Loading...
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
36static unsigned int no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
37
38DECLARE_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 */
116enum lock_space {
117 VIRTUAL,
118 PHYSICAL
119};
120
121static 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
132static 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
138static 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
148struct throttle {
149 struct rw_semaphore lock;
150 unsigned long threshold;
151 bool throttle_applied;
152};
153
154static void throttle_init(struct throttle *t)
155{
156 init_rwsem(&t->lock);
157 t->throttle_applied = false;
158}
159
160static void throttle_work_start(struct throttle *t)
161{
162 t->threshold = jiffies + THROTTLE_THRESHOLD;
163}
164
165static 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
173static 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
181static void throttle_lock(struct throttle *t)
182{
183 down_read(&t->lock);
184}
185
186static 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 */
198struct dm_thin_new_mapping;
199
200/*
201 * The pool runs in various modes. Ordered in degraded order for comparisons.
202 */
203enum 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
216struct 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
225struct thin_c;
226typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
227typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
228typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
229
230#define CELL_SORT_ARRAY_SIZE 8192
231
232struct 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
290static void metadata_operation_failed(struct pool *pool, const char *op, int r);
291
292static enum pool_mode get_pool_mode(struct pool *pool)
293{
294 return pool->pf.mode;
295}
296
297static 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 */
325struct 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 */
339struct 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
367static bool block_size_is_power_of_two(struct pool *pool)
368{
369 return pool->sectors_per_block_shift >= 0;
370}
371
372static 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
381struct discard_op {
382 struct thin_c *tc;
383 struct blk_plug plug;
384 struct bio *parent_bio;
385 struct bio *bio;
386};
387
388static 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
398static 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
407static 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 */
436static void wake_worker(struct pool *pool)
437{
438 queue_work(pool->wq, &pool->worker);
439}
440
441/*----------------------------------------------------------------*/
442
443static 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
464 return r;
465}
466
467static void cell_release(struct pool *pool,
468 struct dm_bio_prison_cell *cell,
469 struct bio_list *bios)
470{
471 dm_cell_release(pool->prison, cell, bios);
472 dm_bio_prison_free_cell(pool->prison, cell);
473}
474
475static void cell_visit_release(struct pool *pool,
476 void (*fn)(void *, struct dm_bio_prison_cell *),
477 void *context,
478 struct dm_bio_prison_cell *cell)
479{
480 dm_cell_visit_release(pool->prison, fn, context, cell);
481 dm_bio_prison_free_cell(pool->prison, cell);
482}
483
484static void cell_release_no_holder(struct pool *pool,
485 struct dm_bio_prison_cell *cell,
486 struct bio_list *bios)
487{
488 dm_cell_release_no_holder(pool->prison, cell, bios);
489 dm_bio_prison_free_cell(pool->prison, cell);
490}
491
492static void cell_error_with_code(struct pool *pool,
493 struct dm_bio_prison_cell *cell, blk_status_t error_code)
494{
495 dm_cell_error(pool->prison, cell, error_code);
496 dm_bio_prison_free_cell(pool->prison, cell);
497}
498
499static blk_status_t get_pool_io_error_code(struct pool *pool)
500{
501 return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
502}
503
504static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
505{
506 cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
507}
508
509static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
510{
511 cell_error_with_code(pool, cell, 0);
512}
513
514static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
515{
516 cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
517}
518
519/*----------------------------------------------------------------*/
520
521/*
522 * A global list of pools that uses a struct mapped_device as a key.
523 */
524static struct dm_thin_pool_table {
525 struct mutex mutex;
526 struct list_head pools;
527} dm_thin_pool_table;
528
529static void pool_table_init(void)
530{
531 mutex_init(&dm_thin_pool_table.mutex);
532 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
533}
534
535static void pool_table_exit(void)
536{
537 mutex_destroy(&dm_thin_pool_table.mutex);
538}
539
540static void __pool_table_insert(struct pool *pool)
541{
542 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
543 list_add(&pool->list, &dm_thin_pool_table.pools);
544}
545
546static void __pool_table_remove(struct pool *pool)
547{
548 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
549 list_del(&pool->list);
550}
551
552static struct pool *__pool_table_lookup(struct mapped_device *md)
553{
554 struct pool *pool = NULL, *tmp;
555
556 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
557
558 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
559 if (tmp->pool_md == md) {
560 pool = tmp;
561 break;
562 }
563 }
564
565 return pool;
566}
567
568static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
569{
570 struct pool *pool = NULL, *tmp;
571
572 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
573
574 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
575 if (tmp->md_dev == md_dev) {
576 pool = tmp;
577 break;
578 }
579 }
580
581 return pool;
582}
583
584/*----------------------------------------------------------------*/
585
586struct dm_thin_endio_hook {
587 struct thin_c *tc;
588 struct dm_deferred_entry *shared_read_entry;
589 struct dm_deferred_entry *all_io_entry;
590 struct dm_thin_new_mapping *overwrite_mapping;
591 struct rb_node rb_node;
592 struct dm_bio_prison_cell *cell;
593};
594
595static void error_bio_list(struct bio_list *bios, blk_status_t error)
596{
597 struct bio *bio;
598
599 while ((bio = bio_list_pop(bios))) {
600 bio->bi_status = error;
601 bio_endio(bio);
602 }
603}
604
605static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
606 blk_status_t error)
607{
608 struct bio_list bios;
609
610 bio_list_init(&bios);
611
612 spin_lock_irq(&tc->lock);
613 bio_list_merge_init(&bios, master);
614 spin_unlock_irq(&tc->lock);
615
616 error_bio_list(&bios, error);
617}
618
619static void requeue_deferred_cells(struct thin_c *tc)
620{
621 struct pool *pool = tc->pool;
622 struct list_head cells;
623 struct dm_bio_prison_cell *cell, *tmp;
624
625 INIT_LIST_HEAD(&cells);
626
627 spin_lock_irq(&tc->lock);
628 list_splice_init(&tc->deferred_cells, &cells);
629 spin_unlock_irq(&tc->lock);
630
631 list_for_each_entry_safe(cell, tmp, &cells, user_list)
632 cell_requeue(pool, cell);
633}
634
635static void requeue_io(struct thin_c *tc)
636{
637 struct bio_list bios;
638
639 bio_list_init(&bios);
640
641 spin_lock_irq(&tc->lock);
642 bio_list_merge_init(&bios, &tc->deferred_bio_list);
643 bio_list_merge_init(&bios, &tc->retry_on_resume_list);
644 spin_unlock_irq(&tc->lock);
645
646 error_bio_list(&bios, BLK_STS_DM_REQUEUE);
647 requeue_deferred_cells(tc);
648}
649
650static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
651{
652 struct thin_c *tc;
653
654 rcu_read_lock();
655 list_for_each_entry_rcu(tc, &pool->active_thins, list)
656 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
657 rcu_read_unlock();
658}
659
660static void error_retry_list(struct pool *pool)
661{
662 error_retry_list_with_code(pool, get_pool_io_error_code(pool));
663}
664
665/*
666 * This section of code contains the logic for processing a thin device's IO.
667 * Much of the code depends on pool object resources (lists, workqueues, etc)
668 * but most is exclusively called from the thin target rather than the thin-pool
669 * target.
670 */
671
672static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
673{
674 struct pool *pool = tc->pool;
675 sector_t block_nr = bio->bi_iter.bi_sector;
676
677 if (block_size_is_power_of_two(pool))
678 block_nr >>= pool->sectors_per_block_shift;
679 else
680 (void) sector_div(block_nr, pool->sectors_per_block);
681
682 return block_nr;
683}
684
685/*
686 * Returns the _complete_ blocks that this bio covers.
687 */
688static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
689 dm_block_t *begin, dm_block_t *end)
690{
691 struct pool *pool = tc->pool;
692 sector_t b = bio->bi_iter.bi_sector;
693 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
694
695 b += pool->sectors_per_block - 1ull; /* so we round up */
696
697 if (block_size_is_power_of_two(pool)) {
698 b >>= pool->sectors_per_block_shift;
699 e >>= pool->sectors_per_block_shift;
700 } else {
701 (void) sector_div(b, pool->sectors_per_block);
702 (void) sector_div(e, pool->sectors_per_block);
703 }
704
705 if (e < b) {
706 /* Can happen if the bio is within a single block. */
707 e = b;
708 }
709
710 *begin = b;
711 *end = e;
712}
713
714static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
715{
716 struct pool *pool = tc->pool;
717 sector_t bi_sector = bio->bi_iter.bi_sector;
718
719 bio_set_dev(bio, tc->pool_dev->bdev);
720 if (block_size_is_power_of_two(pool)) {
721 bio->bi_iter.bi_sector =
722 (block << pool->sectors_per_block_shift) |
723 (bi_sector & (pool->sectors_per_block - 1));
724 } else {
725 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
726 sector_div(bi_sector, pool->sectors_per_block);
727 }
728}
729
730static void remap_to_origin(struct thin_c *tc, struct bio *bio)
731{
732 bio_set_dev(bio, tc->origin_dev->bdev);
733}
734
735static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
736{
737 return op_is_flush(bio->bi_opf) &&
738 dm_thin_changed_this_transaction(tc->td);
739}
740
741static void inc_all_io_entry(struct pool *pool, struct bio *bio)
742{
743 struct dm_thin_endio_hook *h;
744
745 if (bio_op(bio) == REQ_OP_DISCARD)
746 return;
747
748 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
749 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
750}
751
752static void issue(struct thin_c *tc, struct bio *bio)
753{
754 struct pool *pool = tc->pool;
755
756 if (!bio_triggers_commit(tc, bio)) {
757 dm_submit_bio_remap(bio, NULL);
758 return;
759 }
760
761 /*
762 * Complete bio with an error if earlier I/O caused changes to
763 * the metadata that can't be committed e.g, due to I/O errors
764 * on the metadata device.
765 */
766 if (dm_thin_aborted_changes(tc->td)) {
767 bio_io_error(bio);
768 return;
769 }
770
771 /*
772 * Batch together any bios that trigger commits and then issue a
773 * single commit for them in process_deferred_bios().
774 */
775 spin_lock_irq(&pool->lock);
776 bio_list_add(&pool->deferred_flush_bios, bio);
777 spin_unlock_irq(&pool->lock);
778}
779
780static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
781{
782 remap_to_origin(tc, bio);
783 issue(tc, bio);
784}
785
786static void remap_and_issue(struct thin_c *tc, struct bio *bio,
787 dm_block_t block)
788{
789 remap(tc, bio, block);
790 issue(tc, bio);
791}
792
793/*----------------------------------------------------------------*/
794
795/*
796 * Bio endio functions.
797 */
798struct dm_thin_new_mapping {
799 struct list_head list;
800
801 bool pass_discard:1;
802 bool maybe_shared:1;
803
804 /*
805 * Track quiescing, copying and zeroing preparation actions. When this
806 * counter hits zero the block is prepared and can be inserted into the
807 * btree.
808 */
809 atomic_t prepare_actions;
810
811 blk_status_t status;
812 struct thin_c *tc;
813 dm_block_t virt_begin, virt_end;
814 dm_block_t data_block;
815 struct dm_bio_prison_cell *cell;
816
817 /*
818 * If the bio covers the whole area of a block then we can avoid
819 * zeroing or copying. Instead this bio is hooked. The bio will
820 * still be in the cell, so care has to be taken to avoid issuing
821 * the bio twice.
822 */
823 struct bio *bio;
824 bio_end_io_t *saved_bi_end_io;
825};
826
827static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
828{
829 struct pool *pool = m->tc->pool;
830
831 if (atomic_dec_and_test(&m->prepare_actions)) {
832 list_add_tail(&m->list, &pool->prepared_mappings);
833 wake_worker(pool);
834 }
835}
836
837static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
838{
839 unsigned long flags;
840 struct pool *pool = m->tc->pool;
841
842 spin_lock_irqsave(&pool->lock, flags);
843 __complete_mapping_preparation(m);
844 spin_unlock_irqrestore(&pool->lock, flags);
845}
846
847static void copy_complete(int read_err, unsigned long write_err, void *context)
848{
849 struct dm_thin_new_mapping *m = context;
850
851 m->status = read_err || write_err ? BLK_STS_IOERR : 0;
852 complete_mapping_preparation(m);
853}
854
855static void overwrite_endio(struct bio *bio)
856{
857 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
858 struct dm_thin_new_mapping *m = h->overwrite_mapping;
859
860 bio->bi_end_io = m->saved_bi_end_io;
861
862 m->status = bio->bi_status;
863 complete_mapping_preparation(m);
864}
865
866/*----------------------------------------------------------------*/
867
868/*
869 * Workqueue.
870 */
871
872/*
873 * Prepared mapping jobs.
874 */
875
876/*
877 * This sends the bios in the cell, except the original holder, back
878 * to the deferred_bios list.
879 */
880static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
881{
882 struct pool *pool = tc->pool;
883 unsigned long flags;
884 struct bio_list bios;
885
886 bio_list_init(&bios);
887 cell_release_no_holder(pool, cell, &bios);
888
889 if (!bio_list_empty(&bios)) {
890 spin_lock_irqsave(&tc->lock, flags);
891 bio_list_merge(&tc->deferred_bio_list, &bios);
892 spin_unlock_irqrestore(&tc->lock, flags);
893 wake_worker(pool);
894 }
895}
896
897static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
898
899struct remap_info {
900 struct thin_c *tc;
901 struct bio_list defer_bios;
902 struct bio_list issue_bios;
903};
904
905static void __inc_remap_and_issue_cell(void *context,
906 struct dm_bio_prison_cell *cell)
907{
908 struct remap_info *info = context;
909 struct bio *bio;
910
911 while ((bio = bio_list_pop(&cell->bios))) {
912 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
913 bio_list_add(&info->defer_bios, bio);
914 else {
915 inc_all_io_entry(info->tc->pool, bio);
916
917 /*
918 * We can't issue the bios with the bio prison lock
919 * held, so we add them to a list to issue on
920 * return from this function.
921 */
922 bio_list_add(&info->issue_bios, bio);
923 }
924 }
925}
926
927static void inc_remap_and_issue_cell(struct thin_c *tc,
928 struct dm_bio_prison_cell *cell,
929 dm_block_t block)
930{
931 struct bio *bio;
932 struct remap_info info;
933
934 info.tc = tc;
935 bio_list_init(&info.defer_bios);
936 bio_list_init(&info.issue_bios);
937
938 /*
939 * We have to be careful to inc any bios we're about to issue
940 * before the cell is released, and avoid a race with new bios
941 * being added to the cell.
942 */
943 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
944 &info, cell);
945
946 while ((bio = bio_list_pop(&info.defer_bios)))
947 thin_defer_bio(tc, bio);
948
949 while ((bio = bio_list_pop(&info.issue_bios)))
950 remap_and_issue(info.tc, bio, block);
951}
952
953static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
954{
955 cell_error(m->tc->pool, m->cell);
956 list_del(&m->list);
957 mempool_free(m, &m->tc->pool->mapping_pool);
958}
959
960static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
961{
962 struct pool *pool = tc->pool;
963
964 /*
965 * If the bio has the REQ_FUA flag set we must commit the metadata
966 * before signaling its completion.
967 */
968 if (!bio_triggers_commit(tc, bio)) {
969 bio_endio(bio);
970 return;
971 }
972
973 /*
974 * Complete bio with an error if earlier I/O caused changes to the
975 * metadata that can't be committed, e.g, due to I/O errors on the
976 * metadata device.
977 */
978 if (dm_thin_aborted_changes(tc->td)) {
979 bio_io_error(bio);
980 return;
981 }
982
983 /*
984 * Batch together any bios that trigger commits and then issue a
985 * single commit for them in process_deferred_bios().
986 */
987 spin_lock_irq(&pool->lock);
988 bio_list_add(&pool->deferred_flush_completions, bio);
989 spin_unlock_irq(&pool->lock);
990}
991
992static void process_prepared_mapping(struct dm_thin_new_mapping *m)
993{
994 struct thin_c *tc = m->tc;
995 struct pool *pool = tc->pool;
996 struct bio *bio = m->bio;
997 int r;
998
999 if (m->status) {
1000 cell_error(pool, m->cell);
1001 goto out;
1002 }
1003
1004 /*
1005 * Commit the prepared block into the mapping btree.
1006 * Any I/O for this block arriving after this point will get
1007 * remapped to it directly.
1008 */
1009 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1010 if (r) {
1011 metadata_operation_failed(pool, "dm_thin_insert_block", r);
1012 cell_error(pool, m->cell);
1013 goto out;
1014 }
1015
1016 /*
1017 * Release any bios held while the block was being provisioned.
1018 * If we are processing a write bio that completely covers the block,
1019 * we already processed it so can ignore it now when processing
1020 * the bios in the cell.
1021 */
1022 if (bio) {
1023 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1024 complete_overwrite_bio(tc, bio);
1025 } else {
1026 inc_all_io_entry(tc->pool, m->cell->holder);
1027 remap_and_issue(tc, m->cell->holder, m->data_block);
1028 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1029 }
1030
1031out:
1032 list_del(&m->list);
1033 mempool_free(m, &pool->mapping_pool);
1034}
1035
1036/*----------------------------------------------------------------*/
1037
1038static void free_discard_mapping(struct dm_thin_new_mapping *m)
1039{
1040 struct thin_c *tc = m->tc;
1041
1042 if (m->cell)
1043 cell_defer_no_holder(tc, m->cell);
1044 mempool_free(m, &tc->pool->mapping_pool);
1045}
1046
1047static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1048{
1049 bio_io_error(m->bio);
1050 free_discard_mapping(m);
1051}
1052
1053static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1054{
1055 bio_endio(m->bio);
1056 free_discard_mapping(m);
1057}
1058
1059static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1060{
1061 int r;
1062 struct thin_c *tc = m->tc;
1063
1064 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1065 if (r) {
1066 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1067 bio_io_error(m->bio);
1068 } else
1069 bio_endio(m->bio);
1070
1071 cell_defer_no_holder(tc, m->cell);
1072 mempool_free(m, &tc->pool->mapping_pool);
1073}
1074
1075/*----------------------------------------------------------------*/
1076
1077static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1078 struct bio *discard_parent)
1079{
1080 /*
1081 * We've already unmapped this range of blocks, but before we
1082 * passdown we have to check that these blocks are now unused.
1083 */
1084 int r = 0;
1085 bool shared = true;
1086 struct thin_c *tc = m->tc;
1087 struct pool *pool = tc->pool;
1088 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1089 struct discard_op op;
1090
1091 begin_discard(&op, tc, discard_parent);
1092 while (b != end) {
1093 /* find start of unmapped run */
1094 for (; b < end; b++) {
1095 r = dm_pool_block_is_shared(pool->pmd, b, &shared);
1096 if (r)
1097 goto out;
1098
1099 if (!shared)
1100 break;
1101 }
1102
1103 if (b == end)
1104 break;
1105
1106 /* find end of run */
1107 for (e = b + 1; e != end; e++) {
1108 r = dm_pool_block_is_shared(pool->pmd, e, &shared);
1109 if (r)
1110 goto out;
1111
1112 if (shared)
1113 break;
1114 }
1115
1116 r = issue_discard(&op, b, e);
1117 if (r)
1118 goto out;
1119
1120 b = e;
1121 }
1122out:
1123 end_discard(&op, r);
1124}
1125
1126static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1127{
1128 unsigned long flags;
1129 struct pool *pool = m->tc->pool;
1130
1131 spin_lock_irqsave(&pool->lock, flags);
1132 list_add_tail(&m->list, &pool->prepared_discards_pt2);
1133 spin_unlock_irqrestore(&pool->lock, flags);
1134 wake_worker(pool);
1135}
1136
1137static void passdown_endio(struct bio *bio)
1138{
1139 /*
1140 * It doesn't matter if the passdown discard failed, we still want
1141 * to unmap (we ignore err).
1142 */
1143 queue_passdown_pt2(bio->bi_private);
1144 bio_put(bio);
1145}
1146
1147static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1148{
1149 int r;
1150 struct thin_c *tc = m->tc;
1151 struct pool *pool = tc->pool;
1152 struct bio *discard_parent;
1153 dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1154
1155 /*
1156 * Only this thread allocates blocks, so we can be sure that the
1157 * newly unmapped blocks will not be allocated before the end of
1158 * the function.
1159 */
1160 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1161 if (r) {
1162 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1163 bio_io_error(m->bio);
1164 cell_defer_no_holder(tc, m->cell);
1165 mempool_free(m, &pool->mapping_pool);
1166 return;
1167 }
1168
1169 /*
1170 * Increment the unmapped blocks. This prevents a race between the
1171 * passdown io and reallocation of freed blocks.
1172 */
1173 r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1174 if (r) {
1175 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1176 bio_io_error(m->bio);
1177 cell_defer_no_holder(tc, m->cell);
1178 mempool_free(m, &pool->mapping_pool);
1179 return;
1180 }
1181
1182 discard_parent = bio_alloc(NULL, 1, 0, GFP_NOIO);
1183 discard_parent->bi_end_io = passdown_endio;
1184 discard_parent->bi_private = m;
1185 if (m->maybe_shared)
1186 passdown_double_checking_shared_status(m, discard_parent);
1187 else {
1188 struct discard_op op;
1189
1190 begin_discard(&op, tc, discard_parent);
1191 r = issue_discard(&op, m->data_block, data_end);
1192 end_discard(&op, r);
1193 }
1194}
1195
1196static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1197{
1198 int r;
1199 struct thin_c *tc = m->tc;
1200 struct pool *pool = tc->pool;
1201
1202 /*
1203 * The passdown has completed, so now we can decrement all those
1204 * unmapped blocks.
1205 */
1206 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1207 m->data_block + (m->virt_end - m->virt_begin));
1208 if (r) {
1209 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1210 bio_io_error(m->bio);
1211 } else
1212 bio_endio(m->bio);
1213
1214 cell_defer_no_holder(tc, m->cell);
1215 mempool_free(m, &pool->mapping_pool);
1216}
1217
1218static void process_prepared(struct pool *pool, struct list_head *head,
1219 process_mapping_fn *fn)
1220{
1221 struct list_head maps;
1222 struct dm_thin_new_mapping *m, *tmp;
1223
1224 INIT_LIST_HEAD(&maps);
1225 spin_lock_irq(&pool->lock);
1226 list_splice_init(head, &maps);
1227 spin_unlock_irq(&pool->lock);
1228
1229 list_for_each_entry_safe(m, tmp, &maps, list)
1230 (*fn)(m);
1231}
1232
1233/*
1234 * Deferred bio jobs.
1235 */
1236static int io_overlaps_block(struct pool *pool, struct bio *bio)
1237{
1238 return bio->bi_iter.bi_size ==
1239 (pool->sectors_per_block << SECTOR_SHIFT);
1240}
1241
1242static int io_overwrites_block(struct pool *pool, struct bio *bio)
1243{
1244 return (bio_data_dir(bio) == WRITE) &&
1245 io_overlaps_block(pool, bio);
1246}
1247
1248static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1249 bio_end_io_t *fn)
1250{
1251 *save = bio->bi_end_io;
1252 bio->bi_end_io = fn;
1253}
1254
1255static int ensure_next_mapping(struct pool *pool)
1256{
1257 if (pool->next_mapping)
1258 return 0;
1259
1260 pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1261
1262 return pool->next_mapping ? 0 : -ENOMEM;
1263}
1264
1265static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1266{
1267 struct dm_thin_new_mapping *m = pool->next_mapping;
1268
1269 BUG_ON(!pool->next_mapping);
1270
1271 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1272 INIT_LIST_HEAD(&m->list);
1273 m->bio = NULL;
1274
1275 pool->next_mapping = NULL;
1276
1277 return m;
1278}
1279
1280static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1281 sector_t begin, sector_t end)
1282{
1283 struct dm_io_region to;
1284
1285 to.bdev = tc->pool_dev->bdev;
1286 to.sector = begin;
1287 to.count = end - begin;
1288
1289 dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1290}
1291
1292static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1293 dm_block_t data_begin,
1294 struct dm_thin_new_mapping *m)
1295{
1296 struct pool *pool = tc->pool;
1297 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1298
1299 h->overwrite_mapping = m;
1300 m->bio = bio;
1301 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1302 inc_all_io_entry(pool, bio);
1303 remap_and_issue(tc, bio, data_begin);
1304}
1305
1306/*
1307 * A partial copy also needs to zero the uncopied region.
1308 */
1309static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1310 struct dm_dev *origin, dm_block_t data_origin,
1311 dm_block_t data_dest,
1312 struct dm_bio_prison_cell *cell, struct bio *bio,
1313 sector_t len)
1314{
1315 struct pool *pool = tc->pool;
1316 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1317
1318 m->tc = tc;
1319 m->virt_begin = virt_block;
1320 m->virt_end = virt_block + 1u;
1321 m->data_block = data_dest;
1322 m->cell = cell;
1323
1324 /*
1325 * quiesce action + copy action + an extra reference held for the
1326 * duration of this function (we may need to inc later for a
1327 * partial zero).
1328 */
1329 atomic_set(&m->prepare_actions, 3);
1330
1331 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1332 complete_mapping_preparation(m); /* already quiesced */
1333
1334 /*
1335 * IO to pool_dev remaps to the pool target's data_dev.
1336 *
1337 * If the whole block of data is being overwritten, we can issue the
1338 * bio immediately. Otherwise we use kcopyd to clone the data first.
1339 */
1340 if (io_overwrites_block(pool, bio))
1341 remap_and_issue_overwrite(tc, bio, data_dest, m);
1342 else {
1343 struct dm_io_region from, to;
1344
1345 from.bdev = origin->bdev;
1346 from.sector = data_origin * pool->sectors_per_block;
1347 from.count = len;
1348
1349 to.bdev = tc->pool_dev->bdev;
1350 to.sector = data_dest * pool->sectors_per_block;
1351 to.count = len;
1352
1353 dm_kcopyd_copy(pool->copier, &from, 1, &to,
1354 0, copy_complete, m);
1355
1356 /*
1357 * Do we need to zero a tail region?
1358 */
1359 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1360 atomic_inc(&m->prepare_actions);
1361 ll_zero(tc, m,
1362 data_dest * pool->sectors_per_block + len,
1363 (data_dest + 1) * pool->sectors_per_block);
1364 }
1365 }
1366
1367 complete_mapping_preparation(m); /* drop our ref */
1368}
1369
1370static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1371 dm_block_t data_origin, dm_block_t data_dest,
1372 struct dm_bio_prison_cell *cell, struct bio *bio)
1373{
1374 schedule_copy(tc, virt_block, tc->pool_dev,
1375 data_origin, data_dest, cell, bio,
1376 tc->pool->sectors_per_block);
1377}
1378
1379static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1380 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1381 struct bio *bio)
1382{
1383 struct pool *pool = tc->pool;
1384 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1385
1386 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1387 m->tc = tc;
1388 m->virt_begin = virt_block;
1389 m->virt_end = virt_block + 1u;
1390 m->data_block = data_block;
1391 m->cell = cell;
1392
1393 /*
1394 * If the whole block of data is being overwritten or we are not
1395 * zeroing pre-existing data, we can issue the bio immediately.
1396 * Otherwise we use kcopyd to zero the data first.
1397 */
1398 if (pool->pf.zero_new_blocks) {
1399 if (io_overwrites_block(pool, bio))
1400 remap_and_issue_overwrite(tc, bio, data_block, m);
1401 else {
1402 ll_zero(tc, m, data_block * pool->sectors_per_block,
1403 (data_block + 1) * pool->sectors_per_block);
1404 }
1405 } else
1406 process_prepared_mapping(m);
1407}
1408
1409static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1410 dm_block_t data_dest,
1411 struct dm_bio_prison_cell *cell, struct bio *bio)
1412{
1413 struct pool *pool = tc->pool;
1414 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1415 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1416
1417 if (virt_block_end <= tc->origin_size) {
1418 schedule_copy(tc, virt_block, tc->origin_dev,
1419 virt_block, data_dest, cell, bio,
1420 pool->sectors_per_block);
1421
1422 } else if (virt_block_begin < tc->origin_size) {
1423 schedule_copy(tc, virt_block, tc->origin_dev,
1424 virt_block, data_dest, cell, bio,
1425 tc->origin_size - virt_block_begin);
1426
1427 } else
1428 schedule_zero(tc, virt_block, data_dest, cell, bio);
1429}
1430
1431static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1432
1433static void requeue_bios(struct pool *pool);
1434
1435static bool is_read_only_pool_mode(enum pool_mode mode)
1436{
1437 return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1438}
1439
1440static bool is_read_only(struct pool *pool)
1441{
1442 return is_read_only_pool_mode(get_pool_mode(pool));
1443}
1444
1445static void check_for_metadata_space(struct pool *pool)
1446{
1447 int r;
1448 const char *ooms_reason = NULL;
1449 dm_block_t nr_free;
1450
1451 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1452 if (r)
1453 ooms_reason = "Could not get free metadata blocks";
1454 else if (!nr_free)
1455 ooms_reason = "No free metadata blocks";
1456
1457 if (ooms_reason && !is_read_only(pool)) {
1458 DMERR("%s", ooms_reason);
1459 set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1460 }
1461}
1462
1463static void check_for_data_space(struct pool *pool)
1464{
1465 int r;
1466 dm_block_t nr_free;
1467
1468 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1469 return;
1470
1471 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1472 if (r)
1473 return;
1474
1475 if (nr_free) {
1476 set_pool_mode(pool, PM_WRITE);
1477 requeue_bios(pool);
1478 }
1479}
1480
1481/*
1482 * A non-zero return indicates read_only or fail_io mode.
1483 * Many callers don't care about the return value.
1484 */
1485static int commit(struct pool *pool)
1486{
1487 int r;
1488
1489 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1490 return -EINVAL;
1491
1492 r = dm_pool_commit_metadata(pool->pmd);
1493 if (r)
1494 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1495 else {
1496 check_for_metadata_space(pool);
1497 check_for_data_space(pool);
1498 }
1499
1500 return r;
1501}
1502
1503static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1504{
1505 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1506 DMWARN("%s: reached low water mark for data device: sending event.",
1507 dm_device_name(pool->pool_md));
1508 spin_lock_irq(&pool->lock);
1509 pool->low_water_triggered = true;
1510 spin_unlock_irq(&pool->lock);
1511 dm_table_event(pool->ti->table);
1512 }
1513}
1514
1515static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1516{
1517 int r;
1518 dm_block_t free_blocks;
1519 struct pool *pool = tc->pool;
1520
1521 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1522 return -EINVAL;
1523
1524 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1525 if (r) {
1526 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1527 return r;
1528 }
1529
1530 check_low_water_mark(pool, free_blocks);
1531
1532 if (!free_blocks) {
1533 /*
1534 * Try to commit to see if that will free up some
1535 * more space.
1536 */
1537 r = commit(pool);
1538 if (r)
1539 return r;
1540
1541 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1542 if (r) {
1543 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1544 return r;
1545 }
1546
1547 if (!free_blocks) {
1548 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1549 return -ENOSPC;
1550 }
1551 }
1552
1553 r = dm_pool_alloc_data_block(pool->pmd, result);
1554 if (r) {
1555 if (r == -ENOSPC)
1556 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1557 else
1558 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1559 return r;
1560 }
1561
1562 r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1563 if (r) {
1564 metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1565 return r;
1566 }
1567
1568 if (!free_blocks) {
1569 /* Let's commit before we use up the metadata reserve. */
1570 r = commit(pool);
1571 if (r)
1572 return r;
1573 }
1574
1575 return 0;
1576}
1577
1578/*
1579 * If we have run out of space, queue bios until the device is
1580 * resumed, presumably after having been reloaded with more space.
1581 */
1582static void retry_on_resume(struct bio *bio)
1583{
1584 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1585 struct thin_c *tc = h->tc;
1586
1587 spin_lock_irq(&tc->lock);
1588 bio_list_add(&tc->retry_on_resume_list, bio);
1589 spin_unlock_irq(&tc->lock);
1590}
1591
1592static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1593{
1594 enum pool_mode m = get_pool_mode(pool);
1595
1596 switch (m) {
1597 case PM_WRITE:
1598 /* Shouldn't get here */
1599 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1600 return BLK_STS_IOERR;
1601
1602 case PM_OUT_OF_DATA_SPACE:
1603 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1604
1605 case PM_OUT_OF_METADATA_SPACE:
1606 case PM_READ_ONLY:
1607 case PM_FAIL:
1608 return BLK_STS_IOERR;
1609 default:
1610 /* Shouldn't get here */
1611 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1612 return BLK_STS_IOERR;
1613 }
1614}
1615
1616static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1617{
1618 blk_status_t error = should_error_unserviceable_bio(pool);
1619
1620 if (error) {
1621 bio->bi_status = error;
1622 bio_endio(bio);
1623 } else
1624 retry_on_resume(bio);
1625}
1626
1627static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1628{
1629 struct bio *bio;
1630 struct bio_list bios;
1631 blk_status_t error;
1632
1633 error = should_error_unserviceable_bio(pool);
1634 if (error) {
1635 cell_error_with_code(pool, cell, error);
1636 return;
1637 }
1638
1639 bio_list_init(&bios);
1640 cell_release(pool, cell, &bios);
1641
1642 while ((bio = bio_list_pop(&bios)))
1643 retry_on_resume(bio);
1644}
1645
1646static void process_discard_cell_no_passdown(struct thin_c *tc,
1647 struct dm_bio_prison_cell *virt_cell)
1648{
1649 struct pool *pool = tc->pool;
1650 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1651
1652 /*
1653 * We don't need to lock the data blocks, since there's no
1654 * passdown. We only lock data blocks for allocation and breaking sharing.
1655 */
1656 m->tc = tc;
1657 m->virt_begin = virt_cell->key.block_begin;
1658 m->virt_end = virt_cell->key.block_end;
1659 m->cell = virt_cell;
1660 m->bio = virt_cell->holder;
1661
1662 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1663 pool->process_prepared_discard(m);
1664}
1665
1666static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1667 struct bio *bio)
1668{
1669 struct pool *pool = tc->pool;
1670
1671 int r;
1672 bool maybe_shared;
1673 struct dm_cell_key data_key;
1674 struct dm_bio_prison_cell *data_cell;
1675 struct dm_thin_new_mapping *m;
1676 dm_block_t virt_begin, virt_end, data_begin, data_end;
1677 dm_block_t len, next_boundary;
1678
1679 while (begin != end) {
1680 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1681 &data_begin, &maybe_shared);
1682 if (r) {
1683 /*
1684 * Silently fail, letting any mappings we've
1685 * created complete.
1686 */
1687 break;
1688 }
1689
1690 data_end = data_begin + (virt_end - virt_begin);
1691
1692 /*
1693 * Make sure the data region obeys the bio prison restrictions.
1694 */
1695 while (data_begin < data_end) {
1696 r = ensure_next_mapping(pool);
1697 if (r)
1698 return; /* we did our best */
1699
1700 next_boundary = ((data_begin >> BIO_PRISON_MAX_RANGE_SHIFT) + 1)
1701 << BIO_PRISON_MAX_RANGE_SHIFT;
1702 len = min_t(sector_t, data_end - data_begin, next_boundary - data_begin);
1703
1704 /* This key is certainly within range given the above splitting */
1705 (void) build_key(tc->td, PHYSICAL, data_begin, data_begin + len, &data_key);
1706 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1707 /* contention, we'll give up with this range */
1708 data_begin += len;
1709 continue;
1710 }
1711
1712 /*
1713 * IO may still be going to the destination block. We must
1714 * quiesce before we can do the removal.
1715 */
1716 m = get_next_mapping(pool);
1717 m->tc = tc;
1718 m->maybe_shared = maybe_shared;
1719 m->virt_begin = virt_begin;
1720 m->virt_end = virt_begin + len;
1721 m->data_block = data_begin;
1722 m->cell = data_cell;
1723 m->bio = bio;
1724
1725 /*
1726 * The parent bio must not complete before sub discard bios are
1727 * chained to it (see end_discard's bio_chain)!
1728 *
1729 * This per-mapping bi_remaining increment is paired with
1730 * the implicit decrement that occurs via bio_endio() in
1731 * end_discard().
1732 */
1733 bio_inc_remaining(bio);
1734 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1735 pool->process_prepared_discard(m);
1736
1737 virt_begin += len;
1738 data_begin += len;
1739 }
1740
1741 begin = virt_end;
1742 }
1743}
1744
1745static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1746{
1747 struct bio *bio = virt_cell->holder;
1748 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1749
1750 /*
1751 * The virt_cell will only get freed once the origin bio completes.
1752 * This means it will remain locked while all the individual
1753 * passdown bios are in flight.
1754 */
1755 h->cell = virt_cell;
1756 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1757
1758 /*
1759 * We complete the bio now, knowing that the bi_remaining field
1760 * will prevent completion until the sub range discards have
1761 * completed.
1762 */
1763 bio_endio(bio);
1764}
1765
1766static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1767{
1768 dm_block_t begin, end;
1769 struct dm_cell_key virt_key;
1770 struct dm_bio_prison_cell *virt_cell;
1771
1772 get_bio_block_range(tc, bio, &begin, &end);
1773 if (begin == end) {
1774 /*
1775 * The discard covers less than a block.
1776 */
1777 bio_endio(bio);
1778 return;
1779 }
1780
1781 if (unlikely(!build_key(tc->td, VIRTUAL, begin, end, &virt_key))) {
1782 DMERR_LIMIT("Discard doesn't respect bio prison limits");
1783 bio_endio(bio);
1784 return;
1785 }
1786
1787 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell)) {
1788 /*
1789 * Potential starvation issue: We're relying on the
1790 * fs/application being well behaved, and not trying to
1791 * send IO to a region at the same time as discarding it.
1792 * If they do this persistently then it's possible this
1793 * cell will never be granted.
1794 */
1795 return;
1796 }
1797
1798 tc->pool->process_discard_cell(tc, virt_cell);
1799}
1800
1801static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1802 struct dm_cell_key *key,
1803 struct dm_thin_lookup_result *lookup_result,
1804 struct dm_bio_prison_cell *cell)
1805{
1806 int r;
1807 dm_block_t data_block;
1808 struct pool *pool = tc->pool;
1809
1810 r = alloc_data_block(tc, &data_block);
1811 switch (r) {
1812 case 0:
1813 schedule_internal_copy(tc, block, lookup_result->block,
1814 data_block, cell, bio);
1815 break;
1816
1817 case -ENOSPC:
1818 retry_bios_on_resume(pool, cell);
1819 break;
1820
1821 default:
1822 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1823 __func__, r);
1824 cell_error(pool, cell);
1825 break;
1826 }
1827}
1828
1829static void __remap_and_issue_shared_cell(void *context,
1830 struct dm_bio_prison_cell *cell)
1831{
1832 struct remap_info *info = context;
1833 struct bio *bio;
1834
1835 while ((bio = bio_list_pop(&cell->bios))) {
1836 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1837 bio_op(bio) == REQ_OP_DISCARD)
1838 bio_list_add(&info->defer_bios, bio);
1839 else {
1840 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1841
1842 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1843 inc_all_io_entry(info->tc->pool, bio);
1844 bio_list_add(&info->issue_bios, bio);
1845 }
1846 }
1847}
1848
1849static void remap_and_issue_shared_cell(struct thin_c *tc,
1850 struct dm_bio_prison_cell *cell,
1851 dm_block_t block)
1852{
1853 struct bio *bio;
1854 struct remap_info info;
1855
1856 info.tc = tc;
1857 bio_list_init(&info.defer_bios);
1858 bio_list_init(&info.issue_bios);
1859
1860 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1861 &info, cell);
1862
1863 while ((bio = bio_list_pop(&info.defer_bios)))
1864 thin_defer_bio(tc, bio);
1865
1866 while ((bio = bio_list_pop(&info.issue_bios)))
1867 remap_and_issue(tc, bio, block);
1868}
1869
1870static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1871 dm_block_t block,
1872 struct dm_thin_lookup_result *lookup_result,
1873 struct dm_bio_prison_cell *virt_cell)
1874{
1875 struct dm_bio_prison_cell *data_cell;
1876 struct pool *pool = tc->pool;
1877 struct dm_cell_key key;
1878
1879 /*
1880 * If cell is already occupied, then sharing is already in the process
1881 * of being broken so we have nothing further to do here.
1882 */
1883 build_data_key(tc->td, lookup_result->block, &key);
1884 if (bio_detain(pool, &key, bio, &data_cell)) {
1885 cell_defer_no_holder(tc, virt_cell);
1886 return;
1887 }
1888
1889 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1890 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1891 cell_defer_no_holder(tc, virt_cell);
1892 } else {
1893 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1894
1895 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1896 inc_all_io_entry(pool, bio);
1897 remap_and_issue(tc, bio, lookup_result->block);
1898
1899 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1900 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1901 }
1902}
1903
1904static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1905 struct dm_bio_prison_cell *cell)
1906{
1907 int r;
1908 dm_block_t data_block;
1909 struct pool *pool = tc->pool;
1910
1911 /*
1912 * Remap empty bios (flushes) immediately, without provisioning.
1913 */
1914 if (!bio->bi_iter.bi_size) {
1915 inc_all_io_entry(pool, bio);
1916 cell_defer_no_holder(tc, cell);
1917
1918 remap_and_issue(tc, bio, 0);
1919 return;
1920 }
1921
1922 /*
1923 * Fill read bios with zeroes and complete them immediately.
1924 */
1925 if (bio_data_dir(bio) == READ) {
1926 zero_fill_bio(bio);
1927 cell_defer_no_holder(tc, cell);
1928 bio_endio(bio);
1929 return;
1930 }
1931
1932 r = alloc_data_block(tc, &data_block);
1933 switch (r) {
1934 case 0:
1935 if (tc->origin_dev)
1936 schedule_external_copy(tc, block, data_block, cell, bio);
1937 else
1938 schedule_zero(tc, block, data_block, cell, bio);
1939 break;
1940
1941 case -ENOSPC:
1942 retry_bios_on_resume(pool, cell);
1943 break;
1944
1945 default:
1946 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1947 __func__, r);
1948 cell_error(pool, cell);
1949 break;
1950 }
1951}
1952
1953static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1954{
1955 int r;
1956 struct pool *pool = tc->pool;
1957 struct bio *bio = cell->holder;
1958 dm_block_t block = get_bio_block(tc, bio);
1959 struct dm_thin_lookup_result lookup_result;
1960
1961 if (tc->requeue_mode) {
1962 cell_requeue(pool, cell);
1963 return;
1964 }
1965
1966 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1967 switch (r) {
1968 case 0:
1969 if (lookup_result.shared)
1970 process_shared_bio(tc, bio, block, &lookup_result, cell);
1971 else {
1972 inc_all_io_entry(pool, bio);
1973 remap_and_issue(tc, bio, lookup_result.block);
1974 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1975 }
1976 break;
1977
1978 case -ENODATA:
1979 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1980 inc_all_io_entry(pool, bio);
1981 cell_defer_no_holder(tc, cell);
1982
1983 if (bio_end_sector(bio) <= tc->origin_size)
1984 remap_to_origin_and_issue(tc, bio);
1985
1986 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1987 zero_fill_bio(bio);
1988 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1989 remap_to_origin_and_issue(tc, bio);
1990
1991 } else {
1992 zero_fill_bio(bio);
1993 bio_endio(bio);
1994 }
1995 } else
1996 provision_block(tc, bio, block, cell);
1997 break;
1998
1999 default:
2000 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2001 __func__, r);
2002 cell_defer_no_holder(tc, cell);
2003 bio_io_error(bio);
2004 break;
2005 }
2006}
2007
2008static void process_bio(struct thin_c *tc, struct bio *bio)
2009{
2010 struct pool *pool = tc->pool;
2011 dm_block_t block = get_bio_block(tc, bio);
2012 struct dm_bio_prison_cell *cell;
2013 struct dm_cell_key key;
2014
2015 /*
2016 * If cell is already occupied, then the block is already
2017 * being provisioned so we have nothing further to do here.
2018 */
2019 build_virtual_key(tc->td, block, &key);
2020 if (bio_detain(pool, &key, bio, &cell))
2021 return;
2022
2023 process_cell(tc, cell);
2024}
2025
2026static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2027 struct dm_bio_prison_cell *cell)
2028{
2029 int r;
2030 int rw = bio_data_dir(bio);
2031 dm_block_t block = get_bio_block(tc, bio);
2032 struct dm_thin_lookup_result lookup_result;
2033
2034 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2035 switch (r) {
2036 case 0:
2037 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2038 handle_unserviceable_bio(tc->pool, bio);
2039 if (cell)
2040 cell_defer_no_holder(tc, cell);
2041 } else {
2042 inc_all_io_entry(tc->pool, bio);
2043 remap_and_issue(tc, bio, lookup_result.block);
2044 if (cell)
2045 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2046 }
2047 break;
2048
2049 case -ENODATA:
2050 if (cell)
2051 cell_defer_no_holder(tc, cell);
2052 if (rw != READ) {
2053 handle_unserviceable_bio(tc->pool, bio);
2054 break;
2055 }
2056
2057 if (tc->origin_dev) {
2058 inc_all_io_entry(tc->pool, bio);
2059 remap_to_origin_and_issue(tc, bio);
2060 break;
2061 }
2062
2063 zero_fill_bio(bio);
2064 bio_endio(bio);
2065 break;
2066
2067 default:
2068 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2069 __func__, r);
2070 if (cell)
2071 cell_defer_no_holder(tc, cell);
2072 bio_io_error(bio);
2073 break;
2074 }
2075}
2076
2077static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2078{
2079 __process_bio_read_only(tc, bio, NULL);
2080}
2081
2082static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2083{
2084 __process_bio_read_only(tc, cell->holder, cell);
2085}
2086
2087static void process_bio_success(struct thin_c *tc, struct bio *bio)
2088{
2089 bio_endio(bio);
2090}
2091
2092static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2093{
2094 bio_io_error(bio);
2095}
2096
2097static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2098{
2099 cell_success(tc->pool, cell);
2100}
2101
2102static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2103{
2104 cell_error(tc->pool, cell);
2105}
2106
2107/*
2108 * FIXME: should we also commit due to size of transaction, measured in
2109 * metadata blocks?
2110 */
2111static int need_commit_due_to_time(struct pool *pool)
2112{
2113 return !time_in_range(jiffies, pool->last_commit_jiffies,
2114 pool->last_commit_jiffies + COMMIT_PERIOD);
2115}
2116
2117#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2118#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2119
2120static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2121{
2122 struct rb_node **rbp, *parent;
2123 struct dm_thin_endio_hook *pbd;
2124 sector_t bi_sector = bio->bi_iter.bi_sector;
2125
2126 rbp = &tc->sort_bio_list.rb_node;
2127 parent = NULL;
2128 while (*rbp) {
2129 parent = *rbp;
2130 pbd = thin_pbd(parent);
2131
2132 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2133 rbp = &(*rbp)->rb_left;
2134 else
2135 rbp = &(*rbp)->rb_right;
2136 }
2137
2138 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2139 rb_link_node(&pbd->rb_node, parent, rbp);
2140 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2141}
2142
2143static void __extract_sorted_bios(struct thin_c *tc)
2144{
2145 struct rb_node *node;
2146 struct dm_thin_endio_hook *pbd;
2147 struct bio *bio;
2148
2149 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2150 pbd = thin_pbd(node);
2151 bio = thin_bio(pbd);
2152
2153 bio_list_add(&tc->deferred_bio_list, bio);
2154 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2155 }
2156
2157 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2158}
2159
2160static void __sort_thin_deferred_bios(struct thin_c *tc)
2161{
2162 struct bio *bio;
2163 struct bio_list bios;
2164
2165 bio_list_init(&bios);
2166 bio_list_merge(&bios, &tc->deferred_bio_list);
2167 bio_list_init(&tc->deferred_bio_list);
2168
2169 /* Sort deferred_bio_list using rb-tree */
2170 while ((bio = bio_list_pop(&bios)))
2171 __thin_bio_rb_add(tc, bio);
2172
2173 /*
2174 * Transfer the sorted bios in sort_bio_list back to
2175 * deferred_bio_list to allow lockless submission of
2176 * all bios.
2177 */
2178 __extract_sorted_bios(tc);
2179}
2180
2181static void process_thin_deferred_bios(struct thin_c *tc)
2182{
2183 struct pool *pool = tc->pool;
2184 struct bio *bio;
2185 struct bio_list bios;
2186 struct blk_plug plug;
2187 unsigned int count = 0;
2188
2189 if (tc->requeue_mode) {
2190 error_thin_bio_list(tc, &tc->deferred_bio_list,
2191 BLK_STS_DM_REQUEUE);
2192 return;
2193 }
2194
2195 bio_list_init(&bios);
2196
2197 spin_lock_irq(&tc->lock);
2198
2199 if (bio_list_empty(&tc->deferred_bio_list)) {
2200 spin_unlock_irq(&tc->lock);
2201 return;
2202 }
2203
2204 __sort_thin_deferred_bios(tc);
2205
2206 bio_list_merge(&bios, &tc->deferred_bio_list);
2207 bio_list_init(&tc->deferred_bio_list);
2208
2209 spin_unlock_irq(&tc->lock);
2210
2211 blk_start_plug(&plug);
2212 while ((bio = bio_list_pop(&bios))) {
2213 /*
2214 * If we've got no free new_mapping structs, and processing
2215 * this bio might require one, we pause until there are some
2216 * prepared mappings to process.
2217 */
2218 if (ensure_next_mapping(pool)) {
2219 spin_lock_irq(&tc->lock);
2220 bio_list_add(&tc->deferred_bio_list, bio);
2221 bio_list_merge(&tc->deferred_bio_list, &bios);
2222 spin_unlock_irq(&tc->lock);
2223 break;
2224 }
2225
2226 if (bio_op(bio) == REQ_OP_DISCARD)
2227 pool->process_discard(tc, bio);
2228 else
2229 pool->process_bio(tc, bio);
2230
2231 if ((count++ & 127) == 0) {
2232 throttle_work_update(&pool->throttle);
2233 dm_pool_issue_prefetches(pool->pmd);
2234 }
2235 cond_resched();
2236 }
2237 blk_finish_plug(&plug);
2238}
2239
2240static int cmp_cells(const void *lhs, const void *rhs)
2241{
2242 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2243 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2244
2245 BUG_ON(!lhs_cell->holder);
2246 BUG_ON(!rhs_cell->holder);
2247
2248 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2249 return -1;
2250
2251 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2252 return 1;
2253
2254 return 0;
2255}
2256
2257static unsigned int sort_cells(struct pool *pool, struct list_head *cells)
2258{
2259 unsigned int count = 0;
2260 struct dm_bio_prison_cell *cell, *tmp;
2261
2262 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2263 if (count >= CELL_SORT_ARRAY_SIZE)
2264 break;
2265
2266 pool->cell_sort_array[count++] = cell;
2267 list_del(&cell->user_list);
2268 }
2269
2270 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2271
2272 return count;
2273}
2274
2275static void process_thin_deferred_cells(struct thin_c *tc)
2276{
2277 struct pool *pool = tc->pool;
2278 struct list_head cells;
2279 struct dm_bio_prison_cell *cell;
2280 unsigned int i, j, count;
2281
2282 INIT_LIST_HEAD(&cells);
2283
2284 spin_lock_irq(&tc->lock);
2285 list_splice_init(&tc->deferred_cells, &cells);
2286 spin_unlock_irq(&tc->lock);
2287
2288 if (list_empty(&cells))
2289 return;
2290
2291 do {
2292 count = sort_cells(tc->pool, &cells);
2293
2294 for (i = 0; i < count; i++) {
2295 cell = pool->cell_sort_array[i];
2296 BUG_ON(!cell->holder);
2297
2298 /*
2299 * If we've got no free new_mapping structs, and processing
2300 * this bio might require one, we pause until there are some
2301 * prepared mappings to process.
2302 */
2303 if (ensure_next_mapping(pool)) {
2304 for (j = i; j < count; j++)
2305 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2306
2307 spin_lock_irq(&tc->lock);
2308 list_splice(&cells, &tc->deferred_cells);
2309 spin_unlock_irq(&tc->lock);
2310 return;
2311 }
2312
2313 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2314 pool->process_discard_cell(tc, cell);
2315 else
2316 pool->process_cell(tc, cell);
2317 }
2318 cond_resched();
2319 } while (!list_empty(&cells));
2320}
2321
2322static void thin_get(struct thin_c *tc);
2323static void thin_put(struct thin_c *tc);
2324
2325/*
2326 * We can't hold rcu_read_lock() around code that can block. So we
2327 * find a thin with the rcu lock held; bump a refcount; then drop
2328 * the lock.
2329 */
2330static struct thin_c *get_first_thin(struct pool *pool)
2331{
2332 struct thin_c *tc = NULL;
2333
2334 rcu_read_lock();
2335 tc = list_first_or_null_rcu(&pool->active_thins, struct thin_c, list);
2336 if (tc)
2337 thin_get(tc);
2338 rcu_read_unlock();
2339
2340 return tc;
2341}
2342
2343static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2344{
2345 struct thin_c *old_tc = tc;
2346
2347 rcu_read_lock();
2348 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2349 thin_get(tc);
2350 thin_put(old_tc);
2351 rcu_read_unlock();
2352 return tc;
2353 }
2354 thin_put(old_tc);
2355 rcu_read_unlock();
2356
2357 return NULL;
2358}
2359
2360static void process_deferred_bios(struct pool *pool)
2361{
2362 struct bio *bio;
2363 struct bio_list bios, bio_completions;
2364 struct thin_c *tc;
2365
2366 tc = get_first_thin(pool);
2367 while (tc) {
2368 process_thin_deferred_cells(tc);
2369 process_thin_deferred_bios(tc);
2370 tc = get_next_thin(pool, tc);
2371 }
2372
2373 /*
2374 * If there are any deferred flush bios, we must commit the metadata
2375 * before issuing them or signaling their completion.
2376 */
2377 bio_list_init(&bios);
2378 bio_list_init(&bio_completions);
2379
2380 spin_lock_irq(&pool->lock);
2381 bio_list_merge(&bios, &pool->deferred_flush_bios);
2382 bio_list_init(&pool->deferred_flush_bios);
2383
2384 bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2385 bio_list_init(&pool->deferred_flush_completions);
2386 spin_unlock_irq(&pool->lock);
2387
2388 if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2389 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2390 return;
2391
2392 if (commit(pool)) {
2393 bio_list_merge(&bios, &bio_completions);
2394
2395 while ((bio = bio_list_pop(&bios)))
2396 bio_io_error(bio);
2397 return;
2398 }
2399 pool->last_commit_jiffies = jiffies;
2400
2401 while ((bio = bio_list_pop(&bio_completions)))
2402 bio_endio(bio);
2403
2404 while ((bio = bio_list_pop(&bios))) {
2405 /*
2406 * The data device was flushed as part of metadata commit,
2407 * so complete redundant flushes immediately.
2408 */
2409 if (bio->bi_opf & REQ_PREFLUSH)
2410 bio_endio(bio);
2411 else
2412 dm_submit_bio_remap(bio, NULL);
2413 }
2414}
2415
2416static void do_worker(struct work_struct *ws)
2417{
2418 struct pool *pool = container_of(ws, struct pool, worker);
2419
2420 throttle_work_start(&pool->throttle);
2421 dm_pool_issue_prefetches(pool->pmd);
2422 throttle_work_update(&pool->throttle);
2423 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2424 throttle_work_update(&pool->throttle);
2425 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2426 throttle_work_update(&pool->throttle);
2427 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2428 throttle_work_update(&pool->throttle);
2429 process_deferred_bios(pool);
2430 throttle_work_complete(&pool->throttle);
2431}
2432
2433/*
2434 * We want to commit periodically so that not too much
2435 * unwritten data builds up.
2436 */
2437static void do_waker(struct work_struct *ws)
2438{
2439 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2440
2441 wake_worker(pool);
2442 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2443}
2444
2445/*
2446 * We're holding onto IO to allow userland time to react. After the
2447 * timeout either the pool will have been resized (and thus back in
2448 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2449 */
2450static void do_no_space_timeout(struct work_struct *ws)
2451{
2452 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2453 no_space_timeout);
2454
2455 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2456 pool->pf.error_if_no_space = true;
2457 notify_of_pool_mode_change(pool);
2458 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2459 }
2460}
2461
2462/*----------------------------------------------------------------*/
2463
2464struct pool_work {
2465 struct work_struct worker;
2466 struct completion complete;
2467};
2468
2469static struct pool_work *to_pool_work(struct work_struct *ws)
2470{
2471 return container_of(ws, struct pool_work, worker);
2472}
2473
2474static void pool_work_complete(struct pool_work *pw)
2475{
2476 complete(&pw->complete);
2477}
2478
2479static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2480 void (*fn)(struct work_struct *))
2481{
2482 INIT_WORK_ONSTACK(&pw->worker, fn);
2483 init_completion(&pw->complete);
2484 queue_work(pool->wq, &pw->worker);
2485 wait_for_completion(&pw->complete);
2486 destroy_work_on_stack(&pw->worker);
2487}
2488
2489/*----------------------------------------------------------------*/
2490
2491struct noflush_work {
2492 struct pool_work pw;
2493 struct thin_c *tc;
2494};
2495
2496static struct noflush_work *to_noflush(struct work_struct *ws)
2497{
2498 return container_of(to_pool_work(ws), struct noflush_work, pw);
2499}
2500
2501static void do_noflush_start(struct work_struct *ws)
2502{
2503 struct noflush_work *w = to_noflush(ws);
2504
2505 w->tc->requeue_mode = true;
2506 requeue_io(w->tc);
2507 pool_work_complete(&w->pw);
2508}
2509
2510static void do_noflush_stop(struct work_struct *ws)
2511{
2512 struct noflush_work *w = to_noflush(ws);
2513
2514 w->tc->requeue_mode = false;
2515 pool_work_complete(&w->pw);
2516}
2517
2518static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2519{
2520 struct noflush_work w;
2521
2522 w.tc = tc;
2523 pool_work_wait(&w.pw, tc->pool, fn);
2524}
2525
2526/*----------------------------------------------------------------*/
2527
2528static void set_discard_callbacks(struct pool *pool)
2529{
2530 struct pool_c *pt = pool->ti->private;
2531
2532 if (pt->adjusted_pf.discard_passdown) {
2533 pool->process_discard_cell = process_discard_cell_passdown;
2534 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2535 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2536 } else {
2537 pool->process_discard_cell = process_discard_cell_no_passdown;
2538 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2539 }
2540}
2541
2542static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2543{
2544 struct pool_c *pt = pool->ti->private;
2545 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2546 enum pool_mode old_mode = get_pool_mode(pool);
2547 unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2548
2549 /*
2550 * Never allow the pool to transition to PM_WRITE mode if user
2551 * intervention is required to verify metadata and data consistency.
2552 */
2553 if (new_mode == PM_WRITE && needs_check) {
2554 DMERR("%s: unable to switch pool to write mode until repaired.",
2555 dm_device_name(pool->pool_md));
2556 if (old_mode != new_mode)
2557 new_mode = old_mode;
2558 else
2559 new_mode = PM_READ_ONLY;
2560 }
2561 /*
2562 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2563 * not going to recover without a thin_repair. So we never let the
2564 * pool move out of the old mode.
2565 */
2566 if (old_mode == PM_FAIL)
2567 new_mode = old_mode;
2568
2569 switch (new_mode) {
2570 case PM_FAIL:
2571 dm_pool_metadata_read_only(pool->pmd);
2572 pool->process_bio = process_bio_fail;
2573 pool->process_discard = process_bio_fail;
2574 pool->process_cell = process_cell_fail;
2575 pool->process_discard_cell = process_cell_fail;
2576 pool->process_prepared_mapping = process_prepared_mapping_fail;
2577 pool->process_prepared_discard = process_prepared_discard_fail;
2578
2579 error_retry_list(pool);
2580 break;
2581
2582 case PM_OUT_OF_METADATA_SPACE:
2583 case PM_READ_ONLY:
2584 dm_pool_metadata_read_only(pool->pmd);
2585 pool->process_bio = process_bio_read_only;
2586 pool->process_discard = process_bio_success;
2587 pool->process_cell = process_cell_read_only;
2588 pool->process_discard_cell = process_cell_success;
2589 pool->process_prepared_mapping = process_prepared_mapping_fail;
2590 pool->process_prepared_discard = process_prepared_discard_success;
2591
2592 error_retry_list(pool);
2593 break;
2594
2595 case PM_OUT_OF_DATA_SPACE:
2596 /*
2597 * Ideally we'd never hit this state; the low water mark
2598 * would trigger userland to extend the pool before we
2599 * completely run out of data space. However, many small
2600 * IOs to unprovisioned space can consume data space at an
2601 * alarming rate. Adjust your low water mark if you're
2602 * frequently seeing this mode.
2603 */
2604 pool->out_of_data_space = true;
2605 pool->process_bio = process_bio_read_only;
2606 pool->process_discard = process_discard_bio;
2607 pool->process_cell = process_cell_read_only;
2608 pool->process_prepared_mapping = process_prepared_mapping;
2609 set_discard_callbacks(pool);
2610
2611 if (!pool->pf.error_if_no_space && no_space_timeout)
2612 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2613 break;
2614
2615 case PM_WRITE:
2616 if (old_mode == PM_OUT_OF_DATA_SPACE)
2617 cancel_delayed_work_sync(&pool->no_space_timeout);
2618 pool->out_of_data_space = false;
2619 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2620 dm_pool_metadata_read_write(pool->pmd);
2621 pool->process_bio = process_bio;
2622 pool->process_discard = process_discard_bio;
2623 pool->process_cell = process_cell;
2624 pool->process_prepared_mapping = process_prepared_mapping;
2625 set_discard_callbacks(pool);
2626 break;
2627 }
2628
2629 pool->pf.mode = new_mode;
2630 /*
2631 * The pool mode may have changed, sync it so bind_control_target()
2632 * doesn't cause an unexpected mode transition on resume.
2633 */
2634 pt->adjusted_pf.mode = new_mode;
2635
2636 if (old_mode != new_mode)
2637 notify_of_pool_mode_change(pool);
2638}
2639
2640static void abort_transaction(struct pool *pool)
2641{
2642 const char *dev_name = dm_device_name(pool->pool_md);
2643
2644 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2645 if (dm_pool_abort_metadata(pool->pmd)) {
2646 DMERR("%s: failed to abort metadata transaction", dev_name);
2647 set_pool_mode(pool, PM_FAIL);
2648 }
2649
2650 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2651 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2652 set_pool_mode(pool, PM_FAIL);
2653 }
2654}
2655
2656static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2657{
2658 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2659 dm_device_name(pool->pool_md), op, r);
2660
2661 abort_transaction(pool);
2662 set_pool_mode(pool, PM_READ_ONLY);
2663}
2664
2665/*----------------------------------------------------------------*/
2666
2667/*
2668 * Mapping functions.
2669 */
2670
2671/*
2672 * Called only while mapping a thin bio to hand it over to the workqueue.
2673 */
2674static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2675{
2676 struct pool *pool = tc->pool;
2677
2678 spin_lock_irq(&tc->lock);
2679 bio_list_add(&tc->deferred_bio_list, bio);
2680 spin_unlock_irq(&tc->lock);
2681
2682 wake_worker(pool);
2683}
2684
2685static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2686{
2687 struct pool *pool = tc->pool;
2688
2689 throttle_lock(&pool->throttle);
2690 thin_defer_bio(tc, bio);
2691 throttle_unlock(&pool->throttle);
2692}
2693
2694static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2695{
2696 struct pool *pool = tc->pool;
2697
2698 throttle_lock(&pool->throttle);
2699 spin_lock_irq(&tc->lock);
2700 list_add_tail(&cell->user_list, &tc->deferred_cells);
2701 spin_unlock_irq(&tc->lock);
2702 throttle_unlock(&pool->throttle);
2703
2704 wake_worker(pool);
2705}
2706
2707static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2708{
2709 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2710
2711 h->tc = tc;
2712 h->shared_read_entry = NULL;
2713 h->all_io_entry = NULL;
2714 h->overwrite_mapping = NULL;
2715 h->cell = NULL;
2716}
2717
2718/*
2719 * Non-blocking function called from the thin target's map function.
2720 */
2721static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2722{
2723 int r;
2724 struct thin_c *tc = ti->private;
2725 dm_block_t block = get_bio_block(tc, bio);
2726 struct dm_thin_device *td = tc->td;
2727 struct dm_thin_lookup_result result;
2728 struct dm_bio_prison_cell *virt_cell, *data_cell;
2729 struct dm_cell_key key;
2730
2731 thin_hook_bio(tc, bio);
2732
2733 if (tc->requeue_mode) {
2734 bio->bi_status = BLK_STS_DM_REQUEUE;
2735 bio_endio(bio);
2736 return DM_MAPIO_SUBMITTED;
2737 }
2738
2739 if (get_pool_mode(tc->pool) == PM_FAIL) {
2740 bio_io_error(bio);
2741 return DM_MAPIO_SUBMITTED;
2742 }
2743
2744 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2745 thin_defer_bio_with_throttle(tc, bio);
2746 return DM_MAPIO_SUBMITTED;
2747 }
2748
2749 /*
2750 * We must hold the virtual cell before doing the lookup, otherwise
2751 * there's a race with discard.
2752 */
2753 build_virtual_key(tc->td, block, &key);
2754 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2755 return DM_MAPIO_SUBMITTED;
2756
2757 r = dm_thin_find_block(td, block, 0, &result);
2758
2759 /*
2760 * Note that we defer readahead too.
2761 */
2762 switch (r) {
2763 case 0:
2764 if (unlikely(result.shared)) {
2765 /*
2766 * We have a race condition here between the
2767 * result.shared value returned by the lookup and
2768 * snapshot creation, which may cause new
2769 * sharing.
2770 *
2771 * To avoid this always quiesce the origin before
2772 * taking the snap. You want to do this anyway to
2773 * ensure a consistent application view
2774 * (i.e. lockfs).
2775 *
2776 * More distant ancestors are irrelevant. The
2777 * shared flag will be set in their case.
2778 */
2779 thin_defer_cell(tc, virt_cell);
2780 return DM_MAPIO_SUBMITTED;
2781 }
2782
2783 build_data_key(tc->td, result.block, &key);
2784 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2785 cell_defer_no_holder(tc, virt_cell);
2786 return DM_MAPIO_SUBMITTED;
2787 }
2788
2789 inc_all_io_entry(tc->pool, bio);
2790 cell_defer_no_holder(tc, data_cell);
2791 cell_defer_no_holder(tc, virt_cell);
2792
2793 remap(tc, bio, result.block);
2794 return DM_MAPIO_REMAPPED;
2795
2796 case -ENODATA:
2797 case -EWOULDBLOCK:
2798 thin_defer_cell(tc, virt_cell);
2799 return DM_MAPIO_SUBMITTED;
2800
2801 default:
2802 /*
2803 * Must always call bio_io_error on failure.
2804 * dm_thin_find_block can fail with -EINVAL if the
2805 * pool is switched to fail-io mode.
2806 */
2807 bio_io_error(bio);
2808 cell_defer_no_holder(tc, virt_cell);
2809 return DM_MAPIO_SUBMITTED;
2810 }
2811}
2812
2813static void requeue_bios(struct pool *pool)
2814{
2815 struct thin_c *tc;
2816
2817 rcu_read_lock();
2818 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2819 spin_lock_irq(&tc->lock);
2820 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2821 bio_list_init(&tc->retry_on_resume_list);
2822 spin_unlock_irq(&tc->lock);
2823 }
2824 rcu_read_unlock();
2825}
2826
2827/*
2828 *--------------------------------------------------------------
2829 * Binding of control targets to a pool object
2830 *--------------------------------------------------------------
2831 */
2832static bool is_factor(sector_t block_size, uint32_t n)
2833{
2834 return !sector_div(block_size, n);
2835}
2836
2837/*
2838 * If discard_passdown was enabled verify that the data device
2839 * supports discards. Disable discard_passdown if not.
2840 */
2841static void disable_discard_passdown_if_not_supported(struct pool_c *pt)
2842{
2843 struct pool *pool = pt->pool;
2844 struct block_device *data_bdev = pt->data_dev->bdev;
2845 struct queue_limits *data_limits = bdev_limits(data_bdev);
2846 const char *reason = NULL;
2847
2848 if (!pt->adjusted_pf.discard_passdown)
2849 return;
2850
2851 if (!bdev_max_discard_sectors(pt->data_dev->bdev))
2852 reason = "discard unsupported";
2853
2854 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2855 reason = "max discard sectors smaller than a block";
2856
2857 if (reason) {
2858 DMWARN("Data device (%pg) %s: Disabling discard passdown.", data_bdev, reason);
2859 pt->adjusted_pf.discard_passdown = false;
2860 }
2861}
2862
2863static int bind_control_target(struct pool *pool, struct dm_target *ti)
2864{
2865 struct pool_c *pt = ti->private;
2866
2867 /*
2868 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2869 */
2870 enum pool_mode old_mode = get_pool_mode(pool);
2871 enum pool_mode new_mode = pt->adjusted_pf.mode;
2872
2873 /*
2874 * Don't change the pool's mode until set_pool_mode() below.
2875 * Otherwise the pool's process_* function pointers may
2876 * not match the desired pool mode.
2877 */
2878 pt->adjusted_pf.mode = old_mode;
2879
2880 pool->ti = ti;
2881 pool->pf = pt->adjusted_pf;
2882 pool->low_water_blocks = pt->low_water_blocks;
2883
2884 set_pool_mode(pool, new_mode);
2885
2886 return 0;
2887}
2888
2889static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2890{
2891 if (pool->ti == ti)
2892 pool->ti = NULL;
2893}
2894
2895/*
2896 *--------------------------------------------------------------
2897 * Pool creation
2898 *--------------------------------------------------------------
2899 */
2900/* Initialize pool features. */
2901static void pool_features_init(struct pool_features *pf)
2902{
2903 pf->mode = PM_WRITE;
2904 pf->zero_new_blocks = true;
2905 pf->discard_enabled = true;
2906 pf->discard_passdown = true;
2907 pf->error_if_no_space = false;
2908}
2909
2910static void __pool_destroy(struct pool *pool)
2911{
2912 __pool_table_remove(pool);
2913
2914 vfree(pool->cell_sort_array);
2915 if (dm_pool_metadata_close(pool->pmd) < 0)
2916 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2917
2918 dm_bio_prison_destroy(pool->prison);
2919 dm_kcopyd_client_destroy(pool->copier);
2920
2921 cancel_delayed_work_sync(&pool->waker);
2922 cancel_delayed_work_sync(&pool->no_space_timeout);
2923 if (pool->wq)
2924 destroy_workqueue(pool->wq);
2925
2926 if (pool->next_mapping)
2927 mempool_free(pool->next_mapping, &pool->mapping_pool);
2928 mempool_exit(&pool->mapping_pool);
2929 dm_deferred_set_destroy(pool->shared_read_ds);
2930 dm_deferred_set_destroy(pool->all_io_ds);
2931 kfree(pool);
2932}
2933
2934static struct kmem_cache *_new_mapping_cache;
2935
2936static struct pool *pool_create(struct mapped_device *pool_md,
2937 struct block_device *metadata_dev,
2938 struct block_device *data_dev,
2939 unsigned long block_size,
2940 int read_only, char **error)
2941{
2942 int r;
2943 void *err_p;
2944 struct pool *pool;
2945 struct dm_pool_metadata *pmd;
2946 bool format_device = read_only ? false : true;
2947
2948 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2949 if (IS_ERR(pmd)) {
2950 *error = "Error creating metadata object";
2951 return ERR_CAST(pmd);
2952 }
2953
2954 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2955 if (!pool) {
2956 *error = "Error allocating memory for pool";
2957 err_p = ERR_PTR(-ENOMEM);
2958 goto bad_pool;
2959 }
2960
2961 pool->pmd = pmd;
2962 pool->sectors_per_block = block_size;
2963 if (block_size & (block_size - 1))
2964 pool->sectors_per_block_shift = -1;
2965 else
2966 pool->sectors_per_block_shift = __ffs(block_size);
2967 pool->low_water_blocks = 0;
2968 pool_features_init(&pool->pf);
2969 pool->prison = dm_bio_prison_create();
2970 if (!pool->prison) {
2971 *error = "Error creating pool's bio prison";
2972 err_p = ERR_PTR(-ENOMEM);
2973 goto bad_prison;
2974 }
2975
2976 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2977 if (IS_ERR(pool->copier)) {
2978 r = PTR_ERR(pool->copier);
2979 *error = "Error creating pool's kcopyd client";
2980 err_p = ERR_PTR(r);
2981 goto bad_kcopyd_client;
2982 }
2983
2984 /*
2985 * Create singlethreaded workqueue that will service all devices
2986 * that use this metadata.
2987 */
2988 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2989 if (!pool->wq) {
2990 *error = "Error creating pool's workqueue";
2991 err_p = ERR_PTR(-ENOMEM);
2992 goto bad_wq;
2993 }
2994
2995 throttle_init(&pool->throttle);
2996 INIT_WORK(&pool->worker, do_worker);
2997 INIT_DELAYED_WORK(&pool->waker, do_waker);
2998 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2999 spin_lock_init(&pool->lock);
3000 bio_list_init(&pool->deferred_flush_bios);
3001 bio_list_init(&pool->deferred_flush_completions);
3002 INIT_LIST_HEAD(&pool->prepared_mappings);
3003 INIT_LIST_HEAD(&pool->prepared_discards);
3004 INIT_LIST_HEAD(&pool->prepared_discards_pt2);
3005 INIT_LIST_HEAD(&pool->active_thins);
3006 pool->low_water_triggered = false;
3007 pool->suspended = true;
3008 pool->out_of_data_space = false;
3009
3010 pool->shared_read_ds = dm_deferred_set_create();
3011 if (!pool->shared_read_ds) {
3012 *error = "Error creating pool's shared read deferred set";
3013 err_p = ERR_PTR(-ENOMEM);
3014 goto bad_shared_read_ds;
3015 }
3016
3017 pool->all_io_ds = dm_deferred_set_create();
3018 if (!pool->all_io_ds) {
3019 *error = "Error creating pool's all io deferred set";
3020 err_p = ERR_PTR(-ENOMEM);
3021 goto bad_all_io_ds;
3022 }
3023
3024 pool->next_mapping = NULL;
3025 r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
3026 _new_mapping_cache);
3027 if (r) {
3028 *error = "Error creating pool's mapping mempool";
3029 err_p = ERR_PTR(r);
3030 goto bad_mapping_pool;
3031 }
3032
3033 pool->cell_sort_array =
3034 vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3035 sizeof(*pool->cell_sort_array)));
3036 if (!pool->cell_sort_array) {
3037 *error = "Error allocating cell sort array";
3038 err_p = ERR_PTR(-ENOMEM);
3039 goto bad_sort_array;
3040 }
3041
3042 pool->ref_count = 1;
3043 pool->last_commit_jiffies = jiffies;
3044 pool->pool_md = pool_md;
3045 pool->md_dev = metadata_dev;
3046 pool->data_dev = data_dev;
3047 __pool_table_insert(pool);
3048
3049 return pool;
3050
3051bad_sort_array:
3052 mempool_exit(&pool->mapping_pool);
3053bad_mapping_pool:
3054 dm_deferred_set_destroy(pool->all_io_ds);
3055bad_all_io_ds:
3056 dm_deferred_set_destroy(pool->shared_read_ds);
3057bad_shared_read_ds:
3058 destroy_workqueue(pool->wq);
3059bad_wq:
3060 dm_kcopyd_client_destroy(pool->copier);
3061bad_kcopyd_client:
3062 dm_bio_prison_destroy(pool->prison);
3063bad_prison:
3064 kfree(pool);
3065bad_pool:
3066 if (dm_pool_metadata_close(pmd))
3067 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3068
3069 return err_p;
3070}
3071
3072static void __pool_inc(struct pool *pool)
3073{
3074 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3075 pool->ref_count++;
3076}
3077
3078static void __pool_dec(struct pool *pool)
3079{
3080 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3081 BUG_ON(!pool->ref_count);
3082 if (!--pool->ref_count)
3083 __pool_destroy(pool);
3084}
3085
3086static struct pool *__pool_find(struct mapped_device *pool_md,
3087 struct block_device *metadata_dev,
3088 struct block_device *data_dev,
3089 unsigned long block_size, int read_only,
3090 char **error, int *created)
3091{
3092 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3093
3094 if (pool) {
3095 if (pool->pool_md != pool_md) {
3096 *error = "metadata device already in use by a pool";
3097 return ERR_PTR(-EBUSY);
3098 }
3099 if (pool->data_dev != data_dev) {
3100 *error = "data device already in use by a pool";
3101 return ERR_PTR(-EBUSY);
3102 }
3103 __pool_inc(pool);
3104
3105 } else {
3106 pool = __pool_table_lookup(pool_md);
3107 if (pool) {
3108 if (pool->md_dev != metadata_dev || pool->data_dev != data_dev) {
3109 *error = "different pool cannot replace a pool";
3110 return ERR_PTR(-EINVAL);
3111 }
3112 __pool_inc(pool);
3113
3114 } else {
3115 pool = pool_create(pool_md, metadata_dev, data_dev, block_size, read_only, error);
3116 *created = 1;
3117 }
3118 }
3119
3120 return pool;
3121}
3122
3123/*
3124 *--------------------------------------------------------------
3125 * Pool target methods
3126 *--------------------------------------------------------------
3127 */
3128static void pool_dtr(struct dm_target *ti)
3129{
3130 struct pool_c *pt = ti->private;
3131
3132 mutex_lock(&dm_thin_pool_table.mutex);
3133
3134 unbind_control_target(pt->pool, ti);
3135 __pool_dec(pt->pool);
3136 dm_put_device(ti, pt->metadata_dev);
3137 dm_put_device(ti, pt->data_dev);
3138 kfree(pt);
3139
3140 mutex_unlock(&dm_thin_pool_table.mutex);
3141}
3142
3143static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3144 struct dm_target *ti)
3145{
3146 int r;
3147 unsigned int argc;
3148 const char *arg_name;
3149
3150 static const struct dm_arg _args[] = {
3151 {0, 4, "Invalid number of pool feature arguments"},
3152 };
3153
3154 /*
3155 * No feature arguments supplied.
3156 */
3157 if (!as->argc)
3158 return 0;
3159
3160 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3161 if (r)
3162 return -EINVAL;
3163
3164 while (argc && !r) {
3165 arg_name = dm_shift_arg(as);
3166 argc--;
3167
3168 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3169 pf->zero_new_blocks = false;
3170
3171 else if (!strcasecmp(arg_name, "ignore_discard"))
3172 pf->discard_enabled = false;
3173
3174 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3175 pf->discard_passdown = false;
3176
3177 else if (!strcasecmp(arg_name, "read_only"))
3178 pf->mode = PM_READ_ONLY;
3179
3180 else if (!strcasecmp(arg_name, "error_if_no_space"))
3181 pf->error_if_no_space = true;
3182
3183 else {
3184 ti->error = "Unrecognised pool feature requested";
3185 r = -EINVAL;
3186 break;
3187 }
3188 }
3189
3190 return r;
3191}
3192
3193static void metadata_low_callback(void *context)
3194{
3195 struct pool *pool = context;
3196
3197 DMWARN("%s: reached low water mark for metadata device: sending event.",
3198 dm_device_name(pool->pool_md));
3199
3200 dm_table_event(pool->ti->table);
3201}
3202
3203/*
3204 * We need to flush the data device **before** committing the metadata.
3205 *
3206 * This ensures that the data blocks of any newly inserted mappings are
3207 * properly written to non-volatile storage and won't be lost in case of a
3208 * crash.
3209 *
3210 * Failure to do so can result in data corruption in the case of internal or
3211 * external snapshots and in the case of newly provisioned blocks, when block
3212 * zeroing is enabled.
3213 */
3214static int metadata_pre_commit_callback(void *context)
3215{
3216 struct pool *pool = context;
3217
3218 return blkdev_issue_flush(pool->data_dev);
3219}
3220
3221static sector_t get_dev_size(struct block_device *bdev)
3222{
3223 return bdev_nr_sectors(bdev);
3224}
3225
3226static void warn_if_metadata_device_too_big(struct block_device *bdev)
3227{
3228 sector_t metadata_dev_size = get_dev_size(bdev);
3229
3230 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3231 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
3232 bdev, THIN_METADATA_MAX_SECTORS);
3233}
3234
3235static sector_t get_metadata_dev_size(struct block_device *bdev)
3236{
3237 sector_t metadata_dev_size = get_dev_size(bdev);
3238
3239 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3240 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3241
3242 return metadata_dev_size;
3243}
3244
3245static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3246{
3247 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3248
3249 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3250
3251 return metadata_dev_size;
3252}
3253
3254/*
3255 * When a metadata threshold is crossed a dm event is triggered, and
3256 * userland should respond by growing the metadata device. We could let
3257 * userland set the threshold, like we do with the data threshold, but I'm
3258 * not sure they know enough to do this well.
3259 */
3260static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3261{
3262 /*
3263 * 4M is ample for all ops with the possible exception of thin
3264 * device deletion which is harmless if it fails (just retry the
3265 * delete after you've grown the device).
3266 */
3267 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3268
3269 return min((dm_block_t)1024ULL /* 4M */, quarter);
3270}
3271
3272/*
3273 * thin-pool <metadata dev> <data dev>
3274 * <data block size (sectors)>
3275 * <low water mark (blocks)>
3276 * [<#feature args> [<arg>]*]
3277 *
3278 * Optional feature arguments are:
3279 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3280 * ignore_discard: disable discard
3281 * no_discard_passdown: don't pass discards down to the data device
3282 * read_only: Don't allow any changes to be made to the pool metadata.
3283 * error_if_no_space: error IOs, instead of queueing, if no space.
3284 */
3285static int pool_ctr(struct dm_target *ti, unsigned int argc, char **argv)
3286{
3287 int r, pool_created = 0;
3288 struct pool_c *pt;
3289 struct pool *pool;
3290 struct pool_features pf;
3291 struct dm_arg_set as;
3292 struct dm_dev *data_dev;
3293 unsigned long block_size;
3294 dm_block_t low_water_blocks;
3295 struct dm_dev *metadata_dev;
3296 blk_mode_t metadata_mode;
3297
3298 /*
3299 * FIXME Remove validation from scope of lock.
3300 */
3301 mutex_lock(&dm_thin_pool_table.mutex);
3302
3303 if (argc < 4) {
3304 ti->error = "Invalid argument count";
3305 r = -EINVAL;
3306 goto out_unlock;
3307 }
3308
3309 as.argc = argc;
3310 as.argv = argv;
3311
3312 /* make sure metadata and data are different devices */
3313 if (!strcmp(argv[0], argv[1])) {
3314 ti->error = "Error setting metadata or data device";
3315 r = -EINVAL;
3316 goto out_unlock;
3317 }
3318
3319 /*
3320 * Set default pool features.
3321 */
3322 pool_features_init(&pf);
3323
3324 dm_consume_args(&as, 4);
3325 r = parse_pool_features(&as, &pf, ti);
3326 if (r)
3327 goto out_unlock;
3328
3329 metadata_mode = BLK_OPEN_READ |
3330 ((pf.mode == PM_READ_ONLY) ? 0 : BLK_OPEN_WRITE);
3331 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3332 if (r) {
3333 ti->error = "Error opening metadata block device";
3334 goto out_unlock;
3335 }
3336 warn_if_metadata_device_too_big(metadata_dev->bdev);
3337
3338 r = dm_get_device(ti, argv[1], BLK_OPEN_READ | BLK_OPEN_WRITE, &data_dev);
3339 if (r) {
3340 ti->error = "Error getting data device";
3341 goto out_metadata;
3342 }
3343
3344 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3345 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3346 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3347 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3348 ti->error = "Invalid block size";
3349 r = -EINVAL;
3350 goto out;
3351 }
3352
3353 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3354 ti->error = "Invalid low water mark";
3355 r = -EINVAL;
3356 goto out;
3357 }
3358
3359 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3360 if (!pt) {
3361 r = -ENOMEM;
3362 goto out;
3363 }
3364
3365 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev, data_dev->bdev,
3366 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3367 if (IS_ERR(pool)) {
3368 r = PTR_ERR(pool);
3369 goto out_free_pt;
3370 }
3371
3372 /*
3373 * 'pool_created' reflects whether this is the first table load.
3374 * Top level discard support is not allowed to be changed after
3375 * initial load. This would require a pool reload to trigger thin
3376 * device changes.
3377 */
3378 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3379 ti->error = "Discard support cannot be disabled once enabled";
3380 r = -EINVAL;
3381 goto out_flags_changed;
3382 }
3383
3384 pt->pool = pool;
3385 pt->ti = ti;
3386 pt->metadata_dev = metadata_dev;
3387 pt->data_dev = data_dev;
3388 pt->low_water_blocks = low_water_blocks;
3389 pt->adjusted_pf = pt->requested_pf = pf;
3390 ti->num_flush_bios = 1;
3391 ti->limit_swap_bios = true;
3392
3393 /*
3394 * Only need to enable discards if the pool should pass
3395 * them down to the data device. The thin device's discard
3396 * processing will cause mappings to be removed from the btree.
3397 */
3398 if (pf.discard_enabled && pf.discard_passdown) {
3399 ti->num_discard_bios = 1;
3400 /*
3401 * Setting 'discards_supported' circumvents the normal
3402 * stacking of discard limits (this keeps the pool and
3403 * thin devices' discard limits consistent).
3404 */
3405 ti->discards_supported = true;
3406 ti->max_discard_granularity = true;
3407 }
3408 ti->private = pt;
3409
3410 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3411 calc_metadata_threshold(pt),
3412 metadata_low_callback,
3413 pool);
3414 if (r) {
3415 ti->error = "Error registering metadata threshold";
3416 goto out_flags_changed;
3417 }
3418
3419 dm_pool_register_pre_commit_callback(pool->pmd,
3420 metadata_pre_commit_callback, pool);
3421
3422 mutex_unlock(&dm_thin_pool_table.mutex);
3423
3424 return 0;
3425
3426out_flags_changed:
3427 __pool_dec(pool);
3428out_free_pt:
3429 kfree(pt);
3430out:
3431 dm_put_device(ti, data_dev);
3432out_metadata:
3433 dm_put_device(ti, metadata_dev);
3434out_unlock:
3435 mutex_unlock(&dm_thin_pool_table.mutex);
3436
3437 return r;
3438}
3439
3440static int pool_map(struct dm_target *ti, struct bio *bio)
3441{
3442 struct pool_c *pt = ti->private;
3443 struct pool *pool = pt->pool;
3444
3445 /*
3446 * As this is a singleton target, ti->begin is always zero.
3447 */
3448 spin_lock_irq(&pool->lock);
3449 bio_set_dev(bio, pt->data_dev->bdev);
3450 spin_unlock_irq(&pool->lock);
3451
3452 return DM_MAPIO_REMAPPED;
3453}
3454
3455static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3456{
3457 int r;
3458 struct pool_c *pt = ti->private;
3459 struct pool *pool = pt->pool;
3460 sector_t data_size = ti->len;
3461 dm_block_t sb_data_size;
3462
3463 *need_commit = false;
3464
3465 (void) sector_div(data_size, pool->sectors_per_block);
3466
3467 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3468 if (r) {
3469 DMERR("%s: failed to retrieve data device size",
3470 dm_device_name(pool->pool_md));
3471 return r;
3472 }
3473
3474 if (data_size < sb_data_size) {
3475 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3476 dm_device_name(pool->pool_md),
3477 (unsigned long long)data_size, sb_data_size);
3478 return -EINVAL;
3479
3480 } else if (data_size > sb_data_size) {
3481 if (dm_pool_metadata_needs_check(pool->pmd)) {
3482 DMERR("%s: unable to grow the data device until repaired.",
3483 dm_device_name(pool->pool_md));
3484 return 0;
3485 }
3486
3487 if (sb_data_size)
3488 DMINFO("%s: growing the data device from %llu to %llu blocks",
3489 dm_device_name(pool->pool_md),
3490 sb_data_size, (unsigned long long)data_size);
3491 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3492 if (r) {
3493 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3494 return r;
3495 }
3496
3497 *need_commit = true;
3498 }
3499
3500 return 0;
3501}
3502
3503static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3504{
3505 int r;
3506 struct pool_c *pt = ti->private;
3507 struct pool *pool = pt->pool;
3508 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3509
3510 *need_commit = false;
3511
3512 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3513
3514 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3515 if (r) {
3516 DMERR("%s: failed to retrieve metadata device size",
3517 dm_device_name(pool->pool_md));
3518 return r;
3519 }
3520
3521 if (metadata_dev_size < sb_metadata_dev_size) {
3522 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3523 dm_device_name(pool->pool_md),
3524 metadata_dev_size, sb_metadata_dev_size);
3525 return -EINVAL;
3526
3527 } else if (metadata_dev_size > sb_metadata_dev_size) {
3528 if (dm_pool_metadata_needs_check(pool->pmd)) {
3529 DMERR("%s: unable to grow the metadata device until repaired.",
3530 dm_device_name(pool->pool_md));
3531 return 0;
3532 }
3533
3534 warn_if_metadata_device_too_big(pool->md_dev);
3535 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3536 dm_device_name(pool->pool_md),
3537 sb_metadata_dev_size, metadata_dev_size);
3538
3539 if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3540 set_pool_mode(pool, PM_WRITE);
3541
3542 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3543 if (r) {
3544 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3545 return r;
3546 }
3547
3548 *need_commit = true;
3549 }
3550
3551 return 0;
3552}
3553
3554/*
3555 * Retrieves the number of blocks of the data device from
3556 * the superblock and compares it to the actual device size,
3557 * thus resizing the data device in case it has grown.
3558 *
3559 * This both copes with opening preallocated data devices in the ctr
3560 * being followed by a resume
3561 * -and-
3562 * calling the resume method individually after userspace has
3563 * grown the data device in reaction to a table event.
3564 */
3565static int pool_preresume(struct dm_target *ti)
3566{
3567 int r;
3568 bool need_commit1, need_commit2;
3569 struct pool_c *pt = ti->private;
3570 struct pool *pool = pt->pool;
3571
3572 /*
3573 * Take control of the pool object.
3574 */
3575 r = bind_control_target(pool, ti);
3576 if (r)
3577 goto out;
3578
3579 r = maybe_resize_data_dev(ti, &need_commit1);
3580 if (r)
3581 goto out;
3582
3583 r = maybe_resize_metadata_dev(ti, &need_commit2);
3584 if (r)
3585 goto out;
3586
3587 if (need_commit1 || need_commit2)
3588 (void) commit(pool);
3589out:
3590 /*
3591 * When a thin-pool is PM_FAIL, it cannot be rebuilt if
3592 * bio is in deferred list. Therefore need to return 0
3593 * to allow pool_resume() to flush IO.
3594 */
3595 if (r && get_pool_mode(pool) == PM_FAIL)
3596 r = 0;
3597
3598 return r;
3599}
3600
3601static void pool_suspend_active_thins(struct pool *pool)
3602{
3603 struct thin_c *tc;
3604
3605 /* Suspend all active thin devices */
3606 tc = get_first_thin(pool);
3607 while (tc) {
3608 dm_internal_suspend_noflush(tc->thin_md);
3609 tc = get_next_thin(pool, tc);
3610 }
3611}
3612
3613static void pool_resume_active_thins(struct pool *pool)
3614{
3615 struct thin_c *tc;
3616
3617 /* Resume all active thin devices */
3618 tc = get_first_thin(pool);
3619 while (tc) {
3620 dm_internal_resume(tc->thin_md);
3621 tc = get_next_thin(pool, tc);
3622 }
3623}
3624
3625static void pool_resume(struct dm_target *ti)
3626{
3627 struct pool_c *pt = ti->private;
3628 struct pool *pool = pt->pool;
3629
3630 /*
3631 * Must requeue active_thins' bios and then resume
3632 * active_thins _before_ clearing 'suspend' flag.
3633 */
3634 requeue_bios(pool);
3635 pool_resume_active_thins(pool);
3636
3637 spin_lock_irq(&pool->lock);
3638 pool->low_water_triggered = false;
3639 pool->suspended = false;
3640 spin_unlock_irq(&pool->lock);
3641
3642 do_waker(&pool->waker.work);
3643}
3644
3645static void pool_presuspend(struct dm_target *ti)
3646{
3647 struct pool_c *pt = ti->private;
3648 struct pool *pool = pt->pool;
3649
3650 spin_lock_irq(&pool->lock);
3651 pool->suspended = true;
3652 spin_unlock_irq(&pool->lock);
3653
3654 pool_suspend_active_thins(pool);
3655}
3656
3657static void pool_presuspend_undo(struct dm_target *ti)
3658{
3659 struct pool_c *pt = ti->private;
3660 struct pool *pool = pt->pool;
3661
3662 pool_resume_active_thins(pool);
3663
3664 spin_lock_irq(&pool->lock);
3665 pool->suspended = false;
3666 spin_unlock_irq(&pool->lock);
3667}
3668
3669static void pool_postsuspend(struct dm_target *ti)
3670{
3671 struct pool_c *pt = ti->private;
3672 struct pool *pool = pt->pool;
3673
3674 cancel_delayed_work_sync(&pool->waker);
3675 cancel_delayed_work_sync(&pool->no_space_timeout);
3676 flush_workqueue(pool->wq);
3677 (void) commit(pool);
3678}
3679
3680static int check_arg_count(unsigned int argc, unsigned int args_required)
3681{
3682 if (argc != args_required) {
3683 DMWARN("Message received with %u arguments instead of %u.",
3684 argc, args_required);
3685 return -EINVAL;
3686 }
3687
3688 return 0;
3689}
3690
3691static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3692{
3693 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3694 *dev_id <= MAX_DEV_ID)
3695 return 0;
3696
3697 if (warning)
3698 DMWARN("Message received with invalid device id: %s", arg);
3699
3700 return -EINVAL;
3701}
3702
3703static int process_create_thin_mesg(unsigned int argc, char **argv, struct pool *pool)
3704{
3705 dm_thin_id dev_id;
3706 int r;
3707
3708 r = check_arg_count(argc, 2);
3709 if (r)
3710 return r;
3711
3712 r = read_dev_id(argv[1], &dev_id, 1);
3713 if (r)
3714 return r;
3715
3716 r = dm_pool_create_thin(pool->pmd, dev_id);
3717 if (r) {
3718 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3719 argv[1]);
3720 return r;
3721 }
3722
3723 return 0;
3724}
3725
3726static int process_create_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3727{
3728 dm_thin_id dev_id;
3729 dm_thin_id origin_dev_id;
3730 int r;
3731
3732 r = check_arg_count(argc, 3);
3733 if (r)
3734 return r;
3735
3736 r = read_dev_id(argv[1], &dev_id, 1);
3737 if (r)
3738 return r;
3739
3740 r = read_dev_id(argv[2], &origin_dev_id, 1);
3741 if (r)
3742 return r;
3743
3744 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3745 if (r) {
3746 DMWARN("Creation of new snapshot %s of device %s failed.",
3747 argv[1], argv[2]);
3748 return r;
3749 }
3750
3751 return 0;
3752}
3753
3754static int process_delete_mesg(unsigned int argc, char **argv, struct pool *pool)
3755{
3756 dm_thin_id dev_id;
3757 int r;
3758
3759 r = check_arg_count(argc, 2);
3760 if (r)
3761 return r;
3762
3763 r = read_dev_id(argv[1], &dev_id, 1);
3764 if (r)
3765 return r;
3766
3767 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3768 if (r)
3769 DMWARN("Deletion of thin device %s failed.", argv[1]);
3770
3771 return r;
3772}
3773
3774static int process_set_transaction_id_mesg(unsigned int argc, char **argv, struct pool *pool)
3775{
3776 dm_thin_id old_id, new_id;
3777 int r;
3778
3779 r = check_arg_count(argc, 3);
3780 if (r)
3781 return r;
3782
3783 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3784 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3785 return -EINVAL;
3786 }
3787
3788 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3789 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3790 return -EINVAL;
3791 }
3792
3793 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3794 if (r) {
3795 DMWARN("Failed to change transaction id from %s to %s.",
3796 argv[1], argv[2]);
3797 return r;
3798 }
3799
3800 return 0;
3801}
3802
3803static int process_reserve_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3804{
3805 int r;
3806
3807 r = check_arg_count(argc, 1);
3808 if (r)
3809 return r;
3810
3811 (void) commit(pool);
3812
3813 r = dm_pool_reserve_metadata_snap(pool->pmd);
3814 if (r)
3815 DMWARN("reserve_metadata_snap message failed.");
3816
3817 return r;
3818}
3819
3820static int process_release_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3821{
3822 int r;
3823
3824 r = check_arg_count(argc, 1);
3825 if (r)
3826 return r;
3827
3828 r = dm_pool_release_metadata_snap(pool->pmd);
3829 if (r)
3830 DMWARN("release_metadata_snap message failed.");
3831
3832 return r;
3833}
3834
3835/*
3836 * Messages supported:
3837 * create_thin <dev_id>
3838 * create_snap <dev_id> <origin_id>
3839 * delete <dev_id>
3840 * set_transaction_id <current_trans_id> <new_trans_id>
3841 * reserve_metadata_snap
3842 * release_metadata_snap
3843 */
3844static int pool_message(struct dm_target *ti, unsigned int argc, char **argv,
3845 char *result, unsigned int maxlen)
3846{
3847 int r = -EINVAL;
3848 struct pool_c *pt = ti->private;
3849 struct pool *pool = pt->pool;
3850
3851 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE) {
3852 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3853 dm_device_name(pool->pool_md));
3854 return -EOPNOTSUPP;
3855 }
3856
3857 if (!strcasecmp(argv[0], "create_thin"))
3858 r = process_create_thin_mesg(argc, argv, pool);
3859
3860 else if (!strcasecmp(argv[0], "create_snap"))
3861 r = process_create_snap_mesg(argc, argv, pool);
3862
3863 else if (!strcasecmp(argv[0], "delete"))
3864 r = process_delete_mesg(argc, argv, pool);
3865
3866 else if (!strcasecmp(argv[0], "set_transaction_id"))
3867 r = process_set_transaction_id_mesg(argc, argv, pool);
3868
3869 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3870 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3871
3872 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3873 r = process_release_metadata_snap_mesg(argc, argv, pool);
3874
3875 else
3876 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3877
3878 if (!r)
3879 (void) commit(pool);
3880
3881 return r;
3882}
3883
3884static void emit_flags(struct pool_features *pf, char *result,
3885 unsigned int sz, unsigned int maxlen)
3886{
3887 unsigned int count = !pf->zero_new_blocks + !pf->discard_enabled +
3888 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3889 pf->error_if_no_space;
3890 DMEMIT("%u ", count);
3891
3892 if (!pf->zero_new_blocks)
3893 DMEMIT("skip_block_zeroing ");
3894
3895 if (!pf->discard_enabled)
3896 DMEMIT("ignore_discard ");
3897
3898 if (!pf->discard_passdown)
3899 DMEMIT("no_discard_passdown ");
3900
3901 if (pf->mode == PM_READ_ONLY)
3902 DMEMIT("read_only ");
3903
3904 if (pf->error_if_no_space)
3905 DMEMIT("error_if_no_space ");
3906}
3907
3908/*
3909 * Status line is:
3910 * <transaction id> <used metadata sectors>/<total metadata sectors>
3911 * <used data sectors>/<total data sectors> <held metadata root>
3912 * <pool mode> <discard config> <no space config> <needs_check>
3913 */
3914static void pool_status(struct dm_target *ti, status_type_t type,
3915 unsigned int status_flags, char *result, unsigned int maxlen)
3916{
3917 int r;
3918 unsigned int sz = 0;
3919 uint64_t transaction_id;
3920 dm_block_t nr_free_blocks_data;
3921 dm_block_t nr_free_blocks_metadata;
3922 dm_block_t nr_blocks_data;
3923 dm_block_t nr_blocks_metadata;
3924 dm_block_t held_root;
3925 enum pool_mode mode;
3926 char buf[BDEVNAME_SIZE];
3927 char buf2[BDEVNAME_SIZE];
3928 struct pool_c *pt = ti->private;
3929 struct pool *pool = pt->pool;
3930
3931 switch (type) {
3932 case STATUSTYPE_INFO:
3933 if (get_pool_mode(pool) == PM_FAIL) {
3934 DMEMIT("Fail");
3935 break;
3936 }
3937
3938 /* Commit to ensure statistics aren't out-of-date */
3939 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3940 (void) commit(pool);
3941
3942 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3943 if (r) {
3944 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3945 dm_device_name(pool->pool_md), r);
3946 goto err;
3947 }
3948
3949 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3950 if (r) {
3951 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3952 dm_device_name(pool->pool_md), r);
3953 goto err;
3954 }
3955
3956 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3957 if (r) {
3958 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3959 dm_device_name(pool->pool_md), r);
3960 goto err;
3961 }
3962
3963 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3964 if (r) {
3965 DMERR("%s: dm_pool_get_free_block_count returned %d",
3966 dm_device_name(pool->pool_md), r);
3967 goto err;
3968 }
3969
3970 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3971 if (r) {
3972 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3973 dm_device_name(pool->pool_md), r);
3974 goto err;
3975 }
3976
3977 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3978 if (r) {
3979 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3980 dm_device_name(pool->pool_md), r);
3981 goto err;
3982 }
3983
3984 DMEMIT("%llu %llu/%llu %llu/%llu ",
3985 (unsigned long long)transaction_id,
3986 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3987 (unsigned long long)nr_blocks_metadata,
3988 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3989 (unsigned long long)nr_blocks_data);
3990
3991 if (held_root)
3992 DMEMIT("%llu ", held_root);
3993 else
3994 DMEMIT("- ");
3995
3996 mode = get_pool_mode(pool);
3997 if (mode == PM_OUT_OF_DATA_SPACE)
3998 DMEMIT("out_of_data_space ");
3999 else if (is_read_only_pool_mode(mode))
4000 DMEMIT("ro ");
4001 else
4002 DMEMIT("rw ");
4003
4004 if (!pool->pf.discard_enabled)
4005 DMEMIT("ignore_discard ");
4006 else if (pool->pf.discard_passdown)
4007 DMEMIT("discard_passdown ");
4008 else
4009 DMEMIT("no_discard_passdown ");
4010
4011 if (pool->pf.error_if_no_space)
4012 DMEMIT("error_if_no_space ");
4013 else
4014 DMEMIT("queue_if_no_space ");
4015
4016 if (dm_pool_metadata_needs_check(pool->pmd))
4017 DMEMIT("needs_check ");
4018 else
4019 DMEMIT("- ");
4020
4021 DMEMIT("%llu ", (unsigned long long)calc_metadata_threshold(pt));
4022
4023 break;
4024
4025 case STATUSTYPE_TABLE:
4026 DMEMIT("%s %s %lu %llu ",
4027 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
4028 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
4029 (unsigned long)pool->sectors_per_block,
4030 (unsigned long long)pt->low_water_blocks);
4031 emit_flags(&pt->requested_pf, result, sz, maxlen);
4032 break;
4033
4034 case STATUSTYPE_IMA:
4035 *result = '\0';
4036 break;
4037 }
4038 return;
4039
4040err:
4041 DMEMIT("Error");
4042}
4043
4044static int pool_iterate_devices(struct dm_target *ti,
4045 iterate_devices_callout_fn fn, void *data)
4046{
4047 struct pool_c *pt = ti->private;
4048
4049 return fn(ti, pt->data_dev, 0, ti->len, data);
4050}
4051
4052static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
4053{
4054 struct pool_c *pt = ti->private;
4055 struct pool *pool = pt->pool;
4056 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
4057
4058 /*
4059 * If max_sectors is smaller than pool->sectors_per_block adjust it
4060 * to the highest possible power-of-2 factor of pool->sectors_per_block.
4061 * This is especially beneficial when the pool's data device is a RAID
4062 * device that has a full stripe width that matches pool->sectors_per_block
4063 * -- because even though partial RAID stripe-sized IOs will be issued to a
4064 * single RAID stripe; when aggregated they will end on a full RAID stripe
4065 * boundary.. which avoids additional partial RAID stripe writes cascading
4066 */
4067 if (limits->max_sectors < pool->sectors_per_block) {
4068 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
4069 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
4070 limits->max_sectors--;
4071 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
4072 }
4073 }
4074
4075 /*
4076 * If the system-determined stacked limits are compatible with the
4077 * pool's blocksize (io_opt is a factor) do not override them.
4078 */
4079 if (io_opt_sectors < pool->sectors_per_block ||
4080 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
4081 if (is_factor(pool->sectors_per_block, limits->max_sectors))
4082 limits->io_min = limits->max_sectors << SECTOR_SHIFT;
4083 else
4084 limits->io_min = pool->sectors_per_block << SECTOR_SHIFT;
4085 limits->io_opt = pool->sectors_per_block << SECTOR_SHIFT;
4086 }
4087
4088 /*
4089 * pt->adjusted_pf is a staging area for the actual features to use.
4090 * They get transferred to the live pool in bind_control_target()
4091 * called from pool_preresume().
4092 */
4093
4094 if (pt->adjusted_pf.discard_enabled) {
4095 disable_discard_passdown_if_not_supported(pt);
4096 if (!pt->adjusted_pf.discard_passdown)
4097 limits->max_hw_discard_sectors = 0;
4098 /*
4099 * The pool uses the same discard limits as the underlying data
4100 * device. DM core has already set this up.
4101 */
4102 } else {
4103 /*
4104 * Must explicitly disallow stacking discard limits otherwise the
4105 * block layer will stack them if pool's data device has support.
4106 */
4107 limits->discard_granularity = 0;
4108 }
4109}
4110
4111static struct target_type pool_target = {
4112 .name = "thin-pool",
4113 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4114 DM_TARGET_IMMUTABLE,
4115 .version = {1, 23, 0},
4116 .module = THIS_MODULE,
4117 .ctr = pool_ctr,
4118 .dtr = pool_dtr,
4119 .map = pool_map,
4120 .presuspend = pool_presuspend,
4121 .presuspend_undo = pool_presuspend_undo,
4122 .postsuspend = pool_postsuspend,
4123 .preresume = pool_preresume,
4124 .resume = pool_resume,
4125 .message = pool_message,
4126 .status = pool_status,
4127 .iterate_devices = pool_iterate_devices,
4128 .io_hints = pool_io_hints,
4129};
4130
4131/*
4132 *--------------------------------------------------------------
4133 * Thin target methods
4134 *--------------------------------------------------------------
4135 */
4136static void thin_get(struct thin_c *tc)
4137{
4138 refcount_inc(&tc->refcount);
4139}
4140
4141static void thin_put(struct thin_c *tc)
4142{
4143 if (refcount_dec_and_test(&tc->refcount))
4144 complete(&tc->can_destroy);
4145}
4146
4147static void thin_dtr(struct dm_target *ti)
4148{
4149 struct thin_c *tc = ti->private;
4150
4151 spin_lock_irq(&tc->pool->lock);
4152 list_del_rcu(&tc->list);
4153 spin_unlock_irq(&tc->pool->lock);
4154 synchronize_rcu();
4155
4156 thin_put(tc);
4157 wait_for_completion(&tc->can_destroy);
4158
4159 mutex_lock(&dm_thin_pool_table.mutex);
4160
4161 __pool_dec(tc->pool);
4162 dm_pool_close_thin_device(tc->td);
4163 dm_put_device(ti, tc->pool_dev);
4164 if (tc->origin_dev)
4165 dm_put_device(ti, tc->origin_dev);
4166 kfree(tc);
4167
4168 mutex_unlock(&dm_thin_pool_table.mutex);
4169}
4170
4171/*
4172 * Thin target parameters:
4173 *
4174 * <pool_dev> <dev_id> [origin_dev]
4175 *
4176 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4177 * dev_id: the internal device identifier
4178 * origin_dev: a device external to the pool that should act as the origin
4179 *
4180 * If the pool device has discards disabled, they get disabled for the thin
4181 * device as well.
4182 */
4183static int thin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
4184{
4185 int r;
4186 struct thin_c *tc;
4187 struct dm_dev *pool_dev, *origin_dev;
4188 struct mapped_device *pool_md;
4189
4190 mutex_lock(&dm_thin_pool_table.mutex);
4191
4192 if (argc != 2 && argc != 3) {
4193 ti->error = "Invalid argument count";
4194 r = -EINVAL;
4195 goto out_unlock;
4196 }
4197
4198 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4199 if (!tc) {
4200 ti->error = "Out of memory";
4201 r = -ENOMEM;
4202 goto out_unlock;
4203 }
4204 tc->thin_md = dm_table_get_md(ti->table);
4205 spin_lock_init(&tc->lock);
4206 INIT_LIST_HEAD(&tc->deferred_cells);
4207 bio_list_init(&tc->deferred_bio_list);
4208 bio_list_init(&tc->retry_on_resume_list);
4209 tc->sort_bio_list = RB_ROOT;
4210
4211 if (argc == 3) {
4212 if (!strcmp(argv[0], argv[2])) {
4213 ti->error = "Error setting origin device";
4214 r = -EINVAL;
4215 goto bad_origin_dev;
4216 }
4217
4218 r = dm_get_device(ti, argv[2], BLK_OPEN_READ, &origin_dev);
4219 if (r) {
4220 ti->error = "Error opening origin device";
4221 goto bad_origin_dev;
4222 }
4223 tc->origin_dev = origin_dev;
4224 }
4225
4226 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4227 if (r) {
4228 ti->error = "Error opening pool device";
4229 goto bad_pool_dev;
4230 }
4231 tc->pool_dev = pool_dev;
4232
4233 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4234 ti->error = "Invalid device id";
4235 r = -EINVAL;
4236 goto bad_common;
4237 }
4238
4239 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4240 if (!pool_md) {
4241 ti->error = "Couldn't get pool mapped device";
4242 r = -EINVAL;
4243 goto bad_common;
4244 }
4245
4246 tc->pool = __pool_table_lookup(pool_md);
4247 if (!tc->pool) {
4248 ti->error = "Couldn't find pool object";
4249 r = -EINVAL;
4250 goto bad_pool_lookup;
4251 }
4252 __pool_inc(tc->pool);
4253
4254 if (get_pool_mode(tc->pool) == PM_FAIL) {
4255 ti->error = "Couldn't open thin device, Pool is in fail mode";
4256 r = -EINVAL;
4257 goto bad_pool;
4258 }
4259
4260 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4261 if (r) {
4262 ti->error = "Couldn't open thin internal device";
4263 goto bad_pool;
4264 }
4265
4266 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4267 if (r)
4268 goto bad;
4269
4270 ti->num_flush_bios = 1;
4271 ti->limit_swap_bios = true;
4272 ti->flush_supported = true;
4273 ti->accounts_remapped_io = true;
4274 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4275
4276 /* In case the pool supports discards, pass them on. */
4277 if (tc->pool->pf.discard_enabled) {
4278 ti->discards_supported = true;
4279 ti->num_discard_bios = 1;
4280 ti->max_discard_granularity = true;
4281 }
4282
4283 mutex_unlock(&dm_thin_pool_table.mutex);
4284
4285 spin_lock_irq(&tc->pool->lock);
4286 if (tc->pool->suspended) {
4287 spin_unlock_irq(&tc->pool->lock);
4288 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4289 ti->error = "Unable to activate thin device while pool is suspended";
4290 r = -EINVAL;
4291 goto bad;
4292 }
4293 refcount_set(&tc->refcount, 1);
4294 init_completion(&tc->can_destroy);
4295 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4296 spin_unlock_irq(&tc->pool->lock);
4297 /*
4298 * This synchronize_rcu() call is needed here otherwise we risk a
4299 * wake_worker() call finding no bios to process (because the newly
4300 * added tc isn't yet visible). So this reduces latency since we
4301 * aren't then dependent on the periodic commit to wake_worker().
4302 */
4303 synchronize_rcu();
4304
4305 dm_put(pool_md);
4306
4307 return 0;
4308
4309bad:
4310 dm_pool_close_thin_device(tc->td);
4311bad_pool:
4312 __pool_dec(tc->pool);
4313bad_pool_lookup:
4314 dm_put(pool_md);
4315bad_common:
4316 dm_put_device(ti, tc->pool_dev);
4317bad_pool_dev:
4318 if (tc->origin_dev)
4319 dm_put_device(ti, tc->origin_dev);
4320bad_origin_dev:
4321 kfree(tc);
4322out_unlock:
4323 mutex_unlock(&dm_thin_pool_table.mutex);
4324
4325 return r;
4326}
4327
4328static int thin_map(struct dm_target *ti, struct bio *bio)
4329{
4330 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4331
4332 return thin_bio_map(ti, bio);
4333}
4334
4335static int thin_endio(struct dm_target *ti, struct bio *bio,
4336 blk_status_t *err)
4337{
4338 unsigned long flags;
4339 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4340 struct list_head work;
4341 struct dm_thin_new_mapping *m, *tmp;
4342 struct pool *pool = h->tc->pool;
4343
4344 if (h->shared_read_entry) {
4345 INIT_LIST_HEAD(&work);
4346 dm_deferred_entry_dec(h->shared_read_entry, &work);
4347
4348 spin_lock_irqsave(&pool->lock, flags);
4349 list_for_each_entry_safe(m, tmp, &work, list) {
4350 list_del(&m->list);
4351 __complete_mapping_preparation(m);
4352 }
4353 spin_unlock_irqrestore(&pool->lock, flags);
4354 }
4355
4356 if (h->all_io_entry) {
4357 INIT_LIST_HEAD(&work);
4358 dm_deferred_entry_dec(h->all_io_entry, &work);
4359 if (!list_empty(&work)) {
4360 spin_lock_irqsave(&pool->lock, flags);
4361 list_for_each_entry_safe(m, tmp, &work, list)
4362 list_add_tail(&m->list, &pool->prepared_discards);
4363 spin_unlock_irqrestore(&pool->lock, flags);
4364 wake_worker(pool);
4365 }
4366 }
4367
4368 if (h->cell)
4369 cell_defer_no_holder(h->tc, h->cell);
4370
4371 return DM_ENDIO_DONE;
4372}
4373
4374static void thin_presuspend(struct dm_target *ti)
4375{
4376 struct thin_c *tc = ti->private;
4377
4378 if (dm_noflush_suspending(ti))
4379 noflush_work(tc, do_noflush_start);
4380}
4381
4382static void thin_postsuspend(struct dm_target *ti)
4383{
4384 struct thin_c *tc = ti->private;
4385
4386 /*
4387 * The dm_noflush_suspending flag has been cleared by now, so
4388 * unfortunately we must always run this.
4389 */
4390 noflush_work(tc, do_noflush_stop);
4391}
4392
4393static int thin_preresume(struct dm_target *ti)
4394{
4395 struct thin_c *tc = ti->private;
4396
4397 if (tc->origin_dev)
4398 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4399
4400 return 0;
4401}
4402
4403/*
4404 * <nr mapped sectors> <highest mapped sector>
4405 */
4406static void thin_status(struct dm_target *ti, status_type_t type,
4407 unsigned int status_flags, char *result, unsigned int maxlen)
4408{
4409 int r;
4410 ssize_t sz = 0;
4411 dm_block_t mapped, highest;
4412 char buf[BDEVNAME_SIZE];
4413 struct thin_c *tc = ti->private;
4414
4415 if (get_pool_mode(tc->pool) == PM_FAIL) {
4416 DMEMIT("Fail");
4417 return;
4418 }
4419
4420 if (!tc->td)
4421 DMEMIT("-");
4422 else {
4423 switch (type) {
4424 case STATUSTYPE_INFO:
4425 r = dm_thin_get_mapped_count(tc->td, &mapped);
4426 if (r) {
4427 DMERR("dm_thin_get_mapped_count returned %d", r);
4428 goto err;
4429 }
4430
4431 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4432 if (r < 0) {
4433 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4434 goto err;
4435 }
4436
4437 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4438 if (r)
4439 DMEMIT("%llu", ((highest + 1) *
4440 tc->pool->sectors_per_block) - 1);
4441 else
4442 DMEMIT("-");
4443 break;
4444
4445 case STATUSTYPE_TABLE:
4446 DMEMIT("%s %lu",
4447 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4448 (unsigned long) tc->dev_id);
4449 if (tc->origin_dev)
4450 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4451 break;
4452
4453 case STATUSTYPE_IMA:
4454 *result = '\0';
4455 break;
4456 }
4457 }
4458
4459 return;
4460
4461err:
4462 DMEMIT("Error");
4463}
4464
4465static int thin_iterate_devices(struct dm_target *ti,
4466 iterate_devices_callout_fn fn, void *data)
4467{
4468 sector_t blocks;
4469 struct thin_c *tc = ti->private;
4470 struct pool *pool = tc->pool;
4471
4472 /*
4473 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4474 * we follow a more convoluted path through to the pool's target.
4475 */
4476 if (!pool->ti)
4477 return 0; /* nothing is bound */
4478
4479 blocks = pool->ti->len;
4480 (void) sector_div(blocks, pool->sectors_per_block);
4481 if (blocks)
4482 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4483
4484 return 0;
4485}
4486
4487static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4488{
4489 struct thin_c *tc = ti->private;
4490 struct pool *pool = tc->pool;
4491
4492 if (pool->pf.discard_enabled) {
4493 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4494 limits->max_hw_discard_sectors = pool->sectors_per_block * BIO_PRISON_MAX_RANGE;
4495 }
4496}
4497
4498static struct target_type thin_target = {
4499 .name = "thin",
4500 .version = {1, 23, 0},
4501 .module = THIS_MODULE,
4502 .ctr = thin_ctr,
4503 .dtr = thin_dtr,
4504 .map = thin_map,
4505 .end_io = thin_endio,
4506 .preresume = thin_preresume,
4507 .presuspend = thin_presuspend,
4508 .postsuspend = thin_postsuspend,
4509 .status = thin_status,
4510 .iterate_devices = thin_iterate_devices,
4511 .io_hints = thin_io_hints,
4512};
4513
4514/*----------------------------------------------------------------*/
4515
4516static int __init dm_thin_init(void)
4517{
4518 int r = -ENOMEM;
4519
4520 pool_table_init();
4521
4522 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4523 if (!_new_mapping_cache)
4524 return r;
4525
4526 r = dm_register_target(&thin_target);
4527 if (r)
4528 goto bad_new_mapping_cache;
4529
4530 r = dm_register_target(&pool_target);
4531 if (r)
4532 goto bad_thin_target;
4533
4534 return 0;
4535
4536bad_thin_target:
4537 dm_unregister_target(&thin_target);
4538bad_new_mapping_cache:
4539 kmem_cache_destroy(_new_mapping_cache);
4540
4541 return r;
4542}
4543
4544static void dm_thin_exit(void)
4545{
4546 dm_unregister_target(&thin_target);
4547 dm_unregister_target(&pool_target);
4548
4549 kmem_cache_destroy(_new_mapping_cache);
4550
4551 pool_table_exit();
4552}
4553
4554module_init(dm_thin_init);
4555module_exit(dm_thin_exit);
4556
4557module_param_named(no_space_timeout, no_space_timeout_secs, uint, 0644);
4558MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4559
4560MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4561MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
4562MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2011-2012 Red Hat UK.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm-thin-metadata.h"
8#include "dm-bio-prison-v1.h"
9#include "dm.h"
10
11#include <linux/device-mapper.h>
12#include <linux/dm-io.h>
13#include <linux/dm-kcopyd.h>
14#include <linux/jiffies.h>
15#include <linux/log2.h>
16#include <linux/list.h>
17#include <linux/rculist.h>
18#include <linux/init.h>
19#include <linux/module.h>
20#include <linux/slab.h>
21#include <linux/vmalloc.h>
22#include <linux/sort.h>
23#include <linux/rbtree.h>
24
25#define DM_MSG_PREFIX "thin"
26
27/*
28 * Tunable constants
29 */
30#define ENDIO_HOOK_POOL_SIZE 1024
31#define MAPPING_POOL_SIZE 1024
32#define COMMIT_PERIOD HZ
33#define NO_SPACE_TIMEOUT_SECS 60
34
35static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
36
37DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
38 "A percentage of time allocated for copy on write");
39
40/*
41 * The block size of the device holding pool data must be
42 * between 64KB and 1GB.
43 */
44#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
45#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
46
47/*
48 * Device id is restricted to 24 bits.
49 */
50#define MAX_DEV_ID ((1 << 24) - 1)
51
52/*
53 * How do we handle breaking sharing of data blocks?
54 * =================================================
55 *
56 * We use a standard copy-on-write btree to store the mappings for the
57 * devices (note I'm talking about copy-on-write of the metadata here, not
58 * the data). When you take an internal snapshot you clone the root node
59 * of the origin btree. After this there is no concept of an origin or a
60 * snapshot. They are just two device trees that happen to point to the
61 * same data blocks.
62 *
63 * When we get a write in we decide if it's to a shared data block using
64 * some timestamp magic. If it is, we have to break sharing.
65 *
66 * Let's say we write to a shared block in what was the origin. The
67 * steps are:
68 *
69 * i) plug io further to this physical block. (see bio_prison code).
70 *
71 * ii) quiesce any read io to that shared data block. Obviously
72 * including all devices that share this block. (see dm_deferred_set code)
73 *
74 * iii) copy the data block to a newly allocate block. This step can be
75 * missed out if the io covers the block. (schedule_copy).
76 *
77 * iv) insert the new mapping into the origin's btree
78 * (process_prepared_mapping). This act of inserting breaks some
79 * sharing of btree nodes between the two devices. Breaking sharing only
80 * effects the btree of that specific device. Btrees for the other
81 * devices that share the block never change. The btree for the origin
82 * device as it was after the last commit is untouched, ie. we're using
83 * persistent data structures in the functional programming sense.
84 *
85 * v) unplug io to this physical block, including the io that triggered
86 * the breaking of sharing.
87 *
88 * Steps (ii) and (iii) occur in parallel.
89 *
90 * The metadata _doesn't_ need to be committed before the io continues. We
91 * get away with this because the io is always written to a _new_ block.
92 * If there's a crash, then:
93 *
94 * - The origin mapping will point to the old origin block (the shared
95 * one). This will contain the data as it was before the io that triggered
96 * the breaking of sharing came in.
97 *
98 * - The snap mapping still points to the old block. As it would after
99 * the commit.
100 *
101 * The downside of this scheme is the timestamp magic isn't perfect, and
102 * will continue to think that data block in the snapshot device is shared
103 * even after the write to the origin has broken sharing. I suspect data
104 * blocks will typically be shared by many different devices, so we're
105 * breaking sharing n + 1 times, rather than n, where n is the number of
106 * devices that reference this data block. At the moment I think the
107 * benefits far, far outweigh the disadvantages.
108 */
109
110/*----------------------------------------------------------------*/
111
112/*
113 * Key building.
114 */
115enum lock_space {
116 VIRTUAL,
117 PHYSICAL
118};
119
120static void build_key(struct dm_thin_device *td, enum lock_space ls,
121 dm_block_t b, dm_block_t e, struct dm_cell_key *key)
122{
123 key->virtual = (ls == VIRTUAL);
124 key->dev = dm_thin_dev_id(td);
125 key->block_begin = b;
126 key->block_end = e;
127}
128
129static void build_data_key(struct dm_thin_device *td, dm_block_t b,
130 struct dm_cell_key *key)
131{
132 build_key(td, PHYSICAL, b, b + 1llu, key);
133}
134
135static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
136 struct dm_cell_key *key)
137{
138 build_key(td, VIRTUAL, b, b + 1llu, key);
139}
140
141/*----------------------------------------------------------------*/
142
143#define THROTTLE_THRESHOLD (1 * HZ)
144
145struct throttle {
146 struct rw_semaphore lock;
147 unsigned long threshold;
148 bool throttle_applied;
149};
150
151static void throttle_init(struct throttle *t)
152{
153 init_rwsem(&t->lock);
154 t->throttle_applied = false;
155}
156
157static void throttle_work_start(struct throttle *t)
158{
159 t->threshold = jiffies + THROTTLE_THRESHOLD;
160}
161
162static void throttle_work_update(struct throttle *t)
163{
164 if (!t->throttle_applied && jiffies > t->threshold) {
165 down_write(&t->lock);
166 t->throttle_applied = true;
167 }
168}
169
170static void throttle_work_complete(struct throttle *t)
171{
172 if (t->throttle_applied) {
173 t->throttle_applied = false;
174 up_write(&t->lock);
175 }
176}
177
178static void throttle_lock(struct throttle *t)
179{
180 down_read(&t->lock);
181}
182
183static void throttle_unlock(struct throttle *t)
184{
185 up_read(&t->lock);
186}
187
188/*----------------------------------------------------------------*/
189
190/*
191 * A pool device ties together a metadata device and a data device. It
192 * also provides the interface for creating and destroying internal
193 * devices.
194 */
195struct dm_thin_new_mapping;
196
197/*
198 * The pool runs in various modes. Ordered in degraded order for comparisons.
199 */
200enum pool_mode {
201 PM_WRITE, /* metadata may be changed */
202 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
203
204 /*
205 * Like READ_ONLY, except may switch back to WRITE on metadata resize. Reported as READ_ONLY.
206 */
207 PM_OUT_OF_METADATA_SPACE,
208 PM_READ_ONLY, /* metadata may not be changed */
209
210 PM_FAIL, /* all I/O fails */
211};
212
213struct pool_features {
214 enum pool_mode mode;
215
216 bool zero_new_blocks:1;
217 bool discard_enabled:1;
218 bool discard_passdown:1;
219 bool error_if_no_space:1;
220};
221
222struct thin_c;
223typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
224typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
225typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
226
227#define CELL_SORT_ARRAY_SIZE 8192
228
229struct pool {
230 struct list_head list;
231 struct dm_target *ti; /* Only set if a pool target is bound */
232
233 struct mapped_device *pool_md;
234 struct block_device *md_dev;
235 struct dm_pool_metadata *pmd;
236
237 dm_block_t low_water_blocks;
238 uint32_t sectors_per_block;
239 int sectors_per_block_shift;
240
241 struct pool_features pf;
242 bool low_water_triggered:1; /* A dm event has been sent */
243 bool suspended:1;
244 bool out_of_data_space:1;
245
246 struct dm_bio_prison *prison;
247 struct dm_kcopyd_client *copier;
248
249 struct work_struct worker;
250 struct workqueue_struct *wq;
251 struct throttle throttle;
252 struct delayed_work waker;
253 struct delayed_work no_space_timeout;
254
255 unsigned long last_commit_jiffies;
256 unsigned ref_count;
257
258 spinlock_t lock;
259 struct bio_list deferred_flush_bios;
260 struct bio_list deferred_flush_completions;
261 struct list_head prepared_mappings;
262 struct list_head prepared_discards;
263 struct list_head prepared_discards_pt2;
264 struct list_head active_thins;
265
266 struct dm_deferred_set *shared_read_ds;
267 struct dm_deferred_set *all_io_ds;
268
269 struct dm_thin_new_mapping *next_mapping;
270
271 process_bio_fn process_bio;
272 process_bio_fn process_discard;
273
274 process_cell_fn process_cell;
275 process_cell_fn process_discard_cell;
276
277 process_mapping_fn process_prepared_mapping;
278 process_mapping_fn process_prepared_discard;
279 process_mapping_fn process_prepared_discard_pt2;
280
281 struct dm_bio_prison_cell **cell_sort_array;
282
283 mempool_t mapping_pool;
284};
285
286static void metadata_operation_failed(struct pool *pool, const char *op, int r);
287
288static enum pool_mode get_pool_mode(struct pool *pool)
289{
290 return pool->pf.mode;
291}
292
293static void notify_of_pool_mode_change(struct pool *pool)
294{
295 const char *descs[] = {
296 "write",
297 "out-of-data-space",
298 "read-only",
299 "read-only",
300 "fail"
301 };
302 const char *extra_desc = NULL;
303 enum pool_mode mode = get_pool_mode(pool);
304
305 if (mode == PM_OUT_OF_DATA_SPACE) {
306 if (!pool->pf.error_if_no_space)
307 extra_desc = " (queue IO)";
308 else
309 extra_desc = " (error IO)";
310 }
311
312 dm_table_event(pool->ti->table);
313 DMINFO("%s: switching pool to %s%s mode",
314 dm_device_name(pool->pool_md),
315 descs[(int)mode], extra_desc ? : "");
316}
317
318/*
319 * Target context for a pool.
320 */
321struct pool_c {
322 struct dm_target *ti;
323 struct pool *pool;
324 struct dm_dev *data_dev;
325 struct dm_dev *metadata_dev;
326 struct dm_target_callbacks callbacks;
327
328 dm_block_t low_water_blocks;
329 struct pool_features requested_pf; /* Features requested during table load */
330 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
331};
332
333/*
334 * Target context for a thin.
335 */
336struct thin_c {
337 struct list_head list;
338 struct dm_dev *pool_dev;
339 struct dm_dev *origin_dev;
340 sector_t origin_size;
341 dm_thin_id dev_id;
342
343 struct pool *pool;
344 struct dm_thin_device *td;
345 struct mapped_device *thin_md;
346
347 bool requeue_mode:1;
348 spinlock_t lock;
349 struct list_head deferred_cells;
350 struct bio_list deferred_bio_list;
351 struct bio_list retry_on_resume_list;
352 struct rb_root sort_bio_list; /* sorted list of deferred bios */
353
354 /*
355 * Ensures the thin is not destroyed until the worker has finished
356 * iterating the active_thins list.
357 */
358 refcount_t refcount;
359 struct completion can_destroy;
360};
361
362/*----------------------------------------------------------------*/
363
364static bool block_size_is_power_of_two(struct pool *pool)
365{
366 return pool->sectors_per_block_shift >= 0;
367}
368
369static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
370{
371 return block_size_is_power_of_two(pool) ?
372 (b << pool->sectors_per_block_shift) :
373 (b * pool->sectors_per_block);
374}
375
376/*----------------------------------------------------------------*/
377
378struct discard_op {
379 struct thin_c *tc;
380 struct blk_plug plug;
381 struct bio *parent_bio;
382 struct bio *bio;
383};
384
385static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
386{
387 BUG_ON(!parent);
388
389 op->tc = tc;
390 blk_start_plug(&op->plug);
391 op->parent_bio = parent;
392 op->bio = NULL;
393}
394
395static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
396{
397 struct thin_c *tc = op->tc;
398 sector_t s = block_to_sectors(tc->pool, data_b);
399 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
400
401 return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
402 GFP_NOWAIT, 0, &op->bio);
403}
404
405static void end_discard(struct discard_op *op, int r)
406{
407 if (op->bio) {
408 /*
409 * Even if one of the calls to issue_discard failed, we
410 * need to wait for the chain to complete.
411 */
412 bio_chain(op->bio, op->parent_bio);
413 bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
414 submit_bio(op->bio);
415 }
416
417 blk_finish_plug(&op->plug);
418
419 /*
420 * Even if r is set, there could be sub discards in flight that we
421 * need to wait for.
422 */
423 if (r && !op->parent_bio->bi_status)
424 op->parent_bio->bi_status = errno_to_blk_status(r);
425 bio_endio(op->parent_bio);
426}
427
428/*----------------------------------------------------------------*/
429
430/*
431 * wake_worker() is used when new work is queued and when pool_resume is
432 * ready to continue deferred IO processing.
433 */
434static void wake_worker(struct pool *pool)
435{
436 queue_work(pool->wq, &pool->worker);
437}
438
439/*----------------------------------------------------------------*/
440
441static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
442 struct dm_bio_prison_cell **cell_result)
443{
444 int r;
445 struct dm_bio_prison_cell *cell_prealloc;
446
447 /*
448 * Allocate a cell from the prison's mempool.
449 * This might block but it can't fail.
450 */
451 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
452
453 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
454 if (r)
455 /*
456 * We reused an old cell; we can get rid of
457 * the new one.
458 */
459 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
460
461 return r;
462}
463
464static void cell_release(struct pool *pool,
465 struct dm_bio_prison_cell *cell,
466 struct bio_list *bios)
467{
468 dm_cell_release(pool->prison, cell, bios);
469 dm_bio_prison_free_cell(pool->prison, cell);
470}
471
472static void cell_visit_release(struct pool *pool,
473 void (*fn)(void *, struct dm_bio_prison_cell *),
474 void *context,
475 struct dm_bio_prison_cell *cell)
476{
477 dm_cell_visit_release(pool->prison, fn, context, cell);
478 dm_bio_prison_free_cell(pool->prison, cell);
479}
480
481static void cell_release_no_holder(struct pool *pool,
482 struct dm_bio_prison_cell *cell,
483 struct bio_list *bios)
484{
485 dm_cell_release_no_holder(pool->prison, cell, bios);
486 dm_bio_prison_free_cell(pool->prison, cell);
487}
488
489static void cell_error_with_code(struct pool *pool,
490 struct dm_bio_prison_cell *cell, blk_status_t error_code)
491{
492 dm_cell_error(pool->prison, cell, error_code);
493 dm_bio_prison_free_cell(pool->prison, cell);
494}
495
496static blk_status_t get_pool_io_error_code(struct pool *pool)
497{
498 return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
499}
500
501static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
502{
503 cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
504}
505
506static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
507{
508 cell_error_with_code(pool, cell, 0);
509}
510
511static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
512{
513 cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
514}
515
516/*----------------------------------------------------------------*/
517
518/*
519 * A global list of pools that uses a struct mapped_device as a key.
520 */
521static struct dm_thin_pool_table {
522 struct mutex mutex;
523 struct list_head pools;
524} dm_thin_pool_table;
525
526static void pool_table_init(void)
527{
528 mutex_init(&dm_thin_pool_table.mutex);
529 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
530}
531
532static void pool_table_exit(void)
533{
534 mutex_destroy(&dm_thin_pool_table.mutex);
535}
536
537static void __pool_table_insert(struct pool *pool)
538{
539 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
540 list_add(&pool->list, &dm_thin_pool_table.pools);
541}
542
543static void __pool_table_remove(struct pool *pool)
544{
545 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
546 list_del(&pool->list);
547}
548
549static struct pool *__pool_table_lookup(struct mapped_device *md)
550{
551 struct pool *pool = NULL, *tmp;
552
553 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
554
555 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
556 if (tmp->pool_md == md) {
557 pool = tmp;
558 break;
559 }
560 }
561
562 return pool;
563}
564
565static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
566{
567 struct pool *pool = NULL, *tmp;
568
569 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
570
571 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
572 if (tmp->md_dev == md_dev) {
573 pool = tmp;
574 break;
575 }
576 }
577
578 return pool;
579}
580
581/*----------------------------------------------------------------*/
582
583struct dm_thin_endio_hook {
584 struct thin_c *tc;
585 struct dm_deferred_entry *shared_read_entry;
586 struct dm_deferred_entry *all_io_entry;
587 struct dm_thin_new_mapping *overwrite_mapping;
588 struct rb_node rb_node;
589 struct dm_bio_prison_cell *cell;
590};
591
592static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
593{
594 bio_list_merge(bios, master);
595 bio_list_init(master);
596}
597
598static void error_bio_list(struct bio_list *bios, blk_status_t error)
599{
600 struct bio *bio;
601
602 while ((bio = bio_list_pop(bios))) {
603 bio->bi_status = error;
604 bio_endio(bio);
605 }
606}
607
608static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
609 blk_status_t error)
610{
611 struct bio_list bios;
612 unsigned long flags;
613
614 bio_list_init(&bios);
615
616 spin_lock_irqsave(&tc->lock, flags);
617 __merge_bio_list(&bios, master);
618 spin_unlock_irqrestore(&tc->lock, flags);
619
620 error_bio_list(&bios, error);
621}
622
623static void requeue_deferred_cells(struct thin_c *tc)
624{
625 struct pool *pool = tc->pool;
626 unsigned long flags;
627 struct list_head cells;
628 struct dm_bio_prison_cell *cell, *tmp;
629
630 INIT_LIST_HEAD(&cells);
631
632 spin_lock_irqsave(&tc->lock, flags);
633 list_splice_init(&tc->deferred_cells, &cells);
634 spin_unlock_irqrestore(&tc->lock, flags);
635
636 list_for_each_entry_safe(cell, tmp, &cells, user_list)
637 cell_requeue(pool, cell);
638}
639
640static void requeue_io(struct thin_c *tc)
641{
642 struct bio_list bios;
643 unsigned long flags;
644
645 bio_list_init(&bios);
646
647 spin_lock_irqsave(&tc->lock, flags);
648 __merge_bio_list(&bios, &tc->deferred_bio_list);
649 __merge_bio_list(&bios, &tc->retry_on_resume_list);
650 spin_unlock_irqrestore(&tc->lock, flags);
651
652 error_bio_list(&bios, BLK_STS_DM_REQUEUE);
653 requeue_deferred_cells(tc);
654}
655
656static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
657{
658 struct thin_c *tc;
659
660 rcu_read_lock();
661 list_for_each_entry_rcu(tc, &pool->active_thins, list)
662 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
663 rcu_read_unlock();
664}
665
666static void error_retry_list(struct pool *pool)
667{
668 error_retry_list_with_code(pool, get_pool_io_error_code(pool));
669}
670
671/*
672 * This section of code contains the logic for processing a thin device's IO.
673 * Much of the code depends on pool object resources (lists, workqueues, etc)
674 * but most is exclusively called from the thin target rather than the thin-pool
675 * target.
676 */
677
678static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
679{
680 struct pool *pool = tc->pool;
681 sector_t block_nr = bio->bi_iter.bi_sector;
682
683 if (block_size_is_power_of_two(pool))
684 block_nr >>= pool->sectors_per_block_shift;
685 else
686 (void) sector_div(block_nr, pool->sectors_per_block);
687
688 return block_nr;
689}
690
691/*
692 * Returns the _complete_ blocks that this bio covers.
693 */
694static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
695 dm_block_t *begin, dm_block_t *end)
696{
697 struct pool *pool = tc->pool;
698 sector_t b = bio->bi_iter.bi_sector;
699 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
700
701 b += pool->sectors_per_block - 1ull; /* so we round up */
702
703 if (block_size_is_power_of_two(pool)) {
704 b >>= pool->sectors_per_block_shift;
705 e >>= pool->sectors_per_block_shift;
706 } else {
707 (void) sector_div(b, pool->sectors_per_block);
708 (void) sector_div(e, pool->sectors_per_block);
709 }
710
711 if (e < b)
712 /* Can happen if the bio is within a single block. */
713 e = b;
714
715 *begin = b;
716 *end = e;
717}
718
719static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
720{
721 struct pool *pool = tc->pool;
722 sector_t bi_sector = bio->bi_iter.bi_sector;
723
724 bio_set_dev(bio, tc->pool_dev->bdev);
725 if (block_size_is_power_of_two(pool))
726 bio->bi_iter.bi_sector =
727 (block << pool->sectors_per_block_shift) |
728 (bi_sector & (pool->sectors_per_block - 1));
729 else
730 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
731 sector_div(bi_sector, pool->sectors_per_block);
732}
733
734static void remap_to_origin(struct thin_c *tc, struct bio *bio)
735{
736 bio_set_dev(bio, tc->origin_dev->bdev);
737}
738
739static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
740{
741 return op_is_flush(bio->bi_opf) &&
742 dm_thin_changed_this_transaction(tc->td);
743}
744
745static void inc_all_io_entry(struct pool *pool, struct bio *bio)
746{
747 struct dm_thin_endio_hook *h;
748
749 if (bio_op(bio) == REQ_OP_DISCARD)
750 return;
751
752 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
753 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
754}
755
756static void issue(struct thin_c *tc, struct bio *bio)
757{
758 struct pool *pool = tc->pool;
759 unsigned long flags;
760
761 if (!bio_triggers_commit(tc, bio)) {
762 generic_make_request(bio);
763 return;
764 }
765
766 /*
767 * Complete bio with an error if earlier I/O caused changes to
768 * the metadata that can't be committed e.g, due to I/O errors
769 * on the metadata device.
770 */
771 if (dm_thin_aborted_changes(tc->td)) {
772 bio_io_error(bio);
773 return;
774 }
775
776 /*
777 * Batch together any bios that trigger commits and then issue a
778 * single commit for them in process_deferred_bios().
779 */
780 spin_lock_irqsave(&pool->lock, flags);
781 bio_list_add(&pool->deferred_flush_bios, bio);
782 spin_unlock_irqrestore(&pool->lock, flags);
783}
784
785static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
786{
787 remap_to_origin(tc, bio);
788 issue(tc, bio);
789}
790
791static void remap_and_issue(struct thin_c *tc, struct bio *bio,
792 dm_block_t block)
793{
794 remap(tc, bio, block);
795 issue(tc, bio);
796}
797
798/*----------------------------------------------------------------*/
799
800/*
801 * Bio endio functions.
802 */
803struct dm_thin_new_mapping {
804 struct list_head list;
805
806 bool pass_discard:1;
807 bool maybe_shared:1;
808
809 /*
810 * Track quiescing, copying and zeroing preparation actions. When this
811 * counter hits zero the block is prepared and can be inserted into the
812 * btree.
813 */
814 atomic_t prepare_actions;
815
816 blk_status_t status;
817 struct thin_c *tc;
818 dm_block_t virt_begin, virt_end;
819 dm_block_t data_block;
820 struct dm_bio_prison_cell *cell;
821
822 /*
823 * If the bio covers the whole area of a block then we can avoid
824 * zeroing or copying. Instead this bio is hooked. The bio will
825 * still be in the cell, so care has to be taken to avoid issuing
826 * the bio twice.
827 */
828 struct bio *bio;
829 bio_end_io_t *saved_bi_end_io;
830};
831
832static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
833{
834 struct pool *pool = m->tc->pool;
835
836 if (atomic_dec_and_test(&m->prepare_actions)) {
837 list_add_tail(&m->list, &pool->prepared_mappings);
838 wake_worker(pool);
839 }
840}
841
842static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
843{
844 unsigned long flags;
845 struct pool *pool = m->tc->pool;
846
847 spin_lock_irqsave(&pool->lock, flags);
848 __complete_mapping_preparation(m);
849 spin_unlock_irqrestore(&pool->lock, flags);
850}
851
852static void copy_complete(int read_err, unsigned long write_err, void *context)
853{
854 struct dm_thin_new_mapping *m = context;
855
856 m->status = read_err || write_err ? BLK_STS_IOERR : 0;
857 complete_mapping_preparation(m);
858}
859
860static void overwrite_endio(struct bio *bio)
861{
862 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
863 struct dm_thin_new_mapping *m = h->overwrite_mapping;
864
865 bio->bi_end_io = m->saved_bi_end_io;
866
867 m->status = bio->bi_status;
868 complete_mapping_preparation(m);
869}
870
871/*----------------------------------------------------------------*/
872
873/*
874 * Workqueue.
875 */
876
877/*
878 * Prepared mapping jobs.
879 */
880
881/*
882 * This sends the bios in the cell, except the original holder, back
883 * to the deferred_bios list.
884 */
885static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
886{
887 struct pool *pool = tc->pool;
888 unsigned long flags;
889
890 spin_lock_irqsave(&tc->lock, flags);
891 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
892 spin_unlock_irqrestore(&tc->lock, flags);
893
894 wake_worker(pool);
895}
896
897static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
898
899struct remap_info {
900 struct thin_c *tc;
901 struct bio_list defer_bios;
902 struct bio_list issue_bios;
903};
904
905static void __inc_remap_and_issue_cell(void *context,
906 struct dm_bio_prison_cell *cell)
907{
908 struct remap_info *info = context;
909 struct bio *bio;
910
911 while ((bio = bio_list_pop(&cell->bios))) {
912 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
913 bio_list_add(&info->defer_bios, bio);
914 else {
915 inc_all_io_entry(info->tc->pool, bio);
916
917 /*
918 * We can't issue the bios with the bio prison lock
919 * held, so we add them to a list to issue on
920 * return from this function.
921 */
922 bio_list_add(&info->issue_bios, bio);
923 }
924 }
925}
926
927static void inc_remap_and_issue_cell(struct thin_c *tc,
928 struct dm_bio_prison_cell *cell,
929 dm_block_t block)
930{
931 struct bio *bio;
932 struct remap_info info;
933
934 info.tc = tc;
935 bio_list_init(&info.defer_bios);
936 bio_list_init(&info.issue_bios);
937
938 /*
939 * We have to be careful to inc any bios we're about to issue
940 * before the cell is released, and avoid a race with new bios
941 * being added to the cell.
942 */
943 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
944 &info, cell);
945
946 while ((bio = bio_list_pop(&info.defer_bios)))
947 thin_defer_bio(tc, bio);
948
949 while ((bio = bio_list_pop(&info.issue_bios)))
950 remap_and_issue(info.tc, bio, block);
951}
952
953static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
954{
955 cell_error(m->tc->pool, m->cell);
956 list_del(&m->list);
957 mempool_free(m, &m->tc->pool->mapping_pool);
958}
959
960static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
961{
962 struct pool *pool = tc->pool;
963 unsigned long flags;
964
965 /*
966 * If the bio has the REQ_FUA flag set we must commit the metadata
967 * before signaling its completion.
968 */
969 if (!bio_triggers_commit(tc, bio)) {
970 bio_endio(bio);
971 return;
972 }
973
974 /*
975 * Complete bio with an error if earlier I/O caused changes to the
976 * metadata that can't be committed, e.g, due to I/O errors on the
977 * metadata device.
978 */
979 if (dm_thin_aborted_changes(tc->td)) {
980 bio_io_error(bio);
981 return;
982 }
983
984 /*
985 * Batch together any bios that trigger commits and then issue a
986 * single commit for them in process_deferred_bios().
987 */
988 spin_lock_irqsave(&pool->lock, flags);
989 bio_list_add(&pool->deferred_flush_completions, bio);
990 spin_unlock_irqrestore(&pool->lock, flags);
991}
992
993static void process_prepared_mapping(struct dm_thin_new_mapping *m)
994{
995 struct thin_c *tc = m->tc;
996 struct pool *pool = tc->pool;
997 struct bio *bio = m->bio;
998 int r;
999
1000 if (m->status) {
1001 cell_error(pool, m->cell);
1002 goto out;
1003 }
1004
1005 /*
1006 * Commit the prepared block into the mapping btree.
1007 * Any I/O for this block arriving after this point will get
1008 * remapped to it directly.
1009 */
1010 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1011 if (r) {
1012 metadata_operation_failed(pool, "dm_thin_insert_block", r);
1013 cell_error(pool, m->cell);
1014 goto out;
1015 }
1016
1017 /*
1018 * Release any bios held while the block was being provisioned.
1019 * If we are processing a write bio that completely covers the block,
1020 * we already processed it so can ignore it now when processing
1021 * the bios in the cell.
1022 */
1023 if (bio) {
1024 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1025 complete_overwrite_bio(tc, bio);
1026 } else {
1027 inc_all_io_entry(tc->pool, m->cell->holder);
1028 remap_and_issue(tc, m->cell->holder, m->data_block);
1029 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1030 }
1031
1032out:
1033 list_del(&m->list);
1034 mempool_free(m, &pool->mapping_pool);
1035}
1036
1037/*----------------------------------------------------------------*/
1038
1039static void free_discard_mapping(struct dm_thin_new_mapping *m)
1040{
1041 struct thin_c *tc = m->tc;
1042 if (m->cell)
1043 cell_defer_no_holder(tc, m->cell);
1044 mempool_free(m, &tc->pool->mapping_pool);
1045}
1046
1047static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1048{
1049 bio_io_error(m->bio);
1050 free_discard_mapping(m);
1051}
1052
1053static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1054{
1055 bio_endio(m->bio);
1056 free_discard_mapping(m);
1057}
1058
1059static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1060{
1061 int r;
1062 struct thin_c *tc = m->tc;
1063
1064 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1065 if (r) {
1066 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1067 bio_io_error(m->bio);
1068 } else
1069 bio_endio(m->bio);
1070
1071 cell_defer_no_holder(tc, m->cell);
1072 mempool_free(m, &tc->pool->mapping_pool);
1073}
1074
1075/*----------------------------------------------------------------*/
1076
1077static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1078 struct bio *discard_parent)
1079{
1080 /*
1081 * We've already unmapped this range of blocks, but before we
1082 * passdown we have to check that these blocks are now unused.
1083 */
1084 int r = 0;
1085 bool shared = true;
1086 struct thin_c *tc = m->tc;
1087 struct pool *pool = tc->pool;
1088 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1089 struct discard_op op;
1090
1091 begin_discard(&op, tc, discard_parent);
1092 while (b != end) {
1093 /* find start of unmapped run */
1094 for (; b < end; b++) {
1095 r = dm_pool_block_is_shared(pool->pmd, b, &shared);
1096 if (r)
1097 goto out;
1098
1099 if (!shared)
1100 break;
1101 }
1102
1103 if (b == end)
1104 break;
1105
1106 /* find end of run */
1107 for (e = b + 1; e != end; e++) {
1108 r = dm_pool_block_is_shared(pool->pmd, e, &shared);
1109 if (r)
1110 goto out;
1111
1112 if (shared)
1113 break;
1114 }
1115
1116 r = issue_discard(&op, b, e);
1117 if (r)
1118 goto out;
1119
1120 b = e;
1121 }
1122out:
1123 end_discard(&op, r);
1124}
1125
1126static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1127{
1128 unsigned long flags;
1129 struct pool *pool = m->tc->pool;
1130
1131 spin_lock_irqsave(&pool->lock, flags);
1132 list_add_tail(&m->list, &pool->prepared_discards_pt2);
1133 spin_unlock_irqrestore(&pool->lock, flags);
1134 wake_worker(pool);
1135}
1136
1137static void passdown_endio(struct bio *bio)
1138{
1139 /*
1140 * It doesn't matter if the passdown discard failed, we still want
1141 * to unmap (we ignore err).
1142 */
1143 queue_passdown_pt2(bio->bi_private);
1144 bio_put(bio);
1145}
1146
1147static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1148{
1149 int r;
1150 struct thin_c *tc = m->tc;
1151 struct pool *pool = tc->pool;
1152 struct bio *discard_parent;
1153 dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1154
1155 /*
1156 * Only this thread allocates blocks, so we can be sure that the
1157 * newly unmapped blocks will not be allocated before the end of
1158 * the function.
1159 */
1160 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1161 if (r) {
1162 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1163 bio_io_error(m->bio);
1164 cell_defer_no_holder(tc, m->cell);
1165 mempool_free(m, &pool->mapping_pool);
1166 return;
1167 }
1168
1169 /*
1170 * Increment the unmapped blocks. This prevents a race between the
1171 * passdown io and reallocation of freed blocks.
1172 */
1173 r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1174 if (r) {
1175 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1176 bio_io_error(m->bio);
1177 cell_defer_no_holder(tc, m->cell);
1178 mempool_free(m, &pool->mapping_pool);
1179 return;
1180 }
1181
1182 discard_parent = bio_alloc(GFP_NOIO, 1);
1183 if (!discard_parent) {
1184 DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1185 dm_device_name(tc->pool->pool_md));
1186 queue_passdown_pt2(m);
1187
1188 } else {
1189 discard_parent->bi_end_io = passdown_endio;
1190 discard_parent->bi_private = m;
1191
1192 if (m->maybe_shared)
1193 passdown_double_checking_shared_status(m, discard_parent);
1194 else {
1195 struct discard_op op;
1196
1197 begin_discard(&op, tc, discard_parent);
1198 r = issue_discard(&op, m->data_block, data_end);
1199 end_discard(&op, r);
1200 }
1201 }
1202}
1203
1204static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1205{
1206 int r;
1207 struct thin_c *tc = m->tc;
1208 struct pool *pool = tc->pool;
1209
1210 /*
1211 * The passdown has completed, so now we can decrement all those
1212 * unmapped blocks.
1213 */
1214 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1215 m->data_block + (m->virt_end - m->virt_begin));
1216 if (r) {
1217 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1218 bio_io_error(m->bio);
1219 } else
1220 bio_endio(m->bio);
1221
1222 cell_defer_no_holder(tc, m->cell);
1223 mempool_free(m, &pool->mapping_pool);
1224}
1225
1226static void process_prepared(struct pool *pool, struct list_head *head,
1227 process_mapping_fn *fn)
1228{
1229 unsigned long flags;
1230 struct list_head maps;
1231 struct dm_thin_new_mapping *m, *tmp;
1232
1233 INIT_LIST_HEAD(&maps);
1234 spin_lock_irqsave(&pool->lock, flags);
1235 list_splice_init(head, &maps);
1236 spin_unlock_irqrestore(&pool->lock, flags);
1237
1238 list_for_each_entry_safe(m, tmp, &maps, list)
1239 (*fn)(m);
1240}
1241
1242/*
1243 * Deferred bio jobs.
1244 */
1245static int io_overlaps_block(struct pool *pool, struct bio *bio)
1246{
1247 return bio->bi_iter.bi_size ==
1248 (pool->sectors_per_block << SECTOR_SHIFT);
1249}
1250
1251static int io_overwrites_block(struct pool *pool, struct bio *bio)
1252{
1253 return (bio_data_dir(bio) == WRITE) &&
1254 io_overlaps_block(pool, bio);
1255}
1256
1257static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1258 bio_end_io_t *fn)
1259{
1260 *save = bio->bi_end_io;
1261 bio->bi_end_io = fn;
1262}
1263
1264static int ensure_next_mapping(struct pool *pool)
1265{
1266 if (pool->next_mapping)
1267 return 0;
1268
1269 pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1270
1271 return pool->next_mapping ? 0 : -ENOMEM;
1272}
1273
1274static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1275{
1276 struct dm_thin_new_mapping *m = pool->next_mapping;
1277
1278 BUG_ON(!pool->next_mapping);
1279
1280 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1281 INIT_LIST_HEAD(&m->list);
1282 m->bio = NULL;
1283
1284 pool->next_mapping = NULL;
1285
1286 return m;
1287}
1288
1289static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1290 sector_t begin, sector_t end)
1291{
1292 struct dm_io_region to;
1293
1294 to.bdev = tc->pool_dev->bdev;
1295 to.sector = begin;
1296 to.count = end - begin;
1297
1298 dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1299}
1300
1301static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1302 dm_block_t data_begin,
1303 struct dm_thin_new_mapping *m)
1304{
1305 struct pool *pool = tc->pool;
1306 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1307
1308 h->overwrite_mapping = m;
1309 m->bio = bio;
1310 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1311 inc_all_io_entry(pool, bio);
1312 remap_and_issue(tc, bio, data_begin);
1313}
1314
1315/*
1316 * A partial copy also needs to zero the uncopied region.
1317 */
1318static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1319 struct dm_dev *origin, dm_block_t data_origin,
1320 dm_block_t data_dest,
1321 struct dm_bio_prison_cell *cell, struct bio *bio,
1322 sector_t len)
1323{
1324 struct pool *pool = tc->pool;
1325 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1326
1327 m->tc = tc;
1328 m->virt_begin = virt_block;
1329 m->virt_end = virt_block + 1u;
1330 m->data_block = data_dest;
1331 m->cell = cell;
1332
1333 /*
1334 * quiesce action + copy action + an extra reference held for the
1335 * duration of this function (we may need to inc later for a
1336 * partial zero).
1337 */
1338 atomic_set(&m->prepare_actions, 3);
1339
1340 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1341 complete_mapping_preparation(m); /* already quiesced */
1342
1343 /*
1344 * IO to pool_dev remaps to the pool target's data_dev.
1345 *
1346 * If the whole block of data is being overwritten, we can issue the
1347 * bio immediately. Otherwise we use kcopyd to clone the data first.
1348 */
1349 if (io_overwrites_block(pool, bio))
1350 remap_and_issue_overwrite(tc, bio, data_dest, m);
1351 else {
1352 struct dm_io_region from, to;
1353
1354 from.bdev = origin->bdev;
1355 from.sector = data_origin * pool->sectors_per_block;
1356 from.count = len;
1357
1358 to.bdev = tc->pool_dev->bdev;
1359 to.sector = data_dest * pool->sectors_per_block;
1360 to.count = len;
1361
1362 dm_kcopyd_copy(pool->copier, &from, 1, &to,
1363 0, copy_complete, m);
1364
1365 /*
1366 * Do we need to zero a tail region?
1367 */
1368 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1369 atomic_inc(&m->prepare_actions);
1370 ll_zero(tc, m,
1371 data_dest * pool->sectors_per_block + len,
1372 (data_dest + 1) * pool->sectors_per_block);
1373 }
1374 }
1375
1376 complete_mapping_preparation(m); /* drop our ref */
1377}
1378
1379static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1380 dm_block_t data_origin, dm_block_t data_dest,
1381 struct dm_bio_prison_cell *cell, struct bio *bio)
1382{
1383 schedule_copy(tc, virt_block, tc->pool_dev,
1384 data_origin, data_dest, cell, bio,
1385 tc->pool->sectors_per_block);
1386}
1387
1388static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1389 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1390 struct bio *bio)
1391{
1392 struct pool *pool = tc->pool;
1393 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1394
1395 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1396 m->tc = tc;
1397 m->virt_begin = virt_block;
1398 m->virt_end = virt_block + 1u;
1399 m->data_block = data_block;
1400 m->cell = cell;
1401
1402 /*
1403 * If the whole block of data is being overwritten or we are not
1404 * zeroing pre-existing data, we can issue the bio immediately.
1405 * Otherwise we use kcopyd to zero the data first.
1406 */
1407 if (pool->pf.zero_new_blocks) {
1408 if (io_overwrites_block(pool, bio))
1409 remap_and_issue_overwrite(tc, bio, data_block, m);
1410 else
1411 ll_zero(tc, m, data_block * pool->sectors_per_block,
1412 (data_block + 1) * pool->sectors_per_block);
1413 } else
1414 process_prepared_mapping(m);
1415}
1416
1417static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1418 dm_block_t data_dest,
1419 struct dm_bio_prison_cell *cell, struct bio *bio)
1420{
1421 struct pool *pool = tc->pool;
1422 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1423 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1424
1425 if (virt_block_end <= tc->origin_size)
1426 schedule_copy(tc, virt_block, tc->origin_dev,
1427 virt_block, data_dest, cell, bio,
1428 pool->sectors_per_block);
1429
1430 else if (virt_block_begin < tc->origin_size)
1431 schedule_copy(tc, virt_block, tc->origin_dev,
1432 virt_block, data_dest, cell, bio,
1433 tc->origin_size - virt_block_begin);
1434
1435 else
1436 schedule_zero(tc, virt_block, data_dest, cell, bio);
1437}
1438
1439static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1440
1441static void requeue_bios(struct pool *pool);
1442
1443static bool is_read_only_pool_mode(enum pool_mode mode)
1444{
1445 return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1446}
1447
1448static bool is_read_only(struct pool *pool)
1449{
1450 return is_read_only_pool_mode(get_pool_mode(pool));
1451}
1452
1453static void check_for_metadata_space(struct pool *pool)
1454{
1455 int r;
1456 const char *ooms_reason = NULL;
1457 dm_block_t nr_free;
1458
1459 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1460 if (r)
1461 ooms_reason = "Could not get free metadata blocks";
1462 else if (!nr_free)
1463 ooms_reason = "No free metadata blocks";
1464
1465 if (ooms_reason && !is_read_only(pool)) {
1466 DMERR("%s", ooms_reason);
1467 set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1468 }
1469}
1470
1471static void check_for_data_space(struct pool *pool)
1472{
1473 int r;
1474 dm_block_t nr_free;
1475
1476 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1477 return;
1478
1479 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1480 if (r)
1481 return;
1482
1483 if (nr_free) {
1484 set_pool_mode(pool, PM_WRITE);
1485 requeue_bios(pool);
1486 }
1487}
1488
1489/*
1490 * A non-zero return indicates read_only or fail_io mode.
1491 * Many callers don't care about the return value.
1492 */
1493static int commit(struct pool *pool)
1494{
1495 int r;
1496
1497 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1498 return -EINVAL;
1499
1500 r = dm_pool_commit_metadata(pool->pmd);
1501 if (r)
1502 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1503 else {
1504 check_for_metadata_space(pool);
1505 check_for_data_space(pool);
1506 }
1507
1508 return r;
1509}
1510
1511static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1512{
1513 unsigned long flags;
1514
1515 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1516 DMWARN("%s: reached low water mark for data device: sending event.",
1517 dm_device_name(pool->pool_md));
1518 spin_lock_irqsave(&pool->lock, flags);
1519 pool->low_water_triggered = true;
1520 spin_unlock_irqrestore(&pool->lock, flags);
1521 dm_table_event(pool->ti->table);
1522 }
1523}
1524
1525static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1526{
1527 int r;
1528 dm_block_t free_blocks;
1529 struct pool *pool = tc->pool;
1530
1531 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1532 return -EINVAL;
1533
1534 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1535 if (r) {
1536 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1537 return r;
1538 }
1539
1540 check_low_water_mark(pool, free_blocks);
1541
1542 if (!free_blocks) {
1543 /*
1544 * Try to commit to see if that will free up some
1545 * more space.
1546 */
1547 r = commit(pool);
1548 if (r)
1549 return r;
1550
1551 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1552 if (r) {
1553 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1554 return r;
1555 }
1556
1557 if (!free_blocks) {
1558 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1559 return -ENOSPC;
1560 }
1561 }
1562
1563 r = dm_pool_alloc_data_block(pool->pmd, result);
1564 if (r) {
1565 if (r == -ENOSPC)
1566 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1567 else
1568 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1569 return r;
1570 }
1571
1572 r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1573 if (r) {
1574 metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1575 return r;
1576 }
1577
1578 if (!free_blocks) {
1579 /* Let's commit before we use up the metadata reserve. */
1580 r = commit(pool);
1581 if (r)
1582 return r;
1583 }
1584
1585 return 0;
1586}
1587
1588/*
1589 * If we have run out of space, queue bios until the device is
1590 * resumed, presumably after having been reloaded with more space.
1591 */
1592static void retry_on_resume(struct bio *bio)
1593{
1594 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1595 struct thin_c *tc = h->tc;
1596 unsigned long flags;
1597
1598 spin_lock_irqsave(&tc->lock, flags);
1599 bio_list_add(&tc->retry_on_resume_list, bio);
1600 spin_unlock_irqrestore(&tc->lock, flags);
1601}
1602
1603static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1604{
1605 enum pool_mode m = get_pool_mode(pool);
1606
1607 switch (m) {
1608 case PM_WRITE:
1609 /* Shouldn't get here */
1610 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1611 return BLK_STS_IOERR;
1612
1613 case PM_OUT_OF_DATA_SPACE:
1614 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1615
1616 case PM_OUT_OF_METADATA_SPACE:
1617 case PM_READ_ONLY:
1618 case PM_FAIL:
1619 return BLK_STS_IOERR;
1620 default:
1621 /* Shouldn't get here */
1622 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1623 return BLK_STS_IOERR;
1624 }
1625}
1626
1627static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1628{
1629 blk_status_t error = should_error_unserviceable_bio(pool);
1630
1631 if (error) {
1632 bio->bi_status = error;
1633 bio_endio(bio);
1634 } else
1635 retry_on_resume(bio);
1636}
1637
1638static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1639{
1640 struct bio *bio;
1641 struct bio_list bios;
1642 blk_status_t error;
1643
1644 error = should_error_unserviceable_bio(pool);
1645 if (error) {
1646 cell_error_with_code(pool, cell, error);
1647 return;
1648 }
1649
1650 bio_list_init(&bios);
1651 cell_release(pool, cell, &bios);
1652
1653 while ((bio = bio_list_pop(&bios)))
1654 retry_on_resume(bio);
1655}
1656
1657static void process_discard_cell_no_passdown(struct thin_c *tc,
1658 struct dm_bio_prison_cell *virt_cell)
1659{
1660 struct pool *pool = tc->pool;
1661 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1662
1663 /*
1664 * We don't need to lock the data blocks, since there's no
1665 * passdown. We only lock data blocks for allocation and breaking sharing.
1666 */
1667 m->tc = tc;
1668 m->virt_begin = virt_cell->key.block_begin;
1669 m->virt_end = virt_cell->key.block_end;
1670 m->cell = virt_cell;
1671 m->bio = virt_cell->holder;
1672
1673 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1674 pool->process_prepared_discard(m);
1675}
1676
1677static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1678 struct bio *bio)
1679{
1680 struct pool *pool = tc->pool;
1681
1682 int r;
1683 bool maybe_shared;
1684 struct dm_cell_key data_key;
1685 struct dm_bio_prison_cell *data_cell;
1686 struct dm_thin_new_mapping *m;
1687 dm_block_t virt_begin, virt_end, data_begin;
1688
1689 while (begin != end) {
1690 r = ensure_next_mapping(pool);
1691 if (r)
1692 /* we did our best */
1693 return;
1694
1695 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1696 &data_begin, &maybe_shared);
1697 if (r)
1698 /*
1699 * Silently fail, letting any mappings we've
1700 * created complete.
1701 */
1702 break;
1703
1704 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1705 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1706 /* contention, we'll give up with this range */
1707 begin = virt_end;
1708 continue;
1709 }
1710
1711 /*
1712 * IO may still be going to the destination block. We must
1713 * quiesce before we can do the removal.
1714 */
1715 m = get_next_mapping(pool);
1716 m->tc = tc;
1717 m->maybe_shared = maybe_shared;
1718 m->virt_begin = virt_begin;
1719 m->virt_end = virt_end;
1720 m->data_block = data_begin;
1721 m->cell = data_cell;
1722 m->bio = bio;
1723
1724 /*
1725 * The parent bio must not complete before sub discard bios are
1726 * chained to it (see end_discard's bio_chain)!
1727 *
1728 * This per-mapping bi_remaining increment is paired with
1729 * the implicit decrement that occurs via bio_endio() in
1730 * end_discard().
1731 */
1732 bio_inc_remaining(bio);
1733 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1734 pool->process_prepared_discard(m);
1735
1736 begin = virt_end;
1737 }
1738}
1739
1740static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1741{
1742 struct bio *bio = virt_cell->holder;
1743 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1744
1745 /*
1746 * The virt_cell will only get freed once the origin bio completes.
1747 * This means it will remain locked while all the individual
1748 * passdown bios are in flight.
1749 */
1750 h->cell = virt_cell;
1751 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1752
1753 /*
1754 * We complete the bio now, knowing that the bi_remaining field
1755 * will prevent completion until the sub range discards have
1756 * completed.
1757 */
1758 bio_endio(bio);
1759}
1760
1761static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1762{
1763 dm_block_t begin, end;
1764 struct dm_cell_key virt_key;
1765 struct dm_bio_prison_cell *virt_cell;
1766
1767 get_bio_block_range(tc, bio, &begin, &end);
1768 if (begin == end) {
1769 /*
1770 * The discard covers less than a block.
1771 */
1772 bio_endio(bio);
1773 return;
1774 }
1775
1776 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1777 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1778 /*
1779 * Potential starvation issue: We're relying on the
1780 * fs/application being well behaved, and not trying to
1781 * send IO to a region at the same time as discarding it.
1782 * If they do this persistently then it's possible this
1783 * cell will never be granted.
1784 */
1785 return;
1786
1787 tc->pool->process_discard_cell(tc, virt_cell);
1788}
1789
1790static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1791 struct dm_cell_key *key,
1792 struct dm_thin_lookup_result *lookup_result,
1793 struct dm_bio_prison_cell *cell)
1794{
1795 int r;
1796 dm_block_t data_block;
1797 struct pool *pool = tc->pool;
1798
1799 r = alloc_data_block(tc, &data_block);
1800 switch (r) {
1801 case 0:
1802 schedule_internal_copy(tc, block, lookup_result->block,
1803 data_block, cell, bio);
1804 break;
1805
1806 case -ENOSPC:
1807 retry_bios_on_resume(pool, cell);
1808 break;
1809
1810 default:
1811 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1812 __func__, r);
1813 cell_error(pool, cell);
1814 break;
1815 }
1816}
1817
1818static void __remap_and_issue_shared_cell(void *context,
1819 struct dm_bio_prison_cell *cell)
1820{
1821 struct remap_info *info = context;
1822 struct bio *bio;
1823
1824 while ((bio = bio_list_pop(&cell->bios))) {
1825 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1826 bio_op(bio) == REQ_OP_DISCARD)
1827 bio_list_add(&info->defer_bios, bio);
1828 else {
1829 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1830
1831 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1832 inc_all_io_entry(info->tc->pool, bio);
1833 bio_list_add(&info->issue_bios, bio);
1834 }
1835 }
1836}
1837
1838static void remap_and_issue_shared_cell(struct thin_c *tc,
1839 struct dm_bio_prison_cell *cell,
1840 dm_block_t block)
1841{
1842 struct bio *bio;
1843 struct remap_info info;
1844
1845 info.tc = tc;
1846 bio_list_init(&info.defer_bios);
1847 bio_list_init(&info.issue_bios);
1848
1849 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1850 &info, cell);
1851
1852 while ((bio = bio_list_pop(&info.defer_bios)))
1853 thin_defer_bio(tc, bio);
1854
1855 while ((bio = bio_list_pop(&info.issue_bios)))
1856 remap_and_issue(tc, bio, block);
1857}
1858
1859static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1860 dm_block_t block,
1861 struct dm_thin_lookup_result *lookup_result,
1862 struct dm_bio_prison_cell *virt_cell)
1863{
1864 struct dm_bio_prison_cell *data_cell;
1865 struct pool *pool = tc->pool;
1866 struct dm_cell_key key;
1867
1868 /*
1869 * If cell is already occupied, then sharing is already in the process
1870 * of being broken so we have nothing further to do here.
1871 */
1872 build_data_key(tc->td, lookup_result->block, &key);
1873 if (bio_detain(pool, &key, bio, &data_cell)) {
1874 cell_defer_no_holder(tc, virt_cell);
1875 return;
1876 }
1877
1878 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1879 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1880 cell_defer_no_holder(tc, virt_cell);
1881 } else {
1882 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1883
1884 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1885 inc_all_io_entry(pool, bio);
1886 remap_and_issue(tc, bio, lookup_result->block);
1887
1888 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1889 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1890 }
1891}
1892
1893static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1894 struct dm_bio_prison_cell *cell)
1895{
1896 int r;
1897 dm_block_t data_block;
1898 struct pool *pool = tc->pool;
1899
1900 /*
1901 * Remap empty bios (flushes) immediately, without provisioning.
1902 */
1903 if (!bio->bi_iter.bi_size) {
1904 inc_all_io_entry(pool, bio);
1905 cell_defer_no_holder(tc, cell);
1906
1907 remap_and_issue(tc, bio, 0);
1908 return;
1909 }
1910
1911 /*
1912 * Fill read bios with zeroes and complete them immediately.
1913 */
1914 if (bio_data_dir(bio) == READ) {
1915 zero_fill_bio(bio);
1916 cell_defer_no_holder(tc, cell);
1917 bio_endio(bio);
1918 return;
1919 }
1920
1921 r = alloc_data_block(tc, &data_block);
1922 switch (r) {
1923 case 0:
1924 if (tc->origin_dev)
1925 schedule_external_copy(tc, block, data_block, cell, bio);
1926 else
1927 schedule_zero(tc, block, data_block, cell, bio);
1928 break;
1929
1930 case -ENOSPC:
1931 retry_bios_on_resume(pool, cell);
1932 break;
1933
1934 default:
1935 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1936 __func__, r);
1937 cell_error(pool, cell);
1938 break;
1939 }
1940}
1941
1942static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1943{
1944 int r;
1945 struct pool *pool = tc->pool;
1946 struct bio *bio = cell->holder;
1947 dm_block_t block = get_bio_block(tc, bio);
1948 struct dm_thin_lookup_result lookup_result;
1949
1950 if (tc->requeue_mode) {
1951 cell_requeue(pool, cell);
1952 return;
1953 }
1954
1955 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1956 switch (r) {
1957 case 0:
1958 if (lookup_result.shared)
1959 process_shared_bio(tc, bio, block, &lookup_result, cell);
1960 else {
1961 inc_all_io_entry(pool, bio);
1962 remap_and_issue(tc, bio, lookup_result.block);
1963 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1964 }
1965 break;
1966
1967 case -ENODATA:
1968 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1969 inc_all_io_entry(pool, bio);
1970 cell_defer_no_holder(tc, cell);
1971
1972 if (bio_end_sector(bio) <= tc->origin_size)
1973 remap_to_origin_and_issue(tc, bio);
1974
1975 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1976 zero_fill_bio(bio);
1977 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1978 remap_to_origin_and_issue(tc, bio);
1979
1980 } else {
1981 zero_fill_bio(bio);
1982 bio_endio(bio);
1983 }
1984 } else
1985 provision_block(tc, bio, block, cell);
1986 break;
1987
1988 default:
1989 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1990 __func__, r);
1991 cell_defer_no_holder(tc, cell);
1992 bio_io_error(bio);
1993 break;
1994 }
1995}
1996
1997static void process_bio(struct thin_c *tc, struct bio *bio)
1998{
1999 struct pool *pool = tc->pool;
2000 dm_block_t block = get_bio_block(tc, bio);
2001 struct dm_bio_prison_cell *cell;
2002 struct dm_cell_key key;
2003
2004 /*
2005 * If cell is already occupied, then the block is already
2006 * being provisioned so we have nothing further to do here.
2007 */
2008 build_virtual_key(tc->td, block, &key);
2009 if (bio_detain(pool, &key, bio, &cell))
2010 return;
2011
2012 process_cell(tc, cell);
2013}
2014
2015static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2016 struct dm_bio_prison_cell *cell)
2017{
2018 int r;
2019 int rw = bio_data_dir(bio);
2020 dm_block_t block = get_bio_block(tc, bio);
2021 struct dm_thin_lookup_result lookup_result;
2022
2023 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2024 switch (r) {
2025 case 0:
2026 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2027 handle_unserviceable_bio(tc->pool, bio);
2028 if (cell)
2029 cell_defer_no_holder(tc, cell);
2030 } else {
2031 inc_all_io_entry(tc->pool, bio);
2032 remap_and_issue(tc, bio, lookup_result.block);
2033 if (cell)
2034 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2035 }
2036 break;
2037
2038 case -ENODATA:
2039 if (cell)
2040 cell_defer_no_holder(tc, cell);
2041 if (rw != READ) {
2042 handle_unserviceable_bio(tc->pool, bio);
2043 break;
2044 }
2045
2046 if (tc->origin_dev) {
2047 inc_all_io_entry(tc->pool, bio);
2048 remap_to_origin_and_issue(tc, bio);
2049 break;
2050 }
2051
2052 zero_fill_bio(bio);
2053 bio_endio(bio);
2054 break;
2055
2056 default:
2057 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2058 __func__, r);
2059 if (cell)
2060 cell_defer_no_holder(tc, cell);
2061 bio_io_error(bio);
2062 break;
2063 }
2064}
2065
2066static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2067{
2068 __process_bio_read_only(tc, bio, NULL);
2069}
2070
2071static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2072{
2073 __process_bio_read_only(tc, cell->holder, cell);
2074}
2075
2076static void process_bio_success(struct thin_c *tc, struct bio *bio)
2077{
2078 bio_endio(bio);
2079}
2080
2081static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2082{
2083 bio_io_error(bio);
2084}
2085
2086static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2087{
2088 cell_success(tc->pool, cell);
2089}
2090
2091static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2092{
2093 cell_error(tc->pool, cell);
2094}
2095
2096/*
2097 * FIXME: should we also commit due to size of transaction, measured in
2098 * metadata blocks?
2099 */
2100static int need_commit_due_to_time(struct pool *pool)
2101{
2102 return !time_in_range(jiffies, pool->last_commit_jiffies,
2103 pool->last_commit_jiffies + COMMIT_PERIOD);
2104}
2105
2106#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2107#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2108
2109static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2110{
2111 struct rb_node **rbp, *parent;
2112 struct dm_thin_endio_hook *pbd;
2113 sector_t bi_sector = bio->bi_iter.bi_sector;
2114
2115 rbp = &tc->sort_bio_list.rb_node;
2116 parent = NULL;
2117 while (*rbp) {
2118 parent = *rbp;
2119 pbd = thin_pbd(parent);
2120
2121 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2122 rbp = &(*rbp)->rb_left;
2123 else
2124 rbp = &(*rbp)->rb_right;
2125 }
2126
2127 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2128 rb_link_node(&pbd->rb_node, parent, rbp);
2129 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2130}
2131
2132static void __extract_sorted_bios(struct thin_c *tc)
2133{
2134 struct rb_node *node;
2135 struct dm_thin_endio_hook *pbd;
2136 struct bio *bio;
2137
2138 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2139 pbd = thin_pbd(node);
2140 bio = thin_bio(pbd);
2141
2142 bio_list_add(&tc->deferred_bio_list, bio);
2143 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2144 }
2145
2146 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2147}
2148
2149static void __sort_thin_deferred_bios(struct thin_c *tc)
2150{
2151 struct bio *bio;
2152 struct bio_list bios;
2153
2154 bio_list_init(&bios);
2155 bio_list_merge(&bios, &tc->deferred_bio_list);
2156 bio_list_init(&tc->deferred_bio_list);
2157
2158 /* Sort deferred_bio_list using rb-tree */
2159 while ((bio = bio_list_pop(&bios)))
2160 __thin_bio_rb_add(tc, bio);
2161
2162 /*
2163 * Transfer the sorted bios in sort_bio_list back to
2164 * deferred_bio_list to allow lockless submission of
2165 * all bios.
2166 */
2167 __extract_sorted_bios(tc);
2168}
2169
2170static void process_thin_deferred_bios(struct thin_c *tc)
2171{
2172 struct pool *pool = tc->pool;
2173 unsigned long flags;
2174 struct bio *bio;
2175 struct bio_list bios;
2176 struct blk_plug plug;
2177 unsigned count = 0;
2178
2179 if (tc->requeue_mode) {
2180 error_thin_bio_list(tc, &tc->deferred_bio_list,
2181 BLK_STS_DM_REQUEUE);
2182 return;
2183 }
2184
2185 bio_list_init(&bios);
2186
2187 spin_lock_irqsave(&tc->lock, flags);
2188
2189 if (bio_list_empty(&tc->deferred_bio_list)) {
2190 spin_unlock_irqrestore(&tc->lock, flags);
2191 return;
2192 }
2193
2194 __sort_thin_deferred_bios(tc);
2195
2196 bio_list_merge(&bios, &tc->deferred_bio_list);
2197 bio_list_init(&tc->deferred_bio_list);
2198
2199 spin_unlock_irqrestore(&tc->lock, flags);
2200
2201 blk_start_plug(&plug);
2202 while ((bio = bio_list_pop(&bios))) {
2203 /*
2204 * If we've got no free new_mapping structs, and processing
2205 * this bio might require one, we pause until there are some
2206 * prepared mappings to process.
2207 */
2208 if (ensure_next_mapping(pool)) {
2209 spin_lock_irqsave(&tc->lock, flags);
2210 bio_list_add(&tc->deferred_bio_list, bio);
2211 bio_list_merge(&tc->deferred_bio_list, &bios);
2212 spin_unlock_irqrestore(&tc->lock, flags);
2213 break;
2214 }
2215
2216 if (bio_op(bio) == REQ_OP_DISCARD)
2217 pool->process_discard(tc, bio);
2218 else
2219 pool->process_bio(tc, bio);
2220
2221 if ((count++ & 127) == 0) {
2222 throttle_work_update(&pool->throttle);
2223 dm_pool_issue_prefetches(pool->pmd);
2224 }
2225 }
2226 blk_finish_plug(&plug);
2227}
2228
2229static int cmp_cells(const void *lhs, const void *rhs)
2230{
2231 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2232 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2233
2234 BUG_ON(!lhs_cell->holder);
2235 BUG_ON(!rhs_cell->holder);
2236
2237 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2238 return -1;
2239
2240 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2241 return 1;
2242
2243 return 0;
2244}
2245
2246static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2247{
2248 unsigned count = 0;
2249 struct dm_bio_prison_cell *cell, *tmp;
2250
2251 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2252 if (count >= CELL_SORT_ARRAY_SIZE)
2253 break;
2254
2255 pool->cell_sort_array[count++] = cell;
2256 list_del(&cell->user_list);
2257 }
2258
2259 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2260
2261 return count;
2262}
2263
2264static void process_thin_deferred_cells(struct thin_c *tc)
2265{
2266 struct pool *pool = tc->pool;
2267 unsigned long flags;
2268 struct list_head cells;
2269 struct dm_bio_prison_cell *cell;
2270 unsigned i, j, count;
2271
2272 INIT_LIST_HEAD(&cells);
2273
2274 spin_lock_irqsave(&tc->lock, flags);
2275 list_splice_init(&tc->deferred_cells, &cells);
2276 spin_unlock_irqrestore(&tc->lock, flags);
2277
2278 if (list_empty(&cells))
2279 return;
2280
2281 do {
2282 count = sort_cells(tc->pool, &cells);
2283
2284 for (i = 0; i < count; i++) {
2285 cell = pool->cell_sort_array[i];
2286 BUG_ON(!cell->holder);
2287
2288 /*
2289 * If we've got no free new_mapping structs, and processing
2290 * this bio might require one, we pause until there are some
2291 * prepared mappings to process.
2292 */
2293 if (ensure_next_mapping(pool)) {
2294 for (j = i; j < count; j++)
2295 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2296
2297 spin_lock_irqsave(&tc->lock, flags);
2298 list_splice(&cells, &tc->deferred_cells);
2299 spin_unlock_irqrestore(&tc->lock, flags);
2300 return;
2301 }
2302
2303 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2304 pool->process_discard_cell(tc, cell);
2305 else
2306 pool->process_cell(tc, cell);
2307 }
2308 } while (!list_empty(&cells));
2309}
2310
2311static void thin_get(struct thin_c *tc);
2312static void thin_put(struct thin_c *tc);
2313
2314/*
2315 * We can't hold rcu_read_lock() around code that can block. So we
2316 * find a thin with the rcu lock held; bump a refcount; then drop
2317 * the lock.
2318 */
2319static struct thin_c *get_first_thin(struct pool *pool)
2320{
2321 struct thin_c *tc = NULL;
2322
2323 rcu_read_lock();
2324 if (!list_empty(&pool->active_thins)) {
2325 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2326 thin_get(tc);
2327 }
2328 rcu_read_unlock();
2329
2330 return tc;
2331}
2332
2333static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2334{
2335 struct thin_c *old_tc = tc;
2336
2337 rcu_read_lock();
2338 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2339 thin_get(tc);
2340 thin_put(old_tc);
2341 rcu_read_unlock();
2342 return tc;
2343 }
2344 thin_put(old_tc);
2345 rcu_read_unlock();
2346
2347 return NULL;
2348}
2349
2350static void process_deferred_bios(struct pool *pool)
2351{
2352 unsigned long flags;
2353 struct bio *bio;
2354 struct bio_list bios, bio_completions;
2355 struct thin_c *tc;
2356
2357 tc = get_first_thin(pool);
2358 while (tc) {
2359 process_thin_deferred_cells(tc);
2360 process_thin_deferred_bios(tc);
2361 tc = get_next_thin(pool, tc);
2362 }
2363
2364 /*
2365 * If there are any deferred flush bios, we must commit the metadata
2366 * before issuing them or signaling their completion.
2367 */
2368 bio_list_init(&bios);
2369 bio_list_init(&bio_completions);
2370
2371 spin_lock_irqsave(&pool->lock, flags);
2372 bio_list_merge(&bios, &pool->deferred_flush_bios);
2373 bio_list_init(&pool->deferred_flush_bios);
2374
2375 bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2376 bio_list_init(&pool->deferred_flush_completions);
2377 spin_unlock_irqrestore(&pool->lock, flags);
2378
2379 if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2380 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2381 return;
2382
2383 if (commit(pool)) {
2384 bio_list_merge(&bios, &bio_completions);
2385
2386 while ((bio = bio_list_pop(&bios)))
2387 bio_io_error(bio);
2388 return;
2389 }
2390 pool->last_commit_jiffies = jiffies;
2391
2392 while ((bio = bio_list_pop(&bio_completions)))
2393 bio_endio(bio);
2394
2395 while ((bio = bio_list_pop(&bios)))
2396 generic_make_request(bio);
2397}
2398
2399static void do_worker(struct work_struct *ws)
2400{
2401 struct pool *pool = container_of(ws, struct pool, worker);
2402
2403 throttle_work_start(&pool->throttle);
2404 dm_pool_issue_prefetches(pool->pmd);
2405 throttle_work_update(&pool->throttle);
2406 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2407 throttle_work_update(&pool->throttle);
2408 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2409 throttle_work_update(&pool->throttle);
2410 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2411 throttle_work_update(&pool->throttle);
2412 process_deferred_bios(pool);
2413 throttle_work_complete(&pool->throttle);
2414}
2415
2416/*
2417 * We want to commit periodically so that not too much
2418 * unwritten data builds up.
2419 */
2420static void do_waker(struct work_struct *ws)
2421{
2422 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2423 wake_worker(pool);
2424 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2425}
2426
2427/*
2428 * We're holding onto IO to allow userland time to react. After the
2429 * timeout either the pool will have been resized (and thus back in
2430 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2431 */
2432static void do_no_space_timeout(struct work_struct *ws)
2433{
2434 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2435 no_space_timeout);
2436
2437 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2438 pool->pf.error_if_no_space = true;
2439 notify_of_pool_mode_change(pool);
2440 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2441 }
2442}
2443
2444/*----------------------------------------------------------------*/
2445
2446struct pool_work {
2447 struct work_struct worker;
2448 struct completion complete;
2449};
2450
2451static struct pool_work *to_pool_work(struct work_struct *ws)
2452{
2453 return container_of(ws, struct pool_work, worker);
2454}
2455
2456static void pool_work_complete(struct pool_work *pw)
2457{
2458 complete(&pw->complete);
2459}
2460
2461static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2462 void (*fn)(struct work_struct *))
2463{
2464 INIT_WORK_ONSTACK(&pw->worker, fn);
2465 init_completion(&pw->complete);
2466 queue_work(pool->wq, &pw->worker);
2467 wait_for_completion(&pw->complete);
2468}
2469
2470/*----------------------------------------------------------------*/
2471
2472struct noflush_work {
2473 struct pool_work pw;
2474 struct thin_c *tc;
2475};
2476
2477static struct noflush_work *to_noflush(struct work_struct *ws)
2478{
2479 return container_of(to_pool_work(ws), struct noflush_work, pw);
2480}
2481
2482static void do_noflush_start(struct work_struct *ws)
2483{
2484 struct noflush_work *w = to_noflush(ws);
2485 w->tc->requeue_mode = true;
2486 requeue_io(w->tc);
2487 pool_work_complete(&w->pw);
2488}
2489
2490static void do_noflush_stop(struct work_struct *ws)
2491{
2492 struct noflush_work *w = to_noflush(ws);
2493 w->tc->requeue_mode = false;
2494 pool_work_complete(&w->pw);
2495}
2496
2497static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2498{
2499 struct noflush_work w;
2500
2501 w.tc = tc;
2502 pool_work_wait(&w.pw, tc->pool, fn);
2503}
2504
2505/*----------------------------------------------------------------*/
2506
2507static bool passdown_enabled(struct pool_c *pt)
2508{
2509 return pt->adjusted_pf.discard_passdown;
2510}
2511
2512static void set_discard_callbacks(struct pool *pool)
2513{
2514 struct pool_c *pt = pool->ti->private;
2515
2516 if (passdown_enabled(pt)) {
2517 pool->process_discard_cell = process_discard_cell_passdown;
2518 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2519 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2520 } else {
2521 pool->process_discard_cell = process_discard_cell_no_passdown;
2522 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2523 }
2524}
2525
2526static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2527{
2528 struct pool_c *pt = pool->ti->private;
2529 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2530 enum pool_mode old_mode = get_pool_mode(pool);
2531 unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2532
2533 /*
2534 * Never allow the pool to transition to PM_WRITE mode if user
2535 * intervention is required to verify metadata and data consistency.
2536 */
2537 if (new_mode == PM_WRITE && needs_check) {
2538 DMERR("%s: unable to switch pool to write mode until repaired.",
2539 dm_device_name(pool->pool_md));
2540 if (old_mode != new_mode)
2541 new_mode = old_mode;
2542 else
2543 new_mode = PM_READ_ONLY;
2544 }
2545 /*
2546 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2547 * not going to recover without a thin_repair. So we never let the
2548 * pool move out of the old mode.
2549 */
2550 if (old_mode == PM_FAIL)
2551 new_mode = old_mode;
2552
2553 switch (new_mode) {
2554 case PM_FAIL:
2555 dm_pool_metadata_read_only(pool->pmd);
2556 pool->process_bio = process_bio_fail;
2557 pool->process_discard = process_bio_fail;
2558 pool->process_cell = process_cell_fail;
2559 pool->process_discard_cell = process_cell_fail;
2560 pool->process_prepared_mapping = process_prepared_mapping_fail;
2561 pool->process_prepared_discard = process_prepared_discard_fail;
2562
2563 error_retry_list(pool);
2564 break;
2565
2566 case PM_OUT_OF_METADATA_SPACE:
2567 case PM_READ_ONLY:
2568 dm_pool_metadata_read_only(pool->pmd);
2569 pool->process_bio = process_bio_read_only;
2570 pool->process_discard = process_bio_success;
2571 pool->process_cell = process_cell_read_only;
2572 pool->process_discard_cell = process_cell_success;
2573 pool->process_prepared_mapping = process_prepared_mapping_fail;
2574 pool->process_prepared_discard = process_prepared_discard_success;
2575
2576 error_retry_list(pool);
2577 break;
2578
2579 case PM_OUT_OF_DATA_SPACE:
2580 /*
2581 * Ideally we'd never hit this state; the low water mark
2582 * would trigger userland to extend the pool before we
2583 * completely run out of data space. However, many small
2584 * IOs to unprovisioned space can consume data space at an
2585 * alarming rate. Adjust your low water mark if you're
2586 * frequently seeing this mode.
2587 */
2588 pool->out_of_data_space = true;
2589 pool->process_bio = process_bio_read_only;
2590 pool->process_discard = process_discard_bio;
2591 pool->process_cell = process_cell_read_only;
2592 pool->process_prepared_mapping = process_prepared_mapping;
2593 set_discard_callbacks(pool);
2594
2595 if (!pool->pf.error_if_no_space && no_space_timeout)
2596 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2597 break;
2598
2599 case PM_WRITE:
2600 if (old_mode == PM_OUT_OF_DATA_SPACE)
2601 cancel_delayed_work_sync(&pool->no_space_timeout);
2602 pool->out_of_data_space = false;
2603 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2604 dm_pool_metadata_read_write(pool->pmd);
2605 pool->process_bio = process_bio;
2606 pool->process_discard = process_discard_bio;
2607 pool->process_cell = process_cell;
2608 pool->process_prepared_mapping = process_prepared_mapping;
2609 set_discard_callbacks(pool);
2610 break;
2611 }
2612
2613 pool->pf.mode = new_mode;
2614 /*
2615 * The pool mode may have changed, sync it so bind_control_target()
2616 * doesn't cause an unexpected mode transition on resume.
2617 */
2618 pt->adjusted_pf.mode = new_mode;
2619
2620 if (old_mode != new_mode)
2621 notify_of_pool_mode_change(pool);
2622}
2623
2624static void abort_transaction(struct pool *pool)
2625{
2626 const char *dev_name = dm_device_name(pool->pool_md);
2627
2628 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2629 if (dm_pool_abort_metadata(pool->pmd)) {
2630 DMERR("%s: failed to abort metadata transaction", dev_name);
2631 set_pool_mode(pool, PM_FAIL);
2632 }
2633
2634 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2635 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2636 set_pool_mode(pool, PM_FAIL);
2637 }
2638}
2639
2640static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2641{
2642 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2643 dm_device_name(pool->pool_md), op, r);
2644
2645 abort_transaction(pool);
2646 set_pool_mode(pool, PM_READ_ONLY);
2647}
2648
2649/*----------------------------------------------------------------*/
2650
2651/*
2652 * Mapping functions.
2653 */
2654
2655/*
2656 * Called only while mapping a thin bio to hand it over to the workqueue.
2657 */
2658static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2659{
2660 unsigned long flags;
2661 struct pool *pool = tc->pool;
2662
2663 spin_lock_irqsave(&tc->lock, flags);
2664 bio_list_add(&tc->deferred_bio_list, bio);
2665 spin_unlock_irqrestore(&tc->lock, flags);
2666
2667 wake_worker(pool);
2668}
2669
2670static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2671{
2672 struct pool *pool = tc->pool;
2673
2674 throttle_lock(&pool->throttle);
2675 thin_defer_bio(tc, bio);
2676 throttle_unlock(&pool->throttle);
2677}
2678
2679static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2680{
2681 unsigned long flags;
2682 struct pool *pool = tc->pool;
2683
2684 throttle_lock(&pool->throttle);
2685 spin_lock_irqsave(&tc->lock, flags);
2686 list_add_tail(&cell->user_list, &tc->deferred_cells);
2687 spin_unlock_irqrestore(&tc->lock, flags);
2688 throttle_unlock(&pool->throttle);
2689
2690 wake_worker(pool);
2691}
2692
2693static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2694{
2695 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2696
2697 h->tc = tc;
2698 h->shared_read_entry = NULL;
2699 h->all_io_entry = NULL;
2700 h->overwrite_mapping = NULL;
2701 h->cell = NULL;
2702}
2703
2704/*
2705 * Non-blocking function called from the thin target's map function.
2706 */
2707static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2708{
2709 int r;
2710 struct thin_c *tc = ti->private;
2711 dm_block_t block = get_bio_block(tc, bio);
2712 struct dm_thin_device *td = tc->td;
2713 struct dm_thin_lookup_result result;
2714 struct dm_bio_prison_cell *virt_cell, *data_cell;
2715 struct dm_cell_key key;
2716
2717 thin_hook_bio(tc, bio);
2718
2719 if (tc->requeue_mode) {
2720 bio->bi_status = BLK_STS_DM_REQUEUE;
2721 bio_endio(bio);
2722 return DM_MAPIO_SUBMITTED;
2723 }
2724
2725 if (get_pool_mode(tc->pool) == PM_FAIL) {
2726 bio_io_error(bio);
2727 return DM_MAPIO_SUBMITTED;
2728 }
2729
2730 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2731 thin_defer_bio_with_throttle(tc, bio);
2732 return DM_MAPIO_SUBMITTED;
2733 }
2734
2735 /*
2736 * We must hold the virtual cell before doing the lookup, otherwise
2737 * there's a race with discard.
2738 */
2739 build_virtual_key(tc->td, block, &key);
2740 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2741 return DM_MAPIO_SUBMITTED;
2742
2743 r = dm_thin_find_block(td, block, 0, &result);
2744
2745 /*
2746 * Note that we defer readahead too.
2747 */
2748 switch (r) {
2749 case 0:
2750 if (unlikely(result.shared)) {
2751 /*
2752 * We have a race condition here between the
2753 * result.shared value returned by the lookup and
2754 * snapshot creation, which may cause new
2755 * sharing.
2756 *
2757 * To avoid this always quiesce the origin before
2758 * taking the snap. You want to do this anyway to
2759 * ensure a consistent application view
2760 * (i.e. lockfs).
2761 *
2762 * More distant ancestors are irrelevant. The
2763 * shared flag will be set in their case.
2764 */
2765 thin_defer_cell(tc, virt_cell);
2766 return DM_MAPIO_SUBMITTED;
2767 }
2768
2769 build_data_key(tc->td, result.block, &key);
2770 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2771 cell_defer_no_holder(tc, virt_cell);
2772 return DM_MAPIO_SUBMITTED;
2773 }
2774
2775 inc_all_io_entry(tc->pool, bio);
2776 cell_defer_no_holder(tc, data_cell);
2777 cell_defer_no_holder(tc, virt_cell);
2778
2779 remap(tc, bio, result.block);
2780 return DM_MAPIO_REMAPPED;
2781
2782 case -ENODATA:
2783 case -EWOULDBLOCK:
2784 thin_defer_cell(tc, virt_cell);
2785 return DM_MAPIO_SUBMITTED;
2786
2787 default:
2788 /*
2789 * Must always call bio_io_error on failure.
2790 * dm_thin_find_block can fail with -EINVAL if the
2791 * pool is switched to fail-io mode.
2792 */
2793 bio_io_error(bio);
2794 cell_defer_no_holder(tc, virt_cell);
2795 return DM_MAPIO_SUBMITTED;
2796 }
2797}
2798
2799static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2800{
2801 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2802 struct request_queue *q;
2803
2804 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2805 return 1;
2806
2807 q = bdev_get_queue(pt->data_dev->bdev);
2808 return bdi_congested(q->backing_dev_info, bdi_bits);
2809}
2810
2811static void requeue_bios(struct pool *pool)
2812{
2813 unsigned long flags;
2814 struct thin_c *tc;
2815
2816 rcu_read_lock();
2817 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2818 spin_lock_irqsave(&tc->lock, flags);
2819 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2820 bio_list_init(&tc->retry_on_resume_list);
2821 spin_unlock_irqrestore(&tc->lock, flags);
2822 }
2823 rcu_read_unlock();
2824}
2825
2826/*----------------------------------------------------------------
2827 * Binding of control targets to a pool object
2828 *--------------------------------------------------------------*/
2829static bool data_dev_supports_discard(struct pool_c *pt)
2830{
2831 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2832
2833 return q && blk_queue_discard(q);
2834}
2835
2836static bool is_factor(sector_t block_size, uint32_t n)
2837{
2838 return !sector_div(block_size, n);
2839}
2840
2841/*
2842 * If discard_passdown was enabled verify that the data device
2843 * supports discards. Disable discard_passdown if not.
2844 */
2845static void disable_passdown_if_not_supported(struct pool_c *pt)
2846{
2847 struct pool *pool = pt->pool;
2848 struct block_device *data_bdev = pt->data_dev->bdev;
2849 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2850 const char *reason = NULL;
2851 char buf[BDEVNAME_SIZE];
2852
2853 if (!pt->adjusted_pf.discard_passdown)
2854 return;
2855
2856 if (!data_dev_supports_discard(pt))
2857 reason = "discard unsupported";
2858
2859 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2860 reason = "max discard sectors smaller than a block";
2861
2862 if (reason) {
2863 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2864 pt->adjusted_pf.discard_passdown = false;
2865 }
2866}
2867
2868static int bind_control_target(struct pool *pool, struct dm_target *ti)
2869{
2870 struct pool_c *pt = ti->private;
2871
2872 /*
2873 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2874 */
2875 enum pool_mode old_mode = get_pool_mode(pool);
2876 enum pool_mode new_mode = pt->adjusted_pf.mode;
2877
2878 /*
2879 * Don't change the pool's mode until set_pool_mode() below.
2880 * Otherwise the pool's process_* function pointers may
2881 * not match the desired pool mode.
2882 */
2883 pt->adjusted_pf.mode = old_mode;
2884
2885 pool->ti = ti;
2886 pool->pf = pt->adjusted_pf;
2887 pool->low_water_blocks = pt->low_water_blocks;
2888
2889 set_pool_mode(pool, new_mode);
2890
2891 return 0;
2892}
2893
2894static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2895{
2896 if (pool->ti == ti)
2897 pool->ti = NULL;
2898}
2899
2900/*----------------------------------------------------------------
2901 * Pool creation
2902 *--------------------------------------------------------------*/
2903/* Initialize pool features. */
2904static void pool_features_init(struct pool_features *pf)
2905{
2906 pf->mode = PM_WRITE;
2907 pf->zero_new_blocks = true;
2908 pf->discard_enabled = true;
2909 pf->discard_passdown = true;
2910 pf->error_if_no_space = false;
2911}
2912
2913static void __pool_destroy(struct pool *pool)
2914{
2915 __pool_table_remove(pool);
2916
2917 vfree(pool->cell_sort_array);
2918 if (dm_pool_metadata_close(pool->pmd) < 0)
2919 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2920
2921 dm_bio_prison_destroy(pool->prison);
2922 dm_kcopyd_client_destroy(pool->copier);
2923
2924 if (pool->wq)
2925 destroy_workqueue(pool->wq);
2926
2927 if (pool->next_mapping)
2928 mempool_free(pool->next_mapping, &pool->mapping_pool);
2929 mempool_exit(&pool->mapping_pool);
2930 dm_deferred_set_destroy(pool->shared_read_ds);
2931 dm_deferred_set_destroy(pool->all_io_ds);
2932 kfree(pool);
2933}
2934
2935static struct kmem_cache *_new_mapping_cache;
2936
2937static struct pool *pool_create(struct mapped_device *pool_md,
2938 struct block_device *metadata_dev,
2939 unsigned long block_size,
2940 int read_only, char **error)
2941{
2942 int r;
2943 void *err_p;
2944 struct pool *pool;
2945 struct dm_pool_metadata *pmd;
2946 bool format_device = read_only ? false : true;
2947
2948 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2949 if (IS_ERR(pmd)) {
2950 *error = "Error creating metadata object";
2951 return (struct pool *)pmd;
2952 }
2953
2954 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2955 if (!pool) {
2956 *error = "Error allocating memory for pool";
2957 err_p = ERR_PTR(-ENOMEM);
2958 goto bad_pool;
2959 }
2960
2961 pool->pmd = pmd;
2962 pool->sectors_per_block = block_size;
2963 if (block_size & (block_size - 1))
2964 pool->sectors_per_block_shift = -1;
2965 else
2966 pool->sectors_per_block_shift = __ffs(block_size);
2967 pool->low_water_blocks = 0;
2968 pool_features_init(&pool->pf);
2969 pool->prison = dm_bio_prison_create();
2970 if (!pool->prison) {
2971 *error = "Error creating pool's bio prison";
2972 err_p = ERR_PTR(-ENOMEM);
2973 goto bad_prison;
2974 }
2975
2976 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2977 if (IS_ERR(pool->copier)) {
2978 r = PTR_ERR(pool->copier);
2979 *error = "Error creating pool's kcopyd client";
2980 err_p = ERR_PTR(r);
2981 goto bad_kcopyd_client;
2982 }
2983
2984 /*
2985 * Create singlethreaded workqueue that will service all devices
2986 * that use this metadata.
2987 */
2988 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2989 if (!pool->wq) {
2990 *error = "Error creating pool's workqueue";
2991 err_p = ERR_PTR(-ENOMEM);
2992 goto bad_wq;
2993 }
2994
2995 throttle_init(&pool->throttle);
2996 INIT_WORK(&pool->worker, do_worker);
2997 INIT_DELAYED_WORK(&pool->waker, do_waker);
2998 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2999 spin_lock_init(&pool->lock);
3000 bio_list_init(&pool->deferred_flush_bios);
3001 bio_list_init(&pool->deferred_flush_completions);
3002 INIT_LIST_HEAD(&pool->prepared_mappings);
3003 INIT_LIST_HEAD(&pool->prepared_discards);
3004 INIT_LIST_HEAD(&pool->prepared_discards_pt2);
3005 INIT_LIST_HEAD(&pool->active_thins);
3006 pool->low_water_triggered = false;
3007 pool->suspended = true;
3008 pool->out_of_data_space = false;
3009
3010 pool->shared_read_ds = dm_deferred_set_create();
3011 if (!pool->shared_read_ds) {
3012 *error = "Error creating pool's shared read deferred set";
3013 err_p = ERR_PTR(-ENOMEM);
3014 goto bad_shared_read_ds;
3015 }
3016
3017 pool->all_io_ds = dm_deferred_set_create();
3018 if (!pool->all_io_ds) {
3019 *error = "Error creating pool's all io deferred set";
3020 err_p = ERR_PTR(-ENOMEM);
3021 goto bad_all_io_ds;
3022 }
3023
3024 pool->next_mapping = NULL;
3025 r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
3026 _new_mapping_cache);
3027 if (r) {
3028 *error = "Error creating pool's mapping mempool";
3029 err_p = ERR_PTR(r);
3030 goto bad_mapping_pool;
3031 }
3032
3033 pool->cell_sort_array =
3034 vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3035 sizeof(*pool->cell_sort_array)));
3036 if (!pool->cell_sort_array) {
3037 *error = "Error allocating cell sort array";
3038 err_p = ERR_PTR(-ENOMEM);
3039 goto bad_sort_array;
3040 }
3041
3042 pool->ref_count = 1;
3043 pool->last_commit_jiffies = jiffies;
3044 pool->pool_md = pool_md;
3045 pool->md_dev = metadata_dev;
3046 __pool_table_insert(pool);
3047
3048 return pool;
3049
3050bad_sort_array:
3051 mempool_exit(&pool->mapping_pool);
3052bad_mapping_pool:
3053 dm_deferred_set_destroy(pool->all_io_ds);
3054bad_all_io_ds:
3055 dm_deferred_set_destroy(pool->shared_read_ds);
3056bad_shared_read_ds:
3057 destroy_workqueue(pool->wq);
3058bad_wq:
3059 dm_kcopyd_client_destroy(pool->copier);
3060bad_kcopyd_client:
3061 dm_bio_prison_destroy(pool->prison);
3062bad_prison:
3063 kfree(pool);
3064bad_pool:
3065 if (dm_pool_metadata_close(pmd))
3066 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3067
3068 return err_p;
3069}
3070
3071static void __pool_inc(struct pool *pool)
3072{
3073 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3074 pool->ref_count++;
3075}
3076
3077static void __pool_dec(struct pool *pool)
3078{
3079 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3080 BUG_ON(!pool->ref_count);
3081 if (!--pool->ref_count)
3082 __pool_destroy(pool);
3083}
3084
3085static struct pool *__pool_find(struct mapped_device *pool_md,
3086 struct block_device *metadata_dev,
3087 unsigned long block_size, int read_only,
3088 char **error, int *created)
3089{
3090 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3091
3092 if (pool) {
3093 if (pool->pool_md != pool_md) {
3094 *error = "metadata device already in use by a pool";
3095 return ERR_PTR(-EBUSY);
3096 }
3097 __pool_inc(pool);
3098
3099 } else {
3100 pool = __pool_table_lookup(pool_md);
3101 if (pool) {
3102 if (pool->md_dev != metadata_dev) {
3103 *error = "different pool cannot replace a pool";
3104 return ERR_PTR(-EINVAL);
3105 }
3106 __pool_inc(pool);
3107
3108 } else {
3109 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
3110 *created = 1;
3111 }
3112 }
3113
3114 return pool;
3115}
3116
3117/*----------------------------------------------------------------
3118 * Pool target methods
3119 *--------------------------------------------------------------*/
3120static void pool_dtr(struct dm_target *ti)
3121{
3122 struct pool_c *pt = ti->private;
3123
3124 mutex_lock(&dm_thin_pool_table.mutex);
3125
3126 unbind_control_target(pt->pool, ti);
3127 __pool_dec(pt->pool);
3128 dm_put_device(ti, pt->metadata_dev);
3129 dm_put_device(ti, pt->data_dev);
3130 kfree(pt);
3131
3132 mutex_unlock(&dm_thin_pool_table.mutex);
3133}
3134
3135static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3136 struct dm_target *ti)
3137{
3138 int r;
3139 unsigned argc;
3140 const char *arg_name;
3141
3142 static const struct dm_arg _args[] = {
3143 {0, 4, "Invalid number of pool feature arguments"},
3144 };
3145
3146 /*
3147 * No feature arguments supplied.
3148 */
3149 if (!as->argc)
3150 return 0;
3151
3152 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3153 if (r)
3154 return -EINVAL;
3155
3156 while (argc && !r) {
3157 arg_name = dm_shift_arg(as);
3158 argc--;
3159
3160 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3161 pf->zero_new_blocks = false;
3162
3163 else if (!strcasecmp(arg_name, "ignore_discard"))
3164 pf->discard_enabled = false;
3165
3166 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3167 pf->discard_passdown = false;
3168
3169 else if (!strcasecmp(arg_name, "read_only"))
3170 pf->mode = PM_READ_ONLY;
3171
3172 else if (!strcasecmp(arg_name, "error_if_no_space"))
3173 pf->error_if_no_space = true;
3174
3175 else {
3176 ti->error = "Unrecognised pool feature requested";
3177 r = -EINVAL;
3178 break;
3179 }
3180 }
3181
3182 return r;
3183}
3184
3185static void metadata_low_callback(void *context)
3186{
3187 struct pool *pool = context;
3188
3189 DMWARN("%s: reached low water mark for metadata device: sending event.",
3190 dm_device_name(pool->pool_md));
3191
3192 dm_table_event(pool->ti->table);
3193}
3194
3195static sector_t get_dev_size(struct block_device *bdev)
3196{
3197 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3198}
3199
3200static void warn_if_metadata_device_too_big(struct block_device *bdev)
3201{
3202 sector_t metadata_dev_size = get_dev_size(bdev);
3203 char buffer[BDEVNAME_SIZE];
3204
3205 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3206 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3207 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3208}
3209
3210static sector_t get_metadata_dev_size(struct block_device *bdev)
3211{
3212 sector_t metadata_dev_size = get_dev_size(bdev);
3213
3214 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3215 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3216
3217 return metadata_dev_size;
3218}
3219
3220static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3221{
3222 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3223
3224 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3225
3226 return metadata_dev_size;
3227}
3228
3229/*
3230 * When a metadata threshold is crossed a dm event is triggered, and
3231 * userland should respond by growing the metadata device. We could let
3232 * userland set the threshold, like we do with the data threshold, but I'm
3233 * not sure they know enough to do this well.
3234 */
3235static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3236{
3237 /*
3238 * 4M is ample for all ops with the possible exception of thin
3239 * device deletion which is harmless if it fails (just retry the
3240 * delete after you've grown the device).
3241 */
3242 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3243 return min((dm_block_t)1024ULL /* 4M */, quarter);
3244}
3245
3246/*
3247 * thin-pool <metadata dev> <data dev>
3248 * <data block size (sectors)>
3249 * <low water mark (blocks)>
3250 * [<#feature args> [<arg>]*]
3251 *
3252 * Optional feature arguments are:
3253 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3254 * ignore_discard: disable discard
3255 * no_discard_passdown: don't pass discards down to the data device
3256 * read_only: Don't allow any changes to be made to the pool metadata.
3257 * error_if_no_space: error IOs, instead of queueing, if no space.
3258 */
3259static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3260{
3261 int r, pool_created = 0;
3262 struct pool_c *pt;
3263 struct pool *pool;
3264 struct pool_features pf;
3265 struct dm_arg_set as;
3266 struct dm_dev *data_dev;
3267 unsigned long block_size;
3268 dm_block_t low_water_blocks;
3269 struct dm_dev *metadata_dev;
3270 fmode_t metadata_mode;
3271
3272 /*
3273 * FIXME Remove validation from scope of lock.
3274 */
3275 mutex_lock(&dm_thin_pool_table.mutex);
3276
3277 if (argc < 4) {
3278 ti->error = "Invalid argument count";
3279 r = -EINVAL;
3280 goto out_unlock;
3281 }
3282
3283 as.argc = argc;
3284 as.argv = argv;
3285
3286 /* make sure metadata and data are different devices */
3287 if (!strcmp(argv[0], argv[1])) {
3288 ti->error = "Error setting metadata or data device";
3289 r = -EINVAL;
3290 goto out_unlock;
3291 }
3292
3293 /*
3294 * Set default pool features.
3295 */
3296 pool_features_init(&pf);
3297
3298 dm_consume_args(&as, 4);
3299 r = parse_pool_features(&as, &pf, ti);
3300 if (r)
3301 goto out_unlock;
3302
3303 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3304 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3305 if (r) {
3306 ti->error = "Error opening metadata block device";
3307 goto out_unlock;
3308 }
3309 warn_if_metadata_device_too_big(metadata_dev->bdev);
3310
3311 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3312 if (r) {
3313 ti->error = "Error getting data device";
3314 goto out_metadata;
3315 }
3316
3317 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3318 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3319 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3320 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3321 ti->error = "Invalid block size";
3322 r = -EINVAL;
3323 goto out;
3324 }
3325
3326 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3327 ti->error = "Invalid low water mark";
3328 r = -EINVAL;
3329 goto out;
3330 }
3331
3332 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3333 if (!pt) {
3334 r = -ENOMEM;
3335 goto out;
3336 }
3337
3338 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3339 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3340 if (IS_ERR(pool)) {
3341 r = PTR_ERR(pool);
3342 goto out_free_pt;
3343 }
3344
3345 /*
3346 * 'pool_created' reflects whether this is the first table load.
3347 * Top level discard support is not allowed to be changed after
3348 * initial load. This would require a pool reload to trigger thin
3349 * device changes.
3350 */
3351 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3352 ti->error = "Discard support cannot be disabled once enabled";
3353 r = -EINVAL;
3354 goto out_flags_changed;
3355 }
3356
3357 pt->pool = pool;
3358 pt->ti = ti;
3359 pt->metadata_dev = metadata_dev;
3360 pt->data_dev = data_dev;
3361 pt->low_water_blocks = low_water_blocks;
3362 pt->adjusted_pf = pt->requested_pf = pf;
3363 ti->num_flush_bios = 1;
3364
3365 /*
3366 * Only need to enable discards if the pool should pass
3367 * them down to the data device. The thin device's discard
3368 * processing will cause mappings to be removed from the btree.
3369 */
3370 if (pf.discard_enabled && pf.discard_passdown) {
3371 ti->num_discard_bios = 1;
3372
3373 /*
3374 * Setting 'discards_supported' circumvents the normal
3375 * stacking of discard limits (this keeps the pool and
3376 * thin devices' discard limits consistent).
3377 */
3378 ti->discards_supported = true;
3379 }
3380 ti->private = pt;
3381
3382 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3383 calc_metadata_threshold(pt),
3384 metadata_low_callback,
3385 pool);
3386 if (r)
3387 goto out_flags_changed;
3388
3389 pt->callbacks.congested_fn = pool_is_congested;
3390 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3391
3392 mutex_unlock(&dm_thin_pool_table.mutex);
3393
3394 return 0;
3395
3396out_flags_changed:
3397 __pool_dec(pool);
3398out_free_pt:
3399 kfree(pt);
3400out:
3401 dm_put_device(ti, data_dev);
3402out_metadata:
3403 dm_put_device(ti, metadata_dev);
3404out_unlock:
3405 mutex_unlock(&dm_thin_pool_table.mutex);
3406
3407 return r;
3408}
3409
3410static int pool_map(struct dm_target *ti, struct bio *bio)
3411{
3412 int r;
3413 struct pool_c *pt = ti->private;
3414 struct pool *pool = pt->pool;
3415 unsigned long flags;
3416
3417 /*
3418 * As this is a singleton target, ti->begin is always zero.
3419 */
3420 spin_lock_irqsave(&pool->lock, flags);
3421 bio_set_dev(bio, pt->data_dev->bdev);
3422 r = DM_MAPIO_REMAPPED;
3423 spin_unlock_irqrestore(&pool->lock, flags);
3424
3425 return r;
3426}
3427
3428static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3429{
3430 int r;
3431 struct pool_c *pt = ti->private;
3432 struct pool *pool = pt->pool;
3433 sector_t data_size = ti->len;
3434 dm_block_t sb_data_size;
3435
3436 *need_commit = false;
3437
3438 (void) sector_div(data_size, pool->sectors_per_block);
3439
3440 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3441 if (r) {
3442 DMERR("%s: failed to retrieve data device size",
3443 dm_device_name(pool->pool_md));
3444 return r;
3445 }
3446
3447 if (data_size < sb_data_size) {
3448 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3449 dm_device_name(pool->pool_md),
3450 (unsigned long long)data_size, sb_data_size);
3451 return -EINVAL;
3452
3453 } else if (data_size > sb_data_size) {
3454 if (dm_pool_metadata_needs_check(pool->pmd)) {
3455 DMERR("%s: unable to grow the data device until repaired.",
3456 dm_device_name(pool->pool_md));
3457 return 0;
3458 }
3459
3460 if (sb_data_size)
3461 DMINFO("%s: growing the data device from %llu to %llu blocks",
3462 dm_device_name(pool->pool_md),
3463 sb_data_size, (unsigned long long)data_size);
3464 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3465 if (r) {
3466 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3467 return r;
3468 }
3469
3470 *need_commit = true;
3471 }
3472
3473 return 0;
3474}
3475
3476static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3477{
3478 int r;
3479 struct pool_c *pt = ti->private;
3480 struct pool *pool = pt->pool;
3481 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3482
3483 *need_commit = false;
3484
3485 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3486
3487 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3488 if (r) {
3489 DMERR("%s: failed to retrieve metadata device size",
3490 dm_device_name(pool->pool_md));
3491 return r;
3492 }
3493
3494 if (metadata_dev_size < sb_metadata_dev_size) {
3495 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3496 dm_device_name(pool->pool_md),
3497 metadata_dev_size, sb_metadata_dev_size);
3498 return -EINVAL;
3499
3500 } else if (metadata_dev_size > sb_metadata_dev_size) {
3501 if (dm_pool_metadata_needs_check(pool->pmd)) {
3502 DMERR("%s: unable to grow the metadata device until repaired.",
3503 dm_device_name(pool->pool_md));
3504 return 0;
3505 }
3506
3507 warn_if_metadata_device_too_big(pool->md_dev);
3508 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3509 dm_device_name(pool->pool_md),
3510 sb_metadata_dev_size, metadata_dev_size);
3511
3512 if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3513 set_pool_mode(pool, PM_WRITE);
3514
3515 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3516 if (r) {
3517 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3518 return r;
3519 }
3520
3521 *need_commit = true;
3522 }
3523
3524 return 0;
3525}
3526
3527/*
3528 * Retrieves the number of blocks of the data device from
3529 * the superblock and compares it to the actual device size,
3530 * thus resizing the data device in case it has grown.
3531 *
3532 * This both copes with opening preallocated data devices in the ctr
3533 * being followed by a resume
3534 * -and-
3535 * calling the resume method individually after userspace has
3536 * grown the data device in reaction to a table event.
3537 */
3538static int pool_preresume(struct dm_target *ti)
3539{
3540 int r;
3541 bool need_commit1, need_commit2;
3542 struct pool_c *pt = ti->private;
3543 struct pool *pool = pt->pool;
3544
3545 /*
3546 * Take control of the pool object.
3547 */
3548 r = bind_control_target(pool, ti);
3549 if (r)
3550 return r;
3551
3552 r = maybe_resize_data_dev(ti, &need_commit1);
3553 if (r)
3554 return r;
3555
3556 r = maybe_resize_metadata_dev(ti, &need_commit2);
3557 if (r)
3558 return r;
3559
3560 if (need_commit1 || need_commit2)
3561 (void) commit(pool);
3562
3563 return 0;
3564}
3565
3566static void pool_suspend_active_thins(struct pool *pool)
3567{
3568 struct thin_c *tc;
3569
3570 /* Suspend all active thin devices */
3571 tc = get_first_thin(pool);
3572 while (tc) {
3573 dm_internal_suspend_noflush(tc->thin_md);
3574 tc = get_next_thin(pool, tc);
3575 }
3576}
3577
3578static void pool_resume_active_thins(struct pool *pool)
3579{
3580 struct thin_c *tc;
3581
3582 /* Resume all active thin devices */
3583 tc = get_first_thin(pool);
3584 while (tc) {
3585 dm_internal_resume(tc->thin_md);
3586 tc = get_next_thin(pool, tc);
3587 }
3588}
3589
3590static void pool_resume(struct dm_target *ti)
3591{
3592 struct pool_c *pt = ti->private;
3593 struct pool *pool = pt->pool;
3594 unsigned long flags;
3595
3596 /*
3597 * Must requeue active_thins' bios and then resume
3598 * active_thins _before_ clearing 'suspend' flag.
3599 */
3600 requeue_bios(pool);
3601 pool_resume_active_thins(pool);
3602
3603 spin_lock_irqsave(&pool->lock, flags);
3604 pool->low_water_triggered = false;
3605 pool->suspended = false;
3606 spin_unlock_irqrestore(&pool->lock, flags);
3607
3608 do_waker(&pool->waker.work);
3609}
3610
3611static void pool_presuspend(struct dm_target *ti)
3612{
3613 struct pool_c *pt = ti->private;
3614 struct pool *pool = pt->pool;
3615 unsigned long flags;
3616
3617 spin_lock_irqsave(&pool->lock, flags);
3618 pool->suspended = true;
3619 spin_unlock_irqrestore(&pool->lock, flags);
3620
3621 pool_suspend_active_thins(pool);
3622}
3623
3624static void pool_presuspend_undo(struct dm_target *ti)
3625{
3626 struct pool_c *pt = ti->private;
3627 struct pool *pool = pt->pool;
3628 unsigned long flags;
3629
3630 pool_resume_active_thins(pool);
3631
3632 spin_lock_irqsave(&pool->lock, flags);
3633 pool->suspended = false;
3634 spin_unlock_irqrestore(&pool->lock, flags);
3635}
3636
3637static void pool_postsuspend(struct dm_target *ti)
3638{
3639 struct pool_c *pt = ti->private;
3640 struct pool *pool = pt->pool;
3641
3642 cancel_delayed_work_sync(&pool->waker);
3643 cancel_delayed_work_sync(&pool->no_space_timeout);
3644 flush_workqueue(pool->wq);
3645 (void) commit(pool);
3646}
3647
3648static int check_arg_count(unsigned argc, unsigned args_required)
3649{
3650 if (argc != args_required) {
3651 DMWARN("Message received with %u arguments instead of %u.",
3652 argc, args_required);
3653 return -EINVAL;
3654 }
3655
3656 return 0;
3657}
3658
3659static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3660{
3661 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3662 *dev_id <= MAX_DEV_ID)
3663 return 0;
3664
3665 if (warning)
3666 DMWARN("Message received with invalid device id: %s", arg);
3667
3668 return -EINVAL;
3669}
3670
3671static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3672{
3673 dm_thin_id dev_id;
3674 int r;
3675
3676 r = check_arg_count(argc, 2);
3677 if (r)
3678 return r;
3679
3680 r = read_dev_id(argv[1], &dev_id, 1);
3681 if (r)
3682 return r;
3683
3684 r = dm_pool_create_thin(pool->pmd, dev_id);
3685 if (r) {
3686 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3687 argv[1]);
3688 return r;
3689 }
3690
3691 return 0;
3692}
3693
3694static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3695{
3696 dm_thin_id dev_id;
3697 dm_thin_id origin_dev_id;
3698 int r;
3699
3700 r = check_arg_count(argc, 3);
3701 if (r)
3702 return r;
3703
3704 r = read_dev_id(argv[1], &dev_id, 1);
3705 if (r)
3706 return r;
3707
3708 r = read_dev_id(argv[2], &origin_dev_id, 1);
3709 if (r)
3710 return r;
3711
3712 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3713 if (r) {
3714 DMWARN("Creation of new snapshot %s of device %s failed.",
3715 argv[1], argv[2]);
3716 return r;
3717 }
3718
3719 return 0;
3720}
3721
3722static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3723{
3724 dm_thin_id dev_id;
3725 int r;
3726
3727 r = check_arg_count(argc, 2);
3728 if (r)
3729 return r;
3730
3731 r = read_dev_id(argv[1], &dev_id, 1);
3732 if (r)
3733 return r;
3734
3735 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3736 if (r)
3737 DMWARN("Deletion of thin device %s failed.", argv[1]);
3738
3739 return r;
3740}
3741
3742static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3743{
3744 dm_thin_id old_id, new_id;
3745 int r;
3746
3747 r = check_arg_count(argc, 3);
3748 if (r)
3749 return r;
3750
3751 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3752 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3753 return -EINVAL;
3754 }
3755
3756 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3757 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3758 return -EINVAL;
3759 }
3760
3761 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3762 if (r) {
3763 DMWARN("Failed to change transaction id from %s to %s.",
3764 argv[1], argv[2]);
3765 return r;
3766 }
3767
3768 return 0;
3769}
3770
3771static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3772{
3773 int r;
3774
3775 r = check_arg_count(argc, 1);
3776 if (r)
3777 return r;
3778
3779 (void) commit(pool);
3780
3781 r = dm_pool_reserve_metadata_snap(pool->pmd);
3782 if (r)
3783 DMWARN("reserve_metadata_snap message failed.");
3784
3785 return r;
3786}
3787
3788static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3789{
3790 int r;
3791
3792 r = check_arg_count(argc, 1);
3793 if (r)
3794 return r;
3795
3796 r = dm_pool_release_metadata_snap(pool->pmd);
3797 if (r)
3798 DMWARN("release_metadata_snap message failed.");
3799
3800 return r;
3801}
3802
3803/*
3804 * Messages supported:
3805 * create_thin <dev_id>
3806 * create_snap <dev_id> <origin_id>
3807 * delete <dev_id>
3808 * set_transaction_id <current_trans_id> <new_trans_id>
3809 * reserve_metadata_snap
3810 * release_metadata_snap
3811 */
3812static int pool_message(struct dm_target *ti, unsigned argc, char **argv,
3813 char *result, unsigned maxlen)
3814{
3815 int r = -EINVAL;
3816 struct pool_c *pt = ti->private;
3817 struct pool *pool = pt->pool;
3818
3819 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE) {
3820 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3821 dm_device_name(pool->pool_md));
3822 return -EOPNOTSUPP;
3823 }
3824
3825 if (!strcasecmp(argv[0], "create_thin"))
3826 r = process_create_thin_mesg(argc, argv, pool);
3827
3828 else if (!strcasecmp(argv[0], "create_snap"))
3829 r = process_create_snap_mesg(argc, argv, pool);
3830
3831 else if (!strcasecmp(argv[0], "delete"))
3832 r = process_delete_mesg(argc, argv, pool);
3833
3834 else if (!strcasecmp(argv[0], "set_transaction_id"))
3835 r = process_set_transaction_id_mesg(argc, argv, pool);
3836
3837 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3838 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3839
3840 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3841 r = process_release_metadata_snap_mesg(argc, argv, pool);
3842
3843 else
3844 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3845
3846 if (!r)
3847 (void) commit(pool);
3848
3849 return r;
3850}
3851
3852static void emit_flags(struct pool_features *pf, char *result,
3853 unsigned sz, unsigned maxlen)
3854{
3855 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3856 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3857 pf->error_if_no_space;
3858 DMEMIT("%u ", count);
3859
3860 if (!pf->zero_new_blocks)
3861 DMEMIT("skip_block_zeroing ");
3862
3863 if (!pf->discard_enabled)
3864 DMEMIT("ignore_discard ");
3865
3866 if (!pf->discard_passdown)
3867 DMEMIT("no_discard_passdown ");
3868
3869 if (pf->mode == PM_READ_ONLY)
3870 DMEMIT("read_only ");
3871
3872 if (pf->error_if_no_space)
3873 DMEMIT("error_if_no_space ");
3874}
3875
3876/*
3877 * Status line is:
3878 * <transaction id> <used metadata sectors>/<total metadata sectors>
3879 * <used data sectors>/<total data sectors> <held metadata root>
3880 * <pool mode> <discard config> <no space config> <needs_check>
3881 */
3882static void pool_status(struct dm_target *ti, status_type_t type,
3883 unsigned status_flags, char *result, unsigned maxlen)
3884{
3885 int r;
3886 unsigned sz = 0;
3887 uint64_t transaction_id;
3888 dm_block_t nr_free_blocks_data;
3889 dm_block_t nr_free_blocks_metadata;
3890 dm_block_t nr_blocks_data;
3891 dm_block_t nr_blocks_metadata;
3892 dm_block_t held_root;
3893 enum pool_mode mode;
3894 char buf[BDEVNAME_SIZE];
3895 char buf2[BDEVNAME_SIZE];
3896 struct pool_c *pt = ti->private;
3897 struct pool *pool = pt->pool;
3898
3899 switch (type) {
3900 case STATUSTYPE_INFO:
3901 if (get_pool_mode(pool) == PM_FAIL) {
3902 DMEMIT("Fail");
3903 break;
3904 }
3905
3906 /* Commit to ensure statistics aren't out-of-date */
3907 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3908 (void) commit(pool);
3909
3910 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3911 if (r) {
3912 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3913 dm_device_name(pool->pool_md), r);
3914 goto err;
3915 }
3916
3917 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3918 if (r) {
3919 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3920 dm_device_name(pool->pool_md), r);
3921 goto err;
3922 }
3923
3924 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3925 if (r) {
3926 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3927 dm_device_name(pool->pool_md), r);
3928 goto err;
3929 }
3930
3931 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3932 if (r) {
3933 DMERR("%s: dm_pool_get_free_block_count returned %d",
3934 dm_device_name(pool->pool_md), r);
3935 goto err;
3936 }
3937
3938 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3939 if (r) {
3940 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3941 dm_device_name(pool->pool_md), r);
3942 goto err;
3943 }
3944
3945 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3946 if (r) {
3947 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3948 dm_device_name(pool->pool_md), r);
3949 goto err;
3950 }
3951
3952 DMEMIT("%llu %llu/%llu %llu/%llu ",
3953 (unsigned long long)transaction_id,
3954 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3955 (unsigned long long)nr_blocks_metadata,
3956 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3957 (unsigned long long)nr_blocks_data);
3958
3959 if (held_root)
3960 DMEMIT("%llu ", held_root);
3961 else
3962 DMEMIT("- ");
3963
3964 mode = get_pool_mode(pool);
3965 if (mode == PM_OUT_OF_DATA_SPACE)
3966 DMEMIT("out_of_data_space ");
3967 else if (is_read_only_pool_mode(mode))
3968 DMEMIT("ro ");
3969 else
3970 DMEMIT("rw ");
3971
3972 if (!pool->pf.discard_enabled)
3973 DMEMIT("ignore_discard ");
3974 else if (pool->pf.discard_passdown)
3975 DMEMIT("discard_passdown ");
3976 else
3977 DMEMIT("no_discard_passdown ");
3978
3979 if (pool->pf.error_if_no_space)
3980 DMEMIT("error_if_no_space ");
3981 else
3982 DMEMIT("queue_if_no_space ");
3983
3984 if (dm_pool_metadata_needs_check(pool->pmd))
3985 DMEMIT("needs_check ");
3986 else
3987 DMEMIT("- ");
3988
3989 DMEMIT("%llu ", (unsigned long long)calc_metadata_threshold(pt));
3990
3991 break;
3992
3993 case STATUSTYPE_TABLE:
3994 DMEMIT("%s %s %lu %llu ",
3995 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3996 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3997 (unsigned long)pool->sectors_per_block,
3998 (unsigned long long)pt->low_water_blocks);
3999 emit_flags(&pt->requested_pf, result, sz, maxlen);
4000 break;
4001 }
4002 return;
4003
4004err:
4005 DMEMIT("Error");
4006}
4007
4008static int pool_iterate_devices(struct dm_target *ti,
4009 iterate_devices_callout_fn fn, void *data)
4010{
4011 struct pool_c *pt = ti->private;
4012
4013 return fn(ti, pt->data_dev, 0, ti->len, data);
4014}
4015
4016static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
4017{
4018 struct pool_c *pt = ti->private;
4019 struct pool *pool = pt->pool;
4020 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
4021
4022 /*
4023 * If max_sectors is smaller than pool->sectors_per_block adjust it
4024 * to the highest possible power-of-2 factor of pool->sectors_per_block.
4025 * This is especially beneficial when the pool's data device is a RAID
4026 * device that has a full stripe width that matches pool->sectors_per_block
4027 * -- because even though partial RAID stripe-sized IOs will be issued to a
4028 * single RAID stripe; when aggregated they will end on a full RAID stripe
4029 * boundary.. which avoids additional partial RAID stripe writes cascading
4030 */
4031 if (limits->max_sectors < pool->sectors_per_block) {
4032 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
4033 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
4034 limits->max_sectors--;
4035 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
4036 }
4037 }
4038
4039 /*
4040 * If the system-determined stacked limits are compatible with the
4041 * pool's blocksize (io_opt is a factor) do not override them.
4042 */
4043 if (io_opt_sectors < pool->sectors_per_block ||
4044 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
4045 if (is_factor(pool->sectors_per_block, limits->max_sectors))
4046 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
4047 else
4048 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
4049 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
4050 }
4051
4052 /*
4053 * pt->adjusted_pf is a staging area for the actual features to use.
4054 * They get transferred to the live pool in bind_control_target()
4055 * called from pool_preresume().
4056 */
4057 if (!pt->adjusted_pf.discard_enabled) {
4058 /*
4059 * Must explicitly disallow stacking discard limits otherwise the
4060 * block layer will stack them if pool's data device has support.
4061 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
4062 * user to see that, so make sure to set all discard limits to 0.
4063 */
4064 limits->discard_granularity = 0;
4065 return;
4066 }
4067
4068 disable_passdown_if_not_supported(pt);
4069
4070 /*
4071 * The pool uses the same discard limits as the underlying data
4072 * device. DM core has already set this up.
4073 */
4074}
4075
4076static struct target_type pool_target = {
4077 .name = "thin-pool",
4078 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4079 DM_TARGET_IMMUTABLE,
4080 .version = {1, 21, 0},
4081 .module = THIS_MODULE,
4082 .ctr = pool_ctr,
4083 .dtr = pool_dtr,
4084 .map = pool_map,
4085 .presuspend = pool_presuspend,
4086 .presuspend_undo = pool_presuspend_undo,
4087 .postsuspend = pool_postsuspend,
4088 .preresume = pool_preresume,
4089 .resume = pool_resume,
4090 .message = pool_message,
4091 .status = pool_status,
4092 .iterate_devices = pool_iterate_devices,
4093 .io_hints = pool_io_hints,
4094};
4095
4096/*----------------------------------------------------------------
4097 * Thin target methods
4098 *--------------------------------------------------------------*/
4099static void thin_get(struct thin_c *tc)
4100{
4101 refcount_inc(&tc->refcount);
4102}
4103
4104static void thin_put(struct thin_c *tc)
4105{
4106 if (refcount_dec_and_test(&tc->refcount))
4107 complete(&tc->can_destroy);
4108}
4109
4110static void thin_dtr(struct dm_target *ti)
4111{
4112 struct thin_c *tc = ti->private;
4113 unsigned long flags;
4114
4115 spin_lock_irqsave(&tc->pool->lock, flags);
4116 list_del_rcu(&tc->list);
4117 spin_unlock_irqrestore(&tc->pool->lock, flags);
4118 synchronize_rcu();
4119
4120 thin_put(tc);
4121 wait_for_completion(&tc->can_destroy);
4122
4123 mutex_lock(&dm_thin_pool_table.mutex);
4124
4125 __pool_dec(tc->pool);
4126 dm_pool_close_thin_device(tc->td);
4127 dm_put_device(ti, tc->pool_dev);
4128 if (tc->origin_dev)
4129 dm_put_device(ti, tc->origin_dev);
4130 kfree(tc);
4131
4132 mutex_unlock(&dm_thin_pool_table.mutex);
4133}
4134
4135/*
4136 * Thin target parameters:
4137 *
4138 * <pool_dev> <dev_id> [origin_dev]
4139 *
4140 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4141 * dev_id: the internal device identifier
4142 * origin_dev: a device external to the pool that should act as the origin
4143 *
4144 * If the pool device has discards disabled, they get disabled for the thin
4145 * device as well.
4146 */
4147static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4148{
4149 int r;
4150 struct thin_c *tc;
4151 struct dm_dev *pool_dev, *origin_dev;
4152 struct mapped_device *pool_md;
4153 unsigned long flags;
4154
4155 mutex_lock(&dm_thin_pool_table.mutex);
4156
4157 if (argc != 2 && argc != 3) {
4158 ti->error = "Invalid argument count";
4159 r = -EINVAL;
4160 goto out_unlock;
4161 }
4162
4163 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4164 if (!tc) {
4165 ti->error = "Out of memory";
4166 r = -ENOMEM;
4167 goto out_unlock;
4168 }
4169 tc->thin_md = dm_table_get_md(ti->table);
4170 spin_lock_init(&tc->lock);
4171 INIT_LIST_HEAD(&tc->deferred_cells);
4172 bio_list_init(&tc->deferred_bio_list);
4173 bio_list_init(&tc->retry_on_resume_list);
4174 tc->sort_bio_list = RB_ROOT;
4175
4176 if (argc == 3) {
4177 if (!strcmp(argv[0], argv[2])) {
4178 ti->error = "Error setting origin device";
4179 r = -EINVAL;
4180 goto bad_origin_dev;
4181 }
4182
4183 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4184 if (r) {
4185 ti->error = "Error opening origin device";
4186 goto bad_origin_dev;
4187 }
4188 tc->origin_dev = origin_dev;
4189 }
4190
4191 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4192 if (r) {
4193 ti->error = "Error opening pool device";
4194 goto bad_pool_dev;
4195 }
4196 tc->pool_dev = pool_dev;
4197
4198 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4199 ti->error = "Invalid device id";
4200 r = -EINVAL;
4201 goto bad_common;
4202 }
4203
4204 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4205 if (!pool_md) {
4206 ti->error = "Couldn't get pool mapped device";
4207 r = -EINVAL;
4208 goto bad_common;
4209 }
4210
4211 tc->pool = __pool_table_lookup(pool_md);
4212 if (!tc->pool) {
4213 ti->error = "Couldn't find pool object";
4214 r = -EINVAL;
4215 goto bad_pool_lookup;
4216 }
4217 __pool_inc(tc->pool);
4218
4219 if (get_pool_mode(tc->pool) == PM_FAIL) {
4220 ti->error = "Couldn't open thin device, Pool is in fail mode";
4221 r = -EINVAL;
4222 goto bad_pool;
4223 }
4224
4225 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4226 if (r) {
4227 ti->error = "Couldn't open thin internal device";
4228 goto bad_pool;
4229 }
4230
4231 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4232 if (r)
4233 goto bad;
4234
4235 ti->num_flush_bios = 1;
4236 ti->flush_supported = true;
4237 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4238
4239 /* In case the pool supports discards, pass them on. */
4240 if (tc->pool->pf.discard_enabled) {
4241 ti->discards_supported = true;
4242 ti->num_discard_bios = 1;
4243 }
4244
4245 mutex_unlock(&dm_thin_pool_table.mutex);
4246
4247 spin_lock_irqsave(&tc->pool->lock, flags);
4248 if (tc->pool->suspended) {
4249 spin_unlock_irqrestore(&tc->pool->lock, flags);
4250 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4251 ti->error = "Unable to activate thin device while pool is suspended";
4252 r = -EINVAL;
4253 goto bad;
4254 }
4255 refcount_set(&tc->refcount, 1);
4256 init_completion(&tc->can_destroy);
4257 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4258 spin_unlock_irqrestore(&tc->pool->lock, flags);
4259 /*
4260 * This synchronize_rcu() call is needed here otherwise we risk a
4261 * wake_worker() call finding no bios to process (because the newly
4262 * added tc isn't yet visible). So this reduces latency since we
4263 * aren't then dependent on the periodic commit to wake_worker().
4264 */
4265 synchronize_rcu();
4266
4267 dm_put(pool_md);
4268
4269 return 0;
4270
4271bad:
4272 dm_pool_close_thin_device(tc->td);
4273bad_pool:
4274 __pool_dec(tc->pool);
4275bad_pool_lookup:
4276 dm_put(pool_md);
4277bad_common:
4278 dm_put_device(ti, tc->pool_dev);
4279bad_pool_dev:
4280 if (tc->origin_dev)
4281 dm_put_device(ti, tc->origin_dev);
4282bad_origin_dev:
4283 kfree(tc);
4284out_unlock:
4285 mutex_unlock(&dm_thin_pool_table.mutex);
4286
4287 return r;
4288}
4289
4290static int thin_map(struct dm_target *ti, struct bio *bio)
4291{
4292 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4293
4294 return thin_bio_map(ti, bio);
4295}
4296
4297static int thin_endio(struct dm_target *ti, struct bio *bio,
4298 blk_status_t *err)
4299{
4300 unsigned long flags;
4301 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4302 struct list_head work;
4303 struct dm_thin_new_mapping *m, *tmp;
4304 struct pool *pool = h->tc->pool;
4305
4306 if (h->shared_read_entry) {
4307 INIT_LIST_HEAD(&work);
4308 dm_deferred_entry_dec(h->shared_read_entry, &work);
4309
4310 spin_lock_irqsave(&pool->lock, flags);
4311 list_for_each_entry_safe(m, tmp, &work, list) {
4312 list_del(&m->list);
4313 __complete_mapping_preparation(m);
4314 }
4315 spin_unlock_irqrestore(&pool->lock, flags);
4316 }
4317
4318 if (h->all_io_entry) {
4319 INIT_LIST_HEAD(&work);
4320 dm_deferred_entry_dec(h->all_io_entry, &work);
4321 if (!list_empty(&work)) {
4322 spin_lock_irqsave(&pool->lock, flags);
4323 list_for_each_entry_safe(m, tmp, &work, list)
4324 list_add_tail(&m->list, &pool->prepared_discards);
4325 spin_unlock_irqrestore(&pool->lock, flags);
4326 wake_worker(pool);
4327 }
4328 }
4329
4330 if (h->cell)
4331 cell_defer_no_holder(h->tc, h->cell);
4332
4333 return DM_ENDIO_DONE;
4334}
4335
4336static void thin_presuspend(struct dm_target *ti)
4337{
4338 struct thin_c *tc = ti->private;
4339
4340 if (dm_noflush_suspending(ti))
4341 noflush_work(tc, do_noflush_start);
4342}
4343
4344static void thin_postsuspend(struct dm_target *ti)
4345{
4346 struct thin_c *tc = ti->private;
4347
4348 /*
4349 * The dm_noflush_suspending flag has been cleared by now, so
4350 * unfortunately we must always run this.
4351 */
4352 noflush_work(tc, do_noflush_stop);
4353}
4354
4355static int thin_preresume(struct dm_target *ti)
4356{
4357 struct thin_c *tc = ti->private;
4358
4359 if (tc->origin_dev)
4360 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4361
4362 return 0;
4363}
4364
4365/*
4366 * <nr mapped sectors> <highest mapped sector>
4367 */
4368static void thin_status(struct dm_target *ti, status_type_t type,
4369 unsigned status_flags, char *result, unsigned maxlen)
4370{
4371 int r;
4372 ssize_t sz = 0;
4373 dm_block_t mapped, highest;
4374 char buf[BDEVNAME_SIZE];
4375 struct thin_c *tc = ti->private;
4376
4377 if (get_pool_mode(tc->pool) == PM_FAIL) {
4378 DMEMIT("Fail");
4379 return;
4380 }
4381
4382 if (!tc->td)
4383 DMEMIT("-");
4384 else {
4385 switch (type) {
4386 case STATUSTYPE_INFO:
4387 r = dm_thin_get_mapped_count(tc->td, &mapped);
4388 if (r) {
4389 DMERR("dm_thin_get_mapped_count returned %d", r);
4390 goto err;
4391 }
4392
4393 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4394 if (r < 0) {
4395 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4396 goto err;
4397 }
4398
4399 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4400 if (r)
4401 DMEMIT("%llu", ((highest + 1) *
4402 tc->pool->sectors_per_block) - 1);
4403 else
4404 DMEMIT("-");
4405 break;
4406
4407 case STATUSTYPE_TABLE:
4408 DMEMIT("%s %lu",
4409 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4410 (unsigned long) tc->dev_id);
4411 if (tc->origin_dev)
4412 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4413 break;
4414 }
4415 }
4416
4417 return;
4418
4419err:
4420 DMEMIT("Error");
4421}
4422
4423static int thin_iterate_devices(struct dm_target *ti,
4424 iterate_devices_callout_fn fn, void *data)
4425{
4426 sector_t blocks;
4427 struct thin_c *tc = ti->private;
4428 struct pool *pool = tc->pool;
4429
4430 /*
4431 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4432 * we follow a more convoluted path through to the pool's target.
4433 */
4434 if (!pool->ti)
4435 return 0; /* nothing is bound */
4436
4437 blocks = pool->ti->len;
4438 (void) sector_div(blocks, pool->sectors_per_block);
4439 if (blocks)
4440 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4441
4442 return 0;
4443}
4444
4445static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4446{
4447 struct thin_c *tc = ti->private;
4448 struct pool *pool = tc->pool;
4449
4450 if (!pool->pf.discard_enabled)
4451 return;
4452
4453 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4454 limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4455}
4456
4457static struct target_type thin_target = {
4458 .name = "thin",
4459 .version = {1, 21, 0},
4460 .module = THIS_MODULE,
4461 .ctr = thin_ctr,
4462 .dtr = thin_dtr,
4463 .map = thin_map,
4464 .end_io = thin_endio,
4465 .preresume = thin_preresume,
4466 .presuspend = thin_presuspend,
4467 .postsuspend = thin_postsuspend,
4468 .status = thin_status,
4469 .iterate_devices = thin_iterate_devices,
4470 .io_hints = thin_io_hints,
4471};
4472
4473/*----------------------------------------------------------------*/
4474
4475static int __init dm_thin_init(void)
4476{
4477 int r = -ENOMEM;
4478
4479 pool_table_init();
4480
4481 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4482 if (!_new_mapping_cache)
4483 return r;
4484
4485 r = dm_register_target(&thin_target);
4486 if (r)
4487 goto bad_new_mapping_cache;
4488
4489 r = dm_register_target(&pool_target);
4490 if (r)
4491 goto bad_thin_target;
4492
4493 return 0;
4494
4495bad_thin_target:
4496 dm_unregister_target(&thin_target);
4497bad_new_mapping_cache:
4498 kmem_cache_destroy(_new_mapping_cache);
4499
4500 return r;
4501}
4502
4503static void dm_thin_exit(void)
4504{
4505 dm_unregister_target(&thin_target);
4506 dm_unregister_target(&pool_target);
4507
4508 kmem_cache_destroy(_new_mapping_cache);
4509
4510 pool_table_exit();
4511}
4512
4513module_init(dm_thin_init);
4514module_exit(dm_thin_exit);
4515
4516module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4517MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4518
4519MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4520MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4521MODULE_LICENSE("GPL");