xref: /openbmc/linux/drivers/md/dm-thin.c (revision d088d6b6)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011-2012 Red Hat UK.
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include "dm-thin-metadata.h"
9 #include "dm-bio-prison-v1.h"
10 #include "dm.h"
11 
12 #include <linux/device-mapper.h>
13 #include <linux/dm-io.h>
14 #include <linux/dm-kcopyd.h>
15 #include <linux/jiffies.h>
16 #include <linux/log2.h>
17 #include <linux/list.h>
18 #include <linux/rculist.h>
19 #include <linux/init.h>
20 #include <linux/module.h>
21 #include <linux/slab.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sort.h>
24 #include <linux/rbtree.h>
25 
26 #define	DM_MSG_PREFIX	"thin"
27 
28 /*
29  * Tunable constants
30  */
31 #define ENDIO_HOOK_POOL_SIZE 1024
32 #define MAPPING_POOL_SIZE 1024
33 #define COMMIT_PERIOD HZ
34 #define NO_SPACE_TIMEOUT_SECS 60
35 
36 static unsigned int no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
37 
38 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
39 		"A percentage of time allocated for copy on write");
40 
41 /*
42  * The block size of the device holding pool data must be
43  * between 64KB and 1GB.
44  */
45 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
46 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
47 
48 /*
49  * Device id is restricted to 24 bits.
50  */
51 #define MAX_DEV_ID ((1 << 24) - 1)
52 
53 /*
54  * How do we handle breaking sharing of data blocks?
55  * =================================================
56  *
57  * We use a standard copy-on-write btree to store the mappings for the
58  * devices (note I'm talking about copy-on-write of the metadata here, not
59  * the data).  When you take an internal snapshot you clone the root node
60  * of the origin btree.  After this there is no concept of an origin or a
61  * snapshot.  They are just two device trees that happen to point to the
62  * same data blocks.
63  *
64  * When we get a write in we decide if it's to a shared data block using
65  * some timestamp magic.  If it is, we have to break sharing.
66  *
67  * Let's say we write to a shared block in what was the origin.  The
68  * steps are:
69  *
70  * i) plug io further to this physical block. (see bio_prison code).
71  *
72  * ii) quiesce any read io to that shared data block.  Obviously
73  * including all devices that share this block.  (see dm_deferred_set code)
74  *
75  * iii) copy the data block to a newly allocate block.  This step can be
76  * missed out if the io covers the block. (schedule_copy).
77  *
78  * iv) insert the new mapping into the origin's btree
79  * (process_prepared_mapping).  This act of inserting breaks some
80  * sharing of btree nodes between the two devices.  Breaking sharing only
81  * effects the btree of that specific device.  Btrees for the other
82  * devices that share the block never change.  The btree for the origin
83  * device as it was after the last commit is untouched, ie. we're using
84  * persistent data structures in the functional programming sense.
85  *
86  * v) unplug io to this physical block, including the io that triggered
87  * the breaking of sharing.
88  *
89  * Steps (ii) and (iii) occur in parallel.
90  *
91  * The metadata _doesn't_ need to be committed before the io continues.  We
92  * get away with this because the io is always written to a _new_ block.
93  * If there's a crash, then:
94  *
95  * - The origin mapping will point to the old origin block (the shared
96  * one).  This will contain the data as it was before the io that triggered
97  * the breaking of sharing came in.
98  *
99  * - The snap mapping still points to the old block.  As it would after
100  * the commit.
101  *
102  * The downside of this scheme is the timestamp magic isn't perfect, and
103  * will continue to think that data block in the snapshot device is shared
104  * even after the write to the origin has broken sharing.  I suspect data
105  * blocks will typically be shared by many different devices, so we're
106  * breaking sharing n + 1 times, rather than n, where n is the number of
107  * devices that reference this data block.  At the moment I think the
108  * benefits far, far outweigh the disadvantages.
109  */
110 
111 /*----------------------------------------------------------------*/
112 
113 /*
114  * Key building.
115  */
116 enum lock_space {
117 	VIRTUAL,
118 	PHYSICAL
119 };
120 
121 static bool build_key(struct dm_thin_device *td, enum lock_space ls,
122 		      dm_block_t b, dm_block_t e, struct dm_cell_key *key)
123 {
124 	key->virtual = (ls == VIRTUAL);
125 	key->dev = dm_thin_dev_id(td);
126 	key->block_begin = b;
127 	key->block_end = e;
128 
129 	return dm_cell_key_has_valid_range(key);
130 }
131 
132 static void build_data_key(struct dm_thin_device *td, dm_block_t b,
133 			   struct dm_cell_key *key)
134 {
135 	(void) build_key(td, PHYSICAL, b, b + 1llu, key);
136 }
137 
138 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
139 			      struct dm_cell_key *key)
140 {
141 	(void) build_key(td, VIRTUAL, b, b + 1llu, key);
142 }
143 
144 /*----------------------------------------------------------------*/
145 
146 #define THROTTLE_THRESHOLD (1 * HZ)
147 
148 struct throttle {
149 	struct rw_semaphore lock;
150 	unsigned long threshold;
151 	bool throttle_applied;
152 };
153 
154 static void throttle_init(struct throttle *t)
155 {
156 	init_rwsem(&t->lock);
157 	t->throttle_applied = false;
158 }
159 
160 static void throttle_work_start(struct throttle *t)
161 {
162 	t->threshold = jiffies + THROTTLE_THRESHOLD;
163 }
164 
165 static void throttle_work_update(struct throttle *t)
166 {
167 	if (!t->throttle_applied && time_is_before_jiffies(t->threshold)) {
168 		down_write(&t->lock);
169 		t->throttle_applied = true;
170 	}
171 }
172 
173 static void throttle_work_complete(struct throttle *t)
174 {
175 	if (t->throttle_applied) {
176 		t->throttle_applied = false;
177 		up_write(&t->lock);
178 	}
179 }
180 
181 static void throttle_lock(struct throttle *t)
182 {
183 	down_read(&t->lock);
184 }
185 
186 static void throttle_unlock(struct throttle *t)
187 {
188 	up_read(&t->lock);
189 }
190 
191 /*----------------------------------------------------------------*/
192 
193 /*
194  * A pool device ties together a metadata device and a data device.  It
195  * also provides the interface for creating and destroying internal
196  * devices.
197  */
198 struct dm_thin_new_mapping;
199 
200 /*
201  * The pool runs in various modes.  Ordered in degraded order for comparisons.
202  */
203 enum pool_mode {
204 	PM_WRITE,		/* metadata may be changed */
205 	PM_OUT_OF_DATA_SPACE,	/* metadata may be changed, though data may not be allocated */
206 
207 	/*
208 	 * Like READ_ONLY, except may switch back to WRITE on metadata resize. Reported as READ_ONLY.
209 	 */
210 	PM_OUT_OF_METADATA_SPACE,
211 	PM_READ_ONLY,		/* metadata may not be changed */
212 
213 	PM_FAIL,		/* all I/O fails */
214 };
215 
216 struct pool_features {
217 	enum pool_mode mode;
218 
219 	bool zero_new_blocks:1;
220 	bool discard_enabled:1;
221 	bool discard_passdown:1;
222 	bool error_if_no_space:1;
223 };
224 
225 struct thin_c;
226 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
227 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
228 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
229 
230 #define CELL_SORT_ARRAY_SIZE 8192
231 
232 struct pool {
233 	struct list_head list;
234 	struct dm_target *ti;	/* Only set if a pool target is bound */
235 
236 	struct mapped_device *pool_md;
237 	struct block_device *data_dev;
238 	struct block_device *md_dev;
239 	struct dm_pool_metadata *pmd;
240 
241 	dm_block_t low_water_blocks;
242 	uint32_t sectors_per_block;
243 	int sectors_per_block_shift;
244 
245 	struct pool_features pf;
246 	bool low_water_triggered:1;	/* A dm event has been sent */
247 	bool suspended:1;
248 	bool out_of_data_space:1;
249 
250 	struct dm_bio_prison *prison;
251 	struct dm_kcopyd_client *copier;
252 
253 	struct work_struct worker;
254 	struct workqueue_struct *wq;
255 	struct throttle throttle;
256 	struct delayed_work waker;
257 	struct delayed_work no_space_timeout;
258 
259 	unsigned long last_commit_jiffies;
260 	unsigned int ref_count;
261 
262 	spinlock_t lock;
263 	struct bio_list deferred_flush_bios;
264 	struct bio_list deferred_flush_completions;
265 	struct list_head prepared_mappings;
266 	struct list_head prepared_discards;
267 	struct list_head prepared_discards_pt2;
268 	struct list_head active_thins;
269 
270 	struct dm_deferred_set *shared_read_ds;
271 	struct dm_deferred_set *all_io_ds;
272 
273 	struct dm_thin_new_mapping *next_mapping;
274 
275 	process_bio_fn process_bio;
276 	process_bio_fn process_discard;
277 
278 	process_cell_fn process_cell;
279 	process_cell_fn process_discard_cell;
280 
281 	process_mapping_fn process_prepared_mapping;
282 	process_mapping_fn process_prepared_discard;
283 	process_mapping_fn process_prepared_discard_pt2;
284 
285 	struct dm_bio_prison_cell **cell_sort_array;
286 
287 	mempool_t mapping_pool;
288 };
289 
290 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
291 
292 static enum pool_mode get_pool_mode(struct pool *pool)
293 {
294 	return pool->pf.mode;
295 }
296 
297 static void notify_of_pool_mode_change(struct pool *pool)
298 {
299 	static const char *descs[] = {
300 		"write",
301 		"out-of-data-space",
302 		"read-only",
303 		"read-only",
304 		"fail"
305 	};
306 	const char *extra_desc = NULL;
307 	enum pool_mode mode = get_pool_mode(pool);
308 
309 	if (mode == PM_OUT_OF_DATA_SPACE) {
310 		if (!pool->pf.error_if_no_space)
311 			extra_desc = " (queue IO)";
312 		else
313 			extra_desc = " (error IO)";
314 	}
315 
316 	dm_table_event(pool->ti->table);
317 	DMINFO("%s: switching pool to %s%s mode",
318 	       dm_device_name(pool->pool_md),
319 	       descs[(int)mode], extra_desc ? : "");
320 }
321 
322 /*
323  * Target context for a pool.
324  */
325 struct pool_c {
326 	struct dm_target *ti;
327 	struct pool *pool;
328 	struct dm_dev *data_dev;
329 	struct dm_dev *metadata_dev;
330 
331 	dm_block_t low_water_blocks;
332 	struct pool_features requested_pf; /* Features requested during table load */
333 	struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
334 };
335 
336 /*
337  * Target context for a thin.
338  */
339 struct thin_c {
340 	struct list_head list;
341 	struct dm_dev *pool_dev;
342 	struct dm_dev *origin_dev;
343 	sector_t origin_size;
344 	dm_thin_id dev_id;
345 
346 	struct pool *pool;
347 	struct dm_thin_device *td;
348 	struct mapped_device *thin_md;
349 
350 	bool requeue_mode:1;
351 	spinlock_t lock;
352 	struct list_head deferred_cells;
353 	struct bio_list deferred_bio_list;
354 	struct bio_list retry_on_resume_list;
355 	struct rb_root sort_bio_list; /* sorted list of deferred bios */
356 
357 	/*
358 	 * Ensures the thin is not destroyed until the worker has finished
359 	 * iterating the active_thins list.
360 	 */
361 	refcount_t refcount;
362 	struct completion can_destroy;
363 };
364 
365 /*----------------------------------------------------------------*/
366 
367 static bool block_size_is_power_of_two(struct pool *pool)
368 {
369 	return pool->sectors_per_block_shift >= 0;
370 }
371 
372 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
373 {
374 	return block_size_is_power_of_two(pool) ?
375 		(b << pool->sectors_per_block_shift) :
376 		(b * pool->sectors_per_block);
377 }
378 
379 /*----------------------------------------------------------------*/
380 
381 struct discard_op {
382 	struct thin_c *tc;
383 	struct blk_plug plug;
384 	struct bio *parent_bio;
385 	struct bio *bio;
386 };
387 
388 static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
389 {
390 	BUG_ON(!parent);
391 
392 	op->tc = tc;
393 	blk_start_plug(&op->plug);
394 	op->parent_bio = parent;
395 	op->bio = NULL;
396 }
397 
398 static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
399 {
400 	struct thin_c *tc = op->tc;
401 	sector_t s = block_to_sectors(tc->pool, data_b);
402 	sector_t len = block_to_sectors(tc->pool, data_e - data_b);
403 
404 	return __blkdev_issue_discard(tc->pool_dev->bdev, s, len, GFP_NOIO, &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 		op->bio->bi_opf = REQ_OP_DISCARD;
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 		dm_submit_bio_remap(bio, NULL);
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 	struct bio_list bios;
888 
889 	bio_list_init(&bios);
890 	cell_release_no_holder(pool, cell, &bios);
891 
892 	if (!bio_list_empty(&bios)) {
893 		spin_lock_irqsave(&tc->lock, flags);
894 		bio_list_merge(&tc->deferred_bio_list, &bios);
895 		spin_unlock_irqrestore(&tc->lock, flags);
896 		wake_worker(pool);
897 	}
898 }
899 
900 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
901 
902 struct remap_info {
903 	struct thin_c *tc;
904 	struct bio_list defer_bios;
905 	struct bio_list issue_bios;
906 };
907 
908 static void __inc_remap_and_issue_cell(void *context,
909 				       struct dm_bio_prison_cell *cell)
910 {
911 	struct remap_info *info = context;
912 	struct bio *bio;
913 
914 	while ((bio = bio_list_pop(&cell->bios))) {
915 		if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
916 			bio_list_add(&info->defer_bios, bio);
917 		else {
918 			inc_all_io_entry(info->tc->pool, bio);
919 
920 			/*
921 			 * We can't issue the bios with the bio prison lock
922 			 * held, so we add them to a list to issue on
923 			 * return from this function.
924 			 */
925 			bio_list_add(&info->issue_bios, bio);
926 		}
927 	}
928 }
929 
930 static void inc_remap_and_issue_cell(struct thin_c *tc,
931 				     struct dm_bio_prison_cell *cell,
932 				     dm_block_t block)
933 {
934 	struct bio *bio;
935 	struct remap_info info;
936 
937 	info.tc = tc;
938 	bio_list_init(&info.defer_bios);
939 	bio_list_init(&info.issue_bios);
940 
941 	/*
942 	 * We have to be careful to inc any bios we're about to issue
943 	 * before the cell is released, and avoid a race with new bios
944 	 * being added to the cell.
945 	 */
946 	cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
947 			   &info, cell);
948 
949 	while ((bio = bio_list_pop(&info.defer_bios)))
950 		thin_defer_bio(tc, bio);
951 
952 	while ((bio = bio_list_pop(&info.issue_bios)))
953 		remap_and_issue(info.tc, bio, block);
954 }
955 
956 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
957 {
958 	cell_error(m->tc->pool, m->cell);
959 	list_del(&m->list);
960 	mempool_free(m, &m->tc->pool->mapping_pool);
961 }
962 
963 static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
964 {
965 	struct pool *pool = tc->pool;
966 
967 	/*
968 	 * If the bio has the REQ_FUA flag set we must commit the metadata
969 	 * before signaling its completion.
970 	 */
971 	if (!bio_triggers_commit(tc, bio)) {
972 		bio_endio(bio);
973 		return;
974 	}
975 
976 	/*
977 	 * Complete bio with an error if earlier I/O caused changes to the
978 	 * metadata that can't be committed, e.g, due to I/O errors on the
979 	 * metadata device.
980 	 */
981 	if (dm_thin_aborted_changes(tc->td)) {
982 		bio_io_error(bio);
983 		return;
984 	}
985 
986 	/*
987 	 * Batch together any bios that trigger commits and then issue a
988 	 * single commit for them in process_deferred_bios().
989 	 */
990 	spin_lock_irq(&pool->lock);
991 	bio_list_add(&pool->deferred_flush_completions, bio);
992 	spin_unlock_irq(&pool->lock);
993 }
994 
995 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
996 {
997 	struct thin_c *tc = m->tc;
998 	struct pool *pool = tc->pool;
999 	struct bio *bio = m->bio;
1000 	int r;
1001 
1002 	if (m->status) {
1003 		cell_error(pool, m->cell);
1004 		goto out;
1005 	}
1006 
1007 	/*
1008 	 * Commit the prepared block into the mapping btree.
1009 	 * Any I/O for this block arriving after this point will get
1010 	 * remapped to it directly.
1011 	 */
1012 	r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1013 	if (r) {
1014 		metadata_operation_failed(pool, "dm_thin_insert_block", r);
1015 		cell_error(pool, m->cell);
1016 		goto out;
1017 	}
1018 
1019 	/*
1020 	 * Release any bios held while the block was being provisioned.
1021 	 * If we are processing a write bio that completely covers the block,
1022 	 * we already processed it so can ignore it now when processing
1023 	 * the bios in the cell.
1024 	 */
1025 	if (bio) {
1026 		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1027 		complete_overwrite_bio(tc, bio);
1028 	} else {
1029 		inc_all_io_entry(tc->pool, m->cell->holder);
1030 		remap_and_issue(tc, m->cell->holder, m->data_block);
1031 		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1032 	}
1033 
1034 out:
1035 	list_del(&m->list);
1036 	mempool_free(m, &pool->mapping_pool);
1037 }
1038 
1039 /*----------------------------------------------------------------*/
1040 
1041 static void free_discard_mapping(struct dm_thin_new_mapping *m)
1042 {
1043 	struct thin_c *tc = m->tc;
1044 
1045 	if (m->cell)
1046 		cell_defer_no_holder(tc, m->cell);
1047 	mempool_free(m, &tc->pool->mapping_pool);
1048 }
1049 
1050 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1051 {
1052 	bio_io_error(m->bio);
1053 	free_discard_mapping(m);
1054 }
1055 
1056 static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1057 {
1058 	bio_endio(m->bio);
1059 	free_discard_mapping(m);
1060 }
1061 
1062 static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1063 {
1064 	int r;
1065 	struct thin_c *tc = m->tc;
1066 
1067 	r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1068 	if (r) {
1069 		metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1070 		bio_io_error(m->bio);
1071 	} else
1072 		bio_endio(m->bio);
1073 
1074 	cell_defer_no_holder(tc, m->cell);
1075 	mempool_free(m, &tc->pool->mapping_pool);
1076 }
1077 
1078 /*----------------------------------------------------------------*/
1079 
1080 static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1081 						   struct bio *discard_parent)
1082 {
1083 	/*
1084 	 * We've already unmapped this range of blocks, but before we
1085 	 * passdown we have to check that these blocks are now unused.
1086 	 */
1087 	int r = 0;
1088 	bool shared = true;
1089 	struct thin_c *tc = m->tc;
1090 	struct pool *pool = tc->pool;
1091 	dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1092 	struct discard_op op;
1093 
1094 	begin_discard(&op, tc, discard_parent);
1095 	while (b != end) {
1096 		/* find start of unmapped run */
1097 		for (; b < end; b++) {
1098 			r = dm_pool_block_is_shared(pool->pmd, b, &shared);
1099 			if (r)
1100 				goto out;
1101 
1102 			if (!shared)
1103 				break;
1104 		}
1105 
1106 		if (b == end)
1107 			break;
1108 
1109 		/* find end of run */
1110 		for (e = b + 1; e != end; e++) {
1111 			r = dm_pool_block_is_shared(pool->pmd, e, &shared);
1112 			if (r)
1113 				goto out;
1114 
1115 			if (shared)
1116 				break;
1117 		}
1118 
1119 		r = issue_discard(&op, b, e);
1120 		if (r)
1121 			goto out;
1122 
1123 		b = e;
1124 	}
1125 out:
1126 	end_discard(&op, r);
1127 }
1128 
1129 static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1130 {
1131 	unsigned long flags;
1132 	struct pool *pool = m->tc->pool;
1133 
1134 	spin_lock_irqsave(&pool->lock, flags);
1135 	list_add_tail(&m->list, &pool->prepared_discards_pt2);
1136 	spin_unlock_irqrestore(&pool->lock, flags);
1137 	wake_worker(pool);
1138 }
1139 
1140 static void passdown_endio(struct bio *bio)
1141 {
1142 	/*
1143 	 * It doesn't matter if the passdown discard failed, we still want
1144 	 * to unmap (we ignore err).
1145 	 */
1146 	queue_passdown_pt2(bio->bi_private);
1147 	bio_put(bio);
1148 }
1149 
1150 static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1151 {
1152 	int r;
1153 	struct thin_c *tc = m->tc;
1154 	struct pool *pool = tc->pool;
1155 	struct bio *discard_parent;
1156 	dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1157 
1158 	/*
1159 	 * Only this thread allocates blocks, so we can be sure that the
1160 	 * newly unmapped blocks will not be allocated before the end of
1161 	 * the function.
1162 	 */
1163 	r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1164 	if (r) {
1165 		metadata_operation_failed(pool, "dm_thin_remove_range", r);
1166 		bio_io_error(m->bio);
1167 		cell_defer_no_holder(tc, m->cell);
1168 		mempool_free(m, &pool->mapping_pool);
1169 		return;
1170 	}
1171 
1172 	/*
1173 	 * Increment the unmapped blocks.  This prevents a race between the
1174 	 * passdown io and reallocation of freed blocks.
1175 	 */
1176 	r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1177 	if (r) {
1178 		metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1179 		bio_io_error(m->bio);
1180 		cell_defer_no_holder(tc, m->cell);
1181 		mempool_free(m, &pool->mapping_pool);
1182 		return;
1183 	}
1184 
1185 	discard_parent = bio_alloc(NULL, 1, 0, GFP_NOIO);
1186 	discard_parent->bi_end_io = passdown_endio;
1187 	discard_parent->bi_private = m;
1188 	if (m->maybe_shared)
1189 		passdown_double_checking_shared_status(m, discard_parent);
1190 	else {
1191 		struct discard_op op;
1192 
1193 		begin_discard(&op, tc, discard_parent);
1194 		r = issue_discard(&op, m->data_block, data_end);
1195 		end_discard(&op, r);
1196 	}
1197 }
1198 
1199 static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1200 {
1201 	int r;
1202 	struct thin_c *tc = m->tc;
1203 	struct pool *pool = tc->pool;
1204 
1205 	/*
1206 	 * The passdown has completed, so now we can decrement all those
1207 	 * unmapped blocks.
1208 	 */
1209 	r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1210 				   m->data_block + (m->virt_end - m->virt_begin));
1211 	if (r) {
1212 		metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1213 		bio_io_error(m->bio);
1214 	} else
1215 		bio_endio(m->bio);
1216 
1217 	cell_defer_no_holder(tc, m->cell);
1218 	mempool_free(m, &pool->mapping_pool);
1219 }
1220 
1221 static void process_prepared(struct pool *pool, struct list_head *head,
1222 			     process_mapping_fn *fn)
1223 {
1224 	struct list_head maps;
1225 	struct dm_thin_new_mapping *m, *tmp;
1226 
1227 	INIT_LIST_HEAD(&maps);
1228 	spin_lock_irq(&pool->lock);
1229 	list_splice_init(head, &maps);
1230 	spin_unlock_irq(&pool->lock);
1231 
1232 	list_for_each_entry_safe(m, tmp, &maps, list)
1233 		(*fn)(m);
1234 }
1235 
1236 /*
1237  * Deferred bio jobs.
1238  */
1239 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1240 {
1241 	return bio->bi_iter.bi_size ==
1242 		(pool->sectors_per_block << SECTOR_SHIFT);
1243 }
1244 
1245 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1246 {
1247 	return (bio_data_dir(bio) == WRITE) &&
1248 		io_overlaps_block(pool, bio);
1249 }
1250 
1251 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1252 			       bio_end_io_t *fn)
1253 {
1254 	*save = bio->bi_end_io;
1255 	bio->bi_end_io = fn;
1256 }
1257 
1258 static int ensure_next_mapping(struct pool *pool)
1259 {
1260 	if (pool->next_mapping)
1261 		return 0;
1262 
1263 	pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1264 
1265 	return pool->next_mapping ? 0 : -ENOMEM;
1266 }
1267 
1268 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1269 {
1270 	struct dm_thin_new_mapping *m = pool->next_mapping;
1271 
1272 	BUG_ON(!pool->next_mapping);
1273 
1274 	memset(m, 0, sizeof(struct dm_thin_new_mapping));
1275 	INIT_LIST_HEAD(&m->list);
1276 	m->bio = NULL;
1277 
1278 	pool->next_mapping = NULL;
1279 
1280 	return m;
1281 }
1282 
1283 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1284 		    sector_t begin, sector_t end)
1285 {
1286 	struct dm_io_region to;
1287 
1288 	to.bdev = tc->pool_dev->bdev;
1289 	to.sector = begin;
1290 	to.count = end - begin;
1291 
1292 	dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1293 }
1294 
1295 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1296 				      dm_block_t data_begin,
1297 				      struct dm_thin_new_mapping *m)
1298 {
1299 	struct pool *pool = tc->pool;
1300 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1301 
1302 	h->overwrite_mapping = m;
1303 	m->bio = bio;
1304 	save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1305 	inc_all_io_entry(pool, bio);
1306 	remap_and_issue(tc, bio, data_begin);
1307 }
1308 
1309 /*
1310  * A partial copy also needs to zero the uncopied region.
1311  */
1312 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1313 			  struct dm_dev *origin, dm_block_t data_origin,
1314 			  dm_block_t data_dest,
1315 			  struct dm_bio_prison_cell *cell, struct bio *bio,
1316 			  sector_t len)
1317 {
1318 	struct pool *pool = tc->pool;
1319 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1320 
1321 	m->tc = tc;
1322 	m->virt_begin = virt_block;
1323 	m->virt_end = virt_block + 1u;
1324 	m->data_block = data_dest;
1325 	m->cell = cell;
1326 
1327 	/*
1328 	 * quiesce action + copy action + an extra reference held for the
1329 	 * duration of this function (we may need to inc later for a
1330 	 * partial zero).
1331 	 */
1332 	atomic_set(&m->prepare_actions, 3);
1333 
1334 	if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1335 		complete_mapping_preparation(m); /* already quiesced */
1336 
1337 	/*
1338 	 * IO to pool_dev remaps to the pool target's data_dev.
1339 	 *
1340 	 * If the whole block of data is being overwritten, we can issue the
1341 	 * bio immediately. Otherwise we use kcopyd to clone the data first.
1342 	 */
1343 	if (io_overwrites_block(pool, bio))
1344 		remap_and_issue_overwrite(tc, bio, data_dest, m);
1345 	else {
1346 		struct dm_io_region from, to;
1347 
1348 		from.bdev = origin->bdev;
1349 		from.sector = data_origin * pool->sectors_per_block;
1350 		from.count = len;
1351 
1352 		to.bdev = tc->pool_dev->bdev;
1353 		to.sector = data_dest * pool->sectors_per_block;
1354 		to.count = len;
1355 
1356 		dm_kcopyd_copy(pool->copier, &from, 1, &to,
1357 			       0, copy_complete, m);
1358 
1359 		/*
1360 		 * Do we need to zero a tail region?
1361 		 */
1362 		if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1363 			atomic_inc(&m->prepare_actions);
1364 			ll_zero(tc, m,
1365 				data_dest * pool->sectors_per_block + len,
1366 				(data_dest + 1) * pool->sectors_per_block);
1367 		}
1368 	}
1369 
1370 	complete_mapping_preparation(m); /* drop our ref */
1371 }
1372 
1373 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1374 				   dm_block_t data_origin, dm_block_t data_dest,
1375 				   struct dm_bio_prison_cell *cell, struct bio *bio)
1376 {
1377 	schedule_copy(tc, virt_block, tc->pool_dev,
1378 		      data_origin, data_dest, cell, bio,
1379 		      tc->pool->sectors_per_block);
1380 }
1381 
1382 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1383 			  dm_block_t data_block, struct dm_bio_prison_cell *cell,
1384 			  struct bio *bio)
1385 {
1386 	struct pool *pool = tc->pool;
1387 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1388 
1389 	atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1390 	m->tc = tc;
1391 	m->virt_begin = virt_block;
1392 	m->virt_end = virt_block + 1u;
1393 	m->data_block = data_block;
1394 	m->cell = cell;
1395 
1396 	/*
1397 	 * If the whole block of data is being overwritten or we are not
1398 	 * zeroing pre-existing data, we can issue the bio immediately.
1399 	 * Otherwise we use kcopyd to zero the data first.
1400 	 */
1401 	if (pool->pf.zero_new_blocks) {
1402 		if (io_overwrites_block(pool, bio))
1403 			remap_and_issue_overwrite(tc, bio, data_block, m);
1404 		else
1405 			ll_zero(tc, m, data_block * pool->sectors_per_block,
1406 				(data_block + 1) * pool->sectors_per_block);
1407 	} else
1408 		process_prepared_mapping(m);
1409 }
1410 
1411 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1412 				   dm_block_t data_dest,
1413 				   struct dm_bio_prison_cell *cell, struct bio *bio)
1414 {
1415 	struct pool *pool = tc->pool;
1416 	sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1417 	sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1418 
1419 	if (virt_block_end <= tc->origin_size)
1420 		schedule_copy(tc, virt_block, tc->origin_dev,
1421 			      virt_block, data_dest, cell, bio,
1422 			      pool->sectors_per_block);
1423 
1424 	else if (virt_block_begin < tc->origin_size)
1425 		schedule_copy(tc, virt_block, tc->origin_dev,
1426 			      virt_block, data_dest, cell, bio,
1427 			      tc->origin_size - virt_block_begin);
1428 
1429 	else
1430 		schedule_zero(tc, virt_block, data_dest, cell, bio);
1431 }
1432 
1433 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1434 
1435 static void requeue_bios(struct pool *pool);
1436 
1437 static bool is_read_only_pool_mode(enum pool_mode mode)
1438 {
1439 	return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1440 }
1441 
1442 static bool is_read_only(struct pool *pool)
1443 {
1444 	return is_read_only_pool_mode(get_pool_mode(pool));
1445 }
1446 
1447 static void check_for_metadata_space(struct pool *pool)
1448 {
1449 	int r;
1450 	const char *ooms_reason = NULL;
1451 	dm_block_t nr_free;
1452 
1453 	r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1454 	if (r)
1455 		ooms_reason = "Could not get free metadata blocks";
1456 	else if (!nr_free)
1457 		ooms_reason = "No free metadata blocks";
1458 
1459 	if (ooms_reason && !is_read_only(pool)) {
1460 		DMERR("%s", ooms_reason);
1461 		set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1462 	}
1463 }
1464 
1465 static void check_for_data_space(struct pool *pool)
1466 {
1467 	int r;
1468 	dm_block_t nr_free;
1469 
1470 	if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1471 		return;
1472 
1473 	r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1474 	if (r)
1475 		return;
1476 
1477 	if (nr_free) {
1478 		set_pool_mode(pool, PM_WRITE);
1479 		requeue_bios(pool);
1480 	}
1481 }
1482 
1483 /*
1484  * A non-zero return indicates read_only or fail_io mode.
1485  * Many callers don't care about the return value.
1486  */
1487 static int commit(struct pool *pool)
1488 {
1489 	int r;
1490 
1491 	if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1492 		return -EINVAL;
1493 
1494 	r = dm_pool_commit_metadata(pool->pmd);
1495 	if (r)
1496 		metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1497 	else {
1498 		check_for_metadata_space(pool);
1499 		check_for_data_space(pool);
1500 	}
1501 
1502 	return r;
1503 }
1504 
1505 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1506 {
1507 	if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1508 		DMWARN("%s: reached low water mark for data device: sending event.",
1509 		       dm_device_name(pool->pool_md));
1510 		spin_lock_irq(&pool->lock);
1511 		pool->low_water_triggered = true;
1512 		spin_unlock_irq(&pool->lock);
1513 		dm_table_event(pool->ti->table);
1514 	}
1515 }
1516 
1517 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1518 {
1519 	int r;
1520 	dm_block_t free_blocks;
1521 	struct pool *pool = tc->pool;
1522 
1523 	if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1524 		return -EINVAL;
1525 
1526 	r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1527 	if (r) {
1528 		metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1529 		return r;
1530 	}
1531 
1532 	check_low_water_mark(pool, free_blocks);
1533 
1534 	if (!free_blocks) {
1535 		/*
1536 		 * Try to commit to see if that will free up some
1537 		 * more space.
1538 		 */
1539 		r = commit(pool);
1540 		if (r)
1541 			return r;
1542 
1543 		r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1544 		if (r) {
1545 			metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1546 			return r;
1547 		}
1548 
1549 		if (!free_blocks) {
1550 			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1551 			return -ENOSPC;
1552 		}
1553 	}
1554 
1555 	r = dm_pool_alloc_data_block(pool->pmd, result);
1556 	if (r) {
1557 		if (r == -ENOSPC)
1558 			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1559 		else
1560 			metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1561 		return r;
1562 	}
1563 
1564 	r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1565 	if (r) {
1566 		metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1567 		return r;
1568 	}
1569 
1570 	if (!free_blocks) {
1571 		/* Let's commit before we use up the metadata reserve. */
1572 		r = commit(pool);
1573 		if (r)
1574 			return r;
1575 	}
1576 
1577 	return 0;
1578 }
1579 
1580 /*
1581  * If we have run out of space, queue bios until the device is
1582  * resumed, presumably after having been reloaded with more space.
1583  */
1584 static void retry_on_resume(struct bio *bio)
1585 {
1586 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1587 	struct thin_c *tc = h->tc;
1588 
1589 	spin_lock_irq(&tc->lock);
1590 	bio_list_add(&tc->retry_on_resume_list, bio);
1591 	spin_unlock_irq(&tc->lock);
1592 }
1593 
1594 static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1595 {
1596 	enum pool_mode m = get_pool_mode(pool);
1597 
1598 	switch (m) {
1599 	case PM_WRITE:
1600 		/* Shouldn't get here */
1601 		DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1602 		return BLK_STS_IOERR;
1603 
1604 	case PM_OUT_OF_DATA_SPACE:
1605 		return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1606 
1607 	case PM_OUT_OF_METADATA_SPACE:
1608 	case PM_READ_ONLY:
1609 	case PM_FAIL:
1610 		return BLK_STS_IOERR;
1611 	default:
1612 		/* Shouldn't get here */
1613 		DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1614 		return BLK_STS_IOERR;
1615 	}
1616 }
1617 
1618 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1619 {
1620 	blk_status_t error = should_error_unserviceable_bio(pool);
1621 
1622 	if (error) {
1623 		bio->bi_status = error;
1624 		bio_endio(bio);
1625 	} else
1626 		retry_on_resume(bio);
1627 }
1628 
1629 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1630 {
1631 	struct bio *bio;
1632 	struct bio_list bios;
1633 	blk_status_t error;
1634 
1635 	error = should_error_unserviceable_bio(pool);
1636 	if (error) {
1637 		cell_error_with_code(pool, cell, error);
1638 		return;
1639 	}
1640 
1641 	bio_list_init(&bios);
1642 	cell_release(pool, cell, &bios);
1643 
1644 	while ((bio = bio_list_pop(&bios)))
1645 		retry_on_resume(bio);
1646 }
1647 
1648 static void process_discard_cell_no_passdown(struct thin_c *tc,
1649 					     struct dm_bio_prison_cell *virt_cell)
1650 {
1651 	struct pool *pool = tc->pool;
1652 	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1653 
1654 	/*
1655 	 * We don't need to lock the data blocks, since there's no
1656 	 * passdown.  We only lock data blocks for allocation and breaking sharing.
1657 	 */
1658 	m->tc = tc;
1659 	m->virt_begin = virt_cell->key.block_begin;
1660 	m->virt_end = virt_cell->key.block_end;
1661 	m->cell = virt_cell;
1662 	m->bio = virt_cell->holder;
1663 
1664 	if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1665 		pool->process_prepared_discard(m);
1666 }
1667 
1668 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1669 				 struct bio *bio)
1670 {
1671 	struct pool *pool = tc->pool;
1672 
1673 	int r;
1674 	bool maybe_shared;
1675 	struct dm_cell_key data_key;
1676 	struct dm_bio_prison_cell *data_cell;
1677 	struct dm_thin_new_mapping *m;
1678 	dm_block_t virt_begin, virt_end, data_begin, data_end;
1679 	dm_block_t len, next_boundary;
1680 
1681 	while (begin != end) {
1682 		r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1683 					      &data_begin, &maybe_shared);
1684 		if (r) {
1685 			/*
1686 			 * Silently fail, letting any mappings we've
1687 			 * created complete.
1688 			 */
1689 			break;
1690 		}
1691 
1692 		data_end = data_begin + (virt_end - virt_begin);
1693 
1694 		/*
1695 		 * Make sure the data region obeys the bio prison restrictions.
1696 		 */
1697 		while (data_begin < data_end) {
1698 			r = ensure_next_mapping(pool);
1699 			if (r)
1700 				return; /* we did our best */
1701 
1702 			next_boundary = ((data_begin >> BIO_PRISON_MAX_RANGE_SHIFT) + 1)
1703 				<< BIO_PRISON_MAX_RANGE_SHIFT;
1704 			len = min_t(sector_t, data_end - data_begin, next_boundary - data_begin);
1705 
1706 			/* This key is certainly within range given the above splitting */
1707 			(void) build_key(tc->td, PHYSICAL, data_begin, data_begin + len, &data_key);
1708 			if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1709 				/* contention, we'll give up with this range */
1710 				data_begin += len;
1711 				continue;
1712 			}
1713 
1714 			/*
1715 			 * IO may still be going to the destination block.  We must
1716 			 * quiesce before we can do the removal.
1717 			 */
1718 			m = get_next_mapping(pool);
1719 			m->tc = tc;
1720 			m->maybe_shared = maybe_shared;
1721 			m->virt_begin = virt_begin;
1722 			m->virt_end = virt_begin + len;
1723 			m->data_block = data_begin;
1724 			m->cell = data_cell;
1725 			m->bio = bio;
1726 
1727 			/*
1728 			 * The parent bio must not complete before sub discard bios are
1729 			 * chained to it (see end_discard's bio_chain)!
1730 			 *
1731 			 * This per-mapping bi_remaining increment is paired with
1732 			 * the implicit decrement that occurs via bio_endio() in
1733 			 * end_discard().
1734 			 */
1735 			bio_inc_remaining(bio);
1736 			if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1737 				pool->process_prepared_discard(m);
1738 
1739 			virt_begin += len;
1740 			data_begin += len;
1741 		}
1742 
1743 		begin = virt_end;
1744 	}
1745 }
1746 
1747 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1748 {
1749 	struct bio *bio = virt_cell->holder;
1750 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1751 
1752 	/*
1753 	 * The virt_cell will only get freed once the origin bio completes.
1754 	 * This means it will remain locked while all the individual
1755 	 * passdown bios are in flight.
1756 	 */
1757 	h->cell = virt_cell;
1758 	break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1759 
1760 	/*
1761 	 * We complete the bio now, knowing that the bi_remaining field
1762 	 * will prevent completion until the sub range discards have
1763 	 * completed.
1764 	 */
1765 	bio_endio(bio);
1766 }
1767 
1768 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1769 {
1770 	dm_block_t begin, end;
1771 	struct dm_cell_key virt_key;
1772 	struct dm_bio_prison_cell *virt_cell;
1773 
1774 	get_bio_block_range(tc, bio, &begin, &end);
1775 	if (begin == end) {
1776 		/*
1777 		 * The discard covers less than a block.
1778 		 */
1779 		bio_endio(bio);
1780 		return;
1781 	}
1782 
1783 	if (unlikely(!build_key(tc->td, VIRTUAL, begin, end, &virt_key))) {
1784 		DMERR_LIMIT("Discard doesn't respect bio prison limits");
1785 		bio_endio(bio);
1786 		return;
1787 	}
1788 
1789 	if (bio_detain(tc->pool, &virt_key, bio, &virt_cell)) {
1790 		/*
1791 		 * Potential starvation issue: We're relying on the
1792 		 * fs/application being well behaved, and not trying to
1793 		 * send IO to a region at the same time as discarding it.
1794 		 * If they do this persistently then it's possible this
1795 		 * cell will never be granted.
1796 		 */
1797 		return;
1798 	}
1799 
1800 	tc->pool->process_discard_cell(tc, virt_cell);
1801 }
1802 
1803 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1804 			  struct dm_cell_key *key,
1805 			  struct dm_thin_lookup_result *lookup_result,
1806 			  struct dm_bio_prison_cell *cell)
1807 {
1808 	int r;
1809 	dm_block_t data_block;
1810 	struct pool *pool = tc->pool;
1811 
1812 	r = alloc_data_block(tc, &data_block);
1813 	switch (r) {
1814 	case 0:
1815 		schedule_internal_copy(tc, block, lookup_result->block,
1816 				       data_block, cell, bio);
1817 		break;
1818 
1819 	case -ENOSPC:
1820 		retry_bios_on_resume(pool, cell);
1821 		break;
1822 
1823 	default:
1824 		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1825 			    __func__, r);
1826 		cell_error(pool, cell);
1827 		break;
1828 	}
1829 }
1830 
1831 static void __remap_and_issue_shared_cell(void *context,
1832 					  struct dm_bio_prison_cell *cell)
1833 {
1834 	struct remap_info *info = context;
1835 	struct bio *bio;
1836 
1837 	while ((bio = bio_list_pop(&cell->bios))) {
1838 		if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1839 		    bio_op(bio) == REQ_OP_DISCARD)
1840 			bio_list_add(&info->defer_bios, bio);
1841 		else {
1842 			struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1843 
1844 			h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1845 			inc_all_io_entry(info->tc->pool, bio);
1846 			bio_list_add(&info->issue_bios, bio);
1847 		}
1848 	}
1849 }
1850 
1851 static void remap_and_issue_shared_cell(struct thin_c *tc,
1852 					struct dm_bio_prison_cell *cell,
1853 					dm_block_t block)
1854 {
1855 	struct bio *bio;
1856 	struct remap_info info;
1857 
1858 	info.tc = tc;
1859 	bio_list_init(&info.defer_bios);
1860 	bio_list_init(&info.issue_bios);
1861 
1862 	cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1863 			   &info, cell);
1864 
1865 	while ((bio = bio_list_pop(&info.defer_bios)))
1866 		thin_defer_bio(tc, bio);
1867 
1868 	while ((bio = bio_list_pop(&info.issue_bios)))
1869 		remap_and_issue(tc, bio, block);
1870 }
1871 
1872 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1873 			       dm_block_t block,
1874 			       struct dm_thin_lookup_result *lookup_result,
1875 			       struct dm_bio_prison_cell *virt_cell)
1876 {
1877 	struct dm_bio_prison_cell *data_cell;
1878 	struct pool *pool = tc->pool;
1879 	struct dm_cell_key key;
1880 
1881 	/*
1882 	 * If cell is already occupied, then sharing is already in the process
1883 	 * of being broken so we have nothing further to do here.
1884 	 */
1885 	build_data_key(tc->td, lookup_result->block, &key);
1886 	if (bio_detain(pool, &key, bio, &data_cell)) {
1887 		cell_defer_no_holder(tc, virt_cell);
1888 		return;
1889 	}
1890 
1891 	if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1892 		break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1893 		cell_defer_no_holder(tc, virt_cell);
1894 	} else {
1895 		struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1896 
1897 		h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1898 		inc_all_io_entry(pool, bio);
1899 		remap_and_issue(tc, bio, lookup_result->block);
1900 
1901 		remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1902 		remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1903 	}
1904 }
1905 
1906 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1907 			    struct dm_bio_prison_cell *cell)
1908 {
1909 	int r;
1910 	dm_block_t data_block;
1911 	struct pool *pool = tc->pool;
1912 
1913 	/*
1914 	 * Remap empty bios (flushes) immediately, without provisioning.
1915 	 */
1916 	if (!bio->bi_iter.bi_size) {
1917 		inc_all_io_entry(pool, bio);
1918 		cell_defer_no_holder(tc, cell);
1919 
1920 		remap_and_issue(tc, bio, 0);
1921 		return;
1922 	}
1923 
1924 	/*
1925 	 * Fill read bios with zeroes and complete them immediately.
1926 	 */
1927 	if (bio_data_dir(bio) == READ) {
1928 		zero_fill_bio(bio);
1929 		cell_defer_no_holder(tc, cell);
1930 		bio_endio(bio);
1931 		return;
1932 	}
1933 
1934 	r = alloc_data_block(tc, &data_block);
1935 	switch (r) {
1936 	case 0:
1937 		if (tc->origin_dev)
1938 			schedule_external_copy(tc, block, data_block, cell, bio);
1939 		else
1940 			schedule_zero(tc, block, data_block, cell, bio);
1941 		break;
1942 
1943 	case -ENOSPC:
1944 		retry_bios_on_resume(pool, cell);
1945 		break;
1946 
1947 	default:
1948 		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1949 			    __func__, r);
1950 		cell_error(pool, cell);
1951 		break;
1952 	}
1953 }
1954 
1955 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1956 {
1957 	int r;
1958 	struct pool *pool = tc->pool;
1959 	struct bio *bio = cell->holder;
1960 	dm_block_t block = get_bio_block(tc, bio);
1961 	struct dm_thin_lookup_result lookup_result;
1962 
1963 	if (tc->requeue_mode) {
1964 		cell_requeue(pool, cell);
1965 		return;
1966 	}
1967 
1968 	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1969 	switch (r) {
1970 	case 0:
1971 		if (lookup_result.shared)
1972 			process_shared_bio(tc, bio, block, &lookup_result, cell);
1973 		else {
1974 			inc_all_io_entry(pool, bio);
1975 			remap_and_issue(tc, bio, lookup_result.block);
1976 			inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1977 		}
1978 		break;
1979 
1980 	case -ENODATA:
1981 		if (bio_data_dir(bio) == READ && tc->origin_dev) {
1982 			inc_all_io_entry(pool, bio);
1983 			cell_defer_no_holder(tc, cell);
1984 
1985 			if (bio_end_sector(bio) <= tc->origin_size)
1986 				remap_to_origin_and_issue(tc, bio);
1987 
1988 			else if (bio->bi_iter.bi_sector < tc->origin_size) {
1989 				zero_fill_bio(bio);
1990 				bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1991 				remap_to_origin_and_issue(tc, bio);
1992 
1993 			} else {
1994 				zero_fill_bio(bio);
1995 				bio_endio(bio);
1996 			}
1997 		} else
1998 			provision_block(tc, bio, block, cell);
1999 		break;
2000 
2001 	default:
2002 		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2003 			    __func__, r);
2004 		cell_defer_no_holder(tc, cell);
2005 		bio_io_error(bio);
2006 		break;
2007 	}
2008 }
2009 
2010 static void process_bio(struct thin_c *tc, struct bio *bio)
2011 {
2012 	struct pool *pool = tc->pool;
2013 	dm_block_t block = get_bio_block(tc, bio);
2014 	struct dm_bio_prison_cell *cell;
2015 	struct dm_cell_key key;
2016 
2017 	/*
2018 	 * If cell is already occupied, then the block is already
2019 	 * being provisioned so we have nothing further to do here.
2020 	 */
2021 	build_virtual_key(tc->td, block, &key);
2022 	if (bio_detain(pool, &key, bio, &cell))
2023 		return;
2024 
2025 	process_cell(tc, cell);
2026 }
2027 
2028 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2029 				    struct dm_bio_prison_cell *cell)
2030 {
2031 	int r;
2032 	int rw = bio_data_dir(bio);
2033 	dm_block_t block = get_bio_block(tc, bio);
2034 	struct dm_thin_lookup_result lookup_result;
2035 
2036 	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2037 	switch (r) {
2038 	case 0:
2039 		if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2040 			handle_unserviceable_bio(tc->pool, bio);
2041 			if (cell)
2042 				cell_defer_no_holder(tc, cell);
2043 		} else {
2044 			inc_all_io_entry(tc->pool, bio);
2045 			remap_and_issue(tc, bio, lookup_result.block);
2046 			if (cell)
2047 				inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2048 		}
2049 		break;
2050 
2051 	case -ENODATA:
2052 		if (cell)
2053 			cell_defer_no_holder(tc, cell);
2054 		if (rw != READ) {
2055 			handle_unserviceable_bio(tc->pool, bio);
2056 			break;
2057 		}
2058 
2059 		if (tc->origin_dev) {
2060 			inc_all_io_entry(tc->pool, bio);
2061 			remap_to_origin_and_issue(tc, bio);
2062 			break;
2063 		}
2064 
2065 		zero_fill_bio(bio);
2066 		bio_endio(bio);
2067 		break;
2068 
2069 	default:
2070 		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2071 			    __func__, r);
2072 		if (cell)
2073 			cell_defer_no_holder(tc, cell);
2074 		bio_io_error(bio);
2075 		break;
2076 	}
2077 }
2078 
2079 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2080 {
2081 	__process_bio_read_only(tc, bio, NULL);
2082 }
2083 
2084 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2085 {
2086 	__process_bio_read_only(tc, cell->holder, cell);
2087 }
2088 
2089 static void process_bio_success(struct thin_c *tc, struct bio *bio)
2090 {
2091 	bio_endio(bio);
2092 }
2093 
2094 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2095 {
2096 	bio_io_error(bio);
2097 }
2098 
2099 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2100 {
2101 	cell_success(tc->pool, cell);
2102 }
2103 
2104 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2105 {
2106 	cell_error(tc->pool, cell);
2107 }
2108 
2109 /*
2110  * FIXME: should we also commit due to size of transaction, measured in
2111  * metadata blocks?
2112  */
2113 static int need_commit_due_to_time(struct pool *pool)
2114 {
2115 	return !time_in_range(jiffies, pool->last_commit_jiffies,
2116 			      pool->last_commit_jiffies + COMMIT_PERIOD);
2117 }
2118 
2119 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2120 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2121 
2122 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2123 {
2124 	struct rb_node **rbp, *parent;
2125 	struct dm_thin_endio_hook *pbd;
2126 	sector_t bi_sector = bio->bi_iter.bi_sector;
2127 
2128 	rbp = &tc->sort_bio_list.rb_node;
2129 	parent = NULL;
2130 	while (*rbp) {
2131 		parent = *rbp;
2132 		pbd = thin_pbd(parent);
2133 
2134 		if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2135 			rbp = &(*rbp)->rb_left;
2136 		else
2137 			rbp = &(*rbp)->rb_right;
2138 	}
2139 
2140 	pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2141 	rb_link_node(&pbd->rb_node, parent, rbp);
2142 	rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2143 }
2144 
2145 static void __extract_sorted_bios(struct thin_c *tc)
2146 {
2147 	struct rb_node *node;
2148 	struct dm_thin_endio_hook *pbd;
2149 	struct bio *bio;
2150 
2151 	for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2152 		pbd = thin_pbd(node);
2153 		bio = thin_bio(pbd);
2154 
2155 		bio_list_add(&tc->deferred_bio_list, bio);
2156 		rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2157 	}
2158 
2159 	WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2160 }
2161 
2162 static void __sort_thin_deferred_bios(struct thin_c *tc)
2163 {
2164 	struct bio *bio;
2165 	struct bio_list bios;
2166 
2167 	bio_list_init(&bios);
2168 	bio_list_merge(&bios, &tc->deferred_bio_list);
2169 	bio_list_init(&tc->deferred_bio_list);
2170 
2171 	/* Sort deferred_bio_list using rb-tree */
2172 	while ((bio = bio_list_pop(&bios)))
2173 		__thin_bio_rb_add(tc, bio);
2174 
2175 	/*
2176 	 * Transfer the sorted bios in sort_bio_list back to
2177 	 * deferred_bio_list to allow lockless submission of
2178 	 * all bios.
2179 	 */
2180 	__extract_sorted_bios(tc);
2181 }
2182 
2183 static void process_thin_deferred_bios(struct thin_c *tc)
2184 {
2185 	struct pool *pool = tc->pool;
2186 	struct bio *bio;
2187 	struct bio_list bios;
2188 	struct blk_plug plug;
2189 	unsigned int count = 0;
2190 
2191 	if (tc->requeue_mode) {
2192 		error_thin_bio_list(tc, &tc->deferred_bio_list,
2193 				BLK_STS_DM_REQUEUE);
2194 		return;
2195 	}
2196 
2197 	bio_list_init(&bios);
2198 
2199 	spin_lock_irq(&tc->lock);
2200 
2201 	if (bio_list_empty(&tc->deferred_bio_list)) {
2202 		spin_unlock_irq(&tc->lock);
2203 		return;
2204 	}
2205 
2206 	__sort_thin_deferred_bios(tc);
2207 
2208 	bio_list_merge(&bios, &tc->deferred_bio_list);
2209 	bio_list_init(&tc->deferred_bio_list);
2210 
2211 	spin_unlock_irq(&tc->lock);
2212 
2213 	blk_start_plug(&plug);
2214 	while ((bio = bio_list_pop(&bios))) {
2215 		/*
2216 		 * If we've got no free new_mapping structs, and processing
2217 		 * this bio might require one, we pause until there are some
2218 		 * prepared mappings to process.
2219 		 */
2220 		if (ensure_next_mapping(pool)) {
2221 			spin_lock_irq(&tc->lock);
2222 			bio_list_add(&tc->deferred_bio_list, bio);
2223 			bio_list_merge(&tc->deferred_bio_list, &bios);
2224 			spin_unlock_irq(&tc->lock);
2225 			break;
2226 		}
2227 
2228 		if (bio_op(bio) == REQ_OP_DISCARD)
2229 			pool->process_discard(tc, bio);
2230 		else
2231 			pool->process_bio(tc, bio);
2232 
2233 		if ((count++ & 127) == 0) {
2234 			throttle_work_update(&pool->throttle);
2235 			dm_pool_issue_prefetches(pool->pmd);
2236 		}
2237 		cond_resched();
2238 	}
2239 	blk_finish_plug(&plug);
2240 }
2241 
2242 static int cmp_cells(const void *lhs, const void *rhs)
2243 {
2244 	struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2245 	struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2246 
2247 	BUG_ON(!lhs_cell->holder);
2248 	BUG_ON(!rhs_cell->holder);
2249 
2250 	if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2251 		return -1;
2252 
2253 	if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2254 		return 1;
2255 
2256 	return 0;
2257 }
2258 
2259 static unsigned int sort_cells(struct pool *pool, struct list_head *cells)
2260 {
2261 	unsigned int count = 0;
2262 	struct dm_bio_prison_cell *cell, *tmp;
2263 
2264 	list_for_each_entry_safe(cell, tmp, cells, user_list) {
2265 		if (count >= CELL_SORT_ARRAY_SIZE)
2266 			break;
2267 
2268 		pool->cell_sort_array[count++] = cell;
2269 		list_del(&cell->user_list);
2270 	}
2271 
2272 	sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2273 
2274 	return count;
2275 }
2276 
2277 static void process_thin_deferred_cells(struct thin_c *tc)
2278 {
2279 	struct pool *pool = tc->pool;
2280 	struct list_head cells;
2281 	struct dm_bio_prison_cell *cell;
2282 	unsigned int i, j, count;
2283 
2284 	INIT_LIST_HEAD(&cells);
2285 
2286 	spin_lock_irq(&tc->lock);
2287 	list_splice_init(&tc->deferred_cells, &cells);
2288 	spin_unlock_irq(&tc->lock);
2289 
2290 	if (list_empty(&cells))
2291 		return;
2292 
2293 	do {
2294 		count = sort_cells(tc->pool, &cells);
2295 
2296 		for (i = 0; i < count; i++) {
2297 			cell = pool->cell_sort_array[i];
2298 			BUG_ON(!cell->holder);
2299 
2300 			/*
2301 			 * If we've got no free new_mapping structs, and processing
2302 			 * this bio might require one, we pause until there are some
2303 			 * prepared mappings to process.
2304 			 */
2305 			if (ensure_next_mapping(pool)) {
2306 				for (j = i; j < count; j++)
2307 					list_add(&pool->cell_sort_array[j]->user_list, &cells);
2308 
2309 				spin_lock_irq(&tc->lock);
2310 				list_splice(&cells, &tc->deferred_cells);
2311 				spin_unlock_irq(&tc->lock);
2312 				return;
2313 			}
2314 
2315 			if (bio_op(cell->holder) == REQ_OP_DISCARD)
2316 				pool->process_discard_cell(tc, cell);
2317 			else
2318 				pool->process_cell(tc, cell);
2319 		}
2320 		cond_resched();
2321 	} while (!list_empty(&cells));
2322 }
2323 
2324 static void thin_get(struct thin_c *tc);
2325 static void thin_put(struct thin_c *tc);
2326 
2327 /*
2328  * We can't hold rcu_read_lock() around code that can block.  So we
2329  * find a thin with the rcu lock held; bump a refcount; then drop
2330  * the lock.
2331  */
2332 static struct thin_c *get_first_thin(struct pool *pool)
2333 {
2334 	struct thin_c *tc = NULL;
2335 
2336 	rcu_read_lock();
2337 	if (!list_empty(&pool->active_thins)) {
2338 		tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2339 		thin_get(tc);
2340 	}
2341 	rcu_read_unlock();
2342 
2343 	return tc;
2344 }
2345 
2346 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2347 {
2348 	struct thin_c *old_tc = tc;
2349 
2350 	rcu_read_lock();
2351 	list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2352 		thin_get(tc);
2353 		thin_put(old_tc);
2354 		rcu_read_unlock();
2355 		return tc;
2356 	}
2357 	thin_put(old_tc);
2358 	rcu_read_unlock();
2359 
2360 	return NULL;
2361 }
2362 
2363 static void process_deferred_bios(struct pool *pool)
2364 {
2365 	struct bio *bio;
2366 	struct bio_list bios, bio_completions;
2367 	struct thin_c *tc;
2368 
2369 	tc = get_first_thin(pool);
2370 	while (tc) {
2371 		process_thin_deferred_cells(tc);
2372 		process_thin_deferred_bios(tc);
2373 		tc = get_next_thin(pool, tc);
2374 	}
2375 
2376 	/*
2377 	 * If there are any deferred flush bios, we must commit the metadata
2378 	 * before issuing them or signaling their completion.
2379 	 */
2380 	bio_list_init(&bios);
2381 	bio_list_init(&bio_completions);
2382 
2383 	spin_lock_irq(&pool->lock);
2384 	bio_list_merge(&bios, &pool->deferred_flush_bios);
2385 	bio_list_init(&pool->deferred_flush_bios);
2386 
2387 	bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2388 	bio_list_init(&pool->deferred_flush_completions);
2389 	spin_unlock_irq(&pool->lock);
2390 
2391 	if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2392 	    !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2393 		return;
2394 
2395 	if (commit(pool)) {
2396 		bio_list_merge(&bios, &bio_completions);
2397 
2398 		while ((bio = bio_list_pop(&bios)))
2399 			bio_io_error(bio);
2400 		return;
2401 	}
2402 	pool->last_commit_jiffies = jiffies;
2403 
2404 	while ((bio = bio_list_pop(&bio_completions)))
2405 		bio_endio(bio);
2406 
2407 	while ((bio = bio_list_pop(&bios))) {
2408 		/*
2409 		 * The data device was flushed as part of metadata commit,
2410 		 * so complete redundant flushes immediately.
2411 		 */
2412 		if (bio->bi_opf & REQ_PREFLUSH)
2413 			bio_endio(bio);
2414 		else
2415 			dm_submit_bio_remap(bio, NULL);
2416 	}
2417 }
2418 
2419 static void do_worker(struct work_struct *ws)
2420 {
2421 	struct pool *pool = container_of(ws, struct pool, worker);
2422 
2423 	throttle_work_start(&pool->throttle);
2424 	dm_pool_issue_prefetches(pool->pmd);
2425 	throttle_work_update(&pool->throttle);
2426 	process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2427 	throttle_work_update(&pool->throttle);
2428 	process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2429 	throttle_work_update(&pool->throttle);
2430 	process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2431 	throttle_work_update(&pool->throttle);
2432 	process_deferred_bios(pool);
2433 	throttle_work_complete(&pool->throttle);
2434 }
2435 
2436 /*
2437  * We want to commit periodically so that not too much
2438  * unwritten data builds up.
2439  */
2440 static void do_waker(struct work_struct *ws)
2441 {
2442 	struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2443 
2444 	wake_worker(pool);
2445 	queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2446 }
2447 
2448 /*
2449  * We're holding onto IO to allow userland time to react.  After the
2450  * timeout either the pool will have been resized (and thus back in
2451  * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2452  */
2453 static void do_no_space_timeout(struct work_struct *ws)
2454 {
2455 	struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2456 					 no_space_timeout);
2457 
2458 	if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2459 		pool->pf.error_if_no_space = true;
2460 		notify_of_pool_mode_change(pool);
2461 		error_retry_list_with_code(pool, BLK_STS_NOSPC);
2462 	}
2463 }
2464 
2465 /*----------------------------------------------------------------*/
2466 
2467 struct pool_work {
2468 	struct work_struct worker;
2469 	struct completion complete;
2470 };
2471 
2472 static struct pool_work *to_pool_work(struct work_struct *ws)
2473 {
2474 	return container_of(ws, struct pool_work, worker);
2475 }
2476 
2477 static void pool_work_complete(struct pool_work *pw)
2478 {
2479 	complete(&pw->complete);
2480 }
2481 
2482 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2483 			   void (*fn)(struct work_struct *))
2484 {
2485 	INIT_WORK_ONSTACK(&pw->worker, fn);
2486 	init_completion(&pw->complete);
2487 	queue_work(pool->wq, &pw->worker);
2488 	wait_for_completion(&pw->complete);
2489 }
2490 
2491 /*----------------------------------------------------------------*/
2492 
2493 struct noflush_work {
2494 	struct pool_work pw;
2495 	struct thin_c *tc;
2496 };
2497 
2498 static struct noflush_work *to_noflush(struct work_struct *ws)
2499 {
2500 	return container_of(to_pool_work(ws), struct noflush_work, pw);
2501 }
2502 
2503 static void do_noflush_start(struct work_struct *ws)
2504 {
2505 	struct noflush_work *w = to_noflush(ws);
2506 
2507 	w->tc->requeue_mode = true;
2508 	requeue_io(w->tc);
2509 	pool_work_complete(&w->pw);
2510 }
2511 
2512 static void do_noflush_stop(struct work_struct *ws)
2513 {
2514 	struct noflush_work *w = to_noflush(ws);
2515 
2516 	w->tc->requeue_mode = false;
2517 	pool_work_complete(&w->pw);
2518 }
2519 
2520 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2521 {
2522 	struct noflush_work w;
2523 
2524 	w.tc = tc;
2525 	pool_work_wait(&w.pw, tc->pool, fn);
2526 }
2527 
2528 /*----------------------------------------------------------------*/
2529 
2530 static void set_discard_callbacks(struct pool *pool)
2531 {
2532 	struct pool_c *pt = pool->ti->private;
2533 
2534 	if (pt->adjusted_pf.discard_passdown) {
2535 		pool->process_discard_cell = process_discard_cell_passdown;
2536 		pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2537 		pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2538 	} else {
2539 		pool->process_discard_cell = process_discard_cell_no_passdown;
2540 		pool->process_prepared_discard = process_prepared_discard_no_passdown;
2541 	}
2542 }
2543 
2544 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2545 {
2546 	struct pool_c *pt = pool->ti->private;
2547 	bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2548 	enum pool_mode old_mode = get_pool_mode(pool);
2549 	unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2550 
2551 	/*
2552 	 * Never allow the pool to transition to PM_WRITE mode if user
2553 	 * intervention is required to verify metadata and data consistency.
2554 	 */
2555 	if (new_mode == PM_WRITE && needs_check) {
2556 		DMERR("%s: unable to switch pool to write mode until repaired.",
2557 		      dm_device_name(pool->pool_md));
2558 		if (old_mode != new_mode)
2559 			new_mode = old_mode;
2560 		else
2561 			new_mode = PM_READ_ONLY;
2562 	}
2563 	/*
2564 	 * If we were in PM_FAIL mode, rollback of metadata failed.  We're
2565 	 * not going to recover without a thin_repair.	So we never let the
2566 	 * pool move out of the old mode.
2567 	 */
2568 	if (old_mode == PM_FAIL)
2569 		new_mode = old_mode;
2570 
2571 	switch (new_mode) {
2572 	case PM_FAIL:
2573 		dm_pool_metadata_read_only(pool->pmd);
2574 		pool->process_bio = process_bio_fail;
2575 		pool->process_discard = process_bio_fail;
2576 		pool->process_cell = process_cell_fail;
2577 		pool->process_discard_cell = process_cell_fail;
2578 		pool->process_prepared_mapping = process_prepared_mapping_fail;
2579 		pool->process_prepared_discard = process_prepared_discard_fail;
2580 
2581 		error_retry_list(pool);
2582 		break;
2583 
2584 	case PM_OUT_OF_METADATA_SPACE:
2585 	case PM_READ_ONLY:
2586 		dm_pool_metadata_read_only(pool->pmd);
2587 		pool->process_bio = process_bio_read_only;
2588 		pool->process_discard = process_bio_success;
2589 		pool->process_cell = process_cell_read_only;
2590 		pool->process_discard_cell = process_cell_success;
2591 		pool->process_prepared_mapping = process_prepared_mapping_fail;
2592 		pool->process_prepared_discard = process_prepared_discard_success;
2593 
2594 		error_retry_list(pool);
2595 		break;
2596 
2597 	case PM_OUT_OF_DATA_SPACE:
2598 		/*
2599 		 * Ideally we'd never hit this state; the low water mark
2600 		 * would trigger userland to extend the pool before we
2601 		 * completely run out of data space.  However, many small
2602 		 * IOs to unprovisioned space can consume data space at an
2603 		 * alarming rate.  Adjust your low water mark if you're
2604 		 * frequently seeing this mode.
2605 		 */
2606 		pool->out_of_data_space = true;
2607 		pool->process_bio = process_bio_read_only;
2608 		pool->process_discard = process_discard_bio;
2609 		pool->process_cell = process_cell_read_only;
2610 		pool->process_prepared_mapping = process_prepared_mapping;
2611 		set_discard_callbacks(pool);
2612 
2613 		if (!pool->pf.error_if_no_space && no_space_timeout)
2614 			queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2615 		break;
2616 
2617 	case PM_WRITE:
2618 		if (old_mode == PM_OUT_OF_DATA_SPACE)
2619 			cancel_delayed_work_sync(&pool->no_space_timeout);
2620 		pool->out_of_data_space = false;
2621 		pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2622 		dm_pool_metadata_read_write(pool->pmd);
2623 		pool->process_bio = process_bio;
2624 		pool->process_discard = process_discard_bio;
2625 		pool->process_cell = process_cell;
2626 		pool->process_prepared_mapping = process_prepared_mapping;
2627 		set_discard_callbacks(pool);
2628 		break;
2629 	}
2630 
2631 	pool->pf.mode = new_mode;
2632 	/*
2633 	 * The pool mode may have changed, sync it so bind_control_target()
2634 	 * doesn't cause an unexpected mode transition on resume.
2635 	 */
2636 	pt->adjusted_pf.mode = new_mode;
2637 
2638 	if (old_mode != new_mode)
2639 		notify_of_pool_mode_change(pool);
2640 }
2641 
2642 static void abort_transaction(struct pool *pool)
2643 {
2644 	const char *dev_name = dm_device_name(pool->pool_md);
2645 
2646 	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2647 	if (dm_pool_abort_metadata(pool->pmd)) {
2648 		DMERR("%s: failed to abort metadata transaction", dev_name);
2649 		set_pool_mode(pool, PM_FAIL);
2650 	}
2651 
2652 	if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2653 		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2654 		set_pool_mode(pool, PM_FAIL);
2655 	}
2656 }
2657 
2658 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2659 {
2660 	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2661 		    dm_device_name(pool->pool_md), op, r);
2662 
2663 	abort_transaction(pool);
2664 	set_pool_mode(pool, PM_READ_ONLY);
2665 }
2666 
2667 /*----------------------------------------------------------------*/
2668 
2669 /*
2670  * Mapping functions.
2671  */
2672 
2673 /*
2674  * Called only while mapping a thin bio to hand it over to the workqueue.
2675  */
2676 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2677 {
2678 	struct pool *pool = tc->pool;
2679 
2680 	spin_lock_irq(&tc->lock);
2681 	bio_list_add(&tc->deferred_bio_list, bio);
2682 	spin_unlock_irq(&tc->lock);
2683 
2684 	wake_worker(pool);
2685 }
2686 
2687 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2688 {
2689 	struct pool *pool = tc->pool;
2690 
2691 	throttle_lock(&pool->throttle);
2692 	thin_defer_bio(tc, bio);
2693 	throttle_unlock(&pool->throttle);
2694 }
2695 
2696 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2697 {
2698 	struct pool *pool = tc->pool;
2699 
2700 	throttle_lock(&pool->throttle);
2701 	spin_lock_irq(&tc->lock);
2702 	list_add_tail(&cell->user_list, &tc->deferred_cells);
2703 	spin_unlock_irq(&tc->lock);
2704 	throttle_unlock(&pool->throttle);
2705 
2706 	wake_worker(pool);
2707 }
2708 
2709 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2710 {
2711 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2712 
2713 	h->tc = tc;
2714 	h->shared_read_entry = NULL;
2715 	h->all_io_entry = NULL;
2716 	h->overwrite_mapping = NULL;
2717 	h->cell = NULL;
2718 }
2719 
2720 /*
2721  * Non-blocking function called from the thin target's map function.
2722  */
2723 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2724 {
2725 	int r;
2726 	struct thin_c *tc = ti->private;
2727 	dm_block_t block = get_bio_block(tc, bio);
2728 	struct dm_thin_device *td = tc->td;
2729 	struct dm_thin_lookup_result result;
2730 	struct dm_bio_prison_cell *virt_cell, *data_cell;
2731 	struct dm_cell_key key;
2732 
2733 	thin_hook_bio(tc, bio);
2734 
2735 	if (tc->requeue_mode) {
2736 		bio->bi_status = BLK_STS_DM_REQUEUE;
2737 		bio_endio(bio);
2738 		return DM_MAPIO_SUBMITTED;
2739 	}
2740 
2741 	if (get_pool_mode(tc->pool) == PM_FAIL) {
2742 		bio_io_error(bio);
2743 		return DM_MAPIO_SUBMITTED;
2744 	}
2745 
2746 	if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2747 		thin_defer_bio_with_throttle(tc, bio);
2748 		return DM_MAPIO_SUBMITTED;
2749 	}
2750 
2751 	/*
2752 	 * We must hold the virtual cell before doing the lookup, otherwise
2753 	 * there's a race with discard.
2754 	 */
2755 	build_virtual_key(tc->td, block, &key);
2756 	if (bio_detain(tc->pool, &key, bio, &virt_cell))
2757 		return DM_MAPIO_SUBMITTED;
2758 
2759 	r = dm_thin_find_block(td, block, 0, &result);
2760 
2761 	/*
2762 	 * Note that we defer readahead too.
2763 	 */
2764 	switch (r) {
2765 	case 0:
2766 		if (unlikely(result.shared)) {
2767 			/*
2768 			 * We have a race condition here between the
2769 			 * result.shared value returned by the lookup and
2770 			 * snapshot creation, which may cause new
2771 			 * sharing.
2772 			 *
2773 			 * To avoid this always quiesce the origin before
2774 			 * taking the snap.  You want to do this anyway to
2775 			 * ensure a consistent application view
2776 			 * (i.e. lockfs).
2777 			 *
2778 			 * More distant ancestors are irrelevant. The
2779 			 * shared flag will be set in their case.
2780 			 */
2781 			thin_defer_cell(tc, virt_cell);
2782 			return DM_MAPIO_SUBMITTED;
2783 		}
2784 
2785 		build_data_key(tc->td, result.block, &key);
2786 		if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2787 			cell_defer_no_holder(tc, virt_cell);
2788 			return DM_MAPIO_SUBMITTED;
2789 		}
2790 
2791 		inc_all_io_entry(tc->pool, bio);
2792 		cell_defer_no_holder(tc, data_cell);
2793 		cell_defer_no_holder(tc, virt_cell);
2794 
2795 		remap(tc, bio, result.block);
2796 		return DM_MAPIO_REMAPPED;
2797 
2798 	case -ENODATA:
2799 	case -EWOULDBLOCK:
2800 		thin_defer_cell(tc, virt_cell);
2801 		return DM_MAPIO_SUBMITTED;
2802 
2803 	default:
2804 		/*
2805 		 * Must always call bio_io_error on failure.
2806 		 * dm_thin_find_block can fail with -EINVAL if the
2807 		 * pool is switched to fail-io mode.
2808 		 */
2809 		bio_io_error(bio);
2810 		cell_defer_no_holder(tc, virt_cell);
2811 		return DM_MAPIO_SUBMITTED;
2812 	}
2813 }
2814 
2815 static void requeue_bios(struct pool *pool)
2816 {
2817 	struct thin_c *tc;
2818 
2819 	rcu_read_lock();
2820 	list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2821 		spin_lock_irq(&tc->lock);
2822 		bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2823 		bio_list_init(&tc->retry_on_resume_list);
2824 		spin_unlock_irq(&tc->lock);
2825 	}
2826 	rcu_read_unlock();
2827 }
2828 
2829 /*
2830  *--------------------------------------------------------------
2831  * Binding of control targets to a pool object
2832  *--------------------------------------------------------------
2833  */
2834 static bool is_factor(sector_t block_size, uint32_t n)
2835 {
2836 	return !sector_div(block_size, n);
2837 }
2838 
2839 /*
2840  * If discard_passdown was enabled verify that the data device
2841  * supports discards.  Disable discard_passdown if not.
2842  */
2843 static void disable_discard_passdown_if_not_supported(struct pool_c *pt)
2844 {
2845 	struct pool *pool = pt->pool;
2846 	struct block_device *data_bdev = pt->data_dev->bdev;
2847 	struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2848 	const char *reason = NULL;
2849 
2850 	if (!pt->adjusted_pf.discard_passdown)
2851 		return;
2852 
2853 	if (!bdev_max_discard_sectors(pt->data_dev->bdev))
2854 		reason = "discard unsupported";
2855 
2856 	else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2857 		reason = "max discard sectors smaller than a block";
2858 
2859 	if (reason) {
2860 		DMWARN("Data device (%pg) %s: Disabling discard passdown.", data_bdev, reason);
2861 		pt->adjusted_pf.discard_passdown = false;
2862 	}
2863 }
2864 
2865 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2866 {
2867 	struct pool_c *pt = ti->private;
2868 
2869 	/*
2870 	 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2871 	 */
2872 	enum pool_mode old_mode = get_pool_mode(pool);
2873 	enum pool_mode new_mode = pt->adjusted_pf.mode;
2874 
2875 	/*
2876 	 * Don't change the pool's mode until set_pool_mode() below.
2877 	 * Otherwise the pool's process_* function pointers may
2878 	 * not match the desired pool mode.
2879 	 */
2880 	pt->adjusted_pf.mode = old_mode;
2881 
2882 	pool->ti = ti;
2883 	pool->pf = pt->adjusted_pf;
2884 	pool->low_water_blocks = pt->low_water_blocks;
2885 
2886 	set_pool_mode(pool, new_mode);
2887 
2888 	return 0;
2889 }
2890 
2891 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2892 {
2893 	if (pool->ti == ti)
2894 		pool->ti = NULL;
2895 }
2896 
2897 /*
2898  *--------------------------------------------------------------
2899  * Pool creation
2900  *--------------------------------------------------------------
2901  */
2902 /* Initialize pool features. */
2903 static void pool_features_init(struct pool_features *pf)
2904 {
2905 	pf->mode = PM_WRITE;
2906 	pf->zero_new_blocks = true;
2907 	pf->discard_enabled = true;
2908 	pf->discard_passdown = true;
2909 	pf->error_if_no_space = false;
2910 }
2911 
2912 static void __pool_destroy(struct pool *pool)
2913 {
2914 	__pool_table_remove(pool);
2915 
2916 	vfree(pool->cell_sort_array);
2917 	if (dm_pool_metadata_close(pool->pmd) < 0)
2918 		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2919 
2920 	dm_bio_prison_destroy(pool->prison);
2921 	dm_kcopyd_client_destroy(pool->copier);
2922 
2923 	cancel_delayed_work_sync(&pool->waker);
2924 	cancel_delayed_work_sync(&pool->no_space_timeout);
2925 	if (pool->wq)
2926 		destroy_workqueue(pool->wq);
2927 
2928 	if (pool->next_mapping)
2929 		mempool_free(pool->next_mapping, &pool->mapping_pool);
2930 	mempool_exit(&pool->mapping_pool);
2931 	dm_deferred_set_destroy(pool->shared_read_ds);
2932 	dm_deferred_set_destroy(pool->all_io_ds);
2933 	kfree(pool);
2934 }
2935 
2936 static struct kmem_cache *_new_mapping_cache;
2937 
2938 static struct pool *pool_create(struct mapped_device *pool_md,
2939 				struct block_device *metadata_dev,
2940 				struct block_device *data_dev,
2941 				unsigned long block_size,
2942 				int read_only, char **error)
2943 {
2944 	int r;
2945 	void *err_p;
2946 	struct pool *pool;
2947 	struct dm_pool_metadata *pmd;
2948 	bool format_device = read_only ? false : true;
2949 
2950 	pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2951 	if (IS_ERR(pmd)) {
2952 		*error = "Error creating metadata object";
2953 		return (struct pool *)pmd;
2954 	}
2955 
2956 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2957 	if (!pool) {
2958 		*error = "Error allocating memory for pool";
2959 		err_p = ERR_PTR(-ENOMEM);
2960 		goto bad_pool;
2961 	}
2962 
2963 	pool->pmd = pmd;
2964 	pool->sectors_per_block = block_size;
2965 	if (block_size & (block_size - 1))
2966 		pool->sectors_per_block_shift = -1;
2967 	else
2968 		pool->sectors_per_block_shift = __ffs(block_size);
2969 	pool->low_water_blocks = 0;
2970 	pool_features_init(&pool->pf);
2971 	pool->prison = dm_bio_prison_create();
2972 	if (!pool->prison) {
2973 		*error = "Error creating pool's bio prison";
2974 		err_p = ERR_PTR(-ENOMEM);
2975 		goto bad_prison;
2976 	}
2977 
2978 	pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2979 	if (IS_ERR(pool->copier)) {
2980 		r = PTR_ERR(pool->copier);
2981 		*error = "Error creating pool's kcopyd client";
2982 		err_p = ERR_PTR(r);
2983 		goto bad_kcopyd_client;
2984 	}
2985 
2986 	/*
2987 	 * Create singlethreaded workqueue that will service all devices
2988 	 * that use this metadata.
2989 	 */
2990 	pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2991 	if (!pool->wq) {
2992 		*error = "Error creating pool's workqueue";
2993 		err_p = ERR_PTR(-ENOMEM);
2994 		goto bad_wq;
2995 	}
2996 
2997 	throttle_init(&pool->throttle);
2998 	INIT_WORK(&pool->worker, do_worker);
2999 	INIT_DELAYED_WORK(&pool->waker, do_waker);
3000 	INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
3001 	spin_lock_init(&pool->lock);
3002 	bio_list_init(&pool->deferred_flush_bios);
3003 	bio_list_init(&pool->deferred_flush_completions);
3004 	INIT_LIST_HEAD(&pool->prepared_mappings);
3005 	INIT_LIST_HEAD(&pool->prepared_discards);
3006 	INIT_LIST_HEAD(&pool->prepared_discards_pt2);
3007 	INIT_LIST_HEAD(&pool->active_thins);
3008 	pool->low_water_triggered = false;
3009 	pool->suspended = true;
3010 	pool->out_of_data_space = false;
3011 
3012 	pool->shared_read_ds = dm_deferred_set_create();
3013 	if (!pool->shared_read_ds) {
3014 		*error = "Error creating pool's shared read deferred set";
3015 		err_p = ERR_PTR(-ENOMEM);
3016 		goto bad_shared_read_ds;
3017 	}
3018 
3019 	pool->all_io_ds = dm_deferred_set_create();
3020 	if (!pool->all_io_ds) {
3021 		*error = "Error creating pool's all io deferred set";
3022 		err_p = ERR_PTR(-ENOMEM);
3023 		goto bad_all_io_ds;
3024 	}
3025 
3026 	pool->next_mapping = NULL;
3027 	r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
3028 				   _new_mapping_cache);
3029 	if (r) {
3030 		*error = "Error creating pool's mapping mempool";
3031 		err_p = ERR_PTR(r);
3032 		goto bad_mapping_pool;
3033 	}
3034 
3035 	pool->cell_sort_array =
3036 		vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3037 				   sizeof(*pool->cell_sort_array)));
3038 	if (!pool->cell_sort_array) {
3039 		*error = "Error allocating cell sort array";
3040 		err_p = ERR_PTR(-ENOMEM);
3041 		goto bad_sort_array;
3042 	}
3043 
3044 	pool->ref_count = 1;
3045 	pool->last_commit_jiffies = jiffies;
3046 	pool->pool_md = pool_md;
3047 	pool->md_dev = metadata_dev;
3048 	pool->data_dev = data_dev;
3049 	__pool_table_insert(pool);
3050 
3051 	return pool;
3052 
3053 bad_sort_array:
3054 	mempool_exit(&pool->mapping_pool);
3055 bad_mapping_pool:
3056 	dm_deferred_set_destroy(pool->all_io_ds);
3057 bad_all_io_ds:
3058 	dm_deferred_set_destroy(pool->shared_read_ds);
3059 bad_shared_read_ds:
3060 	destroy_workqueue(pool->wq);
3061 bad_wq:
3062 	dm_kcopyd_client_destroy(pool->copier);
3063 bad_kcopyd_client:
3064 	dm_bio_prison_destroy(pool->prison);
3065 bad_prison:
3066 	kfree(pool);
3067 bad_pool:
3068 	if (dm_pool_metadata_close(pmd))
3069 		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3070 
3071 	return err_p;
3072 }
3073 
3074 static void __pool_inc(struct pool *pool)
3075 {
3076 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3077 	pool->ref_count++;
3078 }
3079 
3080 static void __pool_dec(struct pool *pool)
3081 {
3082 	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3083 	BUG_ON(!pool->ref_count);
3084 	if (!--pool->ref_count)
3085 		__pool_destroy(pool);
3086 }
3087 
3088 static struct pool *__pool_find(struct mapped_device *pool_md,
3089 				struct block_device *metadata_dev,
3090 				struct block_device *data_dev,
3091 				unsigned long block_size, int read_only,
3092 				char **error, int *created)
3093 {
3094 	struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3095 
3096 	if (pool) {
3097 		if (pool->pool_md != pool_md) {
3098 			*error = "metadata device already in use by a pool";
3099 			return ERR_PTR(-EBUSY);
3100 		}
3101 		if (pool->data_dev != data_dev) {
3102 			*error = "data device already in use by a pool";
3103 			return ERR_PTR(-EBUSY);
3104 		}
3105 		__pool_inc(pool);
3106 
3107 	} else {
3108 		pool = __pool_table_lookup(pool_md);
3109 		if (pool) {
3110 			if (pool->md_dev != metadata_dev || pool->data_dev != data_dev) {
3111 				*error = "different pool cannot replace a pool";
3112 				return ERR_PTR(-EINVAL);
3113 			}
3114 			__pool_inc(pool);
3115 
3116 		} else {
3117 			pool = pool_create(pool_md, metadata_dev, data_dev, block_size, read_only, error);
3118 			*created = 1;
3119 		}
3120 	}
3121 
3122 	return pool;
3123 }
3124 
3125 /*
3126  *--------------------------------------------------------------
3127  * Pool target methods
3128  *--------------------------------------------------------------
3129  */
3130 static void pool_dtr(struct dm_target *ti)
3131 {
3132 	struct pool_c *pt = ti->private;
3133 
3134 	mutex_lock(&dm_thin_pool_table.mutex);
3135 
3136 	unbind_control_target(pt->pool, ti);
3137 	__pool_dec(pt->pool);
3138 	dm_put_device(ti, pt->metadata_dev);
3139 	dm_put_device(ti, pt->data_dev);
3140 	kfree(pt);
3141 
3142 	mutex_unlock(&dm_thin_pool_table.mutex);
3143 }
3144 
3145 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3146 			       struct dm_target *ti)
3147 {
3148 	int r;
3149 	unsigned int argc;
3150 	const char *arg_name;
3151 
3152 	static const struct dm_arg _args[] = {
3153 		{0, 4, "Invalid number of pool feature arguments"},
3154 	};
3155 
3156 	/*
3157 	 * No feature arguments supplied.
3158 	 */
3159 	if (!as->argc)
3160 		return 0;
3161 
3162 	r = dm_read_arg_group(_args, as, &argc, &ti->error);
3163 	if (r)
3164 		return -EINVAL;
3165 
3166 	while (argc && !r) {
3167 		arg_name = dm_shift_arg(as);
3168 		argc--;
3169 
3170 		if (!strcasecmp(arg_name, "skip_block_zeroing"))
3171 			pf->zero_new_blocks = false;
3172 
3173 		else if (!strcasecmp(arg_name, "ignore_discard"))
3174 			pf->discard_enabled = false;
3175 
3176 		else if (!strcasecmp(arg_name, "no_discard_passdown"))
3177 			pf->discard_passdown = false;
3178 
3179 		else if (!strcasecmp(arg_name, "read_only"))
3180 			pf->mode = PM_READ_ONLY;
3181 
3182 		else if (!strcasecmp(arg_name, "error_if_no_space"))
3183 			pf->error_if_no_space = true;
3184 
3185 		else {
3186 			ti->error = "Unrecognised pool feature requested";
3187 			r = -EINVAL;
3188 			break;
3189 		}
3190 	}
3191 
3192 	return r;
3193 }
3194 
3195 static void metadata_low_callback(void *context)
3196 {
3197 	struct pool *pool = context;
3198 
3199 	DMWARN("%s: reached low water mark for metadata device: sending event.",
3200 	       dm_device_name(pool->pool_md));
3201 
3202 	dm_table_event(pool->ti->table);
3203 }
3204 
3205 /*
3206  * We need to flush the data device **before** committing the metadata.
3207  *
3208  * This ensures that the data blocks of any newly inserted mappings are
3209  * properly written to non-volatile storage and won't be lost in case of a
3210  * crash.
3211  *
3212  * Failure to do so can result in data corruption in the case of internal or
3213  * external snapshots and in the case of newly provisioned blocks, when block
3214  * zeroing is enabled.
3215  */
3216 static int metadata_pre_commit_callback(void *context)
3217 {
3218 	struct pool *pool = context;
3219 
3220 	return blkdev_issue_flush(pool->data_dev);
3221 }
3222 
3223 static sector_t get_dev_size(struct block_device *bdev)
3224 {
3225 	return bdev_nr_sectors(bdev);
3226 }
3227 
3228 static void warn_if_metadata_device_too_big(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_WARNING)
3233 		DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
3234 		       bdev, THIN_METADATA_MAX_SECTORS);
3235 }
3236 
3237 static sector_t get_metadata_dev_size(struct block_device *bdev)
3238 {
3239 	sector_t metadata_dev_size = get_dev_size(bdev);
3240 
3241 	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3242 		metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3243 
3244 	return metadata_dev_size;
3245 }
3246 
3247 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3248 {
3249 	sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3250 
3251 	sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3252 
3253 	return metadata_dev_size;
3254 }
3255 
3256 /*
3257  * When a metadata threshold is crossed a dm event is triggered, and
3258  * userland should respond by growing the metadata device.  We could let
3259  * userland set the threshold, like we do with the data threshold, but I'm
3260  * not sure they know enough to do this well.
3261  */
3262 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3263 {
3264 	/*
3265 	 * 4M is ample for all ops with the possible exception of thin
3266 	 * device deletion which is harmless if it fails (just retry the
3267 	 * delete after you've grown the device).
3268 	 */
3269 	dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3270 
3271 	return min((dm_block_t)1024ULL /* 4M */, quarter);
3272 }
3273 
3274 /*
3275  * thin-pool <metadata dev> <data dev>
3276  *	     <data block size (sectors)>
3277  *	     <low water mark (blocks)>
3278  *	     [<#feature args> [<arg>]*]
3279  *
3280  * Optional feature arguments are:
3281  *	     skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3282  *	     ignore_discard: disable discard
3283  *	     no_discard_passdown: don't pass discards down to the data device
3284  *	     read_only: Don't allow any changes to be made to the pool metadata.
3285  *	     error_if_no_space: error IOs, instead of queueing, if no space.
3286  */
3287 static int pool_ctr(struct dm_target *ti, unsigned int argc, char **argv)
3288 {
3289 	int r, pool_created = 0;
3290 	struct pool_c *pt;
3291 	struct pool *pool;
3292 	struct pool_features pf;
3293 	struct dm_arg_set as;
3294 	struct dm_dev *data_dev;
3295 	unsigned long block_size;
3296 	dm_block_t low_water_blocks;
3297 	struct dm_dev *metadata_dev;
3298 	blk_mode_t metadata_mode;
3299 
3300 	/*
3301 	 * FIXME Remove validation from scope of lock.
3302 	 */
3303 	mutex_lock(&dm_thin_pool_table.mutex);
3304 
3305 	if (argc < 4) {
3306 		ti->error = "Invalid argument count";
3307 		r = -EINVAL;
3308 		goto out_unlock;
3309 	}
3310 
3311 	as.argc = argc;
3312 	as.argv = argv;
3313 
3314 	/* make sure metadata and data are different devices */
3315 	if (!strcmp(argv[0], argv[1])) {
3316 		ti->error = "Error setting metadata or data device";
3317 		r = -EINVAL;
3318 		goto out_unlock;
3319 	}
3320 
3321 	/*
3322 	 * Set default pool features.
3323 	 */
3324 	pool_features_init(&pf);
3325 
3326 	dm_consume_args(&as, 4);
3327 	r = parse_pool_features(&as, &pf, ti);
3328 	if (r)
3329 		goto out_unlock;
3330 
3331 	metadata_mode = BLK_OPEN_READ |
3332 		((pf.mode == PM_READ_ONLY) ? 0 : BLK_OPEN_WRITE);
3333 	r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3334 	if (r) {
3335 		ti->error = "Error opening metadata block device";
3336 		goto out_unlock;
3337 	}
3338 	warn_if_metadata_device_too_big(metadata_dev->bdev);
3339 
3340 	r = dm_get_device(ti, argv[1], BLK_OPEN_READ | BLK_OPEN_WRITE, &data_dev);
3341 	if (r) {
3342 		ti->error = "Error getting data device";
3343 		goto out_metadata;
3344 	}
3345 
3346 	if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3347 	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3348 	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3349 	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3350 		ti->error = "Invalid block size";
3351 		r = -EINVAL;
3352 		goto out;
3353 	}
3354 
3355 	if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3356 		ti->error = "Invalid low water mark";
3357 		r = -EINVAL;
3358 		goto out;
3359 	}
3360 
3361 	pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3362 	if (!pt) {
3363 		r = -ENOMEM;
3364 		goto out;
3365 	}
3366 
3367 	pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev, data_dev->bdev,
3368 			   block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3369 	if (IS_ERR(pool)) {
3370 		r = PTR_ERR(pool);
3371 		goto out_free_pt;
3372 	}
3373 
3374 	/*
3375 	 * 'pool_created' reflects whether this is the first table load.
3376 	 * Top level discard support is not allowed to be changed after
3377 	 * initial load.  This would require a pool reload to trigger thin
3378 	 * device changes.
3379 	 */
3380 	if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3381 		ti->error = "Discard support cannot be disabled once enabled";
3382 		r = -EINVAL;
3383 		goto out_flags_changed;
3384 	}
3385 
3386 	pt->pool = pool;
3387 	pt->ti = ti;
3388 	pt->metadata_dev = metadata_dev;
3389 	pt->data_dev = data_dev;
3390 	pt->low_water_blocks = low_water_blocks;
3391 	pt->adjusted_pf = pt->requested_pf = pf;
3392 	ti->num_flush_bios = 1;
3393 	ti->limit_swap_bios = true;
3394 
3395 	/*
3396 	 * Only need to enable discards if the pool should pass
3397 	 * them down to the data device.  The thin device's discard
3398 	 * processing will cause mappings to be removed from the btree.
3399 	 */
3400 	if (pf.discard_enabled && pf.discard_passdown) {
3401 		ti->num_discard_bios = 1;
3402 		/*
3403 		 * Setting 'discards_supported' circumvents the normal
3404 		 * stacking of discard limits (this keeps the pool and
3405 		 * thin devices' discard limits consistent).
3406 		 */
3407 		ti->discards_supported = true;
3408 		ti->max_discard_granularity = true;
3409 	}
3410 	ti->private = pt;
3411 
3412 	r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3413 						calc_metadata_threshold(pt),
3414 						metadata_low_callback,
3415 						pool);
3416 	if (r) {
3417 		ti->error = "Error registering metadata threshold";
3418 		goto out_flags_changed;
3419 	}
3420 
3421 	dm_pool_register_pre_commit_callback(pool->pmd,
3422 					     metadata_pre_commit_callback, pool);
3423 
3424 	mutex_unlock(&dm_thin_pool_table.mutex);
3425 
3426 	return 0;
3427 
3428 out_flags_changed:
3429 	__pool_dec(pool);
3430 out_free_pt:
3431 	kfree(pt);
3432 out:
3433 	dm_put_device(ti, data_dev);
3434 out_metadata:
3435 	dm_put_device(ti, metadata_dev);
3436 out_unlock:
3437 	mutex_unlock(&dm_thin_pool_table.mutex);
3438 
3439 	return r;
3440 }
3441 
3442 static int pool_map(struct dm_target *ti, struct bio *bio)
3443 {
3444 	struct pool_c *pt = ti->private;
3445 	struct pool *pool = pt->pool;
3446 
3447 	/*
3448 	 * As this is a singleton target, ti->begin is always zero.
3449 	 */
3450 	spin_lock_irq(&pool->lock);
3451 	bio_set_dev(bio, pt->data_dev->bdev);
3452 	spin_unlock_irq(&pool->lock);
3453 
3454 	return DM_MAPIO_REMAPPED;
3455 }
3456 
3457 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3458 {
3459 	int r;
3460 	struct pool_c *pt = ti->private;
3461 	struct pool *pool = pt->pool;
3462 	sector_t data_size = ti->len;
3463 	dm_block_t sb_data_size;
3464 
3465 	*need_commit = false;
3466 
3467 	(void) sector_div(data_size, pool->sectors_per_block);
3468 
3469 	r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3470 	if (r) {
3471 		DMERR("%s: failed to retrieve data device size",
3472 		      dm_device_name(pool->pool_md));
3473 		return r;
3474 	}
3475 
3476 	if (data_size < sb_data_size) {
3477 		DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3478 		      dm_device_name(pool->pool_md),
3479 		      (unsigned long long)data_size, sb_data_size);
3480 		return -EINVAL;
3481 
3482 	} else if (data_size > sb_data_size) {
3483 		if (dm_pool_metadata_needs_check(pool->pmd)) {
3484 			DMERR("%s: unable to grow the data device until repaired.",
3485 			      dm_device_name(pool->pool_md));
3486 			return 0;
3487 		}
3488 
3489 		if (sb_data_size)
3490 			DMINFO("%s: growing the data device from %llu to %llu blocks",
3491 			       dm_device_name(pool->pool_md),
3492 			       sb_data_size, (unsigned long long)data_size);
3493 		r = dm_pool_resize_data_dev(pool->pmd, data_size);
3494 		if (r) {
3495 			metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3496 			return r;
3497 		}
3498 
3499 		*need_commit = true;
3500 	}
3501 
3502 	return 0;
3503 }
3504 
3505 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3506 {
3507 	int r;
3508 	struct pool_c *pt = ti->private;
3509 	struct pool *pool = pt->pool;
3510 	dm_block_t metadata_dev_size, sb_metadata_dev_size;
3511 
3512 	*need_commit = false;
3513 
3514 	metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3515 
3516 	r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3517 	if (r) {
3518 		DMERR("%s: failed to retrieve metadata device size",
3519 		      dm_device_name(pool->pool_md));
3520 		return r;
3521 	}
3522 
3523 	if (metadata_dev_size < sb_metadata_dev_size) {
3524 		DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3525 		      dm_device_name(pool->pool_md),
3526 		      metadata_dev_size, sb_metadata_dev_size);
3527 		return -EINVAL;
3528 
3529 	} else if (metadata_dev_size > sb_metadata_dev_size) {
3530 		if (dm_pool_metadata_needs_check(pool->pmd)) {
3531 			DMERR("%s: unable to grow the metadata device until repaired.",
3532 			      dm_device_name(pool->pool_md));
3533 			return 0;
3534 		}
3535 
3536 		warn_if_metadata_device_too_big(pool->md_dev);
3537 		DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3538 		       dm_device_name(pool->pool_md),
3539 		       sb_metadata_dev_size, metadata_dev_size);
3540 
3541 		if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3542 			set_pool_mode(pool, PM_WRITE);
3543 
3544 		r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3545 		if (r) {
3546 			metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3547 			return r;
3548 		}
3549 
3550 		*need_commit = true;
3551 	}
3552 
3553 	return 0;
3554 }
3555 
3556 /*
3557  * Retrieves the number of blocks of the data device from
3558  * the superblock and compares it to the actual device size,
3559  * thus resizing the data device in case it has grown.
3560  *
3561  * This both copes with opening preallocated data devices in the ctr
3562  * being followed by a resume
3563  * -and-
3564  * calling the resume method individually after userspace has
3565  * grown the data device in reaction to a table event.
3566  */
3567 static int pool_preresume(struct dm_target *ti)
3568 {
3569 	int r;
3570 	bool need_commit1, need_commit2;
3571 	struct pool_c *pt = ti->private;
3572 	struct pool *pool = pt->pool;
3573 
3574 	/*
3575 	 * Take control of the pool object.
3576 	 */
3577 	r = bind_control_target(pool, ti);
3578 	if (r)
3579 		goto out;
3580 
3581 	r = maybe_resize_data_dev(ti, &need_commit1);
3582 	if (r)
3583 		goto out;
3584 
3585 	r = maybe_resize_metadata_dev(ti, &need_commit2);
3586 	if (r)
3587 		goto out;
3588 
3589 	if (need_commit1 || need_commit2)
3590 		(void) commit(pool);
3591 out:
3592 	/*
3593 	 * When a thin-pool is PM_FAIL, it cannot be rebuilt if
3594 	 * bio is in deferred list. Therefore need to return 0
3595 	 * to allow pool_resume() to flush IO.
3596 	 */
3597 	if (r && get_pool_mode(pool) == PM_FAIL)
3598 		r = 0;
3599 
3600 	return r;
3601 }
3602 
3603 static void pool_suspend_active_thins(struct pool *pool)
3604 {
3605 	struct thin_c *tc;
3606 
3607 	/* Suspend all active thin devices */
3608 	tc = get_first_thin(pool);
3609 	while (tc) {
3610 		dm_internal_suspend_noflush(tc->thin_md);
3611 		tc = get_next_thin(pool, tc);
3612 	}
3613 }
3614 
3615 static void pool_resume_active_thins(struct pool *pool)
3616 {
3617 	struct thin_c *tc;
3618 
3619 	/* Resume all active thin devices */
3620 	tc = get_first_thin(pool);
3621 	while (tc) {
3622 		dm_internal_resume(tc->thin_md);
3623 		tc = get_next_thin(pool, tc);
3624 	}
3625 }
3626 
3627 static void pool_resume(struct dm_target *ti)
3628 {
3629 	struct pool_c *pt = ti->private;
3630 	struct pool *pool = pt->pool;
3631 
3632 	/*
3633 	 * Must requeue active_thins' bios and then resume
3634 	 * active_thins _before_ clearing 'suspend' flag.
3635 	 */
3636 	requeue_bios(pool);
3637 	pool_resume_active_thins(pool);
3638 
3639 	spin_lock_irq(&pool->lock);
3640 	pool->low_water_triggered = false;
3641 	pool->suspended = false;
3642 	spin_unlock_irq(&pool->lock);
3643 
3644 	do_waker(&pool->waker.work);
3645 }
3646 
3647 static void pool_presuspend(struct dm_target *ti)
3648 {
3649 	struct pool_c *pt = ti->private;
3650 	struct pool *pool = pt->pool;
3651 
3652 	spin_lock_irq(&pool->lock);
3653 	pool->suspended = true;
3654 	spin_unlock_irq(&pool->lock);
3655 
3656 	pool_suspend_active_thins(pool);
3657 }
3658 
3659 static void pool_presuspend_undo(struct dm_target *ti)
3660 {
3661 	struct pool_c *pt = ti->private;
3662 	struct pool *pool = pt->pool;
3663 
3664 	pool_resume_active_thins(pool);
3665 
3666 	spin_lock_irq(&pool->lock);
3667 	pool->suspended = false;
3668 	spin_unlock_irq(&pool->lock);
3669 }
3670 
3671 static void pool_postsuspend(struct dm_target *ti)
3672 {
3673 	struct pool_c *pt = ti->private;
3674 	struct pool *pool = pt->pool;
3675 
3676 	cancel_delayed_work_sync(&pool->waker);
3677 	cancel_delayed_work_sync(&pool->no_space_timeout);
3678 	flush_workqueue(pool->wq);
3679 	(void) commit(pool);
3680 }
3681 
3682 static int check_arg_count(unsigned int argc, unsigned int args_required)
3683 {
3684 	if (argc != args_required) {
3685 		DMWARN("Message received with %u arguments instead of %u.",
3686 		       argc, args_required);
3687 		return -EINVAL;
3688 	}
3689 
3690 	return 0;
3691 }
3692 
3693 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3694 {
3695 	if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3696 	    *dev_id <= MAX_DEV_ID)
3697 		return 0;
3698 
3699 	if (warning)
3700 		DMWARN("Message received with invalid device id: %s", arg);
3701 
3702 	return -EINVAL;
3703 }
3704 
3705 static int process_create_thin_mesg(unsigned int argc, char **argv, struct pool *pool)
3706 {
3707 	dm_thin_id dev_id;
3708 	int r;
3709 
3710 	r = check_arg_count(argc, 2);
3711 	if (r)
3712 		return r;
3713 
3714 	r = read_dev_id(argv[1], &dev_id, 1);
3715 	if (r)
3716 		return r;
3717 
3718 	r = dm_pool_create_thin(pool->pmd, dev_id);
3719 	if (r) {
3720 		DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3721 		       argv[1]);
3722 		return r;
3723 	}
3724 
3725 	return 0;
3726 }
3727 
3728 static int process_create_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3729 {
3730 	dm_thin_id dev_id;
3731 	dm_thin_id origin_dev_id;
3732 	int r;
3733 
3734 	r = check_arg_count(argc, 3);
3735 	if (r)
3736 		return r;
3737 
3738 	r = read_dev_id(argv[1], &dev_id, 1);
3739 	if (r)
3740 		return r;
3741 
3742 	r = read_dev_id(argv[2], &origin_dev_id, 1);
3743 	if (r)
3744 		return r;
3745 
3746 	r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3747 	if (r) {
3748 		DMWARN("Creation of new snapshot %s of device %s failed.",
3749 		       argv[1], argv[2]);
3750 		return r;
3751 	}
3752 
3753 	return 0;
3754 }
3755 
3756 static int process_delete_mesg(unsigned int argc, char **argv, struct pool *pool)
3757 {
3758 	dm_thin_id dev_id;
3759 	int r;
3760 
3761 	r = check_arg_count(argc, 2);
3762 	if (r)
3763 		return r;
3764 
3765 	r = read_dev_id(argv[1], &dev_id, 1);
3766 	if (r)
3767 		return r;
3768 
3769 	r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3770 	if (r)
3771 		DMWARN("Deletion of thin device %s failed.", argv[1]);
3772 
3773 	return r;
3774 }
3775 
3776 static int process_set_transaction_id_mesg(unsigned int argc, char **argv, struct pool *pool)
3777 {
3778 	dm_thin_id old_id, new_id;
3779 	int r;
3780 
3781 	r = check_arg_count(argc, 3);
3782 	if (r)
3783 		return r;
3784 
3785 	if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3786 		DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3787 		return -EINVAL;
3788 	}
3789 
3790 	if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3791 		DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3792 		return -EINVAL;
3793 	}
3794 
3795 	r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3796 	if (r) {
3797 		DMWARN("Failed to change transaction id from %s to %s.",
3798 		       argv[1], argv[2]);
3799 		return r;
3800 	}
3801 
3802 	return 0;
3803 }
3804 
3805 static int process_reserve_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3806 {
3807 	int r;
3808 
3809 	r = check_arg_count(argc, 1);
3810 	if (r)
3811 		return r;
3812 
3813 	(void) commit(pool);
3814 
3815 	r = dm_pool_reserve_metadata_snap(pool->pmd);
3816 	if (r)
3817 		DMWARN("reserve_metadata_snap message failed.");
3818 
3819 	return r;
3820 }
3821 
3822 static int process_release_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3823 {
3824 	int r;
3825 
3826 	r = check_arg_count(argc, 1);
3827 	if (r)
3828 		return r;
3829 
3830 	r = dm_pool_release_metadata_snap(pool->pmd);
3831 	if (r)
3832 		DMWARN("release_metadata_snap message failed.");
3833 
3834 	return r;
3835 }
3836 
3837 /*
3838  * Messages supported:
3839  *   create_thin	<dev_id>
3840  *   create_snap	<dev_id> <origin_id>
3841  *   delete		<dev_id>
3842  *   set_transaction_id <current_trans_id> <new_trans_id>
3843  *   reserve_metadata_snap
3844  *   release_metadata_snap
3845  */
3846 static int pool_message(struct dm_target *ti, unsigned int argc, char **argv,
3847 			char *result, unsigned int maxlen)
3848 {
3849 	int r = -EINVAL;
3850 	struct pool_c *pt = ti->private;
3851 	struct pool *pool = pt->pool;
3852 
3853 	if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE) {
3854 		DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3855 		      dm_device_name(pool->pool_md));
3856 		return -EOPNOTSUPP;
3857 	}
3858 
3859 	if (!strcasecmp(argv[0], "create_thin"))
3860 		r = process_create_thin_mesg(argc, argv, pool);
3861 
3862 	else if (!strcasecmp(argv[0], "create_snap"))
3863 		r = process_create_snap_mesg(argc, argv, pool);
3864 
3865 	else if (!strcasecmp(argv[0], "delete"))
3866 		r = process_delete_mesg(argc, argv, pool);
3867 
3868 	else if (!strcasecmp(argv[0], "set_transaction_id"))
3869 		r = process_set_transaction_id_mesg(argc, argv, pool);
3870 
3871 	else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3872 		r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3873 
3874 	else if (!strcasecmp(argv[0], "release_metadata_snap"))
3875 		r = process_release_metadata_snap_mesg(argc, argv, pool);
3876 
3877 	else
3878 		DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3879 
3880 	if (!r)
3881 		(void) commit(pool);
3882 
3883 	return r;
3884 }
3885 
3886 static void emit_flags(struct pool_features *pf, char *result,
3887 		       unsigned int sz, unsigned int maxlen)
3888 {
3889 	unsigned int count = !pf->zero_new_blocks + !pf->discard_enabled +
3890 		!pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3891 		pf->error_if_no_space;
3892 	DMEMIT("%u ", count);
3893 
3894 	if (!pf->zero_new_blocks)
3895 		DMEMIT("skip_block_zeroing ");
3896 
3897 	if (!pf->discard_enabled)
3898 		DMEMIT("ignore_discard ");
3899 
3900 	if (!pf->discard_passdown)
3901 		DMEMIT("no_discard_passdown ");
3902 
3903 	if (pf->mode == PM_READ_ONLY)
3904 		DMEMIT("read_only ");
3905 
3906 	if (pf->error_if_no_space)
3907 		DMEMIT("error_if_no_space ");
3908 }
3909 
3910 /*
3911  * Status line is:
3912  *    <transaction id> <used metadata sectors>/<total metadata sectors>
3913  *    <used data sectors>/<total data sectors> <held metadata root>
3914  *    <pool mode> <discard config> <no space config> <needs_check>
3915  */
3916 static void pool_status(struct dm_target *ti, status_type_t type,
3917 			unsigned int status_flags, char *result, unsigned int maxlen)
3918 {
3919 	int r;
3920 	unsigned int sz = 0;
3921 	uint64_t transaction_id;
3922 	dm_block_t nr_free_blocks_data;
3923 	dm_block_t nr_free_blocks_metadata;
3924 	dm_block_t nr_blocks_data;
3925 	dm_block_t nr_blocks_metadata;
3926 	dm_block_t held_root;
3927 	enum pool_mode mode;
3928 	char buf[BDEVNAME_SIZE];
3929 	char buf2[BDEVNAME_SIZE];
3930 	struct pool_c *pt = ti->private;
3931 	struct pool *pool = pt->pool;
3932 
3933 	switch (type) {
3934 	case STATUSTYPE_INFO:
3935 		if (get_pool_mode(pool) == PM_FAIL) {
3936 			DMEMIT("Fail");
3937 			break;
3938 		}
3939 
3940 		/* Commit to ensure statistics aren't out-of-date */
3941 		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3942 			(void) commit(pool);
3943 
3944 		r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3945 		if (r) {
3946 			DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3947 			      dm_device_name(pool->pool_md), r);
3948 			goto err;
3949 		}
3950 
3951 		r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3952 		if (r) {
3953 			DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3954 			      dm_device_name(pool->pool_md), r);
3955 			goto err;
3956 		}
3957 
3958 		r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3959 		if (r) {
3960 			DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3961 			      dm_device_name(pool->pool_md), r);
3962 			goto err;
3963 		}
3964 
3965 		r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3966 		if (r) {
3967 			DMERR("%s: dm_pool_get_free_block_count returned %d",
3968 			      dm_device_name(pool->pool_md), r);
3969 			goto err;
3970 		}
3971 
3972 		r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3973 		if (r) {
3974 			DMERR("%s: dm_pool_get_data_dev_size returned %d",
3975 			      dm_device_name(pool->pool_md), r);
3976 			goto err;
3977 		}
3978 
3979 		r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3980 		if (r) {
3981 			DMERR("%s: dm_pool_get_metadata_snap returned %d",
3982 			      dm_device_name(pool->pool_md), r);
3983 			goto err;
3984 		}
3985 
3986 		DMEMIT("%llu %llu/%llu %llu/%llu ",
3987 		       (unsigned long long)transaction_id,
3988 		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3989 		       (unsigned long long)nr_blocks_metadata,
3990 		       (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3991 		       (unsigned long long)nr_blocks_data);
3992 
3993 		if (held_root)
3994 			DMEMIT("%llu ", held_root);
3995 		else
3996 			DMEMIT("- ");
3997 
3998 		mode = get_pool_mode(pool);
3999 		if (mode == PM_OUT_OF_DATA_SPACE)
4000 			DMEMIT("out_of_data_space ");
4001 		else if (is_read_only_pool_mode(mode))
4002 			DMEMIT("ro ");
4003 		else
4004 			DMEMIT("rw ");
4005 
4006 		if (!pool->pf.discard_enabled)
4007 			DMEMIT("ignore_discard ");
4008 		else if (pool->pf.discard_passdown)
4009 			DMEMIT("discard_passdown ");
4010 		else
4011 			DMEMIT("no_discard_passdown ");
4012 
4013 		if (pool->pf.error_if_no_space)
4014 			DMEMIT("error_if_no_space ");
4015 		else
4016 			DMEMIT("queue_if_no_space ");
4017 
4018 		if (dm_pool_metadata_needs_check(pool->pmd))
4019 			DMEMIT("needs_check ");
4020 		else
4021 			DMEMIT("- ");
4022 
4023 		DMEMIT("%llu ", (unsigned long long)calc_metadata_threshold(pt));
4024 
4025 		break;
4026 
4027 	case STATUSTYPE_TABLE:
4028 		DMEMIT("%s %s %lu %llu ",
4029 		       format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
4030 		       format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
4031 		       (unsigned long)pool->sectors_per_block,
4032 		       (unsigned long long)pt->low_water_blocks);
4033 		emit_flags(&pt->requested_pf, result, sz, maxlen);
4034 		break;
4035 
4036 	case STATUSTYPE_IMA:
4037 		*result = '\0';
4038 		break;
4039 	}
4040 	return;
4041 
4042 err:
4043 	DMEMIT("Error");
4044 }
4045 
4046 static int pool_iterate_devices(struct dm_target *ti,
4047 				iterate_devices_callout_fn fn, void *data)
4048 {
4049 	struct pool_c *pt = ti->private;
4050 
4051 	return fn(ti, pt->data_dev, 0, ti->len, data);
4052 }
4053 
4054 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
4055 {
4056 	struct pool_c *pt = ti->private;
4057 	struct pool *pool = pt->pool;
4058 	sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
4059 
4060 	/*
4061 	 * If max_sectors is smaller than pool->sectors_per_block adjust it
4062 	 * to the highest possible power-of-2 factor of pool->sectors_per_block.
4063 	 * This is especially beneficial when the pool's data device is a RAID
4064 	 * device that has a full stripe width that matches pool->sectors_per_block
4065 	 * -- because even though partial RAID stripe-sized IOs will be issued to a
4066 	 *    single RAID stripe; when aggregated they will end on a full RAID stripe
4067 	 *    boundary.. which avoids additional partial RAID stripe writes cascading
4068 	 */
4069 	if (limits->max_sectors < pool->sectors_per_block) {
4070 		while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
4071 			if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
4072 				limits->max_sectors--;
4073 			limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
4074 		}
4075 	}
4076 
4077 	/*
4078 	 * If the system-determined stacked limits are compatible with the
4079 	 * pool's blocksize (io_opt is a factor) do not override them.
4080 	 */
4081 	if (io_opt_sectors < pool->sectors_per_block ||
4082 	    !is_factor(io_opt_sectors, pool->sectors_per_block)) {
4083 		if (is_factor(pool->sectors_per_block, limits->max_sectors))
4084 			blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
4085 		else
4086 			blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
4087 		blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
4088 	}
4089 
4090 	/*
4091 	 * pt->adjusted_pf is a staging area for the actual features to use.
4092 	 * They get transferred to the live pool in bind_control_target()
4093 	 * called from pool_preresume().
4094 	 */
4095 
4096 	if (pt->adjusted_pf.discard_enabled) {
4097 		disable_discard_passdown_if_not_supported(pt);
4098 		if (!pt->adjusted_pf.discard_passdown)
4099 			limits->max_discard_sectors = 0;
4100 		/*
4101 		 * The pool uses the same discard limits as the underlying data
4102 		 * device.  DM core has already set this up.
4103 		 */
4104 	} else {
4105 		/*
4106 		 * Must explicitly disallow stacking discard limits otherwise the
4107 		 * block layer will stack them if pool's data device has support.
4108 		 */
4109 		limits->discard_granularity = 0;
4110 	}
4111 }
4112 
4113 static struct target_type pool_target = {
4114 	.name = "thin-pool",
4115 	.features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4116 		    DM_TARGET_IMMUTABLE,
4117 	.version = {1, 23, 0},
4118 	.module = THIS_MODULE,
4119 	.ctr = pool_ctr,
4120 	.dtr = pool_dtr,
4121 	.map = pool_map,
4122 	.presuspend = pool_presuspend,
4123 	.presuspend_undo = pool_presuspend_undo,
4124 	.postsuspend = pool_postsuspend,
4125 	.preresume = pool_preresume,
4126 	.resume = pool_resume,
4127 	.message = pool_message,
4128 	.status = pool_status,
4129 	.iterate_devices = pool_iterate_devices,
4130 	.io_hints = pool_io_hints,
4131 };
4132 
4133 /*
4134  *--------------------------------------------------------------
4135  * Thin target methods
4136  *--------------------------------------------------------------
4137  */
4138 static void thin_get(struct thin_c *tc)
4139 {
4140 	refcount_inc(&tc->refcount);
4141 }
4142 
4143 static void thin_put(struct thin_c *tc)
4144 {
4145 	if (refcount_dec_and_test(&tc->refcount))
4146 		complete(&tc->can_destroy);
4147 }
4148 
4149 static void thin_dtr(struct dm_target *ti)
4150 {
4151 	struct thin_c *tc = ti->private;
4152 
4153 	spin_lock_irq(&tc->pool->lock);
4154 	list_del_rcu(&tc->list);
4155 	spin_unlock_irq(&tc->pool->lock);
4156 	synchronize_rcu();
4157 
4158 	thin_put(tc);
4159 	wait_for_completion(&tc->can_destroy);
4160 
4161 	mutex_lock(&dm_thin_pool_table.mutex);
4162 
4163 	__pool_dec(tc->pool);
4164 	dm_pool_close_thin_device(tc->td);
4165 	dm_put_device(ti, tc->pool_dev);
4166 	if (tc->origin_dev)
4167 		dm_put_device(ti, tc->origin_dev);
4168 	kfree(tc);
4169 
4170 	mutex_unlock(&dm_thin_pool_table.mutex);
4171 }
4172 
4173 /*
4174  * Thin target parameters:
4175  *
4176  * <pool_dev> <dev_id> [origin_dev]
4177  *
4178  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4179  * dev_id: the internal device identifier
4180  * origin_dev: a device external to the pool that should act as the origin
4181  *
4182  * If the pool device has discards disabled, they get disabled for the thin
4183  * device as well.
4184  */
4185 static int thin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
4186 {
4187 	int r;
4188 	struct thin_c *tc;
4189 	struct dm_dev *pool_dev, *origin_dev;
4190 	struct mapped_device *pool_md;
4191 
4192 	mutex_lock(&dm_thin_pool_table.mutex);
4193 
4194 	if (argc != 2 && argc != 3) {
4195 		ti->error = "Invalid argument count";
4196 		r = -EINVAL;
4197 		goto out_unlock;
4198 	}
4199 
4200 	tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4201 	if (!tc) {
4202 		ti->error = "Out of memory";
4203 		r = -ENOMEM;
4204 		goto out_unlock;
4205 	}
4206 	tc->thin_md = dm_table_get_md(ti->table);
4207 	spin_lock_init(&tc->lock);
4208 	INIT_LIST_HEAD(&tc->deferred_cells);
4209 	bio_list_init(&tc->deferred_bio_list);
4210 	bio_list_init(&tc->retry_on_resume_list);
4211 	tc->sort_bio_list = RB_ROOT;
4212 
4213 	if (argc == 3) {
4214 		if (!strcmp(argv[0], argv[2])) {
4215 			ti->error = "Error setting origin device";
4216 			r = -EINVAL;
4217 			goto bad_origin_dev;
4218 		}
4219 
4220 		r = dm_get_device(ti, argv[2], BLK_OPEN_READ, &origin_dev);
4221 		if (r) {
4222 			ti->error = "Error opening origin device";
4223 			goto bad_origin_dev;
4224 		}
4225 		tc->origin_dev = origin_dev;
4226 	}
4227 
4228 	r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4229 	if (r) {
4230 		ti->error = "Error opening pool device";
4231 		goto bad_pool_dev;
4232 	}
4233 	tc->pool_dev = pool_dev;
4234 
4235 	if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4236 		ti->error = "Invalid device id";
4237 		r = -EINVAL;
4238 		goto bad_common;
4239 	}
4240 
4241 	pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4242 	if (!pool_md) {
4243 		ti->error = "Couldn't get pool mapped device";
4244 		r = -EINVAL;
4245 		goto bad_common;
4246 	}
4247 
4248 	tc->pool = __pool_table_lookup(pool_md);
4249 	if (!tc->pool) {
4250 		ti->error = "Couldn't find pool object";
4251 		r = -EINVAL;
4252 		goto bad_pool_lookup;
4253 	}
4254 	__pool_inc(tc->pool);
4255 
4256 	if (get_pool_mode(tc->pool) == PM_FAIL) {
4257 		ti->error = "Couldn't open thin device, Pool is in fail mode";
4258 		r = -EINVAL;
4259 		goto bad_pool;
4260 	}
4261 
4262 	r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4263 	if (r) {
4264 		ti->error = "Couldn't open thin internal device";
4265 		goto bad_pool;
4266 	}
4267 
4268 	r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4269 	if (r)
4270 		goto bad;
4271 
4272 	ti->num_flush_bios = 1;
4273 	ti->limit_swap_bios = true;
4274 	ti->flush_supported = true;
4275 	ti->accounts_remapped_io = true;
4276 	ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4277 
4278 	/* In case the pool supports discards, pass them on. */
4279 	if (tc->pool->pf.discard_enabled) {
4280 		ti->discards_supported = true;
4281 		ti->num_discard_bios = 1;
4282 		ti->max_discard_granularity = true;
4283 	}
4284 
4285 	mutex_unlock(&dm_thin_pool_table.mutex);
4286 
4287 	spin_lock_irq(&tc->pool->lock);
4288 	if (tc->pool->suspended) {
4289 		spin_unlock_irq(&tc->pool->lock);
4290 		mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4291 		ti->error = "Unable to activate thin device while pool is suspended";
4292 		r = -EINVAL;
4293 		goto bad;
4294 	}
4295 	refcount_set(&tc->refcount, 1);
4296 	init_completion(&tc->can_destroy);
4297 	list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4298 	spin_unlock_irq(&tc->pool->lock);
4299 	/*
4300 	 * This synchronize_rcu() call is needed here otherwise we risk a
4301 	 * wake_worker() call finding no bios to process (because the newly
4302 	 * added tc isn't yet visible).  So this reduces latency since we
4303 	 * aren't then dependent on the periodic commit to wake_worker().
4304 	 */
4305 	synchronize_rcu();
4306 
4307 	dm_put(pool_md);
4308 
4309 	return 0;
4310 
4311 bad:
4312 	dm_pool_close_thin_device(tc->td);
4313 bad_pool:
4314 	__pool_dec(tc->pool);
4315 bad_pool_lookup:
4316 	dm_put(pool_md);
4317 bad_common:
4318 	dm_put_device(ti, tc->pool_dev);
4319 bad_pool_dev:
4320 	if (tc->origin_dev)
4321 		dm_put_device(ti, tc->origin_dev);
4322 bad_origin_dev:
4323 	kfree(tc);
4324 out_unlock:
4325 	mutex_unlock(&dm_thin_pool_table.mutex);
4326 
4327 	return r;
4328 }
4329 
4330 static int thin_map(struct dm_target *ti, struct bio *bio)
4331 {
4332 	bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4333 
4334 	return thin_bio_map(ti, bio);
4335 }
4336 
4337 static int thin_endio(struct dm_target *ti, struct bio *bio,
4338 		blk_status_t *err)
4339 {
4340 	unsigned long flags;
4341 	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4342 	struct list_head work;
4343 	struct dm_thin_new_mapping *m, *tmp;
4344 	struct pool *pool = h->tc->pool;
4345 
4346 	if (h->shared_read_entry) {
4347 		INIT_LIST_HEAD(&work);
4348 		dm_deferred_entry_dec(h->shared_read_entry, &work);
4349 
4350 		spin_lock_irqsave(&pool->lock, flags);
4351 		list_for_each_entry_safe(m, tmp, &work, list) {
4352 			list_del(&m->list);
4353 			__complete_mapping_preparation(m);
4354 		}
4355 		spin_unlock_irqrestore(&pool->lock, flags);
4356 	}
4357 
4358 	if (h->all_io_entry) {
4359 		INIT_LIST_HEAD(&work);
4360 		dm_deferred_entry_dec(h->all_io_entry, &work);
4361 		if (!list_empty(&work)) {
4362 			spin_lock_irqsave(&pool->lock, flags);
4363 			list_for_each_entry_safe(m, tmp, &work, list)
4364 				list_add_tail(&m->list, &pool->prepared_discards);
4365 			spin_unlock_irqrestore(&pool->lock, flags);
4366 			wake_worker(pool);
4367 		}
4368 	}
4369 
4370 	if (h->cell)
4371 		cell_defer_no_holder(h->tc, h->cell);
4372 
4373 	return DM_ENDIO_DONE;
4374 }
4375 
4376 static void thin_presuspend(struct dm_target *ti)
4377 {
4378 	struct thin_c *tc = ti->private;
4379 
4380 	if (dm_noflush_suspending(ti))
4381 		noflush_work(tc, do_noflush_start);
4382 }
4383 
4384 static void thin_postsuspend(struct dm_target *ti)
4385 {
4386 	struct thin_c *tc = ti->private;
4387 
4388 	/*
4389 	 * The dm_noflush_suspending flag has been cleared by now, so
4390 	 * unfortunately we must always run this.
4391 	 */
4392 	noflush_work(tc, do_noflush_stop);
4393 }
4394 
4395 static int thin_preresume(struct dm_target *ti)
4396 {
4397 	struct thin_c *tc = ti->private;
4398 
4399 	if (tc->origin_dev)
4400 		tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4401 
4402 	return 0;
4403 }
4404 
4405 /*
4406  * <nr mapped sectors> <highest mapped sector>
4407  */
4408 static void thin_status(struct dm_target *ti, status_type_t type,
4409 			unsigned int status_flags, char *result, unsigned int maxlen)
4410 {
4411 	int r;
4412 	ssize_t sz = 0;
4413 	dm_block_t mapped, highest;
4414 	char buf[BDEVNAME_SIZE];
4415 	struct thin_c *tc = ti->private;
4416 
4417 	if (get_pool_mode(tc->pool) == PM_FAIL) {
4418 		DMEMIT("Fail");
4419 		return;
4420 	}
4421 
4422 	if (!tc->td)
4423 		DMEMIT("-");
4424 	else {
4425 		switch (type) {
4426 		case STATUSTYPE_INFO:
4427 			r = dm_thin_get_mapped_count(tc->td, &mapped);
4428 			if (r) {
4429 				DMERR("dm_thin_get_mapped_count returned %d", r);
4430 				goto err;
4431 			}
4432 
4433 			r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4434 			if (r < 0) {
4435 				DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4436 				goto err;
4437 			}
4438 
4439 			DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4440 			if (r)
4441 				DMEMIT("%llu", ((highest + 1) *
4442 						tc->pool->sectors_per_block) - 1);
4443 			else
4444 				DMEMIT("-");
4445 			break;
4446 
4447 		case STATUSTYPE_TABLE:
4448 			DMEMIT("%s %lu",
4449 			       format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4450 			       (unsigned long) tc->dev_id);
4451 			if (tc->origin_dev)
4452 				DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4453 			break;
4454 
4455 		case STATUSTYPE_IMA:
4456 			*result = '\0';
4457 			break;
4458 		}
4459 	}
4460 
4461 	return;
4462 
4463 err:
4464 	DMEMIT("Error");
4465 }
4466 
4467 static int thin_iterate_devices(struct dm_target *ti,
4468 				iterate_devices_callout_fn fn, void *data)
4469 {
4470 	sector_t blocks;
4471 	struct thin_c *tc = ti->private;
4472 	struct pool *pool = tc->pool;
4473 
4474 	/*
4475 	 * We can't call dm_pool_get_data_dev_size() since that blocks.  So
4476 	 * we follow a more convoluted path through to the pool's target.
4477 	 */
4478 	if (!pool->ti)
4479 		return 0;	/* nothing is bound */
4480 
4481 	blocks = pool->ti->len;
4482 	(void) sector_div(blocks, pool->sectors_per_block);
4483 	if (blocks)
4484 		return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4485 
4486 	return 0;
4487 }
4488 
4489 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4490 {
4491 	struct thin_c *tc = ti->private;
4492 	struct pool *pool = tc->pool;
4493 
4494 	if (pool->pf.discard_enabled) {
4495 		limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4496 		limits->max_discard_sectors = pool->sectors_per_block * BIO_PRISON_MAX_RANGE;
4497 	}
4498 }
4499 
4500 static struct target_type thin_target = {
4501 	.name = "thin",
4502 	.version = {1, 23, 0},
4503 	.module	= THIS_MODULE,
4504 	.ctr = thin_ctr,
4505 	.dtr = thin_dtr,
4506 	.map = thin_map,
4507 	.end_io = thin_endio,
4508 	.preresume = thin_preresume,
4509 	.presuspend = thin_presuspend,
4510 	.postsuspend = thin_postsuspend,
4511 	.status = thin_status,
4512 	.iterate_devices = thin_iterate_devices,
4513 	.io_hints = thin_io_hints,
4514 };
4515 
4516 /*----------------------------------------------------------------*/
4517 
4518 static int __init dm_thin_init(void)
4519 {
4520 	int r = -ENOMEM;
4521 
4522 	pool_table_init();
4523 
4524 	_new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4525 	if (!_new_mapping_cache)
4526 		return r;
4527 
4528 	r = dm_register_target(&thin_target);
4529 	if (r)
4530 		goto bad_new_mapping_cache;
4531 
4532 	r = dm_register_target(&pool_target);
4533 	if (r)
4534 		goto bad_thin_target;
4535 
4536 	return 0;
4537 
4538 bad_thin_target:
4539 	dm_unregister_target(&thin_target);
4540 bad_new_mapping_cache:
4541 	kmem_cache_destroy(_new_mapping_cache);
4542 
4543 	return r;
4544 }
4545 
4546 static void dm_thin_exit(void)
4547 {
4548 	dm_unregister_target(&thin_target);
4549 	dm_unregister_target(&pool_target);
4550 
4551 	kmem_cache_destroy(_new_mapping_cache);
4552 
4553 	pool_table_exit();
4554 }
4555 
4556 module_init(dm_thin_init);
4557 module_exit(dm_thin_exit);
4558 
4559 module_param_named(no_space_timeout, no_space_timeout_secs, uint, 0644);
4560 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4561 
4562 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4563 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4564 MODULE_LICENSE("GPL");
4565