(ti))
4357 		noflush_work(tc, do_noflush_start);
4358 }
4359 
4360 static void thin_postsuspend(struct dm_target *ti)
4361 {
4362 	struct thin_c *tc = ti->private;
4363 
4364 	/*
4365 	 * The dm_noflush_suspending flag has been cleared by now, so
4366 	 * unfortunately we must always run this.
4367 	 */
4368 	noflush_work(tc, do_noflush_stop);
4369 }
4370 
4371 static int thin_preresume(struct dm_target *ti)
4372 {
4373 	struct thin_c *tc = ti->private;
4374 
4375 	if (tc->origin_dev)
4376 		tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4377 
4378 	return 0;
4379 }
4380 
4381 /*
4382  * <nr mapped sectors> <highest mapped sector>
4383  */
4384 static void thin_status(struct dm_target *ti, status_type_t type,
4385 			unsigned status_flags, char *result, unsigned maxlen)
4386 {
4387 	int r;
4388 	ssize_t sz = 0;
4389 	dm_block_t mapped, highest;
4390 	char buf[BDEVNAME_SIZE];
4391 	struct thin_c *tc = ti->private;
4392 
4393 	if (get_pool_mode(tc->pool) == PM_FAIL) {
4394 		DMEMIT("Fail");
4395 		return;
4396 	}
4397 
4398 	if (!tc->td)
4399 		DMEMIT("-");
4400 	else {
4401 		switch (type) {
4402 		case STATUSTYPE_INFO:
4403 			r = dm_thin_get_mapped_count(tc->td, &mapped);
4404 			if (r) {
4405 				DMERR("dm_thin_get_mapped_count returned %d", r);
4406 				goto err;
4407 			}
4408 
4409 			r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4410 			if (r < 0) {
4411 				DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4412 				goto err;
4413 			}
4414 
4415 			DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4416 			if (r)
4417 				DMEMIT("%llu", ((highest + 1) *
4418 						tc->pool->sectors_per_block) - 1);
4419 			else
4420 				DMEMIT("-");
4421 			break;
4422 
4423 		case STATUSTYPE_TABLE:
4424 			DMEMIT("%s %lu",
4425 			       format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4426 			       (unsigned long) tc->dev_id);
4427 			if (tc->origin_dev)
4428 				DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4429 			break;
4430 
4431 		case STATUSTYPE_IMA:
4432 			*result = '\0';
4433 			break;
4434 		}
4435 	}
4436 
4437 	return;
4438 
4439 err:
4440 	DMEMIT("Error");
4441 }
4442 
4443 static int thin_iterate_devices(struct dm_target *ti,
4444 				iterate_devices_callout_fn fn, void *data)
4445 {
4446 	sector_t blocks;
4447 	struct thin_c *tc = ti->private;
4448 	struct pool *pool = tc->pool;
4449 
4450 	/*
4451 	 * We can't call dm_pool_get_data_dev_size() since that blocks.  So
4452 	 * we follow a more convoluted path through to the pool's target.
4453 	 */
4454 	if (!pool->ti)
4455 		return 0;	/* nothing is bound */
4456 
4457 	blocks = pool->ti->len;
4458 	(void) sector_div(blocks, pool->sectors_per_block);
4459 	if (blocks)
4460 		return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4461 
4462 	return 0;
4463 }
4464 
4465 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4466 {
4467 	struct thin_c *tc = ti->private;
4468 	struct pool *pool = tc->pool;
4469 
4470 	if (!pool->pf.discard_enabled)
4471 		return;
4472 
4473 	limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4474 	limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4475 }
4476 
4477 static struct target_type thin_target = {
4478 	.name = "thin",
4479 	.version = {1, 22, 0},
4480 	.module	= THIS_MODULE,
4481 	.ctr = thin_ctr,
4482 	.dtr = thin_dtr,
4483 	.map = thin_map,
4484 	.end_io = thin_endio,
4485 	.preresume = thin_preresume,
4486 	.presuspend = thin_presuspend,
4487 	.postsuspend = thin_postsuspend,
4488 	.status = thin_status,
4489 	.iterate_devices = thin_iterate_devices,
4490 	.io_hints = thin_io_hints,
4491 };
4492 
4493 /*----------------------------------------------------------------*/
4494 
4495 static int __init dm_thin_init(void)
4496 {
4497 	int r = -ENOMEM;
4498 
4499 	pool_table_init();
4500 
4501 	_new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4502 	if (!_new_mapping_cache)
4503 		return r;
4504 
4505 	r = dm_register_target(&thin_target);
4506 	if (r)
4507 		goto bad_new_mapping_cache;
4508 
4509 	r = dm_register_target(&pool_target);
4510 	if (r)
4511 		goto bad_thin_target;
4512 
4513 	return 0;
4514 
4515 bad_thin_target:
4516 	dm_unregister_target(&thin_target);
4517 bad_new_mapping_cache:
4518 	kmem_cache_destroy(_new_mapping_cache);
4519 
4520 	return r;
4521 }
4522 
4523 static void dm_thin_exit(void)
4524 {
4525 	dm_unregister_target(&thin_target);
4526 	dm_unregister_target(&pool_target);
4527 
4528 	kmem_cache_destroy(_new_mapping_cache);
4529 
4530 	pool_table_exit();
4531 }
4532 
4533 module_init(dm_thin_init);
4534 module_exit(dm_thin_exit);
4535 
4536 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4537 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4538 
4539 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4540 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4541 MODULE_LICENSE("GPL");
4542