Root/drivers/md/dm-thin.c

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.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/list.h>
15#include <linux/init.h>
16#include <linux/module.h>
17#include <linux/slab.h>
18
19#define DM_MSG_PREFIX "thin"
20
21/*
22 * Tunable constants
23 */
24#define ENDIO_HOOK_POOL_SIZE 1024
25#define MAPPING_POOL_SIZE 1024
26#define PRISON_CELLS 1024
27#define COMMIT_PERIOD HZ
28
29DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
30        "A percentage of time allocated for copy on write");
31
32/*
33 * The block size of the device holding pool data must be
34 * between 64KB and 1GB.
35 */
36#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
37#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
38
39/*
40 * Device id is restricted to 24 bits.
41 */
42#define MAX_DEV_ID ((1 << 24) - 1)
43
44/*
45 * How do we handle breaking sharing of data blocks?
46 * =================================================
47 *
48 * We use a standard copy-on-write btree to store the mappings for the
49 * devices (note I'm talking about copy-on-write of the metadata here, not
50 * the data). When you take an internal snapshot you clone the root node
51 * of the origin btree. After this there is no concept of an origin or a
52 * snapshot. They are just two device trees that happen to point to the
53 * same data blocks.
54 *
55 * When we get a write in we decide if it's to a shared data block using
56 * some timestamp magic. If it is, we have to break sharing.
57 *
58 * Let's say we write to a shared block in what was the origin. The
59 * steps are:
60 *
61 * i) plug io further to this physical block. (see bio_prison code).
62 *
63 * ii) quiesce any read io to that shared data block. Obviously
64 * including all devices that share this block. (see dm_deferred_set code)
65 *
66 * iii) copy the data block to a newly allocate block. This step can be
67 * missed out if the io covers the block. (schedule_copy).
68 *
69 * iv) insert the new mapping into the origin's btree
70 * (process_prepared_mapping). This act of inserting breaks some
71 * sharing of btree nodes between the two devices. Breaking sharing only
72 * effects the btree of that specific device. Btrees for the other
73 * devices that share the block never change. The btree for the origin
74 * device as it was after the last commit is untouched, ie. we're using
75 * persistent data structures in the functional programming sense.
76 *
77 * v) unplug io to this physical block, including the io that triggered
78 * the breaking of sharing.
79 *
80 * Steps (ii) and (iii) occur in parallel.
81 *
82 * The metadata _doesn't_ need to be committed before the io continues. We
83 * get away with this because the io is always written to a _new_ block.
84 * If there's a crash, then:
85 *
86 * - The origin mapping will point to the old origin block (the shared
87 * one). This will contain the data as it was before the io that triggered
88 * the breaking of sharing came in.
89 *
90 * - The snap mapping still points to the old block. As it would after
91 * the commit.
92 *
93 * The downside of this scheme is the timestamp magic isn't perfect, and
94 * will continue to think that data block in the snapshot device is shared
95 * even after the write to the origin has broken sharing. I suspect data
96 * blocks will typically be shared by many different devices, so we're
97 * breaking sharing n + 1 times, rather than n, where n is the number of
98 * devices that reference this data block. At the moment I think the
99 * benefits far, far outweigh the disadvantages.
100 */
101
102/*----------------------------------------------------------------*/
103
104/*
105 * Key building.
106 */
107static void build_data_key(struct dm_thin_device *td,
108               dm_block_t b, struct dm_cell_key *key)
109{
110    key->virtual = 0;
111    key->dev = dm_thin_dev_id(td);
112    key->block = b;
113}
114
115static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
116                  struct dm_cell_key *key)
117{
118    key->virtual = 1;
119    key->dev = dm_thin_dev_id(td);
120    key->block = b;
121}
122
123/*----------------------------------------------------------------*/
124
125/*
126 * A pool device ties together a metadata device and a data device. It
127 * also provides the interface for creating and destroying internal
128 * devices.
129 */
130struct dm_thin_new_mapping;
131
132/*
133 * The pool runs in 3 modes. Ordered in degraded order for comparisons.
134 */
135enum pool_mode {
136    PM_WRITE, /* metadata may be changed */
137    PM_READ_ONLY, /* metadata may not be changed */
138    PM_FAIL, /* all I/O fails */
139};
140
141struct pool_features {
142    enum pool_mode mode;
143
144    bool zero_new_blocks:1;
145    bool discard_enabled:1;
146    bool discard_passdown:1;
147};
148
149struct thin_c;
150typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
151typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
152
153struct pool {
154    struct list_head list;
155    struct dm_target *ti; /* Only set if a pool target is bound */
156
157    struct mapped_device *pool_md;
158    struct block_device *md_dev;
159    struct dm_pool_metadata *pmd;
160
161    dm_block_t low_water_blocks;
162    uint32_t sectors_per_block;
163    int sectors_per_block_shift;
164
165    struct pool_features pf;
166    unsigned low_water_triggered:1; /* A dm event has been sent */
167    unsigned no_free_space:1; /* A -ENOSPC warning has been issued */
168
169    struct dm_bio_prison *prison;
170    struct dm_kcopyd_client *copier;
171
172    struct workqueue_struct *wq;
173    struct work_struct worker;
174    struct delayed_work waker;
175
176    unsigned long last_commit_jiffies;
177    unsigned ref_count;
178
179    spinlock_t lock;
180    struct bio_list deferred_bios;
181    struct bio_list deferred_flush_bios;
182    struct list_head prepared_mappings;
183    struct list_head prepared_discards;
184
185    struct bio_list retry_on_resume_list;
186
187    struct dm_deferred_set *shared_read_ds;
188    struct dm_deferred_set *all_io_ds;
189
190    struct dm_thin_new_mapping *next_mapping;
191    mempool_t *mapping_pool;
192
193    process_bio_fn process_bio;
194    process_bio_fn process_discard;
195
196    process_mapping_fn process_prepared_mapping;
197    process_mapping_fn process_prepared_discard;
198};
199
200static enum pool_mode get_pool_mode(struct pool *pool);
201static void set_pool_mode(struct pool *pool, enum pool_mode mode);
202
203/*
204 * Target context for a pool.
205 */
206struct pool_c {
207    struct dm_target *ti;
208    struct pool *pool;
209    struct dm_dev *data_dev;
210    struct dm_dev *metadata_dev;
211    struct dm_target_callbacks callbacks;
212
213    dm_block_t low_water_blocks;
214    struct pool_features requested_pf; /* Features requested during table load */
215    struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
216};
217
218/*
219 * Target context for a thin.
220 */
221struct thin_c {
222    struct dm_dev *pool_dev;
223    struct dm_dev *origin_dev;
224    dm_thin_id dev_id;
225
226    struct pool *pool;
227    struct dm_thin_device *td;
228};
229
230/*----------------------------------------------------------------*/
231
232/*
233 * wake_worker() is used when new work is queued and when pool_resume is
234 * ready to continue deferred IO processing.
235 */
236static void wake_worker(struct pool *pool)
237{
238    queue_work(pool->wq, &pool->worker);
239}
240
241/*----------------------------------------------------------------*/
242
243static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
244              struct dm_bio_prison_cell **cell_result)
245{
246    int r;
247    struct dm_bio_prison_cell *cell_prealloc;
248
249    /*
250     * Allocate a cell from the prison's mempool.
251     * This might block but it can't fail.
252     */
253    cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
254
255    r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
256    if (r)
257        /*
258         * We reused an old cell; we can get rid of
259         * the new one.
260         */
261        dm_bio_prison_free_cell(pool->prison, cell_prealloc);
262
263    return r;
264}
265
266static void cell_release(struct pool *pool,
267             struct dm_bio_prison_cell *cell,
268             struct bio_list *bios)
269{
270    dm_cell_release(pool->prison, cell, bios);
271    dm_bio_prison_free_cell(pool->prison, cell);
272}
273
274static void cell_release_no_holder(struct pool *pool,
275                   struct dm_bio_prison_cell *cell,
276                   struct bio_list *bios)
277{
278    dm_cell_release_no_holder(pool->prison, cell, bios);
279    dm_bio_prison_free_cell(pool->prison, cell);
280}
281
282static void cell_defer_no_holder_no_free(struct thin_c *tc,
283                     struct dm_bio_prison_cell *cell)
284{
285    struct pool *pool = tc->pool;
286    unsigned long flags;
287
288    spin_lock_irqsave(&pool->lock, flags);
289    dm_cell_release_no_holder(pool->prison, cell, &pool->deferred_bios);
290    spin_unlock_irqrestore(&pool->lock, flags);
291
292    wake_worker(pool);
293}
294
295static void cell_error(struct pool *pool,
296               struct dm_bio_prison_cell *cell)
297{
298    dm_cell_error(pool->prison, cell);
299    dm_bio_prison_free_cell(pool->prison, cell);
300}
301
302/*----------------------------------------------------------------*/
303
304/*
305 * A global list of pools that uses a struct mapped_device as a key.
306 */
307static struct dm_thin_pool_table {
308    struct mutex mutex;
309    struct list_head pools;
310} dm_thin_pool_table;
311
312static void pool_table_init(void)
313{
314    mutex_init(&dm_thin_pool_table.mutex);
315    INIT_LIST_HEAD(&dm_thin_pool_table.pools);
316}
317
318static void __pool_table_insert(struct pool *pool)
319{
320    BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
321    list_add(&pool->list, &dm_thin_pool_table.pools);
322}
323
324static void __pool_table_remove(struct pool *pool)
325{
326    BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
327    list_del(&pool->list);
328}
329
330static struct pool *__pool_table_lookup(struct mapped_device *md)
331{
332    struct pool *pool = NULL, *tmp;
333
334    BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
335
336    list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
337        if (tmp->pool_md == md) {
338            pool = tmp;
339            break;
340        }
341    }
342
343    return pool;
344}
345
346static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
347{
348    struct pool *pool = NULL, *tmp;
349
350    BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
351
352    list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
353        if (tmp->md_dev == md_dev) {
354            pool = tmp;
355            break;
356        }
357    }
358
359    return pool;
360}
361
362/*----------------------------------------------------------------*/
363
364struct dm_thin_endio_hook {
365    struct thin_c *tc;
366    struct dm_deferred_entry *shared_read_entry;
367    struct dm_deferred_entry *all_io_entry;
368    struct dm_thin_new_mapping *overwrite_mapping;
369};
370
371static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
372{
373    struct bio *bio;
374    struct bio_list bios;
375
376    bio_list_init(&bios);
377    bio_list_merge(&bios, master);
378    bio_list_init(master);
379
380    while ((bio = bio_list_pop(&bios))) {
381        struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
382
383        if (h->tc == tc)
384            bio_endio(bio, DM_ENDIO_REQUEUE);
385        else
386            bio_list_add(master, bio);
387    }
388}
389
390static void requeue_io(struct thin_c *tc)
391{
392    struct pool *pool = tc->pool;
393    unsigned long flags;
394
395    spin_lock_irqsave(&pool->lock, flags);
396    __requeue_bio_list(tc, &pool->deferred_bios);
397    __requeue_bio_list(tc, &pool->retry_on_resume_list);
398    spin_unlock_irqrestore(&pool->lock, flags);
399}
400
401/*
402 * This section of code contains the logic for processing a thin device's IO.
403 * Much of the code depends on pool object resources (lists, workqueues, etc)
404 * but most is exclusively called from the thin target rather than the thin-pool
405 * target.
406 */
407
408static bool block_size_is_power_of_two(struct pool *pool)
409{
410    return pool->sectors_per_block_shift >= 0;
411}
412
413static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
414{
415    struct pool *pool = tc->pool;
416    sector_t block_nr = bio->bi_sector;
417
418    if (block_size_is_power_of_two(pool))
419        block_nr >>= pool->sectors_per_block_shift;
420    else
421        (void) sector_div(block_nr, pool->sectors_per_block);
422
423    return block_nr;
424}
425
426static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
427{
428    struct pool *pool = tc->pool;
429    sector_t bi_sector = bio->bi_sector;
430
431    bio->bi_bdev = tc->pool_dev->bdev;
432    if (block_size_is_power_of_two(pool))
433        bio->bi_sector = (block << pool->sectors_per_block_shift) |
434                (bi_sector & (pool->sectors_per_block - 1));
435    else
436        bio->bi_sector = (block * pool->sectors_per_block) +
437                 sector_div(bi_sector, pool->sectors_per_block);
438}
439
440static void remap_to_origin(struct thin_c *tc, struct bio *bio)
441{
442    bio->bi_bdev = tc->origin_dev->bdev;
443}
444
445static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
446{
447    return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
448        dm_thin_changed_this_transaction(tc->td);
449}
450
451static void inc_all_io_entry(struct pool *pool, struct bio *bio)
452{
453    struct dm_thin_endio_hook *h;
454
455    if (bio->bi_rw & REQ_DISCARD)
456        return;
457
458    h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
459    h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
460}
461
462static void issue(struct thin_c *tc, struct bio *bio)
463{
464    struct pool *pool = tc->pool;
465    unsigned long flags;
466
467    if (!bio_triggers_commit(tc, bio)) {
468        generic_make_request(bio);
469        return;
470    }
471
472    /*
473     * Complete bio with an error if earlier I/O caused changes to
474     * the metadata that can't be committed e.g, due to I/O errors
475     * on the metadata device.
476     */
477    if (dm_thin_aborted_changes(tc->td)) {
478        bio_io_error(bio);
479        return;
480    }
481
482    /*
483     * Batch together any bios that trigger commits and then issue a
484     * single commit for them in process_deferred_bios().
485     */
486    spin_lock_irqsave(&pool->lock, flags);
487    bio_list_add(&pool->deferred_flush_bios, bio);
488    spin_unlock_irqrestore(&pool->lock, flags);
489}
490
491static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
492{
493    remap_to_origin(tc, bio);
494    issue(tc, bio);
495}
496
497static void remap_and_issue(struct thin_c *tc, struct bio *bio,
498                dm_block_t block)
499{
500    remap(tc, bio, block);
501    issue(tc, bio);
502}
503
504/*----------------------------------------------------------------*/
505
506/*
507 * Bio endio functions.
508 */
509struct dm_thin_new_mapping {
510    struct list_head list;
511
512    unsigned quiesced:1;
513    unsigned prepared:1;
514    unsigned pass_discard:1;
515
516    struct thin_c *tc;
517    dm_block_t virt_block;
518    dm_block_t data_block;
519    struct dm_bio_prison_cell *cell, *cell2;
520    int err;
521
522    /*
523     * If the bio covers the whole area of a block then we can avoid
524     * zeroing or copying. Instead this bio is hooked. The bio will
525     * still be in the cell, so care has to be taken to avoid issuing
526     * the bio twice.
527     */
528    struct bio *bio;
529    bio_end_io_t *saved_bi_end_io;
530};
531
532static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
533{
534    struct pool *pool = m->tc->pool;
535
536    if (m->quiesced && m->prepared) {
537        list_add(&m->list, &pool->prepared_mappings);
538        wake_worker(pool);
539    }
540}
541
542static void copy_complete(int read_err, unsigned long write_err, void *context)
543{
544    unsigned long flags;
545    struct dm_thin_new_mapping *m = context;
546    struct pool *pool = m->tc->pool;
547
548    m->err = read_err || write_err ? -EIO : 0;
549
550    spin_lock_irqsave(&pool->lock, flags);
551    m->prepared = 1;
552    __maybe_add_mapping(m);
553    spin_unlock_irqrestore(&pool->lock, flags);
554}
555
556static void overwrite_endio(struct bio *bio, int err)
557{
558    unsigned long flags;
559    struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
560    struct dm_thin_new_mapping *m = h->overwrite_mapping;
561    struct pool *pool = m->tc->pool;
562
563    m->err = err;
564
565    spin_lock_irqsave(&pool->lock, flags);
566    m->prepared = 1;
567    __maybe_add_mapping(m);
568    spin_unlock_irqrestore(&pool->lock, flags);
569}
570
571/*----------------------------------------------------------------*/
572
573/*
574 * Workqueue.
575 */
576
577/*
578 * Prepared mapping jobs.
579 */
580
581/*
582 * This sends the bios in the cell back to the deferred_bios list.
583 */
584static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
585{
586    struct pool *pool = tc->pool;
587    unsigned long flags;
588
589    spin_lock_irqsave(&pool->lock, flags);
590    cell_release(pool, cell, &pool->deferred_bios);
591    spin_unlock_irqrestore(&tc->pool->lock, flags);
592
593    wake_worker(pool);
594}
595
596/*
597 * Same as cell_defer above, except it omits the original holder of the cell.
598 */
599static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
600{
601    struct pool *pool = tc->pool;
602    unsigned long flags;
603
604    spin_lock_irqsave(&pool->lock, flags);
605    cell_release_no_holder(pool, cell, &pool->deferred_bios);
606    spin_unlock_irqrestore(&pool->lock, flags);
607
608    wake_worker(pool);
609}
610
611static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
612{
613    if (m->bio)
614        m->bio->bi_end_io = m->saved_bi_end_io;
615    cell_error(m->tc->pool, m->cell);
616    list_del(&m->list);
617    mempool_free(m, m->tc->pool->mapping_pool);
618}
619
620static void process_prepared_mapping(struct dm_thin_new_mapping *m)
621{
622    struct thin_c *tc = m->tc;
623    struct pool *pool = tc->pool;
624    struct bio *bio;
625    int r;
626
627    bio = m->bio;
628    if (bio)
629        bio->bi_end_io = m->saved_bi_end_io;
630
631    if (m->err) {
632        cell_error(pool, m->cell);
633        goto out;
634    }
635
636    /*
637     * Commit the prepared block into the mapping btree.
638     * Any I/O for this block arriving after this point will get
639     * remapped to it directly.
640     */
641    r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
642    if (r) {
643        DMERR_LIMIT("dm_thin_insert_block() failed");
644        cell_error(pool, m->cell);
645        goto out;
646    }
647
648    /*
649     * Release any bios held while the block was being provisioned.
650     * If we are processing a write bio that completely covers the block,
651     * we already processed it so can ignore it now when processing
652     * the bios in the cell.
653     */
654    if (bio) {
655        cell_defer_no_holder(tc, m->cell);
656        bio_endio(bio, 0);
657    } else
658        cell_defer(tc, m->cell);
659
660out:
661    list_del(&m->list);
662    mempool_free(m, pool->mapping_pool);
663}
664
665static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
666{
667    struct thin_c *tc = m->tc;
668
669    bio_io_error(m->bio);
670    cell_defer_no_holder(tc, m->cell);
671    cell_defer_no_holder(tc, m->cell2);
672    mempool_free(m, tc->pool->mapping_pool);
673}
674
675static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
676{
677    struct thin_c *tc = m->tc;
678
679    inc_all_io_entry(tc->pool, m->bio);
680    cell_defer_no_holder(tc, m->cell);
681    cell_defer_no_holder(tc, m->cell2);
682
683    if (m->pass_discard)
684        remap_and_issue(tc, m->bio, m->data_block);
685    else
686        bio_endio(m->bio, 0);
687
688    mempool_free(m, tc->pool->mapping_pool);
689}
690
691static void process_prepared_discard(struct dm_thin_new_mapping *m)
692{
693    int r;
694    struct thin_c *tc = m->tc;
695
696    r = dm_thin_remove_block(tc->td, m->virt_block);
697    if (r)
698        DMERR_LIMIT("dm_thin_remove_block() failed");
699
700    process_prepared_discard_passdown(m);
701}
702
703static void process_prepared(struct pool *pool, struct list_head *head,
704                 process_mapping_fn *fn)
705{
706    unsigned long flags;
707    struct list_head maps;
708    struct dm_thin_new_mapping *m, *tmp;
709
710    INIT_LIST_HEAD(&maps);
711    spin_lock_irqsave(&pool->lock, flags);
712    list_splice_init(head, &maps);
713    spin_unlock_irqrestore(&pool->lock, flags);
714
715    list_for_each_entry_safe(m, tmp, &maps, list)
716        (*fn)(m);
717}
718
719/*
720 * Deferred bio jobs.
721 */
722static int io_overlaps_block(struct pool *pool, struct bio *bio)
723{
724    return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
725}
726
727static int io_overwrites_block(struct pool *pool, struct bio *bio)
728{
729    return (bio_data_dir(bio) == WRITE) &&
730        io_overlaps_block(pool, bio);
731}
732
733static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
734                   bio_end_io_t *fn)
735{
736    *save = bio->bi_end_io;
737    bio->bi_end_io = fn;
738}
739
740static int ensure_next_mapping(struct pool *pool)
741{
742    if (pool->next_mapping)
743        return 0;
744
745    pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
746
747    return pool->next_mapping ? 0 : -ENOMEM;
748}
749
750static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
751{
752    struct dm_thin_new_mapping *r = pool->next_mapping;
753
754    BUG_ON(!pool->next_mapping);
755
756    pool->next_mapping = NULL;
757
758    return r;
759}
760
761static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
762              struct dm_dev *origin, dm_block_t data_origin,
763              dm_block_t data_dest,
764              struct dm_bio_prison_cell *cell, struct bio *bio)
765{
766    int r;
767    struct pool *pool = tc->pool;
768    struct dm_thin_new_mapping *m = get_next_mapping(pool);
769
770    INIT_LIST_HEAD(&m->list);
771    m->quiesced = 0;
772    m->prepared = 0;
773    m->tc = tc;
774    m->virt_block = virt_block;
775    m->data_block = data_dest;
776    m->cell = cell;
777    m->err = 0;
778    m->bio = NULL;
779
780    if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
781        m->quiesced = 1;
782
783    /*
784     * IO to pool_dev remaps to the pool target's data_dev.
785     *
786     * If the whole block of data is being overwritten, we can issue the
787     * bio immediately. Otherwise we use kcopyd to clone the data first.
788     */
789    if (io_overwrites_block(pool, bio)) {
790        struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
791
792        h->overwrite_mapping = m;
793        m->bio = bio;
794        save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
795        inc_all_io_entry(pool, bio);
796        remap_and_issue(tc, bio, data_dest);
797    } else {
798        struct dm_io_region from, to;
799
800        from.bdev = origin->bdev;
801        from.sector = data_origin * pool->sectors_per_block;
802        from.count = pool->sectors_per_block;
803
804        to.bdev = tc->pool_dev->bdev;
805        to.sector = data_dest * pool->sectors_per_block;
806        to.count = pool->sectors_per_block;
807
808        r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
809                   0, copy_complete, m);
810        if (r < 0) {
811            mempool_free(m, pool->mapping_pool);
812            DMERR_LIMIT("dm_kcopyd_copy() failed");
813            cell_error(pool, cell);
814        }
815    }
816}
817
818static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
819                   dm_block_t data_origin, dm_block_t data_dest,
820                   struct dm_bio_prison_cell *cell, struct bio *bio)
821{
822    schedule_copy(tc, virt_block, tc->pool_dev,
823              data_origin, data_dest, cell, bio);
824}
825
826static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
827                   dm_block_t data_dest,
828                   struct dm_bio_prison_cell *cell, struct bio *bio)
829{
830    schedule_copy(tc, virt_block, tc->origin_dev,
831              virt_block, data_dest, cell, bio);
832}
833
834static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
835              dm_block_t data_block, struct dm_bio_prison_cell *cell,
836              struct bio *bio)
837{
838    struct pool *pool = tc->pool;
839    struct dm_thin_new_mapping *m = get_next_mapping(pool);
840
841    INIT_LIST_HEAD(&m->list);
842    m->quiesced = 1;
843    m->prepared = 0;
844    m->tc = tc;
845    m->virt_block = virt_block;
846    m->data_block = data_block;
847    m->cell = cell;
848    m->err = 0;
849    m->bio = NULL;
850
851    /*
852     * If the whole block of data is being overwritten or we are not
853     * zeroing pre-existing data, we can issue the bio immediately.
854     * Otherwise we use kcopyd to zero the data first.
855     */
856    if (!pool->pf.zero_new_blocks)
857        process_prepared_mapping(m);
858
859    else if (io_overwrites_block(pool, bio)) {
860        struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
861
862        h->overwrite_mapping = m;
863        m->bio = bio;
864        save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
865        inc_all_io_entry(pool, bio);
866        remap_and_issue(tc, bio, data_block);
867    } else {
868        int r;
869        struct dm_io_region to;
870
871        to.bdev = tc->pool_dev->bdev;
872        to.sector = data_block * pool->sectors_per_block;
873        to.count = pool->sectors_per_block;
874
875        r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
876        if (r < 0) {
877            mempool_free(m, pool->mapping_pool);
878            DMERR_LIMIT("dm_kcopyd_zero() failed");
879            cell_error(pool, cell);
880        }
881    }
882}
883
884static int commit(struct pool *pool)
885{
886    int r;
887
888    r = dm_pool_commit_metadata(pool->pmd);
889    if (r)
890        DMERR_LIMIT("commit failed: error = %d", r);
891
892    return r;
893}
894
895/*
896 * A non-zero return indicates read_only or fail_io mode.
897 * Many callers don't care about the return value.
898 */
899static int commit_or_fallback(struct pool *pool)
900{
901    int r;
902
903    if (get_pool_mode(pool) != PM_WRITE)
904        return -EINVAL;
905
906    r = commit(pool);
907    if (r)
908        set_pool_mode(pool, PM_READ_ONLY);
909
910    return r;
911}
912
913static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
914{
915    int r;
916    dm_block_t free_blocks;
917    unsigned long flags;
918    struct pool *pool = tc->pool;
919
920    r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
921    if (r)
922        return r;
923
924    if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
925        DMWARN("%s: reached low water mark, sending event.",
926               dm_device_name(pool->pool_md));
927        spin_lock_irqsave(&pool->lock, flags);
928        pool->low_water_triggered = 1;
929        spin_unlock_irqrestore(&pool->lock, flags);
930        dm_table_event(pool->ti->table);
931    }
932
933    if (!free_blocks) {
934        if (pool->no_free_space)
935            return -ENOSPC;
936        else {
937            /*
938             * Try to commit to see if that will free up some
939             * more space.
940             */
941            (void) commit_or_fallback(pool);
942
943            r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
944            if (r)
945                return r;
946
947            /*
948             * If we still have no space we set a flag to avoid
949             * doing all this checking and return -ENOSPC.
950             */
951            if (!free_blocks) {
952                DMWARN("%s: no free space available.",
953                       dm_device_name(pool->pool_md));
954                spin_lock_irqsave(&pool->lock, flags);
955                pool->no_free_space = 1;
956                spin_unlock_irqrestore(&pool->lock, flags);
957                return -ENOSPC;
958            }
959        }
960    }
961
962    r = dm_pool_alloc_data_block(pool->pmd, result);
963    if (r)
964        return r;
965
966    return 0;
967}
968
969/*
970 * If we have run out of space, queue bios until the device is
971 * resumed, presumably after having been reloaded with more space.
972 */
973static void retry_on_resume(struct bio *bio)
974{
975    struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
976    struct thin_c *tc = h->tc;
977    struct pool *pool = tc->pool;
978    unsigned long flags;
979
980    spin_lock_irqsave(&pool->lock, flags);
981    bio_list_add(&pool->retry_on_resume_list, bio);
982    spin_unlock_irqrestore(&pool->lock, flags);
983}
984
985static void no_space(struct pool *pool, struct dm_bio_prison_cell *cell)
986{
987    struct bio *bio;
988    struct bio_list bios;
989
990    bio_list_init(&bios);
991    cell_release(pool, cell, &bios);
992
993    while ((bio = bio_list_pop(&bios)))
994        retry_on_resume(bio);
995}
996
997static void process_discard(struct thin_c *tc, struct bio *bio)
998{
999    int r;
1000    unsigned long flags;
1001    struct pool *pool = tc->pool;
1002    struct dm_bio_prison_cell *cell, *cell2;
1003    struct dm_cell_key key, key2;
1004    dm_block_t block = get_bio_block(tc, bio);
1005    struct dm_thin_lookup_result lookup_result;
1006    struct dm_thin_new_mapping *m;
1007
1008    build_virtual_key(tc->td, block, &key);
1009    if (bio_detain(tc->pool, &key, bio, &cell))
1010        return;
1011
1012    r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1013    switch (r) {
1014    case 0:
1015        /*
1016         * Check nobody is fiddling with this pool block. This can
1017         * happen if someone's in the process of breaking sharing
1018         * on this block.
1019         */
1020        build_data_key(tc->td, lookup_result.block, &key2);
1021        if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1022            cell_defer_no_holder(tc, cell);
1023            break;
1024        }
1025
1026        if (io_overlaps_block(pool, bio)) {
1027            /*
1028             * IO may still be going to the destination block. We must
1029             * quiesce before we can do the removal.
1030             */
1031            m = get_next_mapping(pool);
1032            m->tc = tc;
1033            m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
1034            m->virt_block = block;
1035            m->data_block = lookup_result.block;
1036            m->cell = cell;
1037            m->cell2 = cell2;
1038            m->err = 0;
1039            m->bio = bio;
1040
1041            if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
1042                spin_lock_irqsave(&pool->lock, flags);
1043                list_add(&m->list, &pool->prepared_discards);
1044                spin_unlock_irqrestore(&pool->lock, flags);
1045                wake_worker(pool);
1046            }
1047        } else {
1048            inc_all_io_entry(pool, bio);
1049            cell_defer_no_holder(tc, cell);
1050            cell_defer_no_holder(tc, cell2);
1051
1052            /*
1053             * The DM core makes sure that the discard doesn't span
1054             * a block boundary. So we submit the discard of a
1055             * partial block appropriately.
1056             */
1057            if ((!lookup_result.shared) && pool->pf.discard_passdown)
1058                remap_and_issue(tc, bio, lookup_result.block);
1059            else
1060                bio_endio(bio, 0);
1061        }
1062        break;
1063
1064    case -ENODATA:
1065        /*
1066         * It isn't provisioned, just forget it.
1067         */
1068        cell_defer_no_holder(tc, cell);
1069        bio_endio(bio, 0);
1070        break;
1071
1072    default:
1073        DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1074                __func__, r);
1075        cell_defer_no_holder(tc, cell);
1076        bio_io_error(bio);
1077        break;
1078    }
1079}
1080
1081static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1082              struct dm_cell_key *key,
1083              struct dm_thin_lookup_result *lookup_result,
1084              struct dm_bio_prison_cell *cell)
1085{
1086    int r;
1087    dm_block_t data_block;
1088
1089    r = alloc_data_block(tc, &data_block);
1090    switch (r) {
1091    case 0:
1092        schedule_internal_copy(tc, block, lookup_result->block,
1093                       data_block, cell, bio);
1094        break;
1095
1096    case -ENOSPC:
1097        no_space(tc->pool, cell);
1098        break;
1099
1100    default:
1101        DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1102                __func__, r);
1103        cell_error(tc->pool, cell);
1104        break;
1105    }
1106}
1107
1108static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1109                   dm_block_t block,
1110                   struct dm_thin_lookup_result *lookup_result)
1111{
1112    struct dm_bio_prison_cell *cell;
1113    struct pool *pool = tc->pool;
1114    struct dm_cell_key key;
1115
1116    /*
1117     * If cell is already occupied, then sharing is already in the process
1118     * of being broken so we have nothing further to do here.
1119     */
1120    build_data_key(tc->td, lookup_result->block, &key);
1121    if (bio_detain(pool, &key, bio, &cell))
1122        return;
1123
1124    if (bio_data_dir(bio) == WRITE && bio->bi_size)
1125        break_sharing(tc, bio, block, &key, lookup_result, cell);
1126    else {
1127        struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1128
1129        h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1130        inc_all_io_entry(pool, bio);
1131        cell_defer_no_holder(tc, cell);
1132
1133        remap_and_issue(tc, bio, lookup_result->block);
1134    }
1135}
1136
1137static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1138                struct dm_bio_prison_cell *cell)
1139{
1140    int r;
1141    dm_block_t data_block;
1142    struct pool *pool = tc->pool;
1143
1144    /*
1145     * Remap empty bios (flushes) immediately, without provisioning.
1146     */
1147    if (!bio->bi_size) {
1148        inc_all_io_entry(pool, bio);
1149        cell_defer_no_holder(tc, cell);
1150
1151        remap_and_issue(tc, bio, 0);
1152        return;
1153    }
1154
1155    /*
1156     * Fill read bios with zeroes and complete them immediately.
1157     */
1158    if (bio_data_dir(bio) == READ) {
1159        zero_fill_bio(bio);
1160        cell_defer_no_holder(tc, cell);
1161        bio_endio(bio, 0);
1162        return;
1163    }
1164
1165    r = alloc_data_block(tc, &data_block);
1166    switch (r) {
1167    case 0:
1168        if (tc->origin_dev)
1169            schedule_external_copy(tc, block, data_block, cell, bio);
1170        else
1171            schedule_zero(tc, block, data_block, cell, bio);
1172        break;
1173
1174    case -ENOSPC:
1175        no_space(pool, cell);
1176        break;
1177
1178    default:
1179        DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1180                __func__, r);
1181        set_pool_mode(pool, PM_READ_ONLY);
1182        cell_error(pool, cell);
1183        break;
1184    }
1185}
1186
1187static void process_bio(struct thin_c *tc, struct bio *bio)
1188{
1189    int r;
1190    struct pool *pool = tc->pool;
1191    dm_block_t block = get_bio_block(tc, bio);
1192    struct dm_bio_prison_cell *cell;
1193    struct dm_cell_key key;
1194    struct dm_thin_lookup_result lookup_result;
1195
1196    /*
1197     * If cell is already occupied, then the block is already
1198     * being provisioned so we have nothing further to do here.
1199     */
1200    build_virtual_key(tc->td, block, &key);
1201    if (bio_detain(pool, &key, bio, &cell))
1202        return;
1203
1204    r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1205    switch (r) {
1206    case 0:
1207        if (lookup_result.shared) {
1208            process_shared_bio(tc, bio, block, &lookup_result);
1209            cell_defer_no_holder(tc, cell); /* FIXME: pass this cell into process_shared? */
1210        } else {
1211            inc_all_io_entry(pool, bio);
1212            cell_defer_no_holder(tc, cell);
1213
1214            remap_and_issue(tc, bio, lookup_result.block);
1215        }
1216        break;
1217
1218    case -ENODATA:
1219        if (bio_data_dir(bio) == READ && tc->origin_dev) {
1220            inc_all_io_entry(pool, bio);
1221            cell_defer_no_holder(tc, cell);
1222
1223            remap_to_origin_and_issue(tc, bio);
1224        } else
1225            provision_block(tc, bio, block, cell);
1226        break;
1227
1228    default:
1229        DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1230                __func__, r);
1231        cell_defer_no_holder(tc, cell);
1232        bio_io_error(bio);
1233        break;
1234    }
1235}
1236
1237static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1238{
1239    int r;
1240    int rw = bio_data_dir(bio);
1241    dm_block_t block = get_bio_block(tc, bio);
1242    struct dm_thin_lookup_result lookup_result;
1243
1244    r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1245    switch (r) {
1246    case 0:
1247        if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1248            bio_io_error(bio);
1249        else {
1250            inc_all_io_entry(tc->pool, bio);
1251            remap_and_issue(tc, bio, lookup_result.block);
1252        }
1253        break;
1254
1255    case -ENODATA:
1256        if (rw != READ) {
1257            bio_io_error(bio);
1258            break;
1259        }
1260
1261        if (tc->origin_dev) {
1262            inc_all_io_entry(tc->pool, bio);
1263            remap_to_origin_and_issue(tc, bio);
1264            break;
1265        }
1266
1267        zero_fill_bio(bio);
1268        bio_endio(bio, 0);
1269        break;
1270
1271    default:
1272        DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1273                __func__, r);
1274        bio_io_error(bio);
1275        break;
1276    }
1277}
1278
1279static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1280{
1281    bio_io_error(bio);
1282}
1283
1284static int need_commit_due_to_time(struct pool *pool)
1285{
1286    return jiffies < pool->last_commit_jiffies ||
1287           jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1288}
1289
1290static void process_deferred_bios(struct pool *pool)
1291{
1292    unsigned long flags;
1293    struct bio *bio;
1294    struct bio_list bios;
1295
1296    bio_list_init(&bios);
1297
1298    spin_lock_irqsave(&pool->lock, flags);
1299    bio_list_merge(&bios, &pool->deferred_bios);
1300    bio_list_init(&pool->deferred_bios);
1301    spin_unlock_irqrestore(&pool->lock, flags);
1302
1303    while ((bio = bio_list_pop(&bios))) {
1304        struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1305        struct thin_c *tc = h->tc;
1306
1307        /*
1308         * If we've got no free new_mapping structs, and processing
1309         * this bio might require one, we pause until there are some
1310         * prepared mappings to process.
1311         */
1312        if (ensure_next_mapping(pool)) {
1313            spin_lock_irqsave(&pool->lock, flags);
1314            bio_list_merge(&pool->deferred_bios, &bios);
1315            spin_unlock_irqrestore(&pool->lock, flags);
1316
1317            break;
1318        }
1319
1320        if (bio->bi_rw & REQ_DISCARD)
1321            pool->process_discard(tc, bio);
1322        else
1323            pool->process_bio(tc, bio);
1324    }
1325
1326    /*
1327     * If there are any deferred flush bios, we must commit
1328     * the metadata before issuing them.
1329     */
1330    bio_list_init(&bios);
1331    spin_lock_irqsave(&pool->lock, flags);
1332    bio_list_merge(&bios, &pool->deferred_flush_bios);
1333    bio_list_init(&pool->deferred_flush_bios);
1334    spin_unlock_irqrestore(&pool->lock, flags);
1335
1336    if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1337        return;
1338
1339    if (commit_or_fallback(pool)) {
1340        while ((bio = bio_list_pop(&bios)))
1341            bio_io_error(bio);
1342        return;
1343    }
1344    pool->last_commit_jiffies = jiffies;
1345
1346    while ((bio = bio_list_pop(&bios)))
1347        generic_make_request(bio);
1348}
1349
1350static void do_worker(struct work_struct *ws)
1351{
1352    struct pool *pool = container_of(ws, struct pool, worker);
1353
1354    process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1355    process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1356    process_deferred_bios(pool);
1357}
1358
1359/*
1360 * We want to commit periodically so that not too much
1361 * unwritten data builds up.
1362 */
1363static void do_waker(struct work_struct *ws)
1364{
1365    struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1366    wake_worker(pool);
1367    queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1368}
1369
1370/*----------------------------------------------------------------*/
1371
1372static enum pool_mode get_pool_mode(struct pool *pool)
1373{
1374    return pool->pf.mode;
1375}
1376
1377static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1378{
1379    int r;
1380
1381    pool->pf.mode = mode;
1382
1383    switch (mode) {
1384    case PM_FAIL:
1385        DMERR("switching pool to failure mode");
1386        pool->process_bio = process_bio_fail;
1387        pool->process_discard = process_bio_fail;
1388        pool->process_prepared_mapping = process_prepared_mapping_fail;
1389        pool->process_prepared_discard = process_prepared_discard_fail;
1390        break;
1391
1392    case PM_READ_ONLY:
1393        DMERR("switching pool to read-only mode");
1394        r = dm_pool_abort_metadata(pool->pmd);
1395        if (r) {
1396            DMERR("aborting transaction failed");
1397            set_pool_mode(pool, PM_FAIL);
1398        } else {
1399            dm_pool_metadata_read_only(pool->pmd);
1400            pool->process_bio = process_bio_read_only;
1401            pool->process_discard = process_discard;
1402            pool->process_prepared_mapping = process_prepared_mapping_fail;
1403            pool->process_prepared_discard = process_prepared_discard_passdown;
1404        }
1405        break;
1406
1407    case PM_WRITE:
1408        pool->process_bio = process_bio;
1409        pool->process_discard = process_discard;
1410        pool->process_prepared_mapping = process_prepared_mapping;
1411        pool->process_prepared_discard = process_prepared_discard;
1412        break;
1413    }
1414}
1415
1416/*----------------------------------------------------------------*/
1417
1418/*
1419 * Mapping functions.
1420 */
1421
1422/*
1423 * Called only while mapping a thin bio to hand it over to the workqueue.
1424 */
1425static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1426{
1427    unsigned long flags;
1428    struct pool *pool = tc->pool;
1429
1430    spin_lock_irqsave(&pool->lock, flags);
1431    bio_list_add(&pool->deferred_bios, bio);
1432    spin_unlock_irqrestore(&pool->lock, flags);
1433
1434    wake_worker(pool);
1435}
1436
1437static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
1438{
1439    struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1440
1441    h->tc = tc;
1442    h->shared_read_entry = NULL;
1443    h->all_io_entry = NULL;
1444    h->overwrite_mapping = NULL;
1445}
1446
1447/*
1448 * Non-blocking function called from the thin target's map function.
1449 */
1450static int thin_bio_map(struct dm_target *ti, struct bio *bio)
1451{
1452    int r;
1453    struct thin_c *tc = ti->private;
1454    dm_block_t block = get_bio_block(tc, bio);
1455    struct dm_thin_device *td = tc->td;
1456    struct dm_thin_lookup_result result;
1457    struct dm_bio_prison_cell cell1, cell2;
1458    struct dm_bio_prison_cell *cell_result;
1459    struct dm_cell_key key;
1460
1461    thin_hook_bio(tc, bio);
1462
1463    if (get_pool_mode(tc->pool) == PM_FAIL) {
1464        bio_io_error(bio);
1465        return DM_MAPIO_SUBMITTED;
1466    }
1467
1468    if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1469        thin_defer_bio(tc, bio);
1470        return DM_MAPIO_SUBMITTED;
1471    }
1472
1473    r = dm_thin_find_block(td, block, 0, &result);
1474
1475    /*
1476     * Note that we defer readahead too.
1477     */
1478    switch (r) {
1479    case 0:
1480        if (unlikely(result.shared)) {
1481            /*
1482             * We have a race condition here between the
1483             * result.shared value returned by the lookup and
1484             * snapshot creation, which may cause new
1485             * sharing.
1486             *
1487             * To avoid this always quiesce the origin before
1488             * taking the snap. You want to do this anyway to
1489             * ensure a consistent application view
1490             * (i.e. lockfs).
1491             *
1492             * More distant ancestors are irrelevant. The
1493             * shared flag will be set in their case.
1494             */
1495            thin_defer_bio(tc, bio);
1496            return DM_MAPIO_SUBMITTED;
1497        }
1498
1499        build_virtual_key(tc->td, block, &key);
1500        if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1, &cell_result))
1501            return DM_MAPIO_SUBMITTED;
1502
1503        build_data_key(tc->td, result.block, &key);
1504        if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2, &cell_result)) {
1505            cell_defer_no_holder_no_free(tc, &cell1);
1506            return DM_MAPIO_SUBMITTED;
1507        }
1508
1509        inc_all_io_entry(tc->pool, bio);
1510        cell_defer_no_holder_no_free(tc, &cell2);
1511        cell_defer_no_holder_no_free(tc, &cell1);
1512
1513        remap(tc, bio, result.block);
1514        return DM_MAPIO_REMAPPED;
1515
1516    case -ENODATA:
1517        if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1518            /*
1519             * This block isn't provisioned, and we have no way
1520             * of doing so. Just error it.
1521             */
1522            bio_io_error(bio);
1523            return DM_MAPIO_SUBMITTED;
1524        }
1525        /* fall through */
1526
1527    case -EWOULDBLOCK:
1528        /*
1529         * In future, the failed dm_thin_find_block above could
1530         * provide the hint to load the metadata into cache.
1531         */
1532        thin_defer_bio(tc, bio);
1533        return DM_MAPIO_SUBMITTED;
1534
1535    default:
1536        /*
1537         * Must always call bio_io_error on failure.
1538         * dm_thin_find_block can fail with -EINVAL if the
1539         * pool is switched to fail-io mode.
1540         */
1541        bio_io_error(bio);
1542        return DM_MAPIO_SUBMITTED;
1543    }
1544}
1545
1546static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1547{
1548    int r;
1549    unsigned long flags;
1550    struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1551
1552    spin_lock_irqsave(&pt->pool->lock, flags);
1553    r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1554    spin_unlock_irqrestore(&pt->pool->lock, flags);
1555
1556    if (!r) {
1557        struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1558        r = bdi_congested(&q->backing_dev_info, bdi_bits);
1559    }
1560
1561    return r;
1562}
1563
1564static void __requeue_bios(struct pool *pool)
1565{
1566    bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1567    bio_list_init(&pool->retry_on_resume_list);
1568}
1569
1570/*----------------------------------------------------------------
1571 * Binding of control targets to a pool object
1572 *--------------------------------------------------------------*/
1573static bool data_dev_supports_discard(struct pool_c *pt)
1574{
1575    struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1576
1577    return q && blk_queue_discard(q);
1578}
1579
1580static bool is_factor(sector_t block_size, uint32_t n)
1581{
1582    return !sector_div(block_size, n);
1583}
1584
1585/*
1586 * If discard_passdown was enabled verify that the data device
1587 * supports discards. Disable discard_passdown if not.
1588 */
1589static void disable_passdown_if_not_supported(struct pool_c *pt)
1590{
1591    struct pool *pool = pt->pool;
1592    struct block_device *data_bdev = pt->data_dev->bdev;
1593    struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1594    sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1595    const char *reason = NULL;
1596    char buf[BDEVNAME_SIZE];
1597
1598    if (!pt->adjusted_pf.discard_passdown)
1599        return;
1600
1601    if (!data_dev_supports_discard(pt))
1602        reason = "discard unsupported";
1603
1604    else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1605        reason = "max discard sectors smaller than a block";
1606
1607    else if (data_limits->discard_granularity > block_size)
1608        reason = "discard granularity larger than a block";
1609
1610    else if (!is_factor(block_size, data_limits->discard_granularity))
1611        reason = "discard granularity not a factor of block size";
1612
1613    if (reason) {
1614        DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1615        pt->adjusted_pf.discard_passdown = false;
1616    }
1617}
1618
1619static int bind_control_target(struct pool *pool, struct dm_target *ti)
1620{
1621    struct pool_c *pt = ti->private;
1622
1623    /*
1624     * We want to make sure that degraded pools are never upgraded.
1625     */
1626    enum pool_mode old_mode = pool->pf.mode;
1627    enum pool_mode new_mode = pt->adjusted_pf.mode;
1628
1629    if (old_mode > new_mode)
1630        new_mode = old_mode;
1631
1632    pool->ti = ti;
1633    pool->low_water_blocks = pt->low_water_blocks;
1634    pool->pf = pt->adjusted_pf;
1635
1636    set_pool_mode(pool, new_mode);
1637
1638    return 0;
1639}
1640
1641static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1642{
1643    if (pool->ti == ti)
1644        pool->ti = NULL;
1645}
1646
1647/*----------------------------------------------------------------
1648 * Pool creation
1649 *--------------------------------------------------------------*/
1650/* Initialize pool features. */
1651static void pool_features_init(struct pool_features *pf)
1652{
1653    pf->mode = PM_WRITE;
1654    pf->zero_new_blocks = true;
1655    pf->discard_enabled = true;
1656    pf->discard_passdown = true;
1657}
1658
1659static void __pool_destroy(struct pool *pool)
1660{
1661    __pool_table_remove(pool);
1662
1663    if (dm_pool_metadata_close(pool->pmd) < 0)
1664        DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1665
1666    dm_bio_prison_destroy(pool->prison);
1667    dm_kcopyd_client_destroy(pool->copier);
1668
1669    if (pool->wq)
1670        destroy_workqueue(pool->wq);
1671
1672    if (pool->next_mapping)
1673        mempool_free(pool->next_mapping, pool->mapping_pool);
1674    mempool_destroy(pool->mapping_pool);
1675    dm_deferred_set_destroy(pool->shared_read_ds);
1676    dm_deferred_set_destroy(pool->all_io_ds);
1677    kfree(pool);
1678}
1679
1680static struct kmem_cache *_new_mapping_cache;
1681
1682static struct pool *pool_create(struct mapped_device *pool_md,
1683                struct block_device *metadata_dev,
1684                unsigned long block_size,
1685                int read_only, char **error)
1686{
1687    int r;
1688    void *err_p;
1689    struct pool *pool;
1690    struct dm_pool_metadata *pmd;
1691    bool format_device = read_only ? false : true;
1692
1693    pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1694    if (IS_ERR(pmd)) {
1695        *error = "Error creating metadata object";
1696        return (struct pool *)pmd;
1697    }
1698
1699    pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1700    if (!pool) {
1701        *error = "Error allocating memory for pool";
1702        err_p = ERR_PTR(-ENOMEM);
1703        goto bad_pool;
1704    }
1705
1706    pool->pmd = pmd;
1707    pool->sectors_per_block = block_size;
1708    if (block_size & (block_size - 1))
1709        pool->sectors_per_block_shift = -1;
1710    else
1711        pool->sectors_per_block_shift = __ffs(block_size);
1712    pool->low_water_blocks = 0;
1713    pool_features_init(&pool->pf);
1714    pool->prison = dm_bio_prison_create(PRISON_CELLS);
1715    if (!pool->prison) {
1716        *error = "Error creating pool's bio prison";
1717        err_p = ERR_PTR(-ENOMEM);
1718        goto bad_prison;
1719    }
1720
1721    pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1722    if (IS_ERR(pool->copier)) {
1723        r = PTR_ERR(pool->copier);
1724        *error = "Error creating pool's kcopyd client";
1725        err_p = ERR_PTR(r);
1726        goto bad_kcopyd_client;
1727    }
1728
1729    /*
1730     * Create singlethreaded workqueue that will service all devices
1731     * that use this metadata.
1732     */
1733    pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1734    if (!pool->wq) {
1735        *error = "Error creating pool's workqueue";
1736        err_p = ERR_PTR(-ENOMEM);
1737        goto bad_wq;
1738    }
1739
1740    INIT_WORK(&pool->worker, do_worker);
1741    INIT_DELAYED_WORK(&pool->waker, do_waker);
1742    spin_lock_init(&pool->lock);
1743    bio_list_init(&pool->deferred_bios);
1744    bio_list_init(&pool->deferred_flush_bios);
1745    INIT_LIST_HEAD(&pool->prepared_mappings);
1746    INIT_LIST_HEAD(&pool->prepared_discards);
1747    pool->low_water_triggered = 0;
1748    pool->no_free_space = 0;
1749    bio_list_init(&pool->retry_on_resume_list);
1750
1751    pool->shared_read_ds = dm_deferred_set_create();
1752    if (!pool->shared_read_ds) {
1753        *error = "Error creating pool's shared read deferred set";
1754        err_p = ERR_PTR(-ENOMEM);
1755        goto bad_shared_read_ds;
1756    }
1757
1758    pool->all_io_ds = dm_deferred_set_create();
1759    if (!pool->all_io_ds) {
1760        *error = "Error creating pool's all io deferred set";
1761        err_p = ERR_PTR(-ENOMEM);
1762        goto bad_all_io_ds;
1763    }
1764
1765    pool->next_mapping = NULL;
1766    pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1767                              _new_mapping_cache);
1768    if (!pool->mapping_pool) {
1769        *error = "Error creating pool's mapping mempool";
1770        err_p = ERR_PTR(-ENOMEM);
1771        goto bad_mapping_pool;
1772    }
1773
1774    pool->ref_count = 1;
1775    pool->last_commit_jiffies = jiffies;
1776    pool->pool_md = pool_md;
1777    pool->md_dev = metadata_dev;
1778    __pool_table_insert(pool);
1779
1780    return pool;
1781
1782bad_mapping_pool:
1783    dm_deferred_set_destroy(pool->all_io_ds);
1784bad_all_io_ds:
1785    dm_deferred_set_destroy(pool->shared_read_ds);
1786bad_shared_read_ds:
1787    destroy_workqueue(pool->wq);
1788bad_wq:
1789    dm_kcopyd_client_destroy(pool->copier);
1790bad_kcopyd_client:
1791    dm_bio_prison_destroy(pool->prison);
1792bad_prison:
1793    kfree(pool);
1794bad_pool:
1795    if (dm_pool_metadata_close(pmd))
1796        DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1797
1798    return err_p;
1799}
1800
1801static void __pool_inc(struct pool *pool)
1802{
1803    BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1804    pool->ref_count++;
1805}
1806
1807static void __pool_dec(struct pool *pool)
1808{
1809    BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1810    BUG_ON(!pool->ref_count);
1811    if (!--pool->ref_count)
1812        __pool_destroy(pool);
1813}
1814
1815static struct pool *__pool_find(struct mapped_device *pool_md,
1816                struct block_device *metadata_dev,
1817                unsigned long block_size, int read_only,
1818                char **error, int *created)
1819{
1820    struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1821
1822    if (pool) {
1823        if (pool->pool_md != pool_md) {
1824            *error = "metadata device already in use by a pool";
1825            return ERR_PTR(-EBUSY);
1826        }
1827        __pool_inc(pool);
1828
1829    } else {
1830        pool = __pool_table_lookup(pool_md);
1831        if (pool) {
1832            if (pool->md_dev != metadata_dev) {
1833                *error = "different pool cannot replace a pool";
1834                return ERR_PTR(-EINVAL);
1835            }
1836            __pool_inc(pool);
1837
1838        } else {
1839            pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1840            *created = 1;
1841        }
1842    }
1843
1844    return pool;
1845}
1846
1847/*----------------------------------------------------------------
1848 * Pool target methods
1849 *--------------------------------------------------------------*/
1850static void pool_dtr(struct dm_target *ti)
1851{
1852    struct pool_c *pt = ti->private;
1853
1854    mutex_lock(&dm_thin_pool_table.mutex);
1855
1856    unbind_control_target(pt->pool, ti);
1857    __pool_dec(pt->pool);
1858    dm_put_device(ti, pt->metadata_dev);
1859    dm_put_device(ti, pt->data_dev);
1860    kfree(pt);
1861
1862    mutex_unlock(&dm_thin_pool_table.mutex);
1863}
1864
1865static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1866                   struct dm_target *ti)
1867{
1868    int r;
1869    unsigned argc;
1870    const char *arg_name;
1871
1872    static struct dm_arg _args[] = {
1873        {0, 3, "Invalid number of pool feature arguments"},
1874    };
1875
1876    /*
1877     * No feature arguments supplied.
1878     */
1879    if (!as->argc)
1880        return 0;
1881
1882    r = dm_read_arg_group(_args, as, &argc, &ti->error);
1883    if (r)
1884        return -EINVAL;
1885
1886    while (argc && !r) {
1887        arg_name = dm_shift_arg(as);
1888        argc--;
1889
1890        if (!strcasecmp(arg_name, "skip_block_zeroing"))
1891            pf->zero_new_blocks = false;
1892
1893        else if (!strcasecmp(arg_name, "ignore_discard"))
1894            pf->discard_enabled = false;
1895
1896        else if (!strcasecmp(arg_name, "no_discard_passdown"))
1897            pf->discard_passdown = false;
1898
1899        else if (!strcasecmp(arg_name, "read_only"))
1900            pf->mode = PM_READ_ONLY;
1901
1902        else {
1903            ti->error = "Unrecognised pool feature requested";
1904            r = -EINVAL;
1905            break;
1906        }
1907    }
1908
1909    return r;
1910}
1911
1912/*
1913 * thin-pool <metadata dev> <data dev>
1914 * <data block size (sectors)>
1915 * <low water mark (blocks)>
1916 * [<#feature args> [<arg>]*]
1917 *
1918 * Optional feature arguments are:
1919 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1920 * ignore_discard: disable discard
1921 * no_discard_passdown: don't pass discards down to the data device
1922 */
1923static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1924{
1925    int r, pool_created = 0;
1926    struct pool_c *pt;
1927    struct pool *pool;
1928    struct pool_features pf;
1929    struct dm_arg_set as;
1930    struct dm_dev *data_dev;
1931    unsigned long block_size;
1932    dm_block_t low_water_blocks;
1933    struct dm_dev *metadata_dev;
1934    sector_t metadata_dev_size;
1935    char b[BDEVNAME_SIZE];
1936
1937    /*
1938     * FIXME Remove validation from scope of lock.
1939     */
1940    mutex_lock(&dm_thin_pool_table.mutex);
1941
1942    if (argc < 4) {
1943        ti->error = "Invalid argument count";
1944        r = -EINVAL;
1945        goto out_unlock;
1946    }
1947    as.argc = argc;
1948    as.argv = argv;
1949
1950    r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1951    if (r) {
1952        ti->error = "Error opening metadata block device";
1953        goto out_unlock;
1954    }
1955
1956    metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1957    if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1958        DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1959               bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1960
1961    r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1962    if (r) {
1963        ti->error = "Error getting data device";
1964        goto out_metadata;
1965    }
1966
1967    if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1968        block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1969        block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1970        block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1971        ti->error = "Invalid block size";
1972        r = -EINVAL;
1973        goto out;
1974    }
1975
1976    if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1977        ti->error = "Invalid low water mark";
1978        r = -EINVAL;
1979        goto out;
1980    }
1981
1982    /*
1983     * Set default pool features.
1984     */
1985    pool_features_init(&pf);
1986
1987    dm_consume_args(&as, 4);
1988    r = parse_pool_features(&as, &pf, ti);
1989    if (r)
1990        goto out;
1991
1992    pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1993    if (!pt) {
1994        r = -ENOMEM;
1995        goto out;
1996    }
1997
1998    pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1999               block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2000    if (IS_ERR(pool)) {
2001        r = PTR_ERR(pool);
2002        goto out_free_pt;
2003    }
2004
2005    /*
2006     * 'pool_created' reflects whether this is the first table load.
2007     * Top level discard support is not allowed to be changed after
2008     * initial load. This would require a pool reload to trigger thin
2009     * device changes.
2010     */
2011    if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2012        ti->error = "Discard support cannot be disabled once enabled";
2013        r = -EINVAL;
2014        goto out_flags_changed;
2015    }
2016
2017    pt->pool = pool;
2018    pt->ti = ti;
2019    pt->metadata_dev = metadata_dev;
2020    pt->data_dev = data_dev;
2021    pt->low_water_blocks = low_water_blocks;
2022    pt->adjusted_pf = pt->requested_pf = pf;
2023    ti->num_flush_bios = 1;
2024
2025    /*
2026     * Only need to enable discards if the pool should pass
2027     * them down to the data device. The thin device's discard
2028     * processing will cause mappings to be removed from the btree.
2029     */
2030    if (pf.discard_enabled && pf.discard_passdown) {
2031        ti->num_discard_bios = 1;
2032
2033        /*
2034         * Setting 'discards_supported' circumvents the normal
2035         * stacking of discard limits (this keeps the pool and
2036         * thin devices' discard limits consistent).
2037         */
2038        ti->discards_supported = true;
2039        ti->discard_zeroes_data_unsupported = true;
2040    }
2041    ti->private = pt;
2042
2043    pt->callbacks.congested_fn = pool_is_congested;
2044    dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2045
2046    mutex_unlock(&dm_thin_pool_table.mutex);
2047
2048    return 0;
2049
2050out_flags_changed:
2051    __pool_dec(pool);
2052out_free_pt:
2053    kfree(pt);
2054out:
2055    dm_put_device(ti, data_dev);
2056out_metadata:
2057    dm_put_device(ti, metadata_dev);
2058out_unlock:
2059    mutex_unlock(&dm_thin_pool_table.mutex);
2060
2061    return r;
2062}
2063
2064static int pool_map(struct dm_target *ti, struct bio *bio)
2065{
2066    int r;
2067    struct pool_c *pt = ti->private;
2068    struct pool *pool = pt->pool;
2069    unsigned long flags;
2070
2071    /*
2072     * As this is a singleton target, ti->begin is always zero.
2073     */
2074    spin_lock_irqsave(&pool->lock, flags);
2075    bio->bi_bdev = pt->data_dev->bdev;
2076    r = DM_MAPIO_REMAPPED;
2077    spin_unlock_irqrestore(&pool->lock, flags);
2078
2079    return r;
2080}
2081
2082/*
2083 * Retrieves the number of blocks of the data device from
2084 * the superblock and compares it to the actual device size,
2085 * thus resizing the data device in case it has grown.
2086 *
2087 * This both copes with opening preallocated data devices in the ctr
2088 * being followed by a resume
2089 * -and-
2090 * calling the resume method individually after userspace has
2091 * grown the data device in reaction to a table event.
2092 */
2093static int pool_preresume(struct dm_target *ti)
2094{
2095    int r;
2096    struct pool_c *pt = ti->private;
2097    struct pool *pool = pt->pool;
2098    sector_t data_size = ti->len;
2099    dm_block_t sb_data_size;
2100
2101    /*
2102     * Take control of the pool object.
2103     */
2104    r = bind_control_target(pool, ti);
2105    if (r)
2106        return r;
2107
2108    (void) sector_div(data_size, pool->sectors_per_block);
2109
2110    r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2111    if (r) {
2112        DMERR("failed to retrieve data device size");
2113        return r;
2114    }
2115
2116    if (data_size < sb_data_size) {
2117        DMERR("pool target too small, is %llu blocks (expected %llu)",
2118              (unsigned long long)data_size, sb_data_size);
2119        return -EINVAL;
2120
2121    } else if (data_size > sb_data_size) {
2122        r = dm_pool_resize_data_dev(pool->pmd, data_size);
2123        if (r) {
2124            DMERR("failed to resize data device");
2125            /* FIXME Stricter than necessary: Rollback transaction instead here */
2126            set_pool_mode(pool, PM_READ_ONLY);
2127            return r;
2128        }
2129
2130        (void) commit_or_fallback(pool);
2131    }
2132
2133    return 0;
2134}
2135
2136static void pool_resume(struct dm_target *ti)
2137{
2138    struct pool_c *pt = ti->private;
2139    struct pool *pool = pt->pool;
2140    unsigned long flags;
2141
2142    spin_lock_irqsave(&pool->lock, flags);
2143    pool->low_water_triggered = 0;
2144    pool->no_free_space = 0;
2145    __requeue_bios(pool);
2146    spin_unlock_irqrestore(&pool->lock, flags);
2147
2148    do_waker(&pool->waker.work);
2149}
2150
2151static void pool_postsuspend(struct dm_target *ti)
2152{
2153    struct pool_c *pt = ti->private;
2154    struct pool *pool = pt->pool;
2155
2156    cancel_delayed_work(&pool->waker);
2157    flush_workqueue(pool->wq);
2158    (void) commit_or_fallback(pool);
2159}
2160
2161static int check_arg_count(unsigned argc, unsigned args_required)
2162{
2163    if (argc != args_required) {
2164        DMWARN("Message received with %u arguments instead of %u.",
2165               argc, args_required);
2166        return -EINVAL;
2167    }
2168
2169    return 0;
2170}
2171
2172static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2173{
2174    if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2175        *dev_id <= MAX_DEV_ID)
2176        return 0;
2177
2178    if (warning)
2179        DMWARN("Message received with invalid device id: %s", arg);
2180
2181    return -EINVAL;
2182}
2183
2184static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2185{
2186    dm_thin_id dev_id;
2187    int r;
2188
2189    r = check_arg_count(argc, 2);
2190    if (r)
2191        return r;
2192
2193    r = read_dev_id(argv[1], &dev_id, 1);
2194    if (r)
2195        return r;
2196
2197    r = dm_pool_create_thin(pool->pmd, dev_id);
2198    if (r) {
2199        DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2200               argv[1]);
2201        return r;
2202    }
2203
2204    return 0;
2205}
2206
2207static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2208{
2209    dm_thin_id dev_id;
2210    dm_thin_id origin_dev_id;
2211    int r;
2212
2213    r = check_arg_count(argc, 3);
2214    if (r)
2215        return r;
2216
2217    r = read_dev_id(argv[1], &dev_id, 1);
2218    if (r)
2219        return r;
2220
2221    r = read_dev_id(argv[2], &origin_dev_id, 1);
2222    if (r)
2223        return r;
2224
2225    r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2226    if (r) {
2227        DMWARN("Creation of new snapshot %s of device %s failed.",
2228               argv[1], argv[2]);
2229        return r;
2230    }
2231
2232    return 0;
2233}
2234
2235static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2236{
2237    dm_thin_id dev_id;
2238    int r;
2239
2240    r = check_arg_count(argc, 2);
2241    if (r)
2242        return r;
2243
2244    r = read_dev_id(argv[1], &dev_id, 1);
2245    if (r)
2246        return r;
2247
2248    r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2249    if (r)
2250        DMWARN("Deletion of thin device %s failed.", argv[1]);
2251
2252    return r;
2253}
2254
2255static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2256{
2257    dm_thin_id old_id, new_id;
2258    int r;
2259
2260    r = check_arg_count(argc, 3);
2261    if (r)
2262        return r;
2263
2264    if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2265        DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2266        return -EINVAL;
2267    }
2268
2269    if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2270        DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2271        return -EINVAL;
2272    }
2273
2274    r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2275    if (r) {
2276        DMWARN("Failed to change transaction id from %s to %s.",
2277               argv[1], argv[2]);
2278        return r;
2279    }
2280
2281    return 0;
2282}
2283
2284static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2285{
2286    int r;
2287
2288    r = check_arg_count(argc, 1);
2289    if (r)
2290        return r;
2291
2292    (void) commit_or_fallback(pool);
2293
2294    r = dm_pool_reserve_metadata_snap(pool->pmd);
2295    if (r)
2296        DMWARN("reserve_metadata_snap message failed.");
2297
2298    return r;
2299}
2300
2301static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2302{
2303    int r;
2304
2305    r = check_arg_count(argc, 1);
2306    if (r)
2307        return r;
2308
2309    r = dm_pool_release_metadata_snap(pool->pmd);
2310    if (r)
2311        DMWARN("release_metadata_snap message failed.");
2312
2313    return r;
2314}
2315
2316/*
2317 * Messages supported:
2318 * create_thin <dev_id>
2319 * create_snap <dev_id> <origin_id>
2320 * delete <dev_id>
2321 * trim <dev_id> <new_size_in_sectors>
2322 * set_transaction_id <current_trans_id> <new_trans_id>
2323 * reserve_metadata_snap
2324 * release_metadata_snap
2325 */
2326static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2327{
2328    int r = -EINVAL;
2329    struct pool_c *pt = ti->private;
2330    struct pool *pool = pt->pool;
2331
2332    if (!strcasecmp(argv[0], "create_thin"))
2333        r = process_create_thin_mesg(argc, argv, pool);
2334
2335    else if (!strcasecmp(argv[0], "create_snap"))
2336        r = process_create_snap_mesg(argc, argv, pool);
2337
2338    else if (!strcasecmp(argv[0], "delete"))
2339        r = process_delete_mesg(argc, argv, pool);
2340
2341    else if (!strcasecmp(argv[0], "set_transaction_id"))
2342        r = process_set_transaction_id_mesg(argc, argv, pool);
2343
2344    else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2345        r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2346
2347    else if (!strcasecmp(argv[0], "release_metadata_snap"))
2348        r = process_release_metadata_snap_mesg(argc, argv, pool);
2349
2350    else
2351        DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2352
2353    if (!r)
2354        (void) commit_or_fallback(pool);
2355
2356    return r;
2357}
2358
2359static void emit_flags(struct pool_features *pf, char *result,
2360               unsigned sz, unsigned maxlen)
2361{
2362    unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2363        !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2364    DMEMIT("%u ", count);
2365
2366    if (!pf->zero_new_blocks)
2367        DMEMIT("skip_block_zeroing ");
2368
2369    if (!pf->discard_enabled)
2370        DMEMIT("ignore_discard ");
2371
2372    if (!pf->discard_passdown)
2373        DMEMIT("no_discard_passdown ");
2374
2375    if (pf->mode == PM_READ_ONLY)
2376        DMEMIT("read_only ");
2377}
2378
2379/*
2380 * Status line is:
2381 * <transaction id> <used metadata sectors>/<total metadata sectors>
2382 * <used data sectors>/<total data sectors> <held metadata root>
2383 */
2384static void pool_status(struct dm_target *ti, status_type_t type,
2385            unsigned status_flags, char *result, unsigned maxlen)
2386{
2387    int r;
2388    unsigned sz = 0;
2389    uint64_t transaction_id;
2390    dm_block_t nr_free_blocks_data;
2391    dm_block_t nr_free_blocks_metadata;
2392    dm_block_t nr_blocks_data;
2393    dm_block_t nr_blocks_metadata;
2394    dm_block_t held_root;
2395    char buf[BDEVNAME_SIZE];
2396    char buf2[BDEVNAME_SIZE];
2397    struct pool_c *pt = ti->private;
2398    struct pool *pool = pt->pool;
2399
2400    switch (type) {
2401    case STATUSTYPE_INFO:
2402        if (get_pool_mode(pool) == PM_FAIL) {
2403            DMEMIT("Fail");
2404            break;
2405        }
2406
2407        /* Commit to ensure statistics aren't out-of-date */
2408        if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2409            (void) commit_or_fallback(pool);
2410
2411        r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
2412        if (r) {
2413            DMERR("dm_pool_get_metadata_transaction_id returned %d", r);
2414            goto err;
2415        }
2416
2417        r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
2418        if (r) {
2419            DMERR("dm_pool_get_free_metadata_block_count returned %d", r);
2420            goto err;
2421        }
2422
2423        r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2424        if (r) {
2425            DMERR("dm_pool_get_metadata_dev_size returned %d", r);
2426            goto err;
2427        }
2428
2429        r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
2430        if (r) {
2431            DMERR("dm_pool_get_free_block_count returned %d", r);
2432            goto err;
2433        }
2434
2435        r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2436        if (r) {
2437            DMERR("dm_pool_get_data_dev_size returned %d", r);
2438            goto err;
2439        }
2440
2441        r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2442        if (r) {
2443            DMERR("dm_pool_get_metadata_snap returned %d", r);
2444            goto err;
2445        }
2446
2447        DMEMIT("%llu %llu/%llu %llu/%llu ",
2448               (unsigned long long)transaction_id,
2449               (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2450               (unsigned long long)nr_blocks_metadata,
2451               (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2452               (unsigned long long)nr_blocks_data);
2453
2454        if (held_root)
2455            DMEMIT("%llu ", held_root);
2456        else
2457            DMEMIT("- ");
2458
2459        if (pool->pf.mode == PM_READ_ONLY)
2460            DMEMIT("ro ");
2461        else
2462            DMEMIT("rw ");
2463
2464        if (!pool->pf.discard_enabled)
2465            DMEMIT("ignore_discard");
2466        else if (pool->pf.discard_passdown)
2467            DMEMIT("discard_passdown");
2468        else
2469            DMEMIT("no_discard_passdown");
2470
2471        break;
2472
2473    case STATUSTYPE_TABLE:
2474        DMEMIT("%s %s %lu %llu ",
2475               format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2476               format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2477               (unsigned long)pool->sectors_per_block,
2478               (unsigned long long)pt->low_water_blocks);
2479        emit_flags(&pt->requested_pf, result, sz, maxlen);
2480        break;
2481    }
2482    return;
2483
2484err:
2485    DMEMIT("Error");
2486}
2487
2488static int pool_iterate_devices(struct dm_target *ti,
2489                iterate_devices_callout_fn fn, void *data)
2490{
2491    struct pool_c *pt = ti->private;
2492
2493    return fn(ti, pt->data_dev, 0, ti->len, data);
2494}
2495
2496static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2497              struct bio_vec *biovec, int max_size)
2498{
2499    struct pool_c *pt = ti->private;
2500    struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2501
2502    if (!q->merge_bvec_fn)
2503        return max_size;
2504
2505    bvm->bi_bdev = pt->data_dev->bdev;
2506
2507    return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2508}
2509
2510static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2511{
2512    struct pool *pool = pt->pool;
2513    struct queue_limits *data_limits;
2514
2515    limits->max_discard_sectors = pool->sectors_per_block;
2516
2517    /*
2518     * discard_granularity is just a hint, and not enforced.
2519     */
2520    if (pt->adjusted_pf.discard_passdown) {
2521        data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2522        limits->discard_granularity = data_limits->discard_granularity;
2523    } else
2524        limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2525}
2526
2527static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2528{
2529    struct pool_c *pt = ti->private;
2530    struct pool *pool = pt->pool;
2531
2532    blk_limits_io_min(limits, 0);
2533    blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2534
2535    /*
2536     * pt->adjusted_pf is a staging area for the actual features to use.
2537     * They get transferred to the live pool in bind_control_target()
2538     * called from pool_preresume().
2539     */
2540    if (!pt->adjusted_pf.discard_enabled)
2541        return;
2542
2543    disable_passdown_if_not_supported(pt);
2544
2545    set_discard_limits(pt, limits);
2546}
2547
2548static struct target_type pool_target = {
2549    .name = "thin-pool",
2550    .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2551            DM_TARGET_IMMUTABLE,
2552    .version = {1, 7, 0},
2553    .module = THIS_MODULE,
2554    .ctr = pool_ctr,
2555    .dtr = pool_dtr,
2556    .map = pool_map,
2557    .postsuspend = pool_postsuspend,
2558    .preresume = pool_preresume,
2559    .resume = pool_resume,
2560    .message = pool_message,
2561    .status = pool_status,
2562    .merge = pool_merge,
2563    .iterate_devices = pool_iterate_devices,
2564    .io_hints = pool_io_hints,
2565};
2566
2567/*----------------------------------------------------------------
2568 * Thin target methods
2569 *--------------------------------------------------------------*/
2570static void thin_dtr(struct dm_target *ti)
2571{
2572    struct thin_c *tc = ti->private;
2573
2574    mutex_lock(&dm_thin_pool_table.mutex);
2575
2576    __pool_dec(tc->pool);
2577    dm_pool_close_thin_device(tc->td);
2578    dm_put_device(ti, tc->pool_dev);
2579    if (tc->origin_dev)
2580        dm_put_device(ti, tc->origin_dev);
2581    kfree(tc);
2582
2583    mutex_unlock(&dm_thin_pool_table.mutex);
2584}
2585
2586/*
2587 * Thin target parameters:
2588 *
2589 * <pool_dev> <dev_id> [origin_dev]
2590 *
2591 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2592 * dev_id: the internal device identifier
2593 * origin_dev: a device external to the pool that should act as the origin
2594 *
2595 * If the pool device has discards disabled, they get disabled for the thin
2596 * device as well.
2597 */
2598static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2599{
2600    int r;
2601    struct thin_c *tc;
2602    struct dm_dev *pool_dev, *origin_dev;
2603    struct mapped_device *pool_md;
2604
2605    mutex_lock(&dm_thin_pool_table.mutex);
2606
2607    if (argc != 2 && argc != 3) {
2608        ti->error = "Invalid argument count";
2609        r = -EINVAL;
2610        goto out_unlock;
2611    }
2612
2613    tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2614    if (!tc) {
2615        ti->error = "Out of memory";
2616        r = -ENOMEM;
2617        goto out_unlock;
2618    }
2619
2620    if (argc == 3) {
2621        r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2622        if (r) {
2623            ti->error = "Error opening origin device";
2624            goto bad_origin_dev;
2625        }
2626        tc->origin_dev = origin_dev;
2627    }
2628
2629    r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2630    if (r) {
2631        ti->error = "Error opening pool device";
2632        goto bad_pool_dev;
2633    }
2634    tc->pool_dev = pool_dev;
2635
2636    if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2637        ti->error = "Invalid device id";
2638        r = -EINVAL;
2639        goto bad_common;
2640    }
2641
2642    pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2643    if (!pool_md) {
2644        ti->error = "Couldn't get pool mapped device";
2645        r = -EINVAL;
2646        goto bad_common;
2647    }
2648
2649    tc->pool = __pool_table_lookup(pool_md);
2650    if (!tc->pool) {
2651        ti->error = "Couldn't find pool object";
2652        r = -EINVAL;
2653        goto bad_pool_lookup;
2654    }
2655    __pool_inc(tc->pool);
2656
2657    if (get_pool_mode(tc->pool) == PM_FAIL) {
2658        ti->error = "Couldn't open thin device, Pool is in fail mode";
2659        goto bad_thin_open;
2660    }
2661
2662    r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2663    if (r) {
2664        ti->error = "Couldn't open thin internal device";
2665        goto bad_thin_open;
2666    }
2667
2668    r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2669    if (r)
2670        goto bad_thin_open;
2671
2672    ti->num_flush_bios = 1;
2673    ti->flush_supported = true;
2674    ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
2675
2676    /* In case the pool supports discards, pass them on. */
2677    if (tc->pool->pf.discard_enabled) {
2678        ti->discards_supported = true;
2679        ti->num_discard_bios = 1;
2680        ti->discard_zeroes_data_unsupported = true;
2681        /* Discard bios must be split on a block boundary */
2682        ti->split_discard_bios = true;
2683    }
2684
2685    dm_put(pool_md);
2686
2687    mutex_unlock(&dm_thin_pool_table.mutex);
2688
2689    return 0;
2690
2691bad_thin_open:
2692    __pool_dec(tc->pool);
2693bad_pool_lookup:
2694    dm_put(pool_md);
2695bad_common:
2696    dm_put_device(ti, tc->pool_dev);
2697bad_pool_dev:
2698    if (tc->origin_dev)
2699        dm_put_device(ti, tc->origin_dev);
2700bad_origin_dev:
2701    kfree(tc);
2702out_unlock:
2703    mutex_unlock(&dm_thin_pool_table.mutex);
2704
2705    return r;
2706}
2707
2708static int thin_map(struct dm_target *ti, struct bio *bio)
2709{
2710    bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2711
2712    return thin_bio_map(ti, bio);
2713}
2714
2715static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
2716{
2717    unsigned long flags;
2718    struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2719    struct list_head work;
2720    struct dm_thin_new_mapping *m, *tmp;
2721    struct pool *pool = h->tc->pool;
2722
2723    if (h->shared_read_entry) {
2724        INIT_LIST_HEAD(&work);
2725        dm_deferred_entry_dec(h->shared_read_entry, &work);
2726
2727        spin_lock_irqsave(&pool->lock, flags);
2728        list_for_each_entry_safe(m, tmp, &work, list) {
2729            list_del(&m->list);
2730            m->quiesced = 1;
2731            __maybe_add_mapping(m);
2732        }
2733        spin_unlock_irqrestore(&pool->lock, flags);
2734    }
2735
2736    if (h->all_io_entry) {
2737        INIT_LIST_HEAD(&work);
2738        dm_deferred_entry_dec(h->all_io_entry, &work);
2739        if (!list_empty(&work)) {
2740            spin_lock_irqsave(&pool->lock, flags);
2741            list_for_each_entry_safe(m, tmp, &work, list)
2742                list_add(&m->list, &pool->prepared_discards);
2743            spin_unlock_irqrestore(&pool->lock, flags);
2744            wake_worker(pool);
2745        }
2746    }
2747
2748    return 0;
2749}
2750
2751static void thin_postsuspend(struct dm_target *ti)
2752{
2753    if (dm_noflush_suspending(ti))
2754        requeue_io((struct thin_c *)ti->private);
2755}
2756
2757/*
2758 * <nr mapped sectors> <highest mapped sector>
2759 */
2760static void thin_status(struct dm_target *ti, status_type_t type,
2761            unsigned status_flags, char *result, unsigned maxlen)
2762{
2763    int r;
2764    ssize_t sz = 0;
2765    dm_block_t mapped, highest;
2766    char buf[BDEVNAME_SIZE];
2767    struct thin_c *tc = ti->private;
2768
2769    if (get_pool_mode(tc->pool) == PM_FAIL) {
2770        DMEMIT("Fail");
2771        return;
2772    }
2773
2774    if (!tc->td)
2775        DMEMIT("-");
2776    else {
2777        switch (type) {
2778        case STATUSTYPE_INFO:
2779            r = dm_thin_get_mapped_count(tc->td, &mapped);
2780            if (r) {
2781                DMERR("dm_thin_get_mapped_count returned %d", r);
2782                goto err;
2783            }
2784
2785            r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2786            if (r < 0) {
2787                DMERR("dm_thin_get_highest_mapped_block returned %d", r);
2788                goto err;
2789            }
2790
2791            DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2792            if (r)
2793                DMEMIT("%llu", ((highest + 1) *
2794                        tc->pool->sectors_per_block) - 1);
2795            else
2796                DMEMIT("-");
2797            break;
2798
2799        case STATUSTYPE_TABLE:
2800            DMEMIT("%s %lu",
2801                   format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2802                   (unsigned long) tc->dev_id);
2803            if (tc->origin_dev)
2804                DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2805            break;
2806        }
2807    }
2808
2809    return;
2810
2811err:
2812    DMEMIT("Error");
2813}
2814
2815static int thin_iterate_devices(struct dm_target *ti,
2816                iterate_devices_callout_fn fn, void *data)
2817{
2818    sector_t blocks;
2819    struct thin_c *tc = ti->private;
2820    struct pool *pool = tc->pool;
2821
2822    /*
2823     * We can't call dm_pool_get_data_dev_size() since that blocks. So
2824     * we follow a more convoluted path through to the pool's target.
2825     */
2826    if (!pool->ti)
2827        return 0; /* nothing is bound */
2828
2829    blocks = pool->ti->len;
2830    (void) sector_div(blocks, pool->sectors_per_block);
2831    if (blocks)
2832        return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2833
2834    return 0;
2835}
2836
2837static struct target_type thin_target = {
2838    .name = "thin",
2839    .version = {1, 8, 0},
2840    .module = THIS_MODULE,
2841    .ctr = thin_ctr,
2842    .dtr = thin_dtr,
2843    .map = thin_map,
2844    .end_io = thin_endio,
2845    .postsuspend = thin_postsuspend,
2846    .status = thin_status,
2847    .iterate_devices = thin_iterate_devices,
2848};
2849
2850/*----------------------------------------------------------------*/
2851
2852static int __init dm_thin_init(void)
2853{
2854    int r;
2855
2856    pool_table_init();
2857
2858    r = dm_register_target(&thin_target);
2859    if (r)
2860        return r;
2861
2862    r = dm_register_target(&pool_target);
2863    if (r)
2864        goto bad_pool_target;
2865
2866    r = -ENOMEM;
2867
2868    _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2869    if (!_new_mapping_cache)
2870        goto bad_new_mapping_cache;
2871
2872    return 0;
2873
2874bad_new_mapping_cache:
2875    dm_unregister_target(&pool_target);
2876bad_pool_target:
2877    dm_unregister_target(&thin_target);
2878
2879    return r;
2880}
2881
2882static void dm_thin_exit(void)
2883{
2884    dm_unregister_target(&thin_target);
2885    dm_unregister_target(&pool_target);
2886
2887    kmem_cache_destroy(_new_mapping_cache);
2888}
2889
2890module_init(dm_thin_init);
2891module_exit(dm_thin_exit);
2892
2893MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2894MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2895MODULE_LICENSE("GPL");
2896

Archive Download this file



interactive