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 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
  30                 "A percentage of time allocated for copy on write");
  31
  32 /*
  33  * The block size of the device holding pool data must be
  34  * between 64KB and 1GB.
  35  */
  36 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
  37 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
  38
  39 /*
  40  * Device id is restricted to 24 bits.
  41  */
  42 #define MAX_DEV_ID ((1 << 24) - 1)
  43
  44 /*
  45  * How do we handle breaking sharing of data blocks?
  46  * =================================================
  47  *
  48  * We use a standard copy-on-write btree to store the mappings for the
  49  * devices (note I'm talking about copy-on-write of the metadata here, not
  50  * the data).  When you take an internal snapshot you clone the root node
  51  * of the origin btree.  After this there is no concept of an origin or a
  52  * snapshot.  They are just two device trees that happen to point to the
  53  * same data blocks.
  54  *
  55  * When we get a write in we decide if it's to a shared data block using
  56  * some timestamp magic.  If it is, we have to break sharing.
  57  *
  58  * Let's say we write to a shared block in what was the origin.  The
  59  * steps are:
  60  *
  61  * i) plug io further to this physical block. (see bio_prison code).
  62  *
  63  * ii) quiesce any read io to that shared data block.  Obviously
  64  * including all devices that share this block.  (see dm_deferred_set code)
  65  *
  66  * iii) copy the data block to a newly allocate block.  This step can be
  67  * missed out if the io covers the block. (schedule_copy).
  68  *
  69  * iv) insert the new mapping into the origin's btree
  70  * (process_prepared_mapping).  This act of inserting breaks some
  71  * sharing of btree nodes between the two devices.  Breaking sharing only
  72  * effects the btree of that specific device.  Btrees for the other
  73  * devices that share the block never change.  The btree for the origin
  74  * device as it was after the last commit is untouched, ie. we're using
  75  * persistent data structures in the functional programming sense.
  76  *
  77  * v) unplug io to this physical block, including the io that triggered
  78  * the breaking of sharing.
  79  *
  80  * Steps (ii) and (iii) occur in parallel.
  81  *
  82  * The metadata _doesn't_ need to be committed before the io continues.  We
  83  * get away with this because the io is always written to a _new_ block.
  84  * If there's a crash, then:
  85  *
  86  * - The origin mapping will point to the old origin block (the shared
  87  * one).  This will contain the data as it was before the io that triggered
  88  * the breaking of sharing came in.
  89  *
  90  * - The snap mapping still points to the old block.  As it would after
  91  * the commit.
  92  *
  93  * The downside of this scheme is the timestamp magic isn't perfect, and
  94  * will continue to think that data block in the snapshot device is shared
  95  * even after the write to the origin has broken sharing.  I suspect data
  96  * blocks will typically be shared by many different devices, so we're
  97  * breaking sharing n + 1 times, rather than n, where n is the number of
  98  * devices that reference this data block.  At the moment I think the
  99  * benefits far, far outweigh the disadvantages.
 100  */
 101
 102 /*----------------------------------------------------------------*/
 103
 104 /*
 105  * Key building.
 106  */
 107 static void build_data_key(struct dm_thin_device *td,
 108                            dm_block_t b, struct dm_cell_key *key)
 109 {
 110         key->virtual = 0;
 111         key->dev = dm_thin_dev_id(td);
 112         key->block = b;
 113 }
 114
 115 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
 116                               struct dm_cell_key *key)
 117 {
 118         key->virtual = 1;
 119         key->dev = dm_thin_dev_id(td);
 120         key->block = b;
 121 }
 122
 123 /*----------------------------------------------------------------*/
 124
 125 /*
 126  * A pool device ties together a metadata device and a data device.  It
 127  * also provides the interface for creating and destroying internal
 128  * devices.
 129  */
 130 struct dm_thin_new_mapping;
 131
 132 /*
 133  * The pool runs in 3 modes.  Ordered in degraded order for comparisons.
 134  */
 135 enum pool_mode {
 136         PM_WRITE,               /* metadata may be changed */
 137         PM_READ_ONLY,           /* metadata may not be changed */
 138         PM_FAIL,                /* all I/O fails */
 139 };
 140
 141 struct pool_features {
 142         enum pool_mode mode;
 143
 144         bool zero_new_blocks:1;
 145         bool discard_enabled:1;
 146         bool discard_passdown:1;
 147 };
 148
 149 struct thin_c;
 150 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
 151 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
 152
 153 struct pool {
 154         struct list_head list;
 155         struct dm_target *ti;   /* Only set if a pool target is bound */
 156
 157         struct mapped_device *pool_md;
 158         struct block_device *md_dev;
 159         struct dm_pool_metadata *pmd;
 160
 161         dm_block_t low_water_blocks;
 162         uint32_t sectors_per_block;
 163         int sectors_per_block_shift;
 164
 165         struct pool_features pf;
 166         unsigned low_water_triggered:1; /* A dm event has been sent */
 167         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
 168
 169         struct dm_bio_prison *prison;
 170         struct dm_kcopyd_client *copier;
 171
 172         struct workqueue_struct *wq;
 173         struct work_struct worker;
 174         struct delayed_work waker;
 175
 176         unsigned long last_commit_jiffies;
 177         unsigned ref_count;
 178
 179         spinlock_t lock;
 180         struct bio_list deferred_bios;
 181         struct bio_list deferred_flush_bios;
 182         struct list_head prepared_mappings;
 183         struct list_head prepared_discards;
 184
 185         struct bio_list retry_on_resume_list;
 186
 187         struct dm_deferred_set *shared_read_ds;
 188         struct dm_deferred_set *all_io_ds;
 189
 190         struct dm_thin_new_mapping *next_mapping;
 191         mempool_t *mapping_pool;
 192
 193         process_bio_fn process_bio;
 194         process_bio_fn process_discard;
 195
 196         process_mapping_fn process_prepared_mapping;
 197         process_mapping_fn process_prepared_discard;
 198 };
 199
 200 static enum pool_mode get_pool_mode(struct pool *pool);
 201 static void set_pool_mode(struct pool *pool, enum pool_mode mode);
 202
 203 /*
 204  * Target context for a pool.
 205  */
 206 struct pool_c {
 207         struct dm_target *ti;
 208         struct pool *pool;
 209         struct dm_dev *data_dev;
 210         struct dm_dev *metadata_dev;
 211         struct dm_target_callbacks callbacks;
 212
 213         dm_block_t low_water_blocks;
 214         struct pool_features requested_pf; /* Features requested during table load */
 215         struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
 216 };
 217
 218 /*
 219  * Target context for a thin.
 220  */
 221 struct thin_c {
 222         struct dm_dev *pool_dev;
 223         struct dm_dev *origin_dev;
 224         dm_thin_id dev_id;
 225
 226         struct pool *pool;
 227         struct dm_thin_device *td;
 228 };
 229
 230 /*----------------------------------------------------------------*/
 231
 232 /*
 233  * A global list of pools that uses a struct mapped_device as a key.
 234  */
 235 static struct dm_thin_pool_table {
 236         struct mutex mutex;
 237         struct list_head pools;
 238 } dm_thin_pool_table;
 239
 240 static void pool_table_init(void)
 241 {
 242         mutex_init(&dm_thin_pool_table.mutex);
 243         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
 244 }
 245
 246 static void __pool_table_insert(struct pool *pool)
 247 {
 248         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 249         list_add(&pool->list, &dm_thin_pool_table.pools);
 250 }
 251
 252 static void __pool_table_remove(struct pool *pool)
 253 {
 254         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 255         list_del(&pool->list);
 256 }
 257
 258 static struct pool *__pool_table_lookup(struct mapped_device *md)
 259 {
 260         struct pool *pool = NULL, *tmp;
 261
 262         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 263
 264         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 265                 if (tmp->pool_md == md) {
 266                         pool = tmp;
 267                         break;
 268                 }
 269         }
 270
 271         return pool;
 272 }
 273
 274 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
 275 {
 276         struct pool *pool = NULL, *tmp;
 277
 278         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 279
 280         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 281                 if (tmp->md_dev == md_dev) {
 282                         pool = tmp;
 283                         break;
 284                 }
 285         }
 286
 287         return pool;
 288 }
 289
 290 /*----------------------------------------------------------------*/
 291
 292 struct dm_thin_endio_hook {
 293         struct thin_c *tc;
 294         struct dm_deferred_entry *shared_read_entry;
 295         struct dm_deferred_entry *all_io_entry;
 296         struct dm_thin_new_mapping *overwrite_mapping;
 297 };
 298
 299 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
 300 {
 301         struct bio *bio;
 302         struct bio_list bios;
 303
 304         bio_list_init(&bios);
 305         bio_list_merge(&bios, master);
 306         bio_list_init(master);
 307
 308         while ((bio = bio_list_pop(&bios))) {
 309                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 310
 311                 if (h->tc == tc)
 312                         bio_endio(bio, DM_ENDIO_REQUEUE);
 313                 else
 314                         bio_list_add(master, bio);
 315         }
 316 }
 317
 318 static void requeue_io(struct thin_c *tc)
 319 {
 320         struct pool *pool = tc->pool;
 321         unsigned long flags;
 322
 323         spin_lock_irqsave(&pool->lock, flags);
 324         __requeue_bio_list(tc, &pool->deferred_bios);
 325         __requeue_bio_list(tc, &pool->retry_on_resume_list);
 326         spin_unlock_irqrestore(&pool->lock, flags);
 327 }
 328
 329 /*
 330  * This section of code contains the logic for processing a thin device's IO.
 331  * Much of the code depends on pool object resources (lists, workqueues, etc)
 332  * but most is exclusively called from the thin target rather than the thin-pool
 333  * target.
 334  */
 335
 336 static bool block_size_is_power_of_two(struct pool *pool)
 337 {
 338         return pool->sectors_per_block_shift >= 0;
 339 }
 340
 341 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
 342 {
 343         struct pool *pool = tc->pool;
 344         sector_t block_nr = bio->bi_sector;
 345
 346         if (block_size_is_power_of_two(pool))
 347                 block_nr >>= pool->sectors_per_block_shift;
 348         else
 349                 (void) sector_div(block_nr, pool->sectors_per_block);
 350
 351         return block_nr;
 352 }
 353
 354 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
 355 {
 356         struct pool *pool = tc->pool;
 357         sector_t bi_sector = bio->bi_sector;
 358
 359         bio->bi_bdev = tc->pool_dev->bdev;
 360         if (block_size_is_power_of_two(pool))
 361                 bio->bi_sector = (block << pool->sectors_per_block_shift) |
 362                                 (bi_sector & (pool->sectors_per_block - 1));
 363         else
 364                 bio->bi_sector = (block * pool->sectors_per_block) +
 365                                  sector_div(bi_sector, pool->sectors_per_block);
 366 }
 367
 368 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
 369 {
 370         bio->bi_bdev = tc->origin_dev->bdev;
 371 }
 372
 373 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
 374 {
 375         return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
 376                 dm_thin_changed_this_transaction(tc->td);
 377 }
 378
 379 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
 380 {
 381         struct dm_thin_endio_hook *h;
 382
 383         if (bio->bi_rw & REQ_DISCARD)
 384                 return;
 385
 386         h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 387         h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
 388 }
 389
 390 static void issue(struct thin_c *tc, struct bio *bio)
 391 {
 392         struct pool *pool = tc->pool;
 393         unsigned long flags;
 394
 395         if (!bio_triggers_commit(tc, bio)) {
 396                 generic_make_request(bio);
 397                 return;
 398         }
 399
 400         /*
 401          * Complete bio with an error if earlier I/O caused changes to
 402          * the metadata that can't be committed e.g, due to I/O errors
 403          * on the metadata device.
 404          */
 405         if (dm_thin_aborted_changes(tc->td)) {
 406                 bio_io_error(bio);
 407                 return;
 408         }
 409
 410         /*
 411          * Batch together any bios that trigger commits and then issue a
 412          * single commit for them in process_deferred_bios().
 413          */
 414         spin_lock_irqsave(&pool->lock, flags);
 415         bio_list_add(&pool->deferred_flush_bios, bio);
 416         spin_unlock_irqrestore(&pool->lock, flags);
 417 }
 418
 419 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
 420 {
 421         remap_to_origin(tc, bio);
 422         issue(tc, bio);
 423 }
 424
 425 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
 426                             dm_block_t block)
 427 {
 428         remap(tc, bio, block);
 429         issue(tc, bio);
 430 }
 431
 432 /*
 433  * wake_worker() is used when new work is queued and when pool_resume is
 434  * ready to continue deferred IO processing.
 435  */
 436 static void wake_worker(struct pool *pool)
 437 {
 438         queue_work(pool->wq, &pool->worker);
 439 }
 440
 441 /*----------------------------------------------------------------*/
 442
 443 /*
 444  * Bio endio functions.
 445  */
 446 struct dm_thin_new_mapping {
 447         struct list_head list;
 448
 449         unsigned quiesced:1;
 450         unsigned prepared:1;
 451         unsigned pass_discard:1;
 452
 453         struct thin_c *tc;
 454         dm_block_t virt_block;
 455         dm_block_t data_block;
 456         struct dm_bio_prison_cell *cell, *cell2;
 457         int err;
 458
 459         /*
 460          * If the bio covers the whole area of a block then we can avoid
 461          * zeroing or copying.  Instead this bio is hooked.  The bio will
 462          * still be in the cell, so care has to be taken to avoid issuing
 463          * the bio twice.
 464          */
 465         struct bio *bio;
 466         bio_end_io_t *saved_bi_end_io;
 467 };
 468
 469 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
 470 {
 471         struct pool *pool = m->tc->pool;
 472
 473         if (m->quiesced && m->prepared) {
 474                 list_add(&m->list, &pool->prepared_mappings);
 475                 wake_worker(pool);
 476         }
 477 }
 478
 479 static void copy_complete(int read_err, unsigned long write_err, void *context)
 480 {
 481         unsigned long flags;
 482         struct dm_thin_new_mapping *m = context;
 483         struct pool *pool = m->tc->pool;
 484
 485         m->err = read_err || write_err ? -EIO : 0;
 486
 487         spin_lock_irqsave(&pool->lock, flags);
 488         m->prepared = 1;
 489         __maybe_add_mapping(m);
 490         spin_unlock_irqrestore(&pool->lock, flags);
 491 }
 492
 493 static void overwrite_endio(struct bio *bio, int err)
 494 {
 495         unsigned long flags;
 496         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 497         struct dm_thin_new_mapping *m = h->overwrite_mapping;
 498         struct pool *pool = m->tc->pool;
 499
 500         m->err = err;
 501
 502         spin_lock_irqsave(&pool->lock, flags);
 503         m->prepared = 1;
 504         __maybe_add_mapping(m);
 505         spin_unlock_irqrestore(&pool->lock, flags);
 506 }
 507
 508 /*----------------------------------------------------------------*/
 509
 510 /*
 511  * Workqueue.
 512  */
 513
 514 /*
 515  * Prepared mapping jobs.
 516  */
 517
 518 /*
 519  * This sends the bios in the cell back to the deferred_bios list.
 520  */
 521 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 522 {
 523         struct pool *pool = tc->pool;
 524         unsigned long flags;
 525
 526         spin_lock_irqsave(&pool->lock, flags);
 527         dm_cell_release(cell, &pool->deferred_bios);
 528         spin_unlock_irqrestore(&tc->pool->lock, flags);
 529
 530         wake_worker(pool);
 531 }
 532
 533 /*
 534  * Same as cell_defer except it omits the original holder of the cell.
 535  */
 536 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 537 {
 538         struct pool *pool = tc->pool;
 539         unsigned long flags;
 540
 541         spin_lock_irqsave(&pool->lock, flags);
 542         dm_cell_release_no_holder(cell, &pool->deferred_bios);
 543         spin_unlock_irqrestore(&pool->lock, flags);
 544
 545         wake_worker(pool);
 546 }
 547
 548 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
 549 {
 550         if (m->bio)
 551                 m->bio->bi_end_io = m->saved_bi_end_io;
 552         dm_cell_error(m->cell);
 553         list_del(&m->list);
 554         mempool_free(m, m->tc->pool->mapping_pool);
 555 }
 556 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
 557 {
 558         struct thin_c *tc = m->tc;
 559         struct bio *bio;
 560         int r;
 561
 562         bio = m->bio;
 563         if (bio)
 564                 bio->bi_end_io = m->saved_bi_end_io;
 565
 566         if (m->err) {
 567                 dm_cell_error(m->cell);
 568                 goto out;
 569         }
 570
 571         /*
 572          * Commit the prepared block into the mapping btree.
 573          * Any I/O for this block arriving after this point will get
 574          * remapped to it directly.
 575          */
 576         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
 577         if (r) {
 578                 DMERR_LIMIT("dm_thin_insert_block() failed");
 579                 dm_cell_error(m->cell);
 580                 goto out;
 581         }
 582
 583         /*
 584          * Release any bios held while the block was being provisioned.
 585          * If we are processing a write bio that completely covers the block,
 586          * we already processed it so can ignore it now when processing
 587          * the bios in the cell.
 588          */
 589         if (bio) {
 590                 cell_defer_no_holder(tc, m->cell);
 591                 bio_endio(bio, 0);
 592         } else
 593                 cell_defer(tc, m->cell);
 594
 595 out:
 596         list_del(&m->list);
 597         mempool_free(m, tc->pool->mapping_pool);
 598 }
 599
 600 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
 601 {
 602         struct thin_c *tc = m->tc;
 603
 604         bio_io_error(m->bio);
 605         cell_defer_no_holder(tc, m->cell);
 606         cell_defer_no_holder(tc, m->cell2);
 607         mempool_free(m, tc->pool->mapping_pool);
 608 }
 609
 610 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
 611 {
 612         struct thin_c *tc = m->tc;
 613
 614         inc_all_io_entry(tc->pool, m->bio);
 615         cell_defer_no_holder(tc, m->cell);
 616         cell_defer_no_holder(tc, m->cell2);
 617
 618         if (m->pass_discard)
 619                 remap_and_issue(tc, m->bio, m->data_block);
 620         else
 621                 bio_endio(m->bio, 0);
 622
 623         mempool_free(m, tc->pool->mapping_pool);
 624 }
 625
 626 static void process_prepared_discard(struct dm_thin_new_mapping *m)
 627 {
 628         int r;
 629         struct thin_c *tc = m->tc;
 630
 631         r = dm_thin_remove_block(tc->td, m->virt_block);
 632         if (r)
 633                 DMERR_LIMIT("dm_thin_remove_block() failed");
 634
 635         process_prepared_discard_passdown(m);
 636 }
 637
 638 static void process_prepared(struct pool *pool, struct list_head *head,
 639                              process_mapping_fn *fn)
 640 {
 641         unsigned long flags;
 642         struct list_head maps;
 643         struct dm_thin_new_mapping *m, *tmp;
 644
 645         INIT_LIST_HEAD(&maps);
 646         spin_lock_irqsave(&pool->lock, flags);
 647         list_splice_init(head, &maps);
 648         spin_unlock_irqrestore(&pool->lock, flags);
 649
 650         list_for_each_entry_safe(m, tmp, &maps, list)
 651                 (*fn)(m);
 652 }
 653
 654 /*
 655  * Deferred bio jobs.
 656  */
 657 static int io_overlaps_block(struct pool *pool, struct bio *bio)
 658 {
 659         return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
 660 }
 661
 662 static int io_overwrites_block(struct pool *pool, struct bio *bio)
 663 {
 664         return (bio_data_dir(bio) == WRITE) &&
 665                 io_overlaps_block(pool, bio);
 666 }
 667
 668 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
 669                                bio_end_io_t *fn)
 670 {
 671         *save = bio->bi_end_io;
 672         bio->bi_end_io = fn;
 673 }
 674
 675 static int ensure_next_mapping(struct pool *pool)
 676 {
 677         if (pool->next_mapping)
 678                 return 0;
 679
 680         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
 681
 682         return pool->next_mapping ? 0 : -ENOMEM;
 683 }
 684
 685 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
 686 {
 687         struct dm_thin_new_mapping *r = pool->next_mapping;
 688
 689         BUG_ON(!pool->next_mapping);
 690
 691         pool->next_mapping = NULL;
 692
 693         return r;
 694 }
 695
 696 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
 697                           struct dm_dev *origin, dm_block_t data_origin,
 698                           dm_block_t data_dest,
 699                           struct dm_bio_prison_cell *cell, struct bio *bio)
 700 {
 701         int r;
 702         struct pool *pool = tc->pool;
 703         struct dm_thin_new_mapping *m = get_next_mapping(pool);
 704
 705         INIT_LIST_HEAD(&m->list);
 706         m->quiesced = 0;
 707         m->prepared = 0;
 708         m->tc = tc;
 709         m->virt_block = virt_block;
 710         m->data_block = data_dest;
 711         m->cell = cell;
 712         m->err = 0;
 713         m->bio = NULL;
 714
 715         if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
 716                 m->quiesced = 1;
 717
 718         /*
 719          * IO to pool_dev remaps to the pool target's data_dev.
 720          *
 721          * If the whole block of data is being overwritten, we can issue the
 722          * bio immediately. Otherwise we use kcopyd to clone the data first.
 723          */
 724         if (io_overwrites_block(pool, bio)) {
 725                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 726
 727                 h->overwrite_mapping = m;
 728                 m->bio = bio;
 729                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 730                 inc_all_io_entry(pool, bio);
 731                 remap_and_issue(tc, bio, data_dest);
 732         } else {
 733                 struct dm_io_region from, to;
 734
 735                 from.bdev = origin->bdev;
 736                 from.sector = data_origin * pool->sectors_per_block;
 737                 from.count = pool->sectors_per_block;
 738
 739                 to.bdev = tc->pool_dev->bdev;
 740                 to.sector = data_dest * pool->sectors_per_block;
 741                 to.count = pool->sectors_per_block;
 742
 743                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
 744                                    0, copy_complete, m);
 745                 if (r < 0) {
 746                         mempool_free(m, pool->mapping_pool);
 747                         DMERR_LIMIT("dm_kcopyd_copy() failed");
 748                         dm_cell_error(cell);
 749                 }
 750         }
 751 }
 752
 753 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
 754                                    dm_block_t data_origin, dm_block_t data_dest,
 755                                    struct dm_bio_prison_cell *cell, struct bio *bio)
 756 {
 757         schedule_copy(tc, virt_block, tc->pool_dev,
 758                       data_origin, data_dest, cell, bio);
 759 }
 760
 761 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
 762                                    dm_block_t data_dest,
 763                                    struct dm_bio_prison_cell *cell, struct bio *bio)
 764 {
 765         schedule_copy(tc, virt_block, tc->origin_dev,
 766                       virt_block, data_dest, cell, bio);
 767 }
 768
 769 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
 770                           dm_block_t data_block, struct dm_bio_prison_cell *cell,
 771                           struct bio *bio)
 772 {
 773         struct pool *pool = tc->pool;
 774         struct dm_thin_new_mapping *m = get_next_mapping(pool);
 775
 776         INIT_LIST_HEAD(&m->list);
 777         m->quiesced = 1;
 778         m->prepared = 0;
 779         m->tc = tc;
 780         m->virt_block = virt_block;
 781         m->data_block = data_block;
 782         m->cell = cell;
 783         m->err = 0;
 784         m->bio = NULL;
 785
 786         /*
 787          * If the whole block of data is being overwritten or we are not
 788          * zeroing pre-existing data, we can issue the bio immediately.
 789          * Otherwise we use kcopyd to zero the data first.
 790          */
 791         if (!pool->pf.zero_new_blocks)
 792                 process_prepared_mapping(m);
 793
 794         else if (io_overwrites_block(pool, bio)) {
 795                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 796
 797                 h->overwrite_mapping = m;
 798                 m->bio = bio;
 799                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 800                 inc_all_io_entry(pool, bio);
 801                 remap_and_issue(tc, bio, data_block);
 802         } else {
 803                 int r;
 804                 struct dm_io_region to;
 805
 806                 to.bdev = tc->pool_dev->bdev;
 807                 to.sector = data_block * pool->sectors_per_block;
 808                 to.count = pool->sectors_per_block;
 809
 810                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
 811                 if (r < 0) {
 812                         mempool_free(m, pool->mapping_pool);
 813                         DMERR_LIMIT("dm_kcopyd_zero() failed");
 814                         dm_cell_error(cell);
 815                 }
 816         }
 817 }
 818
 819 static int commit(struct pool *pool)
 820 {
 821         int r;
 822
 823         r = dm_pool_commit_metadata(pool->pmd);
 824         if (r)
 825                 DMERR_LIMIT("commit failed: error = %d", r);
 826
 827         return r;
 828 }
 829
 830 /*
 831  * A non-zero return indicates read_only or fail_io mode.
 832  * Many callers don't care about the return value.
 833  */
 834 static int commit_or_fallback(struct pool *pool)
 835 {
 836         int r;
 837
 838         if (get_pool_mode(pool) != PM_WRITE)
 839                 return -EINVAL;
 840
 841         r = commit(pool);
 842         if (r)
 843                 set_pool_mode(pool, PM_READ_ONLY);
 844
 845         return r;
 846 }
 847
 848 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
 849 {
 850         int r;
 851         dm_block_t free_blocks;
 852         unsigned long flags;
 853         struct pool *pool = tc->pool;
 854
 855         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 856         if (r)
 857                 return r;
 858
 859         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
 860                 DMWARN("%s: reached low water mark, sending event.",
 861                        dm_device_name(pool->pool_md));
 862                 spin_lock_irqsave(&pool->lock, flags);
 863                 pool->low_water_triggered = 1;
 864                 spin_unlock_irqrestore(&pool->lock, flags);
 865                 dm_table_event(pool->ti->table);
 866         }
 867
 868         if (!free_blocks) {
 869                 if (pool->no_free_space)
 870                         return -ENOSPC;
 871                 else {
 872                         /*
 873                          * Try to commit to see if that will free up some
 874                          * more space.
 875                          */
 876                         (void) commit_or_fallback(pool);
 877
 878                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 879                         if (r)
 880                                 return r;
 881
 882                         /*
 883                          * If we still have no space we set a flag to avoid
 884                          * doing all this checking and return -ENOSPC.
 885                          */
 886                         if (!free_blocks) {
 887                                 DMWARN("%s: no free space available.",
 888                                        dm_device_name(pool->pool_md));
 889                                 spin_lock_irqsave(&pool->lock, flags);
 890                                 pool->no_free_space = 1;
 891                                 spin_unlock_irqrestore(&pool->lock, flags);
 892                                 return -ENOSPC;
 893                         }
 894                 }
 895         }
 896
 897         r = dm_pool_alloc_data_block(pool->pmd, result);
 898         if (r)
 899                 return r;
 900
 901         return 0;
 902 }
 903
 904 /*
 905  * If we have run out of space, queue bios until the device is
 906  * resumed, presumably after having been reloaded with more space.
 907  */
 908 static void retry_on_resume(struct bio *bio)
 909 {
 910         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 911         struct thin_c *tc = h->tc;
 912         struct pool *pool = tc->pool;
 913         unsigned long flags;
 914
 915         spin_lock_irqsave(&pool->lock, flags);
 916         bio_list_add(&pool->retry_on_resume_list, bio);
 917         spin_unlock_irqrestore(&pool->lock, flags);
 918 }
 919
 920 static void no_space(struct dm_bio_prison_cell *cell)
 921 {
 922         struct bio *bio;
 923         struct bio_list bios;
 924
 925         bio_list_init(&bios);
 926         dm_cell_release(cell, &bios);
 927
 928         while ((bio = bio_list_pop(&bios)))
 929                 retry_on_resume(bio);
 930 }
 931
 932 static void process_discard(struct thin_c *tc, struct bio *bio)
 933 {
 934         int r;
 935         unsigned long flags;
 936         struct pool *pool = tc->pool;
 937         struct dm_bio_prison_cell *cell, *cell2;
 938         struct dm_cell_key key, key2;
 939         dm_block_t block = get_bio_block(tc, bio);
 940         struct dm_thin_lookup_result lookup_result;
 941         struct dm_thin_new_mapping *m;
 942
 943         build_virtual_key(tc->td, block, &key);
 944         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
 945                 return;
 946
 947         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
 948         switch (r) {
 949         case 0:
 950                 /*
 951                  * Check nobody is fiddling with this pool block.  This can
 952                  * happen if someone's in the process of breaking sharing
 953                  * on this block.
 954                  */
 955                 build_data_key(tc->td, lookup_result.block, &key2);
 956                 if (dm_bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
 957                         cell_defer_no_holder(tc, cell);
 958                         break;
 959                 }
 960
 961                 if (io_overlaps_block(pool, bio)) {
 962                         /*
 963                          * IO may still be going to the destination block.  We must
 964                          * quiesce before we can do the removal.
 965                          */
 966                         m = get_next_mapping(pool);
 967                         m->tc = tc;
 968                         m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
 969                         m->virt_block = block;
 970                         m->data_block = lookup_result.block;
 971                         m->cell = cell;
 972                         m->cell2 = cell2;
 973                         m->err = 0;
 974                         m->bio = bio;
 975
 976                         if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
 977                                 spin_lock_irqsave(&pool->lock, flags);
 978                                 list_add(&m->list, &pool->prepared_discards);
 979                                 spin_unlock_irqrestore(&pool->lock, flags);
 980                                 wake_worker(pool);
 981                         }
 982                 } else {
 983                         inc_all_io_entry(pool, bio);
 984                         cell_defer_no_holder(tc, cell);
 985                         cell_defer_no_holder(tc, cell2);
 986
 987                         /*
 988                          * The DM core makes sure that the discard doesn't span
 989                          * a block boundary.  So we submit the discard of a
 990                          * partial block appropriately.
 991                          */
 992                         if ((!lookup_result.shared) && pool->pf.discard_passdown)
 993                                 remap_and_issue(tc, bio, lookup_result.block);
 994                         else
 995                                 bio_endio(bio, 0);
 996                 }
 997                 break;
 998
 999         case -ENODATA:
1000                 /*
1001                  * It isn't provisioned, just forget it.
1002                  */
1003                 cell_defer_no_holder(tc, cell);
1004                 bio_endio(bio, 0);
1005                 break;
1006
1007         default:
1008                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1009                             __func__, r);
1010                 cell_defer_no_holder(tc, cell);
1011                 bio_io_error(bio);
1012                 break;
1013         }
1014 }
1015
1016 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1017                           struct dm_cell_key *key,
1018                           struct dm_thin_lookup_result *lookup_result,
1019                           struct dm_bio_prison_cell *cell)
1020 {
1021         int r;
1022         dm_block_t data_block;
1023
1024         r = alloc_data_block(tc, &data_block);
1025         switch (r) {
1026         case 0:
1027                 schedule_internal_copy(tc, block, lookup_result->block,
1028                                        data_block, cell, bio);
1029                 break;
1030
1031         case -ENOSPC:
1032                 no_space(cell);
1033                 break;
1034
1035         default:
1036                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1037                             __func__, r);
1038                 dm_cell_error(cell);
1039                 break;
1040         }
1041 }
1042
1043 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1044                                dm_block_t block,
1045                                struct dm_thin_lookup_result *lookup_result)
1046 {
1047         struct dm_bio_prison_cell *cell;
1048         struct pool *pool = tc->pool;
1049         struct dm_cell_key key;
1050
1051         /*
1052          * If cell is already occupied, then sharing is already in the process
1053          * of being broken so we have nothing further to do here.
1054          */
1055         build_data_key(tc->td, lookup_result->block, &key);
1056         if (dm_bio_detain(pool->prison, &key, bio, &cell))
1057                 return;
1058
1059         if (bio_data_dir(bio) == WRITE && bio->bi_size)
1060                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1061         else {
1062                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1063
1064                 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1065                 inc_all_io_entry(pool, bio);
1066                 cell_defer_no_holder(tc, cell);
1067
1068                 remap_and_issue(tc, bio, lookup_result->block);
1069         }
1070 }
1071
1072 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1073                             struct dm_bio_prison_cell *cell)
1074 {
1075         int r;
1076         dm_block_t data_block;
1077
1078         /*
1079          * Remap empty bios (flushes) immediately, without provisioning.
1080          */
1081         if (!bio->bi_size) {
1082                 inc_all_io_entry(tc->pool, bio);
1083                 cell_defer_no_holder(tc, cell);
1084
1085                 remap_and_issue(tc, bio, 0);
1086                 return;
1087         }
1088
1089         /*
1090          * Fill read bios with zeroes and complete them immediately.
1091          */
1092         if (bio_data_dir(bio) == READ) {
1093                 zero_fill_bio(bio);
1094                 cell_defer_no_holder(tc, cell);
1095                 bio_endio(bio, 0);
1096                 return;
1097         }
1098
1099         r = alloc_data_block(tc, &data_block);
1100         switch (r) {
1101         case 0:
1102                 if (tc->origin_dev)
1103                         schedule_external_copy(tc, block, data_block, cell, bio);
1104                 else
1105                         schedule_zero(tc, block, data_block, cell, bio);
1106                 break;
1107
1108         case -ENOSPC:
1109                 no_space(cell);
1110                 break;
1111
1112         default:
1113                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1114                             __func__, r);
1115                 set_pool_mode(tc->pool, PM_READ_ONLY);
1116                 dm_cell_error(cell);
1117                 break;
1118         }
1119 }
1120
1121 static void process_bio(struct thin_c *tc, struct bio *bio)
1122 {
1123         int r;
1124         dm_block_t block = get_bio_block(tc, bio);
1125         struct dm_bio_prison_cell *cell;
1126         struct dm_cell_key key;
1127         struct dm_thin_lookup_result lookup_result;
1128
1129         /*
1130          * If cell is already occupied, then the block is already
1131          * being provisioned so we have nothing further to do here.
1132          */
1133         build_virtual_key(tc->td, block, &key);
1134         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
1135                 return;
1136
1137         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1138         switch (r) {
1139         case 0:
1140                 if (lookup_result.shared) {
1141                         process_shared_bio(tc, bio, block, &lookup_result);
1142                         cell_defer_no_holder(tc, cell);
1143                 } else {
1144                         inc_all_io_entry(tc->pool, bio);
1145                         cell_defer_no_holder(tc, cell);
1146
1147                         remap_and_issue(tc, bio, lookup_result.block);
1148                 }
1149                 break;
1150
1151         case -ENODATA:
1152                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1153                         inc_all_io_entry(tc->pool, bio);
1154                         cell_defer_no_holder(tc, cell);
1155
1156                         remap_to_origin_and_issue(tc, bio);
1157                 } else
1158                         provision_block(tc, bio, block, cell);
1159                 break;
1160
1161         default:
1162                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1163                             __func__, r);
1164                 cell_defer_no_holder(tc, cell);
1165                 bio_io_error(bio);
1166                 break;
1167         }
1168 }
1169
1170 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1171 {
1172         int r;
1173         int rw = bio_data_dir(bio);
1174         dm_block_t block = get_bio_block(tc, bio);
1175         struct dm_thin_lookup_result lookup_result;
1176
1177         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1178         switch (r) {
1179         case 0:
1180                 if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1181                         bio_io_error(bio);
1182                 else {
1183                         inc_all_io_entry(tc->pool, bio);
1184                         remap_and_issue(tc, bio, lookup_result.block);
1185                 }
1186                 break;
1187
1188         case -ENODATA:
1189                 if (rw != READ) {
1190                         bio_io_error(bio);
1191                         break;
1192                 }
1193
1194                 if (tc->origin_dev) {
1195                         inc_all_io_entry(tc->pool, bio);
1196                         remap_to_origin_and_issue(tc, bio);
1197                         break;
1198                 }
1199
1200                 zero_fill_bio(bio);
1201                 bio_endio(bio, 0);
1202                 break;
1203
1204         default:
1205                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1206                             __func__, r);
1207                 bio_io_error(bio);
1208                 break;
1209         }
1210 }
1211
1212 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1213 {
1214         bio_io_error(bio);
1215 }
1216
1217 static int need_commit_due_to_time(struct pool *pool)
1218 {
1219         return jiffies < pool->last_commit_jiffies ||
1220                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1221 }
1222
1223 static void process_deferred_bios(struct pool *pool)
1224 {
1225         unsigned long flags;
1226         struct bio *bio;
1227         struct bio_list bios;
1228
1229         bio_list_init(&bios);
1230
1231         spin_lock_irqsave(&pool->lock, flags);
1232         bio_list_merge(&bios, &pool->deferred_bios);
1233         bio_list_init(&pool->deferred_bios);
1234         spin_unlock_irqrestore(&pool->lock, flags);
1235
1236         while ((bio = bio_list_pop(&bios))) {
1237                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1238                 struct thin_c *tc = h->tc;
1239
1240                 /*
1241                  * If we've got no free new_mapping structs, and processing
1242                  * this bio might require one, we pause until there are some
1243                  * prepared mappings to process.
1244                  */
1245                 if (ensure_next_mapping(pool)) {
1246                         spin_lock_irqsave(&pool->lock, flags);
1247                         bio_list_merge(&pool->deferred_bios, &bios);
1248                         spin_unlock_irqrestore(&pool->lock, flags);
1249
1250                         break;
1251                 }
1252
1253                 if (bio->bi_rw & REQ_DISCARD)
1254                         pool->process_discard(tc, bio);
1255                 else
1256                         pool->process_bio(tc, bio);
1257         }
1258
1259         /*
1260          * If there are any deferred flush bios, we must commit
1261          * the metadata before issuing them.
1262          */
1263         bio_list_init(&bios);
1264         spin_lock_irqsave(&pool->lock, flags);
1265         bio_list_merge(&bios, &pool->deferred_flush_bios);
1266         bio_list_init(&pool->deferred_flush_bios);
1267         spin_unlock_irqrestore(&pool->lock, flags);
1268
1269         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1270                 return;
1271
1272         if (commit_or_fallback(pool)) {
1273                 while ((bio = bio_list_pop(&bios)))
1274                         bio_io_error(bio);
1275                 return;
1276         }
1277         pool->last_commit_jiffies = jiffies;
1278
1279         while ((bio = bio_list_pop(&bios)))
1280                 generic_make_request(bio);
1281 }
1282
1283 static void do_worker(struct work_struct *ws)
1284 {
1285         struct pool *pool = container_of(ws, struct pool, worker);
1286
1287         process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1288         process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1289         process_deferred_bios(pool);
1290 }
1291
1292 /*
1293  * We want to commit periodically so that not too much
1294  * unwritten data builds up.
1295  */
1296 static void do_waker(struct work_struct *ws)
1297 {
1298         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1299         wake_worker(pool);
1300         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1301 }
1302
1303 /*----------------------------------------------------------------*/
1304
1305 static enum pool_mode get_pool_mode(struct pool *pool)
1306 {
1307         return pool->pf.mode;
1308 }
1309
1310 static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1311 {
1312         int r;
1313
1314         pool->pf.mode = mode;
1315
1316         switch (mode) {
1317         case PM_FAIL:
1318                 DMERR("switching pool to failure mode");
1319                 pool->process_bio = process_bio_fail;
1320                 pool->process_discard = process_bio_fail;
1321                 pool->process_prepared_mapping = process_prepared_mapping_fail;
1322                 pool->process_prepared_discard = process_prepared_discard_fail;
1323                 break;
1324
1325         case PM_READ_ONLY:
1326                 DMERR("switching pool to read-only mode");
1327                 r = dm_pool_abort_metadata(pool->pmd);
1328                 if (r) {
1329                         DMERR("aborting transaction failed");
1330                         set_pool_mode(pool, PM_FAIL);
1331                 } else {
1332                         dm_pool_metadata_read_only(pool->pmd);
1333                         pool->process_bio = process_bio_read_only;
1334                         pool->process_discard = process_discard;
1335                         pool->process_prepared_mapping = process_prepared_mapping_fail;
1336                         pool->process_prepared_discard = process_prepared_discard_passdown;
1337                 }
1338                 break;
1339
1340         case PM_WRITE:
1341                 pool->process_bio = process_bio;
1342                 pool->process_discard = process_discard;
1343                 pool->process_prepared_mapping = process_prepared_mapping;
1344                 pool->process_prepared_discard = process_prepared_discard;
1345                 break;
1346         }
1347 }
1348
1349 /*----------------------------------------------------------------*/
1350
1351 /*
1352  * Mapping functions.
1353  */
1354
1355 /*
1356  * Called only while mapping a thin bio to hand it over to the workqueue.
1357  */
1358 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1359 {
1360         unsigned long flags;
1361         struct pool *pool = tc->pool;
1362
1363         spin_lock_irqsave(&pool->lock, flags);
1364         bio_list_add(&pool->deferred_bios, bio);
1365         spin_unlock_irqrestore(&pool->lock, flags);
1366
1367         wake_worker(pool);
1368 }
1369
1370 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
1371 {
1372         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1373
1374         h->tc = tc;
1375         h->shared_read_entry = NULL;
1376         h->all_io_entry = NULL;
1377         h->overwrite_mapping = NULL;
1378 }
1379
1380 /*
1381  * Non-blocking function called from the thin target's map function.
1382  */
1383 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
1384 {
1385         int r;
1386         struct thin_c *tc = ti->private;
1387         dm_block_t block = get_bio_block(tc, bio);
1388         struct dm_thin_device *td = tc->td;
1389         struct dm_thin_lookup_result result;
1390         struct dm_bio_prison_cell *cell1, *cell2;
1391         struct dm_cell_key key;
1392
1393         thin_hook_bio(tc, bio);
1394
1395         if (get_pool_mode(tc->pool) == PM_FAIL) {
1396                 bio_io_error(bio);
1397                 return DM_MAPIO_SUBMITTED;
1398         }
1399
1400         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1401                 thin_defer_bio(tc, bio);
1402                 return DM_MAPIO_SUBMITTED;
1403         }
1404
1405         r = dm_thin_find_block(td, block, 0, &result);
1406
1407         /*
1408          * Note that we defer readahead too.
1409          */
1410         switch (r) {
1411         case 0:
1412                 if (unlikely(result.shared)) {
1413                         /*
1414                          * We have a race condition here between the
1415                          * result.shared value returned by the lookup and
1416                          * snapshot creation, which may cause new
1417                          * sharing.
1418                          *
1419                          * To avoid this always quiesce the origin before
1420                          * taking the snap.  You want to do this anyway to
1421                          * ensure a consistent application view
1422                          * (i.e. lockfs).
1423                          *
1424                          * More distant ancestors are irrelevant. The
1425                          * shared flag will be set in their case.
1426                          */
1427                         thin_defer_bio(tc, bio);
1428                         return DM_MAPIO_SUBMITTED;
1429                 }
1430
1431                 build_virtual_key(tc->td, block, &key);
1432                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1))
1433                         return DM_MAPIO_SUBMITTED;
1434
1435                 build_data_key(tc->td, result.block, &key);
1436                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2)) {
1437                         cell_defer_no_holder(tc, cell1);
1438                         return DM_MAPIO_SUBMITTED;
1439                 }
1440
1441                 inc_all_io_entry(tc->pool, bio);
1442                 cell_defer_no_holder(tc, cell2);
1443                 cell_defer_no_holder(tc, cell1);
1444
1445                 remap(tc, bio, result.block);
1446                 return DM_MAPIO_REMAPPED;
1447
1448         case -ENODATA:
1449                 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1450                         /*
1451                          * This block isn't provisioned, and we have no way
1452                          * of doing so.  Just error it.
1453                          */
1454                         bio_io_error(bio);
1455                         return DM_MAPIO_SUBMITTED;
1456                 }
1457                 /* fall through */
1458
1459         case -EWOULDBLOCK:
1460                 /*
1461                  * In future, the failed dm_thin_find_block above could
1462                  * provide the hint to load the metadata into cache.
1463                  */
1464                 thin_defer_bio(tc, bio);
1465                 return DM_MAPIO_SUBMITTED;
1466
1467         default:
1468                 /*
1469                  * Must always call bio_io_error on failure.
1470                  * dm_thin_find_block can fail with -EINVAL if the
1471                  * pool is switched to fail-io mode.
1472                  */
1473                 bio_io_error(bio);
1474                 return DM_MAPIO_SUBMITTED;
1475         }
1476 }
1477
1478 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1479 {
1480         int r;
1481         unsigned long flags;
1482         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1483
1484         spin_lock_irqsave(&pt->pool->lock, flags);
1485         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1486         spin_unlock_irqrestore(&pt->pool->lock, flags);
1487
1488         if (!r) {
1489                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1490                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1491         }
1492
1493         return r;
1494 }
1495
1496 static void __requeue_bios(struct pool *pool)
1497 {
1498         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1499         bio_list_init(&pool->retry_on_resume_list);
1500 }
1501
1502 /*----------------------------------------------------------------
1503  * Binding of control targets to a pool object
1504  *--------------------------------------------------------------*/
1505 static bool data_dev_supports_discard(struct pool_c *pt)
1506 {
1507         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1508
1509         return q && blk_queue_discard(q);
1510 }
1511
1512 /*
1513  * If discard_passdown was enabled verify that the data device
1514  * supports discards.  Disable discard_passdown if not.
1515  */
1516 static void disable_passdown_if_not_supported(struct pool_c *pt)
1517 {
1518         struct pool *pool = pt->pool;
1519         struct block_device *data_bdev = pt->data_dev->bdev;
1520         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1521         sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1522         const char *reason = NULL;
1523         char buf[BDEVNAME_SIZE];
1524
1525         if (!pt->adjusted_pf.discard_passdown)
1526                 return;
1527
1528         if (!data_dev_supports_discard(pt))
1529                 reason = "discard unsupported";
1530
1531         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1532                 reason = "max discard sectors smaller than a block";
1533
1534         else if (data_limits->discard_granularity > block_size)
1535                 reason = "discard granularity larger than a block";
1536
1537         else if (block_size & (data_limits->discard_granularity - 1))
1538                 reason = "discard granularity not a factor of block size";
1539
1540         if (reason) {
1541                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1542                 pt->adjusted_pf.discard_passdown = false;
1543         }
1544 }
1545
1546 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1547 {
1548         struct pool_c *pt = ti->private;
1549
1550         /*
1551          * We want to make sure that degraded pools are never upgraded.
1552          */
1553         enum pool_mode old_mode = pool->pf.mode;
1554         enum pool_mode new_mode = pt->adjusted_pf.mode;
1555
1556         if (old_mode > new_mode)
1557                 new_mode = old_mode;
1558
1559         pool->ti = ti;
1560         pool->low_water_blocks = pt->low_water_blocks;
1561         pool->pf = pt->adjusted_pf;
1562
1563         set_pool_mode(pool, new_mode);
1564
1565         return 0;
1566 }
1567
1568 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1569 {
1570         if (pool->ti == ti)
1571                 pool->ti = NULL;
1572 }
1573
1574 /*----------------------------------------------------------------
1575  * Pool creation
1576  *--------------------------------------------------------------*/
1577 /* Initialize pool features. */
1578 static void pool_features_init(struct pool_features *pf)
1579 {
1580         pf->mode = PM_WRITE;
1581         pf->zero_new_blocks = true;
1582         pf->discard_enabled = true;
1583         pf->discard_passdown = true;
1584 }
1585
1586 static void __pool_destroy(struct pool *pool)
1587 {
1588         __pool_table_remove(pool);
1589
1590         if (dm_pool_metadata_close(pool->pmd) < 0)
1591                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1592
1593         dm_bio_prison_destroy(pool->prison);
1594         dm_kcopyd_client_destroy(pool->copier);
1595
1596         if (pool->wq)
1597                 destroy_workqueue(pool->wq);
1598
1599         if (pool->next_mapping)
1600                 mempool_free(pool->next_mapping, pool->mapping_pool);
1601         mempool_destroy(pool->mapping_pool);
1602         dm_deferred_set_destroy(pool->shared_read_ds);
1603         dm_deferred_set_destroy(pool->all_io_ds);
1604         kfree(pool);
1605 }
1606
1607 static struct kmem_cache *_new_mapping_cache;
1608
1609 static struct pool *pool_create(struct mapped_device *pool_md,
1610                                 struct block_device *metadata_dev,
1611                                 unsigned long block_size,
1612                                 int read_only, char **error)
1613 {
1614         int r;
1615         void *err_p;
1616         struct pool *pool;
1617         struct dm_pool_metadata *pmd;
1618         bool format_device = read_only ? false : true;
1619
1620         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1621         if (IS_ERR(pmd)) {
1622                 *error = "Error creating metadata object";
1623                 return (struct pool *)pmd;
1624         }
1625
1626         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1627         if (!pool) {
1628                 *error = "Error allocating memory for pool";
1629                 err_p = ERR_PTR(-ENOMEM);
1630                 goto bad_pool;
1631         }
1632
1633         pool->pmd = pmd;
1634         pool->sectors_per_block = block_size;
1635         if (block_size & (block_size - 1))
1636                 pool->sectors_per_block_shift = -1;
1637         else
1638                 pool->sectors_per_block_shift = __ffs(block_size);
1639         pool->low_water_blocks = 0;
1640         pool_features_init(&pool->pf);
1641         pool->prison = dm_bio_prison_create(PRISON_CELLS);
1642         if (!pool->prison) {
1643                 *error = "Error creating pool's bio prison";
1644                 err_p = ERR_PTR(-ENOMEM);
1645                 goto bad_prison;
1646         }
1647
1648         pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1649         if (IS_ERR(pool->copier)) {
1650                 r = PTR_ERR(pool->copier);
1651                 *error = "Error creating pool's kcopyd client";
1652                 err_p = ERR_PTR(r);
1653                 goto bad_kcopyd_client;
1654         }
1655
1656         /*
1657          * Create singlethreaded workqueue that will service all devices
1658          * that use this metadata.
1659          */
1660         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1661         if (!pool->wq) {
1662                 *error = "Error creating pool's workqueue";
1663                 err_p = ERR_PTR(-ENOMEM);
1664                 goto bad_wq;
1665         }
1666
1667         INIT_WORK(&pool->worker, do_worker);
1668         INIT_DELAYED_WORK(&pool->waker, do_waker);
1669         spin_lock_init(&pool->lock);
1670         bio_list_init(&pool->deferred_bios);
1671         bio_list_init(&pool->deferred_flush_bios);
1672         INIT_LIST_HEAD(&pool->prepared_mappings);
1673         INIT_LIST_HEAD(&pool->prepared_discards);
1674         pool->low_water_triggered = 0;
1675         pool->no_free_space = 0;
1676         bio_list_init(&pool->retry_on_resume_list);
1677
1678         pool->shared_read_ds = dm_deferred_set_create();
1679         if (!pool->shared_read_ds) {
1680                 *error = "Error creating pool's shared read deferred set";
1681                 err_p = ERR_PTR(-ENOMEM);
1682                 goto bad_shared_read_ds;
1683         }
1684
1685         pool->all_io_ds = dm_deferred_set_create();
1686         if (!pool->all_io_ds) {
1687                 *error = "Error creating pool's all io deferred set";
1688                 err_p = ERR_PTR(-ENOMEM);
1689                 goto bad_all_io_ds;
1690         }
1691
1692         pool->next_mapping = NULL;
1693         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1694                                                       _new_mapping_cache);
1695         if (!pool->mapping_pool) {
1696                 *error = "Error creating pool's mapping mempool";
1697                 err_p = ERR_PTR(-ENOMEM);
1698                 goto bad_mapping_pool;
1699         }
1700
1701         pool->ref_count = 1;
1702         pool->last_commit_jiffies = jiffies;
1703         pool->pool_md = pool_md;
1704         pool->md_dev = metadata_dev;
1705         __pool_table_insert(pool);
1706
1707         return pool;
1708
1709 bad_mapping_pool:
1710         dm_deferred_set_destroy(pool->all_io_ds);
1711 bad_all_io_ds:
1712         dm_deferred_set_destroy(pool->shared_read_ds);
1713 bad_shared_read_ds:
1714         destroy_workqueue(pool->wq);
1715 bad_wq:
1716         dm_kcopyd_client_destroy(pool->copier);
1717 bad_kcopyd_client:
1718         dm_bio_prison_destroy(pool->prison);
1719 bad_prison:
1720         kfree(pool);
1721 bad_pool:
1722         if (dm_pool_metadata_close(pmd))
1723                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1724
1725         return err_p;
1726 }
1727
1728 static void __pool_inc(struct pool *pool)
1729 {
1730         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1731         pool->ref_count++;
1732 }
1733
1734 static void __pool_dec(struct pool *pool)
1735 {
1736         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1737         BUG_ON(!pool->ref_count);
1738         if (!--pool->ref_count)
1739                 __pool_destroy(pool);
1740 }
1741
1742 static struct pool *__pool_find(struct mapped_device *pool_md,
1743                                 struct block_device *metadata_dev,
1744                                 unsigned long block_size, int read_only,
1745                                 char **error, int *created)
1746 {
1747         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1748
1749         if (pool) {
1750                 if (pool->pool_md != pool_md) {
1751                         *error = "metadata device already in use by a pool";
1752                         return ERR_PTR(-EBUSY);
1753                 }
1754                 __pool_inc(pool);
1755
1756         } else {
1757                 pool = __pool_table_lookup(pool_md);
1758                 if (pool) {
1759                         if (pool->md_dev != metadata_dev) {
1760                                 *error = "different pool cannot replace a pool";
1761                                 return ERR_PTR(-EINVAL);
1762                         }
1763                         __pool_inc(pool);
1764
1765                 } else {
1766                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1767                         *created = 1;
1768                 }
1769         }
1770
1771         return pool;
1772 }
1773
1774 /*----------------------------------------------------------------
1775  * Pool target methods
1776  *--------------------------------------------------------------*/
1777 static void pool_dtr(struct dm_target *ti)
1778 {
1779         struct pool_c *pt = ti->private;
1780
1781         mutex_lock(&dm_thin_pool_table.mutex);
1782
1783         unbind_control_target(pt->pool, ti);
1784         __pool_dec(pt->pool);
1785         dm_put_device(ti, pt->metadata_dev);
1786         dm_put_device(ti, pt->data_dev);
1787         kfree(pt);
1788
1789         mutex_unlock(&dm_thin_pool_table.mutex);
1790 }
1791
1792 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1793                                struct dm_target *ti)
1794 {
1795         int r;
1796         unsigned argc;
1797         const char *arg_name;
1798
1799         static struct dm_arg _args[] = {
1800                 {0, 3, "Invalid number of pool feature arguments"},
1801         };
1802
1803         /*
1804          * No feature arguments supplied.
1805          */
1806         if (!as->argc)
1807                 return 0;
1808
1809         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1810         if (r)
1811                 return -EINVAL;
1812
1813         while (argc && !r) {
1814                 arg_name = dm_shift_arg(as);
1815                 argc--;
1816
1817                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
1818                         pf->zero_new_blocks = false;
1819
1820                 else if (!strcasecmp(arg_name, "ignore_discard"))
1821                         pf->discard_enabled = false;
1822
1823                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
1824                         pf->discard_passdown = false;
1825
1826                 else if (!strcasecmp(arg_name, "read_only"))
1827                         pf->mode = PM_READ_ONLY;
1828
1829                 else {
1830                         ti->error = "Unrecognised pool feature requested";
1831                         r = -EINVAL;
1832                         break;
1833                 }
1834         }
1835
1836         return r;
1837 }
1838
1839 /*
1840  * thin-pool <metadata dev> <data dev>
1841  *           <data block size (sectors)>
1842  *           <low water mark (blocks)>
1843  *           [<#feature args> [<arg>]*]
1844  *
1845  * Optional feature arguments are:
1846  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1847  *           ignore_discard: disable discard
1848  *           no_discard_passdown: don't pass discards down to the data device
1849  */
1850 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1851 {
1852         int r, pool_created = 0;
1853         struct pool_c *pt;
1854         struct pool *pool;
1855         struct pool_features pf;
1856         struct dm_arg_set as;
1857         struct dm_dev *data_dev;
1858         unsigned long block_size;
1859         dm_block_t low_water_blocks;
1860         struct dm_dev *metadata_dev;
1861         sector_t metadata_dev_size;
1862         char b[BDEVNAME_SIZE];
1863
1864         /*
1865          * FIXME Remove validation from scope of lock.
1866          */
1867         mutex_lock(&dm_thin_pool_table.mutex);
1868
1869         if (argc < 4) {
1870                 ti->error = "Invalid argument count";
1871                 r = -EINVAL;
1872                 goto out_unlock;
1873         }
1874         as.argc = argc;
1875         as.argv = argv;
1876
1877         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1878         if (r) {
1879                 ti->error = "Error opening metadata block device";
1880                 goto out_unlock;
1881         }
1882
1883         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1884         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1885                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1886                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1887
1888         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1889         if (r) {
1890                 ti->error = "Error getting data device";
1891                 goto out_metadata;
1892         }
1893
1894         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1895             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1896             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1897             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1898                 ti->error = "Invalid block size";
1899                 r = -EINVAL;
1900                 goto out;
1901         }
1902
1903         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1904                 ti->error = "Invalid low water mark";
1905                 r = -EINVAL;
1906                 goto out;
1907         }
1908
1909         /*
1910          * Set default pool features.
1911          */
1912         pool_features_init(&pf);
1913
1914         dm_consume_args(&as, 4);
1915         r = parse_pool_features(&as, &pf, ti);
1916         if (r)
1917                 goto out;
1918
1919         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1920         if (!pt) {
1921                 r = -ENOMEM;
1922                 goto out;
1923         }
1924
1925         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1926                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
1927         if (IS_ERR(pool)) {
1928                 r = PTR_ERR(pool);
1929                 goto out_free_pt;
1930         }
1931
1932         /*
1933          * 'pool_created' reflects whether this is the first table load.
1934          * Top level discard support is not allowed to be changed after
1935          * initial load.  This would require a pool reload to trigger thin
1936          * device changes.
1937          */
1938         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1939                 ti->error = "Discard support cannot be disabled once enabled";
1940                 r = -EINVAL;
1941                 goto out_flags_changed;
1942         }
1943
1944         pt->pool = pool;
1945         pt->ti = ti;
1946         pt->metadata_dev = metadata_dev;
1947         pt->data_dev = data_dev;
1948         pt->low_water_blocks = low_water_blocks;
1949         pt->adjusted_pf = pt->requested_pf = pf;
1950         ti->num_flush_bios = 1;
1951
1952         /*
1953          * Only need to enable discards if the pool should pass
1954          * them down to the data device.  The thin device's discard
1955          * processing will cause mappings to be removed from the btree.
1956          */
1957         if (pf.discard_enabled && pf.discard_passdown) {
1958                 ti->num_discard_bios = 1;
1959
1960                 /*
1961                  * Setting 'discards_supported' circumvents the normal
1962                  * stacking of discard limits (this keeps the pool and
1963                  * thin devices' discard limits consistent).
1964                  */
1965                 ti->discards_supported = true;
1966                 ti->discard_zeroes_data_unsupported = true;
1967         }
1968         ti->private = pt;
1969
1970         pt->callbacks.congested_fn = pool_is_congested;
1971         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1972
1973         mutex_unlock(&dm_thin_pool_table.mutex);
1974
1975         return 0;
1976
1977 out_flags_changed:
1978         __pool_dec(pool);
1979 out_free_pt:
1980         kfree(pt);
1981 out:
1982         dm_put_device(ti, data_dev);
1983 out_metadata:
1984         dm_put_device(ti, metadata_dev);
1985 out_unlock:
1986         mutex_unlock(&dm_thin_pool_table.mutex);
1987
1988         return r;
1989 }
1990
1991 static int pool_map(struct dm_target *ti, struct bio *bio)
1992 {
1993         int r;
1994         struct pool_c *pt = ti->private;
1995         struct pool *pool = pt->pool;
1996         unsigned long flags;
1997
1998         /*
1999          * As this is a singleton target, ti->begin is always zero.
2000          */
2001         spin_lock_irqsave(&pool->lock, flags);
2002         bio->bi_bdev = pt->data_dev->bdev;
2003         r = DM_MAPIO_REMAPPED;
2004         spin_unlock_irqrestore(&pool->lock, flags);
2005
2006         return r;
2007 }
2008
2009 /*
2010  * Retrieves the number of blocks of the data device from
2011  * the superblock and compares it to the actual device size,
2012  * thus resizing the data device in case it has grown.
2013  *
2014  * This both copes with opening preallocated data devices in the ctr
2015  * being followed by a resume
2016  * -and-
2017  * calling the resume method individually after userspace has
2018  * grown the data device in reaction to a table event.
2019  */
2020 static int pool_preresume(struct dm_target *ti)
2021 {
2022         int r;
2023         struct pool_c *pt = ti->private;
2024         struct pool *pool = pt->pool;
2025         sector_t data_size = ti->len;
2026         dm_block_t sb_data_size;
2027
2028         /*
2029          * Take control of the pool object.
2030          */
2031         r = bind_control_target(pool, ti);
2032         if (r)
2033                 return r;
2034
2035         (void) sector_div(data_size, pool->sectors_per_block);
2036
2037         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2038         if (r) {
2039                 DMERR("failed to retrieve data device size");
2040                 return r;
2041         }
2042
2043         if (data_size < sb_data_size) {
2044                 DMERR("pool target too small, is %llu blocks (expected %llu)",
2045                       (unsigned long long)data_size, sb_data_size);
2046                 return -EINVAL;
2047
2048         } else if (data_size > sb_data_size) {
2049                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2050                 if (r) {
2051                         DMERR("failed to resize data device");
2052                         /* FIXME Stricter than necessary: Rollback transaction instead here */
2053                         set_pool_mode(pool, PM_READ_ONLY);
2054                         return r;
2055                 }
2056
2057                 (void) commit_or_fallback(pool);
2058         }
2059
2060         return 0;
2061 }
2062
2063 static void pool_resume(struct dm_target *ti)
2064 {
2065         struct pool_c *pt = ti->private;
2066         struct pool *pool = pt->pool;
2067         unsigned long flags;
2068
2069         spin_lock_irqsave(&pool->lock, flags);
2070         pool->low_water_triggered = 0;
2071         pool->no_free_space = 0;
2072         __requeue_bios(pool);
2073         spin_unlock_irqrestore(&pool->lock, flags);
2074
2075         do_waker(&pool->waker.work);
2076 }
2077
2078 static void pool_postsuspend(struct dm_target *ti)
2079 {
2080         struct pool_c *pt = ti->private;
2081         struct pool *pool = pt->pool;
2082
2083         cancel_delayed_work(&pool->waker);
2084         flush_workqueue(pool->wq);
2085         (void) commit_or_fallback(pool);
2086 }
2087
2088 static int check_arg_count(unsigned argc, unsigned args_required)
2089 {
2090         if (argc != args_required) {
2091                 DMWARN("Message received with %u arguments instead of %u.",
2092                        argc, args_required);
2093                 return -EINVAL;
2094         }
2095
2096         return 0;
2097 }
2098
2099 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2100 {
2101         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2102             *dev_id <= MAX_DEV_ID)
2103                 return 0;
2104
2105         if (warning)
2106                 DMWARN("Message received with invalid device id: %s", arg);
2107
2108         return -EINVAL;
2109 }
2110
2111 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2112 {
2113         dm_thin_id dev_id;
2114         int r;
2115
2116         r = check_arg_count(argc, 2);
2117         if (r)
2118                 return r;
2119
2120         r = read_dev_id(argv[1], &dev_id, 1);
2121         if (r)
2122                 return r;
2123
2124         r = dm_pool_create_thin(pool->pmd, dev_id);
2125         if (r) {
2126                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2127                        argv[1]);
2128                 return r;
2129         }
2130
2131         return 0;
2132 }
2133
2134 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2135 {
2136         dm_thin_id dev_id;
2137         dm_thin_id origin_dev_id;
2138         int r;
2139
2140         r = check_arg_count(argc, 3);
2141         if (r)
2142                 return r;
2143
2144         r = read_dev_id(argv[1], &dev_id, 1);
2145         if (r)
2146                 return r;
2147
2148         r = read_dev_id(argv[2], &origin_dev_id, 1);
2149         if (r)
2150                 return r;
2151
2152         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2153         if (r) {
2154                 DMWARN("Creation of new snapshot %s of device %s failed.",
2155                        argv[1], argv[2]);
2156                 return r;
2157         }
2158
2159         return 0;
2160 }
2161
2162 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2163 {
2164         dm_thin_id dev_id;
2165         int r;
2166
2167         r = check_arg_count(argc, 2);
2168         if (r)
2169                 return r;
2170
2171         r = read_dev_id(argv[1], &dev_id, 1);
2172         if (r)
2173                 return r;
2174
2175         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2176         if (r)
2177                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2178
2179         return r;
2180 }
2181
2182 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2183 {
2184         dm_thin_id old_id, new_id;
2185         int r;
2186
2187         r = check_arg_count(argc, 3);
2188         if (r)
2189                 return r;
2190
2191         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2192                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2193                 return -EINVAL;
2194         }
2195
2196         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2197                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2198                 return -EINVAL;
2199         }
2200
2201         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2202         if (r) {
2203                 DMWARN("Failed to change transaction id from %s to %s.",
2204                        argv[1], argv[2]);
2205                 return r;
2206         }
2207
2208         return 0;
2209 }
2210
2211 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2212 {
2213         int r;
2214
2215         r = check_arg_count(argc, 1);
2216         if (r)
2217                 return r;
2218
2219         (void) commit_or_fallback(pool);
2220
2221         r = dm_pool_reserve_metadata_snap(pool->pmd);
2222         if (r)
2223                 DMWARN("reserve_metadata_snap message failed.");
2224
2225         return r;
2226 }
2227
2228 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2229 {
2230         int r;
2231
2232         r = check_arg_count(argc, 1);
2233         if (r)
2234                 return r;
2235
2236         r = dm_pool_release_metadata_snap(pool->pmd);
2237         if (r)
2238                 DMWARN("release_metadata_snap message failed.");
2239
2240         return r;
2241 }
2242
2243 /*
2244  * Messages supported:
2245  *   create_thin        <dev_id>
2246  *   create_snap        <dev_id> <origin_id>
2247  *   delete             <dev_id>
2248  *   trim               <dev_id> <new_size_in_sectors>
2249  *   set_transaction_id <current_trans_id> <new_trans_id>
2250  *   reserve_metadata_snap
2251  *   release_metadata_snap
2252  */
2253 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2254 {
2255         int r = -EINVAL;
2256         struct pool_c *pt = ti->private;
2257         struct pool *pool = pt->pool;
2258
2259         if (!strcasecmp(argv[0], "create_thin"))
2260                 r = process_create_thin_mesg(argc, argv, pool);
2261
2262         else if (!strcasecmp(argv[0], "create_snap"))
2263                 r = process_create_snap_mesg(argc, argv, pool);
2264
2265         else if (!strcasecmp(argv[0], "delete"))
2266                 r = process_delete_mesg(argc, argv, pool);
2267
2268         else if (!strcasecmp(argv[0], "set_transaction_id"))
2269                 r = process_set_transaction_id_mesg(argc, argv, pool);
2270
2271         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2272                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2273
2274         else if (!strcasecmp(argv[0], "release_metadata_snap"))
2275                 r = process_release_metadata_snap_mesg(argc, argv, pool);
2276
2277         else
2278                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2279
2280         if (!r)
2281                 (void) commit_or_fallback(pool);
2282
2283         return r;
2284 }
2285
2286 static void emit_flags(struct pool_features *pf, char *result,
2287                        unsigned sz, unsigned maxlen)
2288 {
2289         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2290                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2291         DMEMIT("%u ", count);
2292
2293         if (!pf->zero_new_blocks)
2294                 DMEMIT("skip_block_zeroing ");
2295
2296         if (!pf->discard_enabled)
2297                 DMEMIT("ignore_discard ");
2298
2299         if (!pf->discard_passdown)
2300                 DMEMIT("no_discard_passdown ");
2301
2302         if (pf->mode == PM_READ_ONLY)
2303                 DMEMIT("read_only ");
2304 }
2305
2306 /*
2307  * Status line is:
2308  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2309  *    <used data sectors>/<total data sectors> <held metadata root>
2310  */
2311 static void pool_status(struct dm_target *ti, status_type_t type,
2312                         unsigned status_flags, char *result, unsigned maxlen)
2313 {
2314         int r;
2315         unsigned sz = 0;
2316         uint64_t transaction_id;
2317         dm_block_t nr_free_blocks_data;
2318         dm_block_t nr_free_blocks_metadata;
2319         dm_block_t nr_blocks_data;
2320         dm_block_t nr_blocks_metadata;
2321         dm_block_t held_root;
2322         char buf[BDEVNAME_SIZE];
2323         char buf2[BDEVNAME_SIZE];
2324         struct pool_c *pt = ti->private;
2325         struct pool *pool = pt->pool;
2326
2327         switch (type) {
2328         case STATUSTYPE_INFO:
2329                 if (get_pool_mode(pool) == PM_FAIL) {
2330                         DMEMIT("Fail");
2331                         break;
2332                 }
2333
2334                 /* Commit to ensure statistics aren't out-of-date */
2335                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2336                         (void) commit_or_fallback(pool);
2337
2338                 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
2339                 if (r) {
2340                         DMERR("dm_pool_get_metadata_transaction_id returned %d", r);
2341                         goto err;
2342                 }
2343
2344                 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
2345                 if (r) {
2346                         DMERR("dm_pool_get_free_metadata_block_count returned %d", r);
2347                         goto err;
2348                 }
2349
2350                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2351                 if (r) {
2352                         DMERR("dm_pool_get_metadata_dev_size returned %d", r);
2353                         goto err;
2354                 }
2355
2356                 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
2357                 if (r) {
2358                         DMERR("dm_pool_get_free_block_count returned %d", r);
2359                         goto err;
2360                 }
2361
2362                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2363                 if (r) {
2364                         DMERR("dm_pool_get_data_dev_size returned %d", r);
2365                         goto err;
2366                 }
2367
2368                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2369                 if (r) {
2370                         DMERR("dm_pool_get_metadata_snap returned %d", r);
2371                         goto err;
2372                 }
2373
2374                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2375                        (unsigned long long)transaction_id,
2376                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2377                        (unsigned long long)nr_blocks_metadata,
2378                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2379                        (unsigned long long)nr_blocks_data);
2380
2381                 if (held_root)
2382                         DMEMIT("%llu ", held_root);
2383                 else
2384                         DMEMIT("- ");
2385
2386                 if (pool->pf.mode == PM_READ_ONLY)
2387                         DMEMIT("ro ");
2388                 else
2389                         DMEMIT("rw ");
2390
2391                 if (!pool->pf.discard_enabled)
2392                         DMEMIT("ignore_discard");
2393                 else if (pool->pf.discard_passdown)
2394                         DMEMIT("discard_passdown");
2395                 else
2396                         DMEMIT("no_discard_passdown");
2397
2398                 break;
2399
2400         case STATUSTYPE_TABLE:
2401                 DMEMIT("%s %s %lu %llu ",
2402                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2403                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2404                        (unsigned long)pool->sectors_per_block,
2405                        (unsigned long long)pt->low_water_blocks);
2406                 emit_flags(&pt->requested_pf, result, sz, maxlen);
2407                 break;
2408         }
2409         return;
2410
2411 err:
2412         DMEMIT("Error");
2413 }
2414
2415 static int pool_iterate_devices(struct dm_target *ti,
2416                                 iterate_devices_callout_fn fn, void *data)
2417 {
2418         struct pool_c *pt = ti->private;
2419
2420         return fn(ti, pt->data_dev, 0, ti->len, data);
2421 }
2422
2423 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2424                       struct bio_vec *biovec, int max_size)
2425 {
2426         struct pool_c *pt = ti->private;
2427         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2428
2429         if (!q->merge_bvec_fn)
2430                 return max_size;
2431
2432         bvm->bi_bdev = pt->data_dev->bdev;
2433
2434         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2435 }
2436
2437 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2438 {
2439         struct pool *pool = pt->pool;
2440         struct queue_limits *data_limits;
2441
2442         limits->max_discard_sectors = pool->sectors_per_block;
2443
2444         /*
2445          * discard_granularity is just a hint, and not enforced.
2446          */
2447         if (pt->adjusted_pf.discard_passdown) {
2448                 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2449                 limits->discard_granularity = data_limits->discard_granularity;
2450         } else
2451                 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2452 }
2453
2454 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2455 {
2456         struct pool_c *pt = ti->private;
2457         struct pool *pool = pt->pool;
2458
2459         blk_limits_io_min(limits, 0);
2460         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2461
2462         /*
2463          * pt->adjusted_pf is a staging area for the actual features to use.
2464          * They get transferred to the live pool in bind_control_target()
2465          * called from pool_preresume().
2466          */
2467         if (!pt->adjusted_pf.discard_enabled)
2468                 return;
2469
2470         disable_passdown_if_not_supported(pt);
2471
2472         set_discard_limits(pt, limits);
2473 }
2474
2475 static struct target_type pool_target = {
2476         .name = "thin-pool",
2477         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2478                     DM_TARGET_IMMUTABLE,
2479         .version = {1, 6, 1},
2480         .module = THIS_MODULE,
2481         .ctr = pool_ctr,
2482         .dtr = pool_dtr,
2483         .map = pool_map,
2484         .postsuspend = pool_postsuspend,
2485         .preresume = pool_preresume,
2486         .resume = pool_resume,
2487         .message = pool_message,
2488         .status = pool_status,
2489         .merge = pool_merge,
2490         .iterate_devices = pool_iterate_devices,
2491         .io_hints = pool_io_hints,
2492 };
2493
2494 /*----------------------------------------------------------------
2495  * Thin target methods
2496  *--------------------------------------------------------------*/
2497 static void thin_dtr(struct dm_target *ti)
2498 {
2499         struct thin_c *tc = ti->private;
2500
2501         mutex_lock(&dm_thin_pool_table.mutex);
2502
2503         __pool_dec(tc->pool);
2504         dm_pool_close_thin_device(tc->td);
2505         dm_put_device(ti, tc->pool_dev);
2506         if (tc->origin_dev)
2507                 dm_put_device(ti, tc->origin_dev);
2508         kfree(tc);
2509
2510         mutex_unlock(&dm_thin_pool_table.mutex);
2511 }
2512
2513 /*
2514  * Thin target parameters:
2515  *
2516  * <pool_dev> <dev_id> [origin_dev]
2517  *
2518  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2519  * dev_id: the internal device identifier
2520  * origin_dev: a device external to the pool that should act as the origin
2521  *
2522  * If the pool device has discards disabled, they get disabled for the thin
2523  * device as well.
2524  */
2525 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2526 {
2527         int r;
2528         struct thin_c *tc;
2529         struct dm_dev *pool_dev, *origin_dev;
2530         struct mapped_device *pool_md;
2531
2532         mutex_lock(&dm_thin_pool_table.mutex);
2533
2534         if (argc != 2 && argc != 3) {
2535                 ti->error = "Invalid argument count";
2536                 r = -EINVAL;
2537                 goto out_unlock;
2538         }
2539
2540         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2541         if (!tc) {
2542                 ti->error = "Out of memory";
2543                 r = -ENOMEM;
2544                 goto out_unlock;
2545         }
2546
2547         if (argc == 3) {
2548                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2549                 if (r) {
2550                         ti->error = "Error opening origin device";
2551                         goto bad_origin_dev;
2552                 }
2553                 tc->origin_dev = origin_dev;
2554         }
2555
2556         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2557         if (r) {
2558                 ti->error = "Error opening pool device";
2559                 goto bad_pool_dev;
2560         }
2561         tc->pool_dev = pool_dev;
2562
2563         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2564                 ti->error = "Invalid device id";
2565                 r = -EINVAL;
2566                 goto bad_common;
2567         }
2568
2569         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2570         if (!pool_md) {
2571                 ti->error = "Couldn't get pool mapped device";
2572                 r = -EINVAL;
2573                 goto bad_common;
2574         }
2575
2576         tc->pool = __pool_table_lookup(pool_md);
2577         if (!tc->pool) {
2578                 ti->error = "Couldn't find pool object";
2579                 r = -EINVAL;
2580                 goto bad_pool_lookup;
2581         }
2582         __pool_inc(tc->pool);
2583
2584         if (get_pool_mode(tc->pool) == PM_FAIL) {
2585                 ti->error = "Couldn't open thin device, Pool is in fail mode";
2586                 goto bad_thin_open;
2587         }
2588
2589         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2590         if (r) {
2591                 ti->error = "Couldn't open thin internal device";
2592                 goto bad_thin_open;
2593         }
2594
2595         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2596         if (r)
2597                 goto bad_thin_open;
2598
2599         ti->num_flush_bios = 1;
2600         ti->flush_supported = true;
2601         ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
2602
2603         /* In case the pool supports discards, pass them on. */
2604         if (tc->pool->pf.discard_enabled) {
2605                 ti->discards_supported = true;
2606                 ti->num_discard_bios = 1;
2607                 ti->discard_zeroes_data_unsupported = true;
2608                 /* Discard bios must be split on a block boundary */
2609                 ti->split_discard_bios = true;
2610         }
2611
2612         dm_put(pool_md);
2613
2614         mutex_unlock(&dm_thin_pool_table.mutex);
2615
2616         return 0;
2617
2618 bad_thin_open:
2619         __pool_dec(tc->pool);
2620 bad_pool_lookup:
2621         dm_put(pool_md);
2622 bad_common:
2623         dm_put_device(ti, tc->pool_dev);
2624 bad_pool_dev:
2625         if (tc->origin_dev)
2626                 dm_put_device(ti, tc->origin_dev);
2627 bad_origin_dev:
2628         kfree(tc);
2629 out_unlock:
2630         mutex_unlock(&dm_thin_pool_table.mutex);
2631
2632         return r;
2633 }
2634
2635 static int thin_map(struct dm_target *ti, struct bio *bio)
2636 {
2637         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2638
2639         return thin_bio_map(ti, bio);
2640 }
2641
2642 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
2643 {
2644         unsigned long flags;
2645         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2646         struct list_head work;
2647         struct dm_thin_new_mapping *m, *tmp;
2648         struct pool *pool = h->tc->pool;
2649
2650         if (h->shared_read_entry) {
2651                 INIT_LIST_HEAD(&work);
2652                 dm_deferred_entry_dec(h->shared_read_entry, &work);
2653
2654                 spin_lock_irqsave(&pool->lock, flags);
2655                 list_for_each_entry_safe(m, tmp, &work, list) {
2656                         list_del(&m->list);
2657                         m->quiesced = 1;
2658                         __maybe_add_mapping(m);
2659                 }
2660                 spin_unlock_irqrestore(&pool->lock, flags);
2661         }
2662
2663         if (h->all_io_entry) {
2664                 INIT_LIST_HEAD(&work);
2665                 dm_deferred_entry_dec(h->all_io_entry, &work);
2666                 if (!list_empty(&work)) {
2667                         spin_lock_irqsave(&pool->lock, flags);
2668                         list_for_each_entry_safe(m, tmp, &work, list)
2669                                 list_add(&m->list, &pool->prepared_discards);
2670                         spin_unlock_irqrestore(&pool->lock, flags);
2671                         wake_worker(pool);
2672                 }
2673         }
2674
2675         return 0;
2676 }
2677
2678 static void thin_postsuspend(struct dm_target *ti)
2679 {
2680         if (dm_noflush_suspending(ti))
2681                 requeue_io((struct thin_c *)ti->private);
2682 }
2683
2684 /*
2685  * <nr mapped sectors> <highest mapped sector>
2686  */
2687 static void thin_status(struct dm_target *ti, status_type_t type,
2688                         unsigned status_flags, char *result, unsigned maxlen)
2689 {
2690         int r;
2691         ssize_t sz = 0;
2692         dm_block_t mapped, highest;
2693         char buf[BDEVNAME_SIZE];
2694         struct thin_c *tc = ti->private;
2695
2696         if (get_pool_mode(tc->pool) == PM_FAIL) {
2697                 DMEMIT("Fail");
2698                 return;
2699         }
2700
2701         if (!tc->td)
2702                 DMEMIT("-");
2703         else {
2704                 switch (type) {
2705                 case STATUSTYPE_INFO:
2706                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2707                         if (r) {
2708                                 DMERR("dm_thin_get_mapped_count returned %d", r);
2709                                 goto err;
2710                         }
2711
2712                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2713                         if (r < 0) {
2714                                 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
2715                                 goto err;
2716                         }
2717
2718                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2719                         if (r)
2720                                 DMEMIT("%llu", ((highest + 1) *
2721                                                 tc->pool->sectors_per_block) - 1);
2722                         else
2723                                 DMEMIT("-");
2724                         break;
2725
2726                 case STATUSTYPE_TABLE:
2727                         DMEMIT("%s %lu",
2728                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2729                                (unsigned long) tc->dev_id);
2730                         if (tc->origin_dev)
2731                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2732                         break;
2733                 }
2734         }
2735
2736         return;
2737
2738 err:
2739         DMEMIT("Error");
2740 }
2741
2742 static int thin_iterate_devices(struct dm_target *ti,
2743                                 iterate_devices_callout_fn fn, void *data)
2744 {
2745         sector_t blocks;
2746         struct thin_c *tc = ti->private;
2747         struct pool *pool = tc->pool;
2748
2749         /*
2750          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2751          * we follow a more convoluted path through to the pool's target.
2752          */
2753         if (!pool->ti)
2754                 return 0;       /* nothing is bound */
2755
2756         blocks = pool->ti->len;
2757         (void) sector_div(blocks, pool->sectors_per_block);
2758         if (blocks)
2759                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2760
2761         return 0;
2762 }
2763
2764 static struct target_type thin_target = {
2765         .name = "thin",
2766         .version = {1, 7, 1},
2767         .module = THIS_MODULE,
2768         .ctr = thin_ctr,
2769         .dtr = thin_dtr,
2770         .map = thin_map,
2771         .end_io = thin_endio,
2772         .postsuspend = thin_postsuspend,
2773         .status = thin_status,
2774         .iterate_devices = thin_iterate_devices,
2775 };
2776
2777 /*----------------------------------------------------------------*/
2778
2779 static int __init dm_thin_init(void)
2780 {
2781         int r;
2782
2783         pool_table_init();
2784
2785         r = dm_register_target(&thin_target);
2786         if (r)
2787                 return r;
2788
2789         r = dm_register_target(&pool_target);
2790         if (r)
2791                 goto bad_pool_target;
2792
2793         r = -ENOMEM;
2794
2795         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2796         if (!_new_mapping_cache)
2797                 goto bad_new_mapping_cache;
2798
2799         return 0;
2800
2801 bad_new_mapping_cache:
2802         dm_unregister_target(&pool_target);
2803 bad_pool_target:
2804         dm_unregister_target(&thin_target);
2805
2806         return r;
2807 }
2808
2809 static void dm_thin_exit(void)
2810 {
2811         dm_unregister_target(&thin_target);
2812         dm_unregister_target(&pool_target);
2813
2814         kmem_cache_destroy(_new_mapping_cache);
2815 }
2816
2817 module_init(dm_thin_init);
2818 module_exit(dm_thin_exit);
2819
2820 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2821 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2822 MODULE_LICENSE("GPL");