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
  29 /*
  30  * The block size of the device holding pool data must be
  31  * between 64KB and 1GB.
  32  */
  33 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
  34 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
  35
  36 /*
  37  * Device id is restricted to 24 bits.
  38  */
  39 #define MAX_DEV_ID ((1 << 24) - 1)
  40
  41 /*
  42  * How do we handle breaking sharing of data blocks?
  43  * =================================================
  44  *
  45  * We use a standard copy-on-write btree to store the mappings for the
  46  * devices (note I'm talking about copy-on-write of the metadata here, not
  47  * the data).  When you take an internal snapshot you clone the root node
  48  * of the origin btree.  After this there is no concept of an origin or a
  49  * snapshot.  They are just two device trees that happen to point to the
  50  * same data blocks.
  51  *
  52  * When we get a write in we decide if it's to a shared data block using
  53  * some timestamp magic.  If it is, we have to break sharing.
  54  *
  55  * Let's say we write to a shared block in what was the origin.  The
  56  * steps are:
  57  *
  58  * i) plug io further to this physical block. (see bio_prison code).
  59  *
  60  * ii) quiesce any read io to that shared data block.  Obviously
  61  * including all devices that share this block.  (see dm_deferred_set code)
  62  *
  63  * iii) copy the data block to a newly allocate block.  This step can be
  64  * missed out if the io covers the block. (schedule_copy).
  65  *
  66  * iv) insert the new mapping into the origin's btree
  67  * (process_prepared_mapping).  This act of inserting breaks some
  68  * sharing of btree nodes between the two devices.  Breaking sharing only
  69  * effects the btree of that specific device.  Btrees for the other
  70  * devices that share the block never change.  The btree for the origin
  71  * device as it was after the last commit is untouched, ie. we're using
  72  * persistent data structures in the functional programming sense.
  73  *
  74  * v) unplug io to this physical block, including the io that triggered
  75  * the breaking of sharing.
  76  *
  77  * Steps (ii) and (iii) occur in parallel.
  78  *
  79  * The metadata _doesn't_ need to be committed before the io continues.  We
  80  * get away with this because the io is always written to a _new_ block.
  81  * If there's a crash, then:
  82  *
  83  * - The origin mapping will point to the old origin block (the shared
  84  * one).  This will contain the data as it was before the io that triggered
  85  * the breaking of sharing came in.
  86  *
  87  * - The snap mapping still points to the old block.  As it would after
  88  * the commit.
  89  *
  90  * The downside of this scheme is the timestamp magic isn't perfect, and
  91  * will continue to think that data block in the snapshot device is shared
  92  * even after the write to the origin has broken sharing.  I suspect data
  93  * blocks will typically be shared by many different devices, so we're
  94  * breaking sharing n + 1 times, rather than n, where n is the number of
  95  * devices that reference this data block.  At the moment I think the
  96  * benefits far, far outweigh the disadvantages.
  97  */
  98
  99 /*----------------------------------------------------------------*/
 100
 101 /*
 102  * Key building.
 103  */
 104 static void build_data_key(struct dm_thin_device *td,
 105                            dm_block_t b, struct dm_cell_key *key)
 106 {
 107         key->virtual = 0;
 108         key->dev = dm_thin_dev_id(td);
 109         key->block = b;
 110 }
 111
 112 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
 113                               struct dm_cell_key *key)
 114 {
 115         key->virtual = 1;
 116         key->dev = dm_thin_dev_id(td);
 117         key->block = b;
 118 }
 119
 120 /*----------------------------------------------------------------*/
 121
 122 /*
 123  * A pool device ties together a metadata device and a data device.  It
 124  * also provides the interface for creating and destroying internal
 125  * devices.
 126  */
 127 struct dm_thin_new_mapping;
 128
 129 /*
 130  * The pool runs in 3 modes.  Ordered in degraded order for comparisons.
 131  */
 132 enum pool_mode {
 133         PM_WRITE,               /* metadata may be changed */
 134         PM_READ_ONLY,           /* metadata may not be changed */
 135         PM_FAIL,                /* all I/O fails */
 136 };
 137
 138 struct pool_features {
 139         enum pool_mode mode;
 140
 141         bool zero_new_blocks:1;
 142         bool discard_enabled:1;
 143         bool discard_passdown:1;
 144 };
 145
 146 struct thin_c;
 147 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
 148 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
 149
 150 struct pool {
 151         struct list_head list;
 152         struct dm_target *ti;   /* Only set if a pool target is bound */
 153
 154         struct mapped_device *pool_md;
 155         struct block_device *md_dev;
 156         struct dm_pool_metadata *pmd;
 157
 158         dm_block_t low_water_blocks;
 159         uint32_t sectors_per_block;
 160         int sectors_per_block_shift;
 161
 162         struct pool_features pf;
 163         unsigned low_water_triggered:1; /* A dm event has been sent */
 164         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
 165
 166         struct dm_bio_prison *prison;
 167         struct dm_kcopyd_client *copier;
 168
 169         struct workqueue_struct *wq;
 170         struct work_struct worker;
 171         struct delayed_work waker;
 172
 173         unsigned long last_commit_jiffies;
 174         unsigned ref_count;
 175
 176         spinlock_t lock;
 177         struct bio_list deferred_bios;
 178         struct bio_list deferred_flush_bios;
 179         struct list_head prepared_mappings;
 180         struct list_head prepared_discards;
 181
 182         struct bio_list retry_on_resume_list;
 183
 184         struct dm_deferred_set *shared_read_ds;
 185         struct dm_deferred_set *all_io_ds;
 186
 187         struct dm_thin_new_mapping *next_mapping;
 188         mempool_t *mapping_pool;
 189         mempool_t *endio_hook_pool;
 190
 191         process_bio_fn process_bio;
 192         process_bio_fn process_discard;
 193
 194         process_mapping_fn process_prepared_mapping;
 195         process_mapping_fn process_prepared_discard;
 196 };
 197
 198 static enum pool_mode get_pool_mode(struct pool *pool);
 199 static void set_pool_mode(struct pool *pool, enum pool_mode mode);
 200
 201 /*
 202  * Target context for a pool.
 203  */
 204 struct pool_c {
 205         struct dm_target *ti;
 206         struct pool *pool;
 207         struct dm_dev *data_dev;
 208         struct dm_dev *metadata_dev;
 209         struct dm_target_callbacks callbacks;
 210
 211         dm_block_t low_water_blocks;
 212         struct pool_features requested_pf; /* Features requested during table load */
 213         struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
 214 };
 215
 216 /*
 217  * Target context for a thin.
 218  */
 219 struct thin_c {
 220         struct dm_dev *pool_dev;
 221         struct dm_dev *origin_dev;
 222         dm_thin_id dev_id;
 223
 224         struct pool *pool;
 225         struct dm_thin_device *td;
 226 };
 227
 228 /*----------------------------------------------------------------*/
 229
 230 /*
 231  * A global list of pools that uses a struct mapped_device as a key.
 232  */
 233 static struct dm_thin_pool_table {
 234         struct mutex mutex;
 235         struct list_head pools;
 236 } dm_thin_pool_table;
 237
 238 static void pool_table_init(void)
 239 {
 240         mutex_init(&dm_thin_pool_table.mutex);
 241         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
 242 }
 243
 244 static void __pool_table_insert(struct pool *pool)
 245 {
 246         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 247         list_add(&pool->list, &dm_thin_pool_table.pools);
 248 }
 249
 250 static void __pool_table_remove(struct pool *pool)
 251 {
 252         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 253         list_del(&pool->list);
 254 }
 255
 256 static struct pool *__pool_table_lookup(struct mapped_device *md)
 257 {
 258         struct pool *pool = NULL, *tmp;
 259
 260         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 261
 262         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 263                 if (tmp->pool_md == md) {
 264                         pool = tmp;
 265                         break;
 266                 }
 267         }
 268
 269         return pool;
 270 }
 271
 272 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
 273 {
 274         struct pool *pool = NULL, *tmp;
 275
 276         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 277
 278         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 279                 if (tmp->md_dev == md_dev) {
 280                         pool = tmp;
 281                         break;
 282                 }
 283         }
 284
 285         return pool;
 286 }
 287
 288 /*----------------------------------------------------------------*/
 289
 290 struct dm_thin_endio_hook {
 291         struct thin_c *tc;
 292         struct dm_deferred_entry *shared_read_entry;
 293         struct dm_deferred_entry *all_io_entry;
 294         struct dm_thin_new_mapping *overwrite_mapping;
 295 };
 296
 297 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
 298 {
 299         struct bio *bio;
 300         struct bio_list bios;
 301
 302         bio_list_init(&bios);
 303         bio_list_merge(&bios, master);
 304         bio_list_init(master);
 305
 306         while ((bio = bio_list_pop(&bios))) {
 307                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 308
 309                 if (h->tc == tc)
 310                         bio_endio(bio, DM_ENDIO_REQUEUE);
 311                 else
 312                         bio_list_add(master, bio);
 313         }
 314 }
 315
 316 static void requeue_io(struct thin_c *tc)
 317 {
 318         struct pool *pool = tc->pool;
 319         unsigned long flags;
 320
 321         spin_lock_irqsave(&pool->lock, flags);
 322         __requeue_bio_list(tc, &pool->deferred_bios);
 323         __requeue_bio_list(tc, &pool->retry_on_resume_list);
 324         spin_unlock_irqrestore(&pool->lock, flags);
 325 }
 326
 327 /*
 328  * This section of code contains the logic for processing a thin device's IO.
 329  * Much of the code depends on pool object resources (lists, workqueues, etc)
 330  * but most is exclusively called from the thin target rather than the thin-pool
 331  * target.
 332  */
 333
 334 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
 335 {
 336         sector_t block_nr = bio->bi_sector;
 337
 338         if (tc->pool->sectors_per_block_shift < 0)
 339                 (void) sector_div(block_nr, tc->pool->sectors_per_block);
 340         else
 341                 block_nr >>= tc->pool->sectors_per_block_shift;
 342
 343         return block_nr;
 344 }
 345
 346 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
 347 {
 348         struct pool *pool = tc->pool;
 349         sector_t bi_sector = bio->bi_sector;
 350
 351         bio->bi_bdev = tc->pool_dev->bdev;
 352         if (tc->pool->sectors_per_block_shift < 0)
 353                 bio->bi_sector = (block * pool->sectors_per_block) +
 354                                  sector_div(bi_sector, pool->sectors_per_block);
 355         else
 356                 bio->bi_sector = (block << pool->sectors_per_block_shift) |
 357                                 (bi_sector & (pool->sectors_per_block - 1));
 358 }
 359
 360 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
 361 {
 362         bio->bi_bdev = tc->origin_dev->bdev;
 363 }
 364
 365 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
 366 {
 367         return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
 368                 dm_thin_changed_this_transaction(tc->td);
 369 }
 370
 371 static void issue(struct thin_c *tc, struct bio *bio)
 372 {
 373         struct pool *pool = tc->pool;
 374         unsigned long flags;
 375
 376         if (!bio_triggers_commit(tc, bio)) {
 377                 generic_make_request(bio);
 378                 return;
 379         }
 380
 381         /*
 382          * Complete bio with an error if earlier I/O caused changes to
 383          * the metadata that can't be committed e.g, due to I/O errors
 384          * on the metadata device.
 385          */
 386         if (dm_thin_aborted_changes(tc->td)) {
 387                 bio_io_error(bio);
 388                 return;
 389         }
 390
 391         /*
 392          * Batch together any bios that trigger commits and then issue a
 393          * single commit for them in process_deferred_bios().
 394          */
 395         spin_lock_irqsave(&pool->lock, flags);
 396         bio_list_add(&pool->deferred_flush_bios, bio);
 397         spin_unlock_irqrestore(&pool->lock, flags);
 398 }
 399
 400 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
 401 {
 402         remap_to_origin(tc, bio);
 403         issue(tc, bio);
 404 }
 405
 406 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
 407                             dm_block_t block)
 408 {
 409         remap(tc, bio, block);
 410         issue(tc, bio);
 411 }
 412
 413 /*
 414  * wake_worker() is used when new work is queued and when pool_resume is
 415  * ready to continue deferred IO processing.
 416  */
 417 static void wake_worker(struct pool *pool)
 418 {
 419         queue_work(pool->wq, &pool->worker);
 420 }
 421
 422 /*----------------------------------------------------------------*/
 423
 424 /*
 425  * Bio endio functions.
 426  */
 427 struct dm_thin_new_mapping {
 428         struct list_head list;
 429
 430         unsigned quiesced:1;
 431         unsigned prepared:1;
 432         unsigned pass_discard:1;
 433
 434         struct thin_c *tc;
 435         dm_block_t virt_block;
 436         dm_block_t data_block;
 437         struct dm_bio_prison_cell *cell, *cell2;
 438         int err;
 439
 440         /*
 441          * If the bio covers the whole area of a block then we can avoid
 442          * zeroing or copying.  Instead this bio is hooked.  The bio will
 443          * still be in the cell, so care has to be taken to avoid issuing
 444          * the bio twice.
 445          */
 446         struct bio *bio;
 447         bio_end_io_t *saved_bi_end_io;
 448 };
 449
 450 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
 451 {
 452         struct pool *pool = m->tc->pool;
 453
 454         if (m->quiesced && m->prepared) {
 455                 list_add(&m->list, &pool->prepared_mappings);
 456                 wake_worker(pool);
 457         }
 458 }
 459
 460 static void copy_complete(int read_err, unsigned long write_err, void *context)
 461 {
 462         unsigned long flags;
 463         struct dm_thin_new_mapping *m = context;
 464         struct pool *pool = m->tc->pool;
 465
 466         m->err = read_err || write_err ? -EIO : 0;
 467
 468         spin_lock_irqsave(&pool->lock, flags);
 469         m->prepared = 1;
 470         __maybe_add_mapping(m);
 471         spin_unlock_irqrestore(&pool->lock, flags);
 472 }
 473
 474 static void overwrite_endio(struct bio *bio, int err)
 475 {
 476         unsigned long flags;
 477         struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 478         struct dm_thin_new_mapping *m = h->overwrite_mapping;
 479         struct pool *pool = m->tc->pool;
 480
 481         m->err = err;
 482
 483         spin_lock_irqsave(&pool->lock, flags);
 484         m->prepared = 1;
 485         __maybe_add_mapping(m);
 486         spin_unlock_irqrestore(&pool->lock, flags);
 487 }
 488
 489 /*----------------------------------------------------------------*/
 490
 491 /*
 492  * Workqueue.
 493  */
 494
 495 /*
 496  * Prepared mapping jobs.
 497  */
 498
 499 /*
 500  * This sends the bios in the cell back to the deferred_bios list.
 501  */
 502 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell,
 503                        dm_block_t data_block)
 504 {
 505         struct pool *pool = tc->pool;
 506         unsigned long flags;
 507
 508         spin_lock_irqsave(&pool->lock, flags);
 509         dm_cell_release(cell, &pool->deferred_bios);
 510         spin_unlock_irqrestore(&tc->pool->lock, flags);
 511
 512         wake_worker(pool);
 513 }
 514
 515 /*
 516  * Same as cell_defer except it omits the original holder of the cell.
 517  */
 518 static void cell_defer_except(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 519 {
 520         struct bio_list bios;
 521         struct pool *pool = tc->pool;
 522         unsigned long flags;
 523
 524         bio_list_init(&bios);
 525
 526         spin_lock_irqsave(&pool->lock, flags);
 527         dm_cell_release_no_holder(cell, &pool->deferred_bios);
 528         spin_unlock_irqrestore(&pool->lock, flags);
 529
 530         wake_worker(pool);
 531 }
 532
 533 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
 534 {
 535         if (m->bio)
 536                 m->bio->bi_end_io = m->saved_bi_end_io;
 537         dm_cell_error(m->cell);
 538         list_del(&m->list);
 539         mempool_free(m, m->tc->pool->mapping_pool);
 540 }
 541 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
 542 {
 543         struct thin_c *tc = m->tc;
 544         struct bio *bio;
 545         int r;
 546
 547         bio = m->bio;
 548         if (bio)
 549                 bio->bi_end_io = m->saved_bi_end_io;
 550
 551         if (m->err) {
 552                 dm_cell_error(m->cell);
 553                 goto out;
 554         }
 555
 556         /*
 557          * Commit the prepared block into the mapping btree.
 558          * Any I/O for this block arriving after this point will get
 559          * remapped to it directly.
 560          */
 561         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
 562         if (r) {
 563                 DMERR("dm_thin_insert_block() failed");
 564                 dm_cell_error(m->cell);
 565                 goto out;
 566         }
 567
 568         /*
 569          * Release any bios held while the block was being provisioned.
 570          * If we are processing a write bio that completely covers the block,
 571          * we already processed it so can ignore it now when processing
 572          * the bios in the cell.
 573          */
 574         if (bio) {
 575                 cell_defer_except(tc, m->cell);
 576                 bio_endio(bio, 0);
 577         } else
 578                 cell_defer(tc, m->cell, m->data_block);
 579
 580 out:
 581         list_del(&m->list);
 582         mempool_free(m, tc->pool->mapping_pool);
 583 }
 584
 585 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
 586 {
 587         struct thin_c *tc = m->tc;
 588
 589         bio_io_error(m->bio);
 590         cell_defer_except(tc, m->cell);
 591         cell_defer_except(tc, m->cell2);
 592         mempool_free(m, tc->pool->mapping_pool);
 593 }
 594
 595 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
 596 {
 597         struct thin_c *tc = m->tc;
 598
 599         if (m->pass_discard)
 600                 remap_and_issue(tc, m->bio, m->data_block);
 601         else
 602                 bio_endio(m->bio, 0);
 603
 604         cell_defer_except(tc, m->cell);
 605         cell_defer_except(tc, m->cell2);
 606         mempool_free(m, tc->pool->mapping_pool);
 607 }
 608
 609 static void process_prepared_discard(struct dm_thin_new_mapping *m)
 610 {
 611         int r;
 612         struct thin_c *tc = m->tc;
 613
 614         r = dm_thin_remove_block(tc->td, m->virt_block);
 615         if (r)
 616                 DMERR("dm_thin_remove_block() failed");
 617
 618         process_prepared_discard_passdown(m);
 619 }
 620
 621 static void process_prepared(struct pool *pool, struct list_head *head,
 622                              process_mapping_fn *fn)
 623 {
 624         unsigned long flags;
 625         struct list_head maps;
 626         struct dm_thin_new_mapping *m, *tmp;
 627
 628         INIT_LIST_HEAD(&maps);
 629         spin_lock_irqsave(&pool->lock, flags);
 630         list_splice_init(head, &maps);
 631         spin_unlock_irqrestore(&pool->lock, flags);
 632
 633         list_for_each_entry_safe(m, tmp, &maps, list)
 634                 (*fn)(m);
 635 }
 636
 637 /*
 638  * Deferred bio jobs.
 639  */
 640 static int io_overlaps_block(struct pool *pool, struct bio *bio)
 641 {
 642         return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
 643 }
 644
 645 static int io_overwrites_block(struct pool *pool, struct bio *bio)
 646 {
 647         return (bio_data_dir(bio) == WRITE) &&
 648                 io_overlaps_block(pool, bio);
 649 }
 650
 651 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
 652                                bio_end_io_t *fn)
 653 {
 654         *save = bio->bi_end_io;
 655         bio->bi_end_io = fn;
 656 }
 657
 658 static int ensure_next_mapping(struct pool *pool)
 659 {
 660         if (pool->next_mapping)
 661                 return 0;
 662
 663         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
 664
 665         return pool->next_mapping ? 0 : -ENOMEM;
 666 }
 667
 668 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
 669 {
 670         struct dm_thin_new_mapping *r = pool->next_mapping;
 671
 672         BUG_ON(!pool->next_mapping);
 673
 674         pool->next_mapping = NULL;
 675
 676         return r;
 677 }
 678
 679 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
 680                           struct dm_dev *origin, dm_block_t data_origin,
 681                           dm_block_t data_dest,
 682                           struct dm_bio_prison_cell *cell, struct bio *bio)
 683 {
 684         int r;
 685         struct pool *pool = tc->pool;
 686         struct dm_thin_new_mapping *m = get_next_mapping(pool);
 687
 688         INIT_LIST_HEAD(&m->list);
 689         m->quiesced = 0;
 690         m->prepared = 0;
 691         m->tc = tc;
 692         m->virt_block = virt_block;
 693         m->data_block = data_dest;
 694         m->cell = cell;
 695         m->err = 0;
 696         m->bio = NULL;
 697
 698         if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
 699                 m->quiesced = 1;
 700
 701         /*
 702          * IO to pool_dev remaps to the pool target's data_dev.
 703          *
 704          * If the whole block of data is being overwritten, we can issue the
 705          * bio immediately. Otherwise we use kcopyd to clone the data first.
 706          */
 707         if (io_overwrites_block(pool, bio)) {
 708                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 709
 710                 h->overwrite_mapping = m;
 711                 m->bio = bio;
 712                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 713                 remap_and_issue(tc, bio, data_dest);
 714         } else {
 715                 struct dm_io_region from, to;
 716
 717                 from.bdev = origin->bdev;
 718                 from.sector = data_origin * pool->sectors_per_block;
 719                 from.count = pool->sectors_per_block;
 720
 721                 to.bdev = tc->pool_dev->bdev;
 722                 to.sector = data_dest * pool->sectors_per_block;
 723                 to.count = pool->sectors_per_block;
 724
 725                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
 726                                    0, copy_complete, m);
 727                 if (r < 0) {
 728                         mempool_free(m, pool->mapping_pool);
 729                         DMERR("dm_kcopyd_copy() failed");
 730                         dm_cell_error(cell);
 731                 }
 732         }
 733 }
 734
 735 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
 736                                    dm_block_t data_origin, dm_block_t data_dest,
 737                                    struct dm_bio_prison_cell *cell, struct bio *bio)
 738 {
 739         schedule_copy(tc, virt_block, tc->pool_dev,
 740                       data_origin, data_dest, cell, bio);
 741 }
 742
 743 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
 744                                    dm_block_t data_dest,
 745                                    struct dm_bio_prison_cell *cell, struct bio *bio)
 746 {
 747         schedule_copy(tc, virt_block, tc->origin_dev,
 748                       virt_block, data_dest, cell, bio);
 749 }
 750
 751 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
 752                           dm_block_t data_block, struct dm_bio_prison_cell *cell,
 753                           struct bio *bio)
 754 {
 755         struct pool *pool = tc->pool;
 756         struct dm_thin_new_mapping *m = get_next_mapping(pool);
 757
 758         INIT_LIST_HEAD(&m->list);
 759         m->quiesced = 1;
 760         m->prepared = 0;
 761         m->tc = tc;
 762         m->virt_block = virt_block;
 763         m->data_block = data_block;
 764         m->cell = cell;
 765         m->err = 0;
 766         m->bio = NULL;
 767
 768         /*
 769          * If the whole block of data is being overwritten or we are not
 770          * zeroing pre-existing data, we can issue the bio immediately.
 771          * Otherwise we use kcopyd to zero the data first.
 772          */
 773         if (!pool->pf.zero_new_blocks)
 774                 process_prepared_mapping(m);
 775
 776         else if (io_overwrites_block(pool, bio)) {
 777                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 778
 779                 h->overwrite_mapping = m;
 780                 m->bio = bio;
 781                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 782                 remap_and_issue(tc, bio, data_block);
 783         } else {
 784                 int r;
 785                 struct dm_io_region to;
 786
 787                 to.bdev = tc->pool_dev->bdev;
 788                 to.sector = data_block * pool->sectors_per_block;
 789                 to.count = pool->sectors_per_block;
 790
 791                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
 792                 if (r < 0) {
 793                         mempool_free(m, pool->mapping_pool);
 794                         DMERR("dm_kcopyd_zero() failed");
 795                         dm_cell_error(cell);
 796                 }
 797         }
 798 }
 799
 800 static int commit(struct pool *pool)
 801 {
 802         int r;
 803
 804         r = dm_pool_commit_metadata(pool->pmd);
 805         if (r)
 806                 DMERR("commit failed, error = %d", r);
 807
 808         return r;
 809 }
 810
 811 /*
 812  * A non-zero return indicates read_only or fail_io mode.
 813  * Many callers don't care about the return value.
 814  */
 815 static int commit_or_fallback(struct pool *pool)
 816 {
 817         int r;
 818
 819         if (get_pool_mode(pool) != PM_WRITE)
 820                 return -EINVAL;
 821
 822         r = commit(pool);
 823         if (r)
 824                 set_pool_mode(pool, PM_READ_ONLY);
 825
 826         return r;
 827 }
 828
 829 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
 830 {
 831         int r;
 832         dm_block_t free_blocks;
 833         unsigned long flags;
 834         struct pool *pool = tc->pool;
 835
 836         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 837         if (r)
 838                 return r;
 839
 840         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
 841                 DMWARN("%s: reached low water mark, sending event.",
 842                        dm_device_name(pool->pool_md));
 843                 spin_lock_irqsave(&pool->lock, flags);
 844                 pool->low_water_triggered = 1;
 845                 spin_unlock_irqrestore(&pool->lock, flags);
 846                 dm_table_event(pool->ti->table);
 847         }
 848
 849         if (!free_blocks) {
 850                 if (pool->no_free_space)
 851                         return -ENOSPC;
 852                 else {
 853                         /*
 854                          * Try to commit to see if that will free up some
 855                          * more space.
 856                          */
 857                         (void) commit_or_fallback(pool);
 858
 859                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 860                         if (r)
 861                                 return r;
 862
 863                         /*
 864                          * If we still have no space we set a flag to avoid
 865                          * doing all this checking and return -ENOSPC.
 866                          */
 867                         if (!free_blocks) {
 868                                 DMWARN("%s: no free space available.",
 869                                        dm_device_name(pool->pool_md));
 870                                 spin_lock_irqsave(&pool->lock, flags);
 871                                 pool->no_free_space = 1;
 872                                 spin_unlock_irqrestore(&pool->lock, flags);
 873                                 return -ENOSPC;
 874                         }
 875                 }
 876         }
 877
 878         r = dm_pool_alloc_data_block(pool->pmd, result);
 879         if (r)
 880                 return r;
 881
 882         return 0;
 883 }
 884
 885 /*
 886  * If we have run out of space, queue bios until the device is
 887  * resumed, presumably after having been reloaded with more space.
 888  */
 889 static void retry_on_resume(struct bio *bio)
 890 {
 891         struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 892         struct thin_c *tc = h->tc;
 893         struct pool *pool = tc->pool;
 894         unsigned long flags;
 895
 896         spin_lock_irqsave(&pool->lock, flags);
 897         bio_list_add(&pool->retry_on_resume_list, bio);
 898         spin_unlock_irqrestore(&pool->lock, flags);
 899 }
 900
 901 static void no_space(struct dm_bio_prison_cell *cell)
 902 {
 903         struct bio *bio;
 904         struct bio_list bios;
 905
 906         bio_list_init(&bios);
 907         dm_cell_release(cell, &bios);
 908
 909         while ((bio = bio_list_pop(&bios)))
 910                 retry_on_resume(bio);
 911 }
 912
 913 static void process_discard(struct thin_c *tc, struct bio *bio)
 914 {
 915         int r;
 916         unsigned long flags;
 917         struct pool *pool = tc->pool;
 918         struct dm_bio_prison_cell *cell, *cell2;
 919         struct dm_cell_key key, key2;
 920         dm_block_t block = get_bio_block(tc, bio);
 921         struct dm_thin_lookup_result lookup_result;
 922         struct dm_thin_new_mapping *m;
 923
 924         build_virtual_key(tc->td, block, &key);
 925         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
 926                 return;
 927
 928         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
 929         switch (r) {
 930         case 0:
 931                 /*
 932                  * Check nobody is fiddling with this pool block.  This can
 933                  * happen if someone's in the process of breaking sharing
 934                  * on this block.
 935                  */
 936                 build_data_key(tc->td, lookup_result.block, &key2);
 937                 if (dm_bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
 938                         cell_defer_except(tc, cell);
 939                         break;
 940                 }
 941
 942                 if (io_overlaps_block(pool, bio)) {
 943                         /*
 944                          * IO may still be going to the destination block.  We must
 945                          * quiesce before we can do the removal.
 946                          */
 947                         m = get_next_mapping(pool);
 948                         m->tc = tc;
 949                         m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
 950                         m->virt_block = block;
 951                         m->data_block = lookup_result.block;
 952                         m->cell = cell;
 953                         m->cell2 = cell2;
 954                         m->err = 0;
 955                         m->bio = bio;
 956
 957                         if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
 958                                 spin_lock_irqsave(&pool->lock, flags);
 959                                 list_add(&m->list, &pool->prepared_discards);
 960                                 spin_unlock_irqrestore(&pool->lock, flags);
 961                                 wake_worker(pool);
 962                         }
 963                 } else {
 964                         /*
 965                          * The DM core makes sure that the discard doesn't span
 966                          * a block boundary.  So we submit the discard of a
 967                          * partial block appropriately.
 968                          */
 969                         cell_defer_except(tc, cell);
 970                         cell_defer_except(tc, cell2);
 971                         if ((!lookup_result.shared) && pool->pf.discard_passdown)
 972                                 remap_and_issue(tc, bio, lookup_result.block);
 973                         else
 974                                 bio_endio(bio, 0);
 975                 }
 976                 break;
 977
 978         case -ENODATA:
 979                 /*
 980                  * It isn't provisioned, just forget it.
 981                  */
 982                 cell_defer_except(tc, cell);
 983                 bio_endio(bio, 0);
 984                 break;
 985
 986         default:
 987                 DMERR("discard: find block unexpectedly returned %d", r);
 988                 cell_defer_except(tc, cell);
 989                 bio_io_error(bio);
 990                 break;
 991         }
 992 }
 993
 994 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
 995                           struct dm_cell_key *key,
 996                           struct dm_thin_lookup_result *lookup_result,
 997                           struct dm_bio_prison_cell *cell)
 998 {
 999         int r;
1000         dm_block_t data_block;
1001
1002         r = alloc_data_block(tc, &data_block);
1003         switch (r) {
1004         case 0:
1005                 schedule_internal_copy(tc, block, lookup_result->block,
1006                                        data_block, cell, bio);
1007                 break;
1008
1009         case -ENOSPC:
1010                 no_space(cell);
1011                 break;
1012
1013         default:
1014                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1015                 dm_cell_error(cell);
1016                 break;
1017         }
1018 }
1019
1020 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1021                                dm_block_t block,
1022                                struct dm_thin_lookup_result *lookup_result)
1023 {
1024         struct dm_bio_prison_cell *cell;
1025         struct pool *pool = tc->pool;
1026         struct dm_cell_key key;
1027
1028         /*
1029          * If cell is already occupied, then sharing is already in the process
1030          * of being broken so we have nothing further to do here.
1031          */
1032         build_data_key(tc->td, lookup_result->block, &key);
1033         if (dm_bio_detain(pool->prison, &key, bio, &cell))
1034                 return;
1035
1036         if (bio_data_dir(bio) == WRITE && bio->bi_size)
1037                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1038         else {
1039                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1040
1041                 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1042
1043                 cell_defer_except(tc, cell);
1044                 remap_and_issue(tc, bio, lookup_result->block);
1045         }
1046 }
1047
1048 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1049                             struct dm_bio_prison_cell *cell)
1050 {
1051         int r;
1052         dm_block_t data_block;
1053
1054         /*
1055          * Remap empty bios (flushes) immediately, without provisioning.
1056          */
1057         if (!bio->bi_size) {
1058                 cell_defer_except(tc, cell);
1059                 remap_and_issue(tc, bio, 0);
1060                 return;
1061         }
1062
1063         /*
1064          * Fill read bios with zeroes and complete them immediately.
1065          */
1066         if (bio_data_dir(bio) == READ) {
1067                 zero_fill_bio(bio);
1068                 cell_defer_except(tc, cell);
1069                 bio_endio(bio, 0);
1070                 return;
1071         }
1072
1073         r = alloc_data_block(tc, &data_block);
1074         switch (r) {
1075         case 0:
1076                 if (tc->origin_dev)
1077                         schedule_external_copy(tc, block, data_block, cell, bio);
1078                 else
1079                         schedule_zero(tc, block, data_block, cell, bio);
1080                 break;
1081
1082         case -ENOSPC:
1083                 no_space(cell);
1084                 break;
1085
1086         default:
1087                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1088                 set_pool_mode(tc->pool, PM_READ_ONLY);
1089                 dm_cell_error(cell);
1090                 break;
1091         }
1092 }
1093
1094 static void process_bio(struct thin_c *tc, struct bio *bio)
1095 {
1096         int r;
1097         dm_block_t block = get_bio_block(tc, bio);
1098         struct dm_bio_prison_cell *cell;
1099         struct dm_cell_key key;
1100         struct dm_thin_lookup_result lookup_result;
1101
1102         /*
1103          * If cell is already occupied, then the block is already
1104          * being provisioned so we have nothing further to do here.
1105          */
1106         build_virtual_key(tc->td, block, &key);
1107         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
1108                 return;
1109
1110         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1111         switch (r) {
1112         case 0:
1113                 /*
1114                  * We can release this cell now.  This thread is the only
1115                  * one that puts bios into a cell, and we know there were
1116                  * no preceding bios.
1117                  */
1118                 /*
1119                  * TODO: this will probably have to change when discard goes
1120                  * back in.
1121                  */
1122                 cell_defer_except(tc, cell);
1123
1124                 if (lookup_result.shared)
1125                         process_shared_bio(tc, bio, block, &lookup_result);
1126                 else
1127                         remap_and_issue(tc, bio, lookup_result.block);
1128                 break;
1129
1130         case -ENODATA:
1131                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1132                         cell_defer_except(tc, cell);
1133                         remap_to_origin_and_issue(tc, bio);
1134                 } else
1135                         provision_block(tc, bio, block, cell);
1136                 break;
1137
1138         default:
1139                 DMERR("dm_thin_find_block() failed, error = %d", r);
1140                 cell_defer_except(tc, cell);
1141                 bio_io_error(bio);
1142                 break;
1143         }
1144 }
1145
1146 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1147 {
1148         int r;
1149         int rw = bio_data_dir(bio);
1150         dm_block_t block = get_bio_block(tc, bio);
1151         struct dm_thin_lookup_result lookup_result;
1152
1153         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1154         switch (r) {
1155         case 0:
1156                 if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1157                         bio_io_error(bio);
1158                 else
1159                         remap_and_issue(tc, bio, lookup_result.block);
1160                 break;
1161
1162         case -ENODATA:
1163                 if (rw != READ) {
1164                         bio_io_error(bio);
1165                         break;
1166                 }
1167
1168                 if (tc->origin_dev) {
1169                         remap_to_origin_and_issue(tc, bio);
1170                         break;
1171                 }
1172
1173                 zero_fill_bio(bio);
1174                 bio_endio(bio, 0);
1175                 break;
1176
1177         default:
1178                 DMERR("dm_thin_find_block() failed, error = %d", r);
1179                 bio_io_error(bio);
1180                 break;
1181         }
1182 }
1183
1184 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1185 {
1186         bio_io_error(bio);
1187 }
1188
1189 static int need_commit_due_to_time(struct pool *pool)
1190 {
1191         return jiffies < pool->last_commit_jiffies ||
1192                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1193 }
1194
1195 static void process_deferred_bios(struct pool *pool)
1196 {
1197         unsigned long flags;
1198         struct bio *bio;
1199         struct bio_list bios;
1200
1201         bio_list_init(&bios);
1202
1203         spin_lock_irqsave(&pool->lock, flags);
1204         bio_list_merge(&bios, &pool->deferred_bios);
1205         bio_list_init(&pool->deferred_bios);
1206         spin_unlock_irqrestore(&pool->lock, flags);
1207
1208         while ((bio = bio_list_pop(&bios))) {
1209                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1210                 struct thin_c *tc = h->tc;
1211
1212                 /*
1213                  * If we've got no free new_mapping structs, and processing
1214                  * this bio might require one, we pause until there are some
1215                  * prepared mappings to process.
1216                  */
1217                 if (ensure_next_mapping(pool)) {
1218                         spin_lock_irqsave(&pool->lock, flags);
1219                         bio_list_merge(&pool->deferred_bios, &bios);
1220                         spin_unlock_irqrestore(&pool->lock, flags);
1221
1222                         break;
1223                 }
1224
1225                 if (bio->bi_rw & REQ_DISCARD)
1226                         pool->process_discard(tc, bio);
1227                 else
1228                         pool->process_bio(tc, bio);
1229         }
1230
1231         /*
1232          * If there are any deferred flush bios, we must commit
1233          * the metadata before issuing them.
1234          */
1235         bio_list_init(&bios);
1236         spin_lock_irqsave(&pool->lock, flags);
1237         bio_list_merge(&bios, &pool->deferred_flush_bios);
1238         bio_list_init(&pool->deferred_flush_bios);
1239         spin_unlock_irqrestore(&pool->lock, flags);
1240
1241         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1242                 return;
1243
1244         if (commit_or_fallback(pool)) {
1245                 while ((bio = bio_list_pop(&bios)))
1246                         bio_io_error(bio);
1247                 return;
1248         }
1249         pool->last_commit_jiffies = jiffies;
1250
1251         while ((bio = bio_list_pop(&bios)))
1252                 generic_make_request(bio);
1253 }
1254
1255 static void do_worker(struct work_struct *ws)
1256 {
1257         struct pool *pool = container_of(ws, struct pool, worker);
1258
1259         process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1260         process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1261         process_deferred_bios(pool);
1262 }
1263
1264 /*
1265  * We want to commit periodically so that not too much
1266  * unwritten data builds up.
1267  */
1268 static void do_waker(struct work_struct *ws)
1269 {
1270         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1271         wake_worker(pool);
1272         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1273 }
1274
1275 /*----------------------------------------------------------------*/
1276
1277 static enum pool_mode get_pool_mode(struct pool *pool)
1278 {
1279         return pool->pf.mode;
1280 }
1281
1282 static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1283 {
1284         int r;
1285
1286         pool->pf.mode = mode;
1287
1288         switch (mode) {
1289         case PM_FAIL:
1290                 DMERR("switching pool to failure mode");
1291                 pool->process_bio = process_bio_fail;
1292                 pool->process_discard = process_bio_fail;
1293                 pool->process_prepared_mapping = process_prepared_mapping_fail;
1294                 pool->process_prepared_discard = process_prepared_discard_fail;
1295                 break;
1296
1297         case PM_READ_ONLY:
1298                 DMERR("switching pool to read-only mode");
1299                 r = dm_pool_abort_metadata(pool->pmd);
1300                 if (r) {
1301                         DMERR("aborting transaction failed");
1302                         set_pool_mode(pool, PM_FAIL);
1303                 } else {
1304                         dm_pool_metadata_read_only(pool->pmd);
1305                         pool->process_bio = process_bio_read_only;
1306                         pool->process_discard = process_discard;
1307                         pool->process_prepared_mapping = process_prepared_mapping_fail;
1308                         pool->process_prepared_discard = process_prepared_discard_passdown;
1309                 }
1310                 break;
1311
1312         case PM_WRITE:
1313                 pool->process_bio = process_bio;
1314                 pool->process_discard = process_discard;
1315                 pool->process_prepared_mapping = process_prepared_mapping;
1316                 pool->process_prepared_discard = process_prepared_discard;
1317                 break;
1318         }
1319 }
1320
1321 /*----------------------------------------------------------------*/
1322
1323 /*
1324  * Mapping functions.
1325  */
1326
1327 /*
1328  * Called only while mapping a thin bio to hand it over to the workqueue.
1329  */
1330 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1331 {
1332         unsigned long flags;
1333         struct pool *pool = tc->pool;
1334
1335         spin_lock_irqsave(&pool->lock, flags);
1336         bio_list_add(&pool->deferred_bios, bio);
1337         spin_unlock_irqrestore(&pool->lock, flags);
1338
1339         wake_worker(pool);
1340 }
1341
1342 static struct dm_thin_endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1343 {
1344         struct pool *pool = tc->pool;
1345         struct dm_thin_endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1346
1347         h->tc = tc;
1348         h->shared_read_entry = NULL;
1349         h->all_io_entry = bio->bi_rw & REQ_DISCARD ? NULL : dm_deferred_entry_inc(pool->all_io_ds);
1350         h->overwrite_mapping = NULL;
1351
1352         return h;
1353 }
1354
1355 /*
1356  * Non-blocking function called from the thin target's map function.
1357  */
1358 static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1359                         union map_info *map_context)
1360 {
1361         int r;
1362         struct thin_c *tc = ti->private;
1363         dm_block_t block = get_bio_block(tc, bio);
1364         struct dm_thin_device *td = tc->td;
1365         struct dm_thin_lookup_result result;
1366
1367         map_context->ptr = thin_hook_bio(tc, bio);
1368
1369         if (get_pool_mode(tc->pool) == PM_FAIL) {
1370                 bio_io_error(bio);
1371                 return DM_MAPIO_SUBMITTED;
1372         }
1373
1374         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1375                 thin_defer_bio(tc, bio);
1376                 return DM_MAPIO_SUBMITTED;
1377         }
1378
1379         r = dm_thin_find_block(td, block, 0, &result);
1380
1381         /*
1382          * Note that we defer readahead too.
1383          */
1384         switch (r) {
1385         case 0:
1386                 if (unlikely(result.shared)) {
1387                         /*
1388                          * We have a race condition here between the
1389                          * result.shared value returned by the lookup and
1390                          * snapshot creation, which may cause new
1391                          * sharing.
1392                          *
1393                          * To avoid this always quiesce the origin before
1394                          * taking the snap.  You want to do this anyway to
1395                          * ensure a consistent application view
1396                          * (i.e. lockfs).
1397                          *
1398                          * More distant ancestors are irrelevant. The
1399                          * shared flag will be set in their case.
1400                          */
1401                         thin_defer_bio(tc, bio);
1402                         r = DM_MAPIO_SUBMITTED;
1403                 } else {
1404                         remap(tc, bio, result.block);
1405                         r = DM_MAPIO_REMAPPED;
1406                 }
1407                 break;
1408
1409         case -ENODATA:
1410                 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1411                         /*
1412                          * This block isn't provisioned, and we have no way
1413                          * of doing so.  Just error it.
1414                          */
1415                         bio_io_error(bio);
1416                         r = DM_MAPIO_SUBMITTED;
1417                         break;
1418                 }
1419                 /* fall through */
1420
1421         case -EWOULDBLOCK:
1422                 /*
1423                  * In future, the failed dm_thin_find_block above could
1424                  * provide the hint to load the metadata into cache.
1425                  */
1426                 thin_defer_bio(tc, bio);
1427                 r = DM_MAPIO_SUBMITTED;
1428                 break;
1429
1430         default:
1431                 /*
1432                  * Must always call bio_io_error on failure.
1433                  * dm_thin_find_block can fail with -EINVAL if the
1434                  * pool is switched to fail-io mode.
1435                  */
1436                 bio_io_error(bio);
1437                 r = DM_MAPIO_SUBMITTED;
1438                 break;
1439         }
1440
1441         return r;
1442 }
1443
1444 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1445 {
1446         int r;
1447         unsigned long flags;
1448         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1449
1450         spin_lock_irqsave(&pt->pool->lock, flags);
1451         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1452         spin_unlock_irqrestore(&pt->pool->lock, flags);
1453
1454         if (!r) {
1455                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1456                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1457         }
1458
1459         return r;
1460 }
1461
1462 static void __requeue_bios(struct pool *pool)
1463 {
1464         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1465         bio_list_init(&pool->retry_on_resume_list);
1466 }
1467
1468 /*----------------------------------------------------------------
1469  * Binding of control targets to a pool object
1470  *--------------------------------------------------------------*/
1471 static bool data_dev_supports_discard(struct pool_c *pt)
1472 {
1473         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1474
1475         return q && blk_queue_discard(q);
1476 }
1477
1478 /*
1479  * If discard_passdown was enabled verify that the data device
1480  * supports discards.  Disable discard_passdown if not.
1481  */
1482 static void disable_passdown_if_not_supported(struct pool_c *pt)
1483 {
1484         struct pool *pool = pt->pool;
1485         struct block_device *data_bdev = pt->data_dev->bdev;
1486         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1487         sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1488         const char *reason = NULL;
1489         char buf[BDEVNAME_SIZE];
1490
1491         if (!pt->adjusted_pf.discard_passdown)
1492                 return;
1493
1494         if (!data_dev_supports_discard(pt))
1495                 reason = "discard unsupported";
1496
1497         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1498                 reason = "max discard sectors smaller than a block";
1499
1500         else if (data_limits->discard_granularity > block_size)
1501                 reason = "discard granularity larger than a block";
1502
1503         else if (block_size & (data_limits->discard_granularity - 1))
1504                 reason = "discard granularity not a factor of block size";
1505
1506         if (reason) {
1507                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1508                 pt->adjusted_pf.discard_passdown = false;
1509         }
1510 }
1511
1512 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1513 {
1514         struct pool_c *pt = ti->private;
1515
1516         /*
1517          * We want to make sure that degraded pools are never upgraded.
1518          */
1519         enum pool_mode old_mode = pool->pf.mode;
1520         enum pool_mode new_mode = pt->adjusted_pf.mode;
1521
1522         if (old_mode > new_mode)
1523                 new_mode = old_mode;
1524
1525         pool->ti = ti;
1526         pool->low_water_blocks = pt->low_water_blocks;
1527         pool->pf = pt->adjusted_pf;
1528
1529         set_pool_mode(pool, new_mode);
1530
1531         return 0;
1532 }
1533
1534 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1535 {
1536         if (pool->ti == ti)
1537                 pool->ti = NULL;
1538 }
1539
1540 /*----------------------------------------------------------------
1541  * Pool creation
1542  *--------------------------------------------------------------*/
1543 /* Initialize pool features. */
1544 static void pool_features_init(struct pool_features *pf)
1545 {
1546         pf->mode = PM_WRITE;
1547         pf->zero_new_blocks = true;
1548         pf->discard_enabled = true;
1549         pf->discard_passdown = true;
1550 }
1551
1552 static void __pool_destroy(struct pool *pool)
1553 {
1554         __pool_table_remove(pool);
1555
1556         if (dm_pool_metadata_close(pool->pmd) < 0)
1557                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1558
1559         dm_bio_prison_destroy(pool->prison);
1560         dm_kcopyd_client_destroy(pool->copier);
1561
1562         if (pool->wq)
1563                 destroy_workqueue(pool->wq);
1564
1565         if (pool->next_mapping)
1566                 mempool_free(pool->next_mapping, pool->mapping_pool);
1567         mempool_destroy(pool->mapping_pool);
1568         mempool_destroy(pool->endio_hook_pool);
1569         dm_deferred_set_destroy(pool->shared_read_ds);
1570         dm_deferred_set_destroy(pool->all_io_ds);
1571         kfree(pool);
1572 }
1573
1574 static struct kmem_cache *_new_mapping_cache;
1575 static struct kmem_cache *_endio_hook_cache;
1576
1577 static struct pool *pool_create(struct mapped_device *pool_md,
1578                                 struct block_device *metadata_dev,
1579                                 unsigned long block_size,
1580                                 int read_only, char **error)
1581 {
1582         int r;
1583         void *err_p;
1584         struct pool *pool;
1585         struct dm_pool_metadata *pmd;
1586         bool format_device = read_only ? false : true;
1587
1588         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1589         if (IS_ERR(pmd)) {
1590                 *error = "Error creating metadata object";
1591                 return (struct pool *)pmd;
1592         }
1593
1594         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1595         if (!pool) {
1596                 *error = "Error allocating memory for pool";
1597                 err_p = ERR_PTR(-ENOMEM);
1598                 goto bad_pool;
1599         }
1600
1601         pool->pmd = pmd;
1602         pool->sectors_per_block = block_size;
1603         if (block_size & (block_size - 1))
1604                 pool->sectors_per_block_shift = -1;
1605         else
1606                 pool->sectors_per_block_shift = __ffs(block_size);
1607         pool->low_water_blocks = 0;
1608         pool_features_init(&pool->pf);
1609         pool->prison = dm_bio_prison_create(PRISON_CELLS);
1610         if (!pool->prison) {
1611                 *error = "Error creating pool's bio prison";
1612                 err_p = ERR_PTR(-ENOMEM);
1613                 goto bad_prison;
1614         }
1615
1616         pool->copier = dm_kcopyd_client_create();
1617         if (IS_ERR(pool->copier)) {
1618                 r = PTR_ERR(pool->copier);
1619                 *error = "Error creating pool's kcopyd client";
1620                 err_p = ERR_PTR(r);
1621                 goto bad_kcopyd_client;
1622         }
1623
1624         /*
1625          * Create singlethreaded workqueue that will service all devices
1626          * that use this metadata.
1627          */
1628         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1629         if (!pool->wq) {
1630                 *error = "Error creating pool's workqueue";
1631                 err_p = ERR_PTR(-ENOMEM);
1632                 goto bad_wq;
1633         }
1634
1635         INIT_WORK(&pool->worker, do_worker);
1636         INIT_DELAYED_WORK(&pool->waker, do_waker);
1637         spin_lock_init(&pool->lock);
1638         bio_list_init(&pool->deferred_bios);
1639         bio_list_init(&pool->deferred_flush_bios);
1640         INIT_LIST_HEAD(&pool->prepared_mappings);
1641         INIT_LIST_HEAD(&pool->prepared_discards);
1642         pool->low_water_triggered = 0;
1643         pool->no_free_space = 0;
1644         bio_list_init(&pool->retry_on_resume_list);
1645
1646         pool->shared_read_ds = dm_deferred_set_create();
1647         if (!pool->shared_read_ds) {
1648                 *error = "Error creating pool's shared read deferred set";
1649                 err_p = ERR_PTR(-ENOMEM);
1650                 goto bad_shared_read_ds;
1651         }
1652
1653         pool->all_io_ds = dm_deferred_set_create();
1654         if (!pool->all_io_ds) {
1655                 *error = "Error creating pool's all io deferred set";
1656                 err_p = ERR_PTR(-ENOMEM);
1657                 goto bad_all_io_ds;
1658         }
1659
1660         pool->next_mapping = NULL;
1661         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1662                                                       _new_mapping_cache);
1663         if (!pool->mapping_pool) {
1664                 *error = "Error creating pool's mapping mempool";
1665                 err_p = ERR_PTR(-ENOMEM);
1666                 goto bad_mapping_pool;
1667         }
1668
1669         pool->endio_hook_pool = mempool_create_slab_pool(ENDIO_HOOK_POOL_SIZE,
1670                                                          _endio_hook_cache);
1671         if (!pool->endio_hook_pool) {
1672                 *error = "Error creating pool's endio_hook mempool";
1673                 err_p = ERR_PTR(-ENOMEM);
1674                 goto bad_endio_hook_pool;
1675         }
1676         pool->ref_count = 1;
1677         pool->last_commit_jiffies = jiffies;
1678         pool->pool_md = pool_md;
1679         pool->md_dev = metadata_dev;
1680         __pool_table_insert(pool);
1681
1682         return pool;
1683
1684 bad_endio_hook_pool:
1685         mempool_destroy(pool->mapping_pool);
1686 bad_mapping_pool:
1687         dm_deferred_set_destroy(pool->all_io_ds);
1688 bad_all_io_ds:
1689         dm_deferred_set_destroy(pool->shared_read_ds);
1690 bad_shared_read_ds:
1691         destroy_workqueue(pool->wq);
1692 bad_wq:
1693         dm_kcopyd_client_destroy(pool->copier);
1694 bad_kcopyd_client:
1695         dm_bio_prison_destroy(pool->prison);
1696 bad_prison:
1697         kfree(pool);
1698 bad_pool:
1699         if (dm_pool_metadata_close(pmd))
1700                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1701
1702         return err_p;
1703 }
1704
1705 static void __pool_inc(struct pool *pool)
1706 {
1707         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1708         pool->ref_count++;
1709 }
1710
1711 static void __pool_dec(struct pool *pool)
1712 {
1713         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1714         BUG_ON(!pool->ref_count);
1715         if (!--pool->ref_count)
1716                 __pool_destroy(pool);
1717 }
1718
1719 static struct pool *__pool_find(struct mapped_device *pool_md,
1720                                 struct block_device *metadata_dev,
1721                                 unsigned long block_size, int read_only,
1722                                 char **error, int *created)
1723 {
1724         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1725
1726         if (pool) {
1727                 if (pool->pool_md != pool_md) {
1728                         *error = "metadata device already in use by a pool";
1729                         return ERR_PTR(-EBUSY);
1730                 }
1731                 __pool_inc(pool);
1732
1733         } else {
1734                 pool = __pool_table_lookup(pool_md);
1735                 if (pool) {
1736                         if (pool->md_dev != metadata_dev) {
1737                                 *error = "different pool cannot replace a pool";
1738                                 return ERR_PTR(-EINVAL);
1739                         }
1740                         __pool_inc(pool);
1741
1742                 } else {
1743                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1744                         *created = 1;
1745                 }
1746         }
1747
1748         return pool;
1749 }
1750
1751 /*----------------------------------------------------------------
1752  * Pool target methods
1753  *--------------------------------------------------------------*/
1754 static void pool_dtr(struct dm_target *ti)
1755 {
1756         struct pool_c *pt = ti->private;
1757
1758         mutex_lock(&dm_thin_pool_table.mutex);
1759
1760         unbind_control_target(pt->pool, ti);
1761         __pool_dec(pt->pool);
1762         dm_put_device(ti, pt->metadata_dev);
1763         dm_put_device(ti, pt->data_dev);
1764         kfree(pt);
1765
1766         mutex_unlock(&dm_thin_pool_table.mutex);
1767 }
1768
1769 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1770                                struct dm_target *ti)
1771 {
1772         int r;
1773         unsigned argc;
1774         const char *arg_name;
1775
1776         static struct dm_arg _args[] = {
1777                 {0, 3, "Invalid number of pool feature arguments"},
1778         };
1779
1780         /*
1781          * No feature arguments supplied.
1782          */
1783         if (!as->argc)
1784                 return 0;
1785
1786         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1787         if (r)
1788                 return -EINVAL;
1789
1790         while (argc && !r) {
1791                 arg_name = dm_shift_arg(as);
1792                 argc--;
1793
1794                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
1795                         pf->zero_new_blocks = false;
1796
1797                 else if (!strcasecmp(arg_name, "ignore_discard"))
1798                         pf->discard_enabled = false;
1799
1800                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
1801                         pf->discard_passdown = false;
1802
1803                 else if (!strcasecmp(arg_name, "read_only"))
1804                         pf->mode = PM_READ_ONLY;
1805
1806                 else {
1807                         ti->error = "Unrecognised pool feature requested";
1808                         r = -EINVAL;
1809                         break;
1810                 }
1811         }
1812
1813         return r;
1814 }
1815
1816 /*
1817  * thin-pool <metadata dev> <data dev>
1818  *           <data block size (sectors)>
1819  *           <low water mark (blocks)>
1820  *           [<#feature args> [<arg>]*]
1821  *
1822  * Optional feature arguments are:
1823  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1824  *           ignore_discard: disable discard
1825  *           no_discard_passdown: don't pass discards down to the data device
1826  */
1827 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1828 {
1829         int r, pool_created = 0;
1830         struct pool_c *pt;
1831         struct pool *pool;
1832         struct pool_features pf;
1833         struct dm_arg_set as;
1834         struct dm_dev *data_dev;
1835         unsigned long block_size;
1836         dm_block_t low_water_blocks;
1837         struct dm_dev *metadata_dev;
1838         sector_t metadata_dev_size;
1839         char b[BDEVNAME_SIZE];
1840
1841         /*
1842          * FIXME Remove validation from scope of lock.
1843          */
1844         mutex_lock(&dm_thin_pool_table.mutex);
1845
1846         if (argc < 4) {
1847                 ti->error = "Invalid argument count";
1848                 r = -EINVAL;
1849                 goto out_unlock;
1850         }
1851         as.argc = argc;
1852         as.argv = argv;
1853
1854         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1855         if (r) {
1856                 ti->error = "Error opening metadata block device";
1857                 goto out_unlock;
1858         }
1859
1860         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1861         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1862                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1863                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1864
1865         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1866         if (r) {
1867                 ti->error = "Error getting data device";
1868                 goto out_metadata;
1869         }
1870
1871         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1872             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1873             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1874             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1875                 ti->error = "Invalid block size";
1876                 r = -EINVAL;
1877                 goto out;
1878         }
1879
1880         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1881                 ti->error = "Invalid low water mark";
1882                 r = -EINVAL;
1883                 goto out;
1884         }
1885
1886         /*
1887          * Set default pool features.
1888          */
1889         pool_features_init(&pf);
1890
1891         dm_consume_args(&as, 4);
1892         r = parse_pool_features(&as, &pf, ti);
1893         if (r)
1894                 goto out;
1895
1896         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1897         if (!pt) {
1898                 r = -ENOMEM;
1899                 goto out;
1900         }
1901
1902         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1903                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
1904         if (IS_ERR(pool)) {
1905                 r = PTR_ERR(pool);
1906                 goto out_free_pt;
1907         }
1908
1909         /*
1910          * 'pool_created' reflects whether this is the first table load.
1911          * Top level discard support is not allowed to be changed after
1912          * initial load.  This would require a pool reload to trigger thin
1913          * device changes.
1914          */
1915         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1916                 ti->error = "Discard support cannot be disabled once enabled";
1917                 r = -EINVAL;
1918                 goto out_flags_changed;
1919         }
1920
1921         pt->pool = pool;
1922         pt->ti = ti;
1923         pt->metadata_dev = metadata_dev;
1924         pt->data_dev = data_dev;
1925         pt->low_water_blocks = low_water_blocks;
1926         pt->adjusted_pf = pt->requested_pf = pf;
1927         ti->num_flush_requests = 1;
1928
1929         /*
1930          * Only need to enable discards if the pool should pass
1931          * them down to the data device.  The thin device's discard
1932          * processing will cause mappings to be removed from the btree.
1933          */
1934         if (pf.discard_enabled && pf.discard_passdown) {
1935                 ti->num_discard_requests = 1;
1936
1937                 /*
1938                  * Setting 'discards_supported' circumvents the normal
1939                  * stacking of discard limits (this keeps the pool and
1940                  * thin devices' discard limits consistent).
1941                  */
1942                 ti->discards_supported = true;
1943                 ti->discard_zeroes_data_unsupported = true;
1944         }
1945         ti->private = pt;
1946
1947         pt->callbacks.congested_fn = pool_is_congested;
1948         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1949
1950         mutex_unlock(&dm_thin_pool_table.mutex);
1951
1952         return 0;
1953
1954 out_flags_changed:
1955         __pool_dec(pool);
1956 out_free_pt:
1957         kfree(pt);
1958 out:
1959         dm_put_device(ti, data_dev);
1960 out_metadata:
1961         dm_put_device(ti, metadata_dev);
1962 out_unlock:
1963         mutex_unlock(&dm_thin_pool_table.mutex);
1964
1965         return r;
1966 }
1967
1968 static int pool_map(struct dm_target *ti, struct bio *bio,
1969                     union map_info *map_context)
1970 {
1971         int r;
1972         struct pool_c *pt = ti->private;
1973         struct pool *pool = pt->pool;
1974         unsigned long flags;
1975
1976         /*
1977          * As this is a singleton target, ti->begin is always zero.
1978          */
1979         spin_lock_irqsave(&pool->lock, flags);
1980         bio->bi_bdev = pt->data_dev->bdev;
1981         r = DM_MAPIO_REMAPPED;
1982         spin_unlock_irqrestore(&pool->lock, flags);
1983
1984         return r;
1985 }
1986
1987 /*
1988  * Retrieves the number of blocks of the data device from
1989  * the superblock and compares it to the actual device size,
1990  * thus resizing the data device in case it has grown.
1991  *
1992  * This both copes with opening preallocated data devices in the ctr
1993  * being followed by a resume
1994  * -and-
1995  * calling the resume method individually after userspace has
1996  * grown the data device in reaction to a table event.
1997  */
1998 static int pool_preresume(struct dm_target *ti)
1999 {
2000         int r;
2001         struct pool_c *pt = ti->private;
2002         struct pool *pool = pt->pool;
2003         sector_t data_size = ti->len;
2004         dm_block_t sb_data_size;
2005
2006         /*
2007          * Take control of the pool object.
2008          */
2009         r = bind_control_target(pool, ti);
2010         if (r)
2011                 return r;
2012
2013         (void) sector_div(data_size, pool->sectors_per_block);
2014
2015         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2016         if (r) {
2017                 DMERR("failed to retrieve data device size");
2018                 return r;
2019         }
2020
2021         if (data_size < sb_data_size) {
2022                 DMERR("pool target too small, is %llu blocks (expected %llu)",
2023                       (unsigned long long)data_size, sb_data_size);
2024                 return -EINVAL;
2025
2026         } else if (data_size > sb_data_size) {
2027                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2028                 if (r) {
2029                         DMERR("failed to resize data device");
2030                         /* FIXME Stricter than necessary: Rollback transaction instead here */
2031                         set_pool_mode(pool, PM_READ_ONLY);
2032                         return r;
2033                 }
2034
2035                 (void) commit_or_fallback(pool);
2036         }
2037
2038         return 0;
2039 }
2040
2041 static void pool_resume(struct dm_target *ti)
2042 {
2043         struct pool_c *pt = ti->private;
2044         struct pool *pool = pt->pool;
2045         unsigned long flags;
2046
2047         spin_lock_irqsave(&pool->lock, flags);
2048         pool->low_water_triggered = 0;
2049         pool->no_free_space = 0;
2050         __requeue_bios(pool);
2051         spin_unlock_irqrestore(&pool->lock, flags);
2052
2053         do_waker(&pool->waker.work);
2054 }
2055
2056 static void pool_postsuspend(struct dm_target *ti)
2057 {
2058         struct pool_c *pt = ti->private;
2059         struct pool *pool = pt->pool;
2060
2061         cancel_delayed_work(&pool->waker);
2062         flush_workqueue(pool->wq);
2063         (void) commit_or_fallback(pool);
2064 }
2065
2066 static int check_arg_count(unsigned argc, unsigned args_required)
2067 {
2068         if (argc != args_required) {
2069                 DMWARN("Message received with %u arguments instead of %u.",
2070                        argc, args_required);
2071                 return -EINVAL;
2072         }
2073
2074         return 0;
2075 }
2076
2077 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2078 {
2079         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2080             *dev_id <= MAX_DEV_ID)
2081                 return 0;
2082
2083         if (warning)
2084                 DMWARN("Message received with invalid device id: %s", arg);
2085
2086         return -EINVAL;
2087 }
2088
2089 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2090 {
2091         dm_thin_id dev_id;
2092         int r;
2093
2094         r = check_arg_count(argc, 2);
2095         if (r)
2096                 return r;
2097
2098         r = read_dev_id(argv[1], &dev_id, 1);
2099         if (r)
2100                 return r;
2101
2102         r = dm_pool_create_thin(pool->pmd, dev_id);
2103         if (r) {
2104                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2105                        argv[1]);
2106                 return r;
2107         }
2108
2109         return 0;
2110 }
2111
2112 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2113 {
2114         dm_thin_id dev_id;
2115         dm_thin_id origin_dev_id;
2116         int r;
2117
2118         r = check_arg_count(argc, 3);
2119         if (r)
2120                 return r;
2121
2122         r = read_dev_id(argv[1], &dev_id, 1);
2123         if (r)
2124                 return r;
2125
2126         r = read_dev_id(argv[2], &origin_dev_id, 1);
2127         if (r)
2128                 return r;
2129
2130         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2131         if (r) {
2132                 DMWARN("Creation of new snapshot %s of device %s failed.",
2133                        argv[1], argv[2]);
2134                 return r;
2135         }
2136
2137         return 0;
2138 }
2139
2140 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2141 {
2142         dm_thin_id dev_id;
2143         int r;
2144
2145         r = check_arg_count(argc, 2);
2146         if (r)
2147                 return r;
2148
2149         r = read_dev_id(argv[1], &dev_id, 1);
2150         if (r)
2151                 return r;
2152
2153         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2154         if (r)
2155                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2156
2157         return r;
2158 }
2159
2160 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2161 {
2162         dm_thin_id old_id, new_id;
2163         int r;
2164
2165         r = check_arg_count(argc, 3);
2166         if (r)
2167                 return r;
2168
2169         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2170                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2171                 return -EINVAL;
2172         }
2173
2174         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2175                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2176                 return -EINVAL;
2177         }
2178
2179         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2180         if (r) {
2181                 DMWARN("Failed to change transaction id from %s to %s.",
2182                        argv[1], argv[2]);
2183                 return r;
2184         }
2185
2186         return 0;
2187 }
2188
2189 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2190 {
2191         int r;
2192
2193         r = check_arg_count(argc, 1);
2194         if (r)
2195                 return r;
2196
2197         (void) commit_or_fallback(pool);
2198
2199         r = dm_pool_reserve_metadata_snap(pool->pmd);
2200         if (r)
2201                 DMWARN("reserve_metadata_snap message failed.");
2202
2203         return r;
2204 }
2205
2206 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2207 {
2208         int r;
2209
2210         r = check_arg_count(argc, 1);
2211         if (r)
2212                 return r;
2213
2214         r = dm_pool_release_metadata_snap(pool->pmd);
2215         if (r)
2216                 DMWARN("release_metadata_snap message failed.");
2217
2218         return r;
2219 }
2220
2221 /*
2222  * Messages supported:
2223  *   create_thin        <dev_id>
2224  *   create_snap        <dev_id> <origin_id>
2225  *   delete             <dev_id>
2226  *   trim               <dev_id> <new_size_in_sectors>
2227  *   set_transaction_id <current_trans_id> <new_trans_id>
2228  *   reserve_metadata_snap
2229  *   release_metadata_snap
2230  */
2231 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2232 {
2233         int r = -EINVAL;
2234         struct pool_c *pt = ti->private;
2235         struct pool *pool = pt->pool;
2236
2237         if (!strcasecmp(argv[0], "create_thin"))
2238                 r = process_create_thin_mesg(argc, argv, pool);
2239
2240         else if (!strcasecmp(argv[0], "create_snap"))
2241                 r = process_create_snap_mesg(argc, argv, pool);
2242
2243         else if (!strcasecmp(argv[0], "delete"))
2244                 r = process_delete_mesg(argc, argv, pool);
2245
2246         else if (!strcasecmp(argv[0], "set_transaction_id"))
2247                 r = process_set_transaction_id_mesg(argc, argv, pool);
2248
2249         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2250                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2251
2252         else if (!strcasecmp(argv[0], "release_metadata_snap"))
2253                 r = process_release_metadata_snap_mesg(argc, argv, pool);
2254
2255         else
2256                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2257
2258         if (!r)
2259                 (void) commit_or_fallback(pool);
2260
2261         return r;
2262 }
2263
2264 static void emit_flags(struct pool_features *pf, char *result,
2265                        unsigned sz, unsigned maxlen)
2266 {
2267         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2268                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2269         DMEMIT("%u ", count);
2270
2271         if (!pf->zero_new_blocks)
2272                 DMEMIT("skip_block_zeroing ");
2273
2274         if (!pf->discard_enabled)
2275                 DMEMIT("ignore_discard ");
2276
2277         if (!pf->discard_passdown)
2278                 DMEMIT("no_discard_passdown ");
2279
2280         if (pf->mode == PM_READ_ONLY)
2281                 DMEMIT("read_only ");
2282 }
2283
2284 /*
2285  * Status line is:
2286  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2287  *    <used data sectors>/<total data sectors> <held metadata root>
2288  */
2289 static int pool_status(struct dm_target *ti, status_type_t type,
2290                        unsigned status_flags, char *result, unsigned maxlen)
2291 {
2292         int r;
2293         unsigned sz = 0;
2294         uint64_t transaction_id;
2295         dm_block_t nr_free_blocks_data;
2296         dm_block_t nr_free_blocks_metadata;
2297         dm_block_t nr_blocks_data;
2298         dm_block_t nr_blocks_metadata;
2299         dm_block_t held_root;
2300         char buf[BDEVNAME_SIZE];
2301         char buf2[BDEVNAME_SIZE];
2302         struct pool_c *pt = ti->private;
2303         struct pool *pool = pt->pool;
2304
2305         switch (type) {
2306         case STATUSTYPE_INFO:
2307                 if (get_pool_mode(pool) == PM_FAIL) {
2308                         DMEMIT("Fail");
2309                         break;
2310                 }
2311
2312                 /* Commit to ensure statistics aren't out-of-date */
2313                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2314                         (void) commit_or_fallback(pool);
2315
2316                 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2317                                                         &transaction_id);
2318                 if (r)
2319                         return r;
2320
2321                 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2322                                                           &nr_free_blocks_metadata);
2323                 if (r)
2324                         return r;
2325
2326                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2327                 if (r)
2328                         return r;
2329
2330                 r = dm_pool_get_free_block_count(pool->pmd,
2331                                                  &nr_free_blocks_data);
2332                 if (r)
2333                         return r;
2334
2335                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2336                 if (r)
2337                         return r;
2338
2339                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2340                 if (r)
2341                         return r;
2342
2343                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2344                        (unsigned long long)transaction_id,
2345                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2346                        (unsigned long long)nr_blocks_metadata,
2347                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2348                        (unsigned long long)nr_blocks_data);
2349
2350                 if (held_root)
2351                         DMEMIT("%llu ", held_root);
2352                 else
2353                         DMEMIT("- ");
2354
2355                 if (pool->pf.mode == PM_READ_ONLY)
2356                         DMEMIT("ro ");
2357                 else
2358                         DMEMIT("rw ");
2359
2360                 if (pool->pf.discard_enabled && pool->pf.discard_passdown)
2361                         DMEMIT("discard_passdown");
2362                 else
2363                         DMEMIT("no_discard_passdown");
2364
2365                 break;
2366
2367         case STATUSTYPE_TABLE:
2368                 DMEMIT("%s %s %lu %llu ",
2369                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2370                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2371                        (unsigned long)pool->sectors_per_block,
2372                        (unsigned long long)pt->low_water_blocks);
2373                 emit_flags(&pt->requested_pf, result, sz, maxlen);
2374                 break;
2375         }
2376
2377         return 0;
2378 }
2379
2380 static int pool_iterate_devices(struct dm_target *ti,
2381                                 iterate_devices_callout_fn fn, void *data)
2382 {
2383         struct pool_c *pt = ti->private;
2384
2385         return fn(ti, pt->data_dev, 0, ti->len, data);
2386 }
2387
2388 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2389                       struct bio_vec *biovec, int max_size)
2390 {
2391         struct pool_c *pt = ti->private;
2392         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2393
2394         if (!q->merge_bvec_fn)
2395                 return max_size;
2396
2397         bvm->bi_bdev = pt->data_dev->bdev;
2398
2399         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2400 }
2401
2402 static bool block_size_is_power_of_two(struct pool *pool)
2403 {
2404         return pool->sectors_per_block_shift >= 0;
2405 }
2406
2407 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2408 {
2409         struct pool *pool = pt->pool;
2410         struct queue_limits *data_limits;
2411
2412         limits->max_discard_sectors = pool->sectors_per_block;
2413
2414         /*
2415          * discard_granularity is just a hint, and not enforced.
2416          */
2417         if (pt->adjusted_pf.discard_passdown) {
2418                 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2419                 limits->discard_granularity = data_limits->discard_granularity;
2420         } else if (block_size_is_power_of_two(pool))
2421                 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2422         else
2423                 /*
2424                  * Use largest power of 2 that is a factor of sectors_per_block
2425                  * but at least DATA_DEV_BLOCK_SIZE_MIN_SECTORS.
2426                  */
2427                 limits->discard_granularity = max(1 << (ffs(pool->sectors_per_block) - 1),
2428                                                   DATA_DEV_BLOCK_SIZE_MIN_SECTORS) << SECTOR_SHIFT;
2429 }
2430
2431 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2432 {
2433         struct pool_c *pt = ti->private;
2434         struct pool *pool = pt->pool;
2435
2436         blk_limits_io_min(limits, 0);
2437         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2438
2439         /*
2440          * pt->adjusted_pf is a staging area for the actual features to use.
2441          * They get transferred to the live pool in bind_control_target()
2442          * called from pool_preresume().
2443          */
2444         if (!pt->adjusted_pf.discard_enabled)
2445                 return;
2446
2447         disable_passdown_if_not_supported(pt);
2448
2449         set_discard_limits(pt, limits);
2450 }
2451
2452 static struct target_type pool_target = {
2453         .name = "thin-pool",
2454         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2455                     DM_TARGET_IMMUTABLE,
2456         .version = {1, 5, 0},
2457         .module = THIS_MODULE,
2458         .ctr = pool_ctr,
2459         .dtr = pool_dtr,
2460         .map = pool_map,
2461         .postsuspend = pool_postsuspend,
2462         .preresume = pool_preresume,
2463         .resume = pool_resume,
2464         .message = pool_message,
2465         .status = pool_status,
2466         .merge = pool_merge,
2467         .iterate_devices = pool_iterate_devices,
2468         .io_hints = pool_io_hints,
2469 };
2470
2471 /*----------------------------------------------------------------
2472  * Thin target methods
2473  *--------------------------------------------------------------*/
2474 static void thin_dtr(struct dm_target *ti)
2475 {
2476         struct thin_c *tc = ti->private;
2477
2478         mutex_lock(&dm_thin_pool_table.mutex);
2479
2480         __pool_dec(tc->pool);
2481         dm_pool_close_thin_device(tc->td);
2482         dm_put_device(ti, tc->pool_dev);
2483         if (tc->origin_dev)
2484                 dm_put_device(ti, tc->origin_dev);
2485         kfree(tc);
2486
2487         mutex_unlock(&dm_thin_pool_table.mutex);
2488 }
2489
2490 /*
2491  * Thin target parameters:
2492  *
2493  * <pool_dev> <dev_id> [origin_dev]
2494  *
2495  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2496  * dev_id: the internal device identifier
2497  * origin_dev: a device external to the pool that should act as the origin
2498  *
2499  * If the pool device has discards disabled, they get disabled for the thin
2500  * device as well.
2501  */
2502 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2503 {
2504         int r;
2505         struct thin_c *tc;
2506         struct dm_dev *pool_dev, *origin_dev;
2507         struct mapped_device *pool_md;
2508
2509         mutex_lock(&dm_thin_pool_table.mutex);
2510
2511         if (argc != 2 && argc != 3) {
2512                 ti->error = "Invalid argument count";
2513                 r = -EINVAL;
2514                 goto out_unlock;
2515         }
2516
2517         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2518         if (!tc) {
2519                 ti->error = "Out of memory";
2520                 r = -ENOMEM;
2521                 goto out_unlock;
2522         }
2523
2524         if (argc == 3) {
2525                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2526                 if (r) {
2527                         ti->error = "Error opening origin device";
2528                         goto bad_origin_dev;
2529                 }
2530                 tc->origin_dev = origin_dev;
2531         }
2532
2533         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2534         if (r) {
2535                 ti->error = "Error opening pool device";
2536                 goto bad_pool_dev;
2537         }
2538         tc->pool_dev = pool_dev;
2539
2540         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2541                 ti->error = "Invalid device id";
2542                 r = -EINVAL;
2543                 goto bad_common;
2544         }
2545
2546         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2547         if (!pool_md) {
2548                 ti->error = "Couldn't get pool mapped device";
2549                 r = -EINVAL;
2550                 goto bad_common;
2551         }
2552
2553         tc->pool = __pool_table_lookup(pool_md);
2554         if (!tc->pool) {
2555                 ti->error = "Couldn't find pool object";
2556                 r = -EINVAL;
2557                 goto bad_pool_lookup;
2558         }
2559         __pool_inc(tc->pool);
2560
2561         if (get_pool_mode(tc->pool) == PM_FAIL) {
2562                 ti->error = "Couldn't open thin device, Pool is in fail mode";
2563                 goto bad_thin_open;
2564         }
2565
2566         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2567         if (r) {
2568                 ti->error = "Couldn't open thin internal device";
2569                 goto bad_thin_open;
2570         }
2571
2572         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2573         if (r)
2574                 goto bad_thin_open;
2575
2576         ti->num_flush_requests = 1;
2577         ti->flush_supported = true;
2578
2579         /* In case the pool supports discards, pass them on. */
2580         if (tc->pool->pf.discard_enabled) {
2581                 ti->discards_supported = true;
2582                 ti->num_discard_requests = 1;
2583                 ti->discard_zeroes_data_unsupported = true;
2584                 /* Discard requests must be split on a block boundary */
2585                 ti->split_discard_requests = true;
2586         }
2587
2588         dm_put(pool_md);
2589
2590         mutex_unlock(&dm_thin_pool_table.mutex);
2591
2592         return 0;
2593
2594 bad_thin_open:
2595         __pool_dec(tc->pool);
2596 bad_pool_lookup:
2597         dm_put(pool_md);
2598 bad_common:
2599         dm_put_device(ti, tc->pool_dev);
2600 bad_pool_dev:
2601         if (tc->origin_dev)
2602                 dm_put_device(ti, tc->origin_dev);
2603 bad_origin_dev:
2604         kfree(tc);
2605 out_unlock:
2606         mutex_unlock(&dm_thin_pool_table.mutex);
2607
2608         return r;
2609 }
2610
2611 static int thin_map(struct dm_target *ti, struct bio *bio,
2612                     union map_info *map_context)
2613 {
2614         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2615
2616         return thin_bio_map(ti, bio, map_context);
2617 }
2618
2619 static int thin_endio(struct dm_target *ti,
2620                       struct bio *bio, int err,
2621                       union map_info *map_context)
2622 {
2623         unsigned long flags;
2624         struct dm_thin_endio_hook *h = map_context->ptr;
2625         struct list_head work;
2626         struct dm_thin_new_mapping *m, *tmp;
2627         struct pool *pool = h->tc->pool;
2628
2629         if (h->shared_read_entry) {
2630                 INIT_LIST_HEAD(&work);
2631                 dm_deferred_entry_dec(h->shared_read_entry, &work);
2632
2633                 spin_lock_irqsave(&pool->lock, flags);
2634                 list_for_each_entry_safe(m, tmp, &work, list) {
2635                         list_del(&m->list);
2636                         m->quiesced = 1;
2637                         __maybe_add_mapping(m);
2638                 }
2639                 spin_unlock_irqrestore(&pool->lock, flags);
2640         }
2641
2642         if (h->all_io_entry) {
2643                 INIT_LIST_HEAD(&work);
2644                 dm_deferred_entry_dec(h->all_io_entry, &work);
2645                 spin_lock_irqsave(&pool->lock, flags);
2646                 list_for_each_entry_safe(m, tmp, &work, list)
2647                         list_add(&m->list, &pool->prepared_discards);
2648                 spin_unlock_irqrestore(&pool->lock, flags);
2649         }
2650
2651         mempool_free(h, pool->endio_hook_pool);
2652
2653         return 0;
2654 }
2655
2656 static void thin_postsuspend(struct dm_target *ti)
2657 {
2658         if (dm_noflush_suspending(ti))
2659                 requeue_io((struct thin_c *)ti->private);
2660 }
2661
2662 /*
2663  * <nr mapped sectors> <highest mapped sector>
2664  */
2665 static int thin_status(struct dm_target *ti, status_type_t type,
2666                        unsigned status_flags, char *result, unsigned maxlen)
2667 {
2668         int r;
2669         ssize_t sz = 0;
2670         dm_block_t mapped, highest;
2671         char buf[BDEVNAME_SIZE];
2672         struct thin_c *tc = ti->private;
2673
2674         if (get_pool_mode(tc->pool) == PM_FAIL) {
2675                 DMEMIT("Fail");
2676                 return 0;
2677         }
2678
2679         if (!tc->td)
2680                 DMEMIT("-");
2681         else {
2682                 switch (type) {
2683                 case STATUSTYPE_INFO:
2684                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2685                         if (r)
2686                                 return r;
2687
2688                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2689                         if (r < 0)
2690                                 return r;
2691
2692                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2693                         if (r)
2694                                 DMEMIT("%llu", ((highest + 1) *
2695                                                 tc->pool->sectors_per_block) - 1);
2696                         else
2697                                 DMEMIT("-");
2698                         break;
2699
2700                 case STATUSTYPE_TABLE:
2701                         DMEMIT("%s %lu",
2702                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2703                                (unsigned long) tc->dev_id);
2704                         if (tc->origin_dev)
2705                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2706                         break;
2707                 }
2708         }
2709
2710         return 0;
2711 }
2712
2713 static int thin_iterate_devices(struct dm_target *ti,
2714                                 iterate_devices_callout_fn fn, void *data)
2715 {
2716         sector_t blocks;
2717         struct thin_c *tc = ti->private;
2718         struct pool *pool = tc->pool;
2719
2720         /*
2721          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2722          * we follow a more convoluted path through to the pool's target.
2723          */
2724         if (!pool->ti)
2725                 return 0;       /* nothing is bound */
2726
2727         blocks = pool->ti->len;
2728         (void) sector_div(blocks, pool->sectors_per_block);
2729         if (blocks)
2730                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2731
2732         return 0;
2733 }
2734
2735 /*
2736  * A thin device always inherits its queue limits from its pool.
2737  */
2738 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2739 {
2740         struct thin_c *tc = ti->private;
2741
2742         *limits = bdev_get_queue(tc->pool_dev->bdev)->limits;
2743 }
2744
2745 static struct target_type thin_target = {
2746         .name = "thin",
2747         .version = {1, 5, 0},
2748         .module = THIS_MODULE,
2749         .ctr = thin_ctr,
2750         .dtr = thin_dtr,
2751         .map = thin_map,
2752         .end_io = thin_endio,
2753         .postsuspend = thin_postsuspend,
2754         .status = thin_status,
2755         .iterate_devices = thin_iterate_devices,
2756         .io_hints = thin_io_hints,
2757 };
2758
2759 /*----------------------------------------------------------------*/
2760
2761 static int __init dm_thin_init(void)
2762 {
2763         int r;
2764
2765         pool_table_init();
2766
2767         r = dm_register_target(&thin_target);
2768         if (r)
2769                 return r;
2770
2771         r = dm_register_target(&pool_target);
2772         if (r)
2773                 goto bad_pool_target;
2774
2775         r = -ENOMEM;
2776
2777         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2778         if (!_new_mapping_cache)
2779                 goto bad_new_mapping_cache;
2780
2781         _endio_hook_cache = KMEM_CACHE(dm_thin_endio_hook, 0);
2782         if (!_endio_hook_cache)
2783                 goto bad_endio_hook_cache;
2784
2785         return 0;
2786
2787 bad_endio_hook_cache:
2788         kmem_cache_destroy(_new_mapping_cache);
2789 bad_new_mapping_cache:
2790         dm_unregister_target(&pool_target);
2791 bad_pool_target:
2792         dm_unregister_target(&thin_target);
2793
2794         return r;
2795 }
2796
2797 static void dm_thin_exit(void)
2798 {
2799         dm_unregister_target(&thin_target);
2800         dm_unregister_target(&pool_target);
2801
2802         kmem_cache_destroy(_new_mapping_cache);
2803         kmem_cache_destroy(_endio_hook_cache);
2804 }
2805
2806 module_init(dm_thin_init);
2807 module_exit(dm_thin_exit);
2808
2809 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2810 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2811 MODULE_LICENSE("GPL");