drivers/md/dm-thin.c

   1 /*
   2  * Copyright (C) 2011-2012 Red Hat UK.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-thin-metadata.h"
   8 #include "dm-bio-prison.h"
   9 #include "dm.h"
  10
  11 #include <linux/device-mapper.h>
  12 #include <linux/dm-io.h>
  13 #include <linux/dm-kcopyd.h>
  14 #include <linux/list.h>
  15 #include <linux/init.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18
  19 #define DM_MSG_PREFIX   "thin"
  20
  21 /*
  22  * Tunable constants
  23  */
  24 #define ENDIO_HOOK_POOL_SIZE 1024
  25 #define MAPPING_POOL_SIZE 1024
  26 #define PRISON_CELLS 1024
  27 #define COMMIT_PERIOD HZ
  28
  29 /*
  30  * The block size of the device holding pool data must be
  31  * between 64KB and 1GB.
  32  */
  33 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
  34 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
  35
  36 /*
  37  * Device id is restricted to 24 bits.
  38  */
  39 #define MAX_DEV_ID ((1 << 24) - 1)
  40
  41 /*
  42  * How do we handle breaking sharing of data blocks?
  43  * =================================================
  44  *
  45  * We use a standard copy-on-write btree to store the mappings for the
  46  * devices (note I'm talking about copy-on-write of the metadata here, not
  47  * the data).  When you take an internal snapshot you clone the root node
  48  * of the origin btree.  After this there is no concept of an origin or a
  49  * snapshot.  They are just two device trees that happen to point to the
  50  * same data blocks.
  51  *
  52  * When we get a write in we decide if it's to a shared data block using
  53  * some timestamp magic.  If it is, we have to break sharing.
  54  *
  55  * Let's say we write to a shared block in what was the origin.  The
  56  * steps are:
  57  *
  58  * i) plug io further to this physical block. (see bio_prison code).
  59  *
  60  * ii) quiesce any read io to that shared data block.  Obviously
  61  * including all devices that share this block.  (see dm_deferred_set code)
  62  *
  63  * iii) copy the data block to a newly allocate block.  This step can be
  64  * missed out if the io covers the block. (schedule_copy).
  65  *
  66  * iv) insert the new mapping into the origin's btree
  67  * (process_prepared_mapping).  This act of inserting breaks some
  68  * sharing of btree nodes between the two devices.  Breaking sharing only
  69  * effects the btree of that specific device.  Btrees for the other
  70  * devices that share the block never change.  The btree for the origin
  71  * device as it was after the last commit is untouched, ie. we're using
  72  * persistent data structures in the functional programming sense.
  73  *
  74  * v) unplug io to this physical block, including the io that triggered
  75  * the breaking of sharing.
  76  *
  77  * Steps (ii) and (iii) occur in parallel.
  78  *
  79  * The metadata _doesn't_ need to be committed before the io continues.  We
  80  * get away with this because the io is always written to a _new_ block.
  81  * If there's a crash, then:
  82  *
  83  * - The origin mapping will point to the old origin block (the shared
  84  * one).  This will contain the data as it was before the io that triggered
  85  * the breaking of sharing came in.
  86  *
  87  * - The snap mapping still points to the old block.  As it would after
  88  * the commit.
  89  *
  90  * The downside of this scheme is the timestamp magic isn't perfect, and
  91  * will continue to think that data block in the snapshot device is shared
  92  * even after the write to the origin has broken sharing.  I suspect data
  93  * blocks will typically be shared by many different devices, so we're
  94  * breaking sharing n + 1 times, rather than n, where n is the number of
  95  * devices that reference this data block.  At the moment I think the
  96  * benefits far, far outweigh the disadvantages.
  97  */
  98
  99 /*----------------------------------------------------------------*/
 100
 101 /*
 102  * Key building.
 103  */
 104 static void build_data_key(struct dm_thin_device *td,
 105                            dm_block_t b, struct dm_cell_key *key)
 106 {
 107         key->virtual = 0;
 108         key->dev = dm_thin_dev_id(td);
 109         key->block = b;
 110 }
 111
 112 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
 113                               struct dm_cell_key *key)
 114 {
 115         key->virtual = 1;
 116         key->dev = dm_thin_dev_id(td);
 117         key->block = b;
 118 }
 119
 120 /*----------------------------------------------------------------*/
 121
 122 /*
 123  * A pool device ties together a metadata device and a data device.  It
 124  * also provides the interface for creating and destroying internal
 125  * devices.
 126  */
 127 struct dm_thin_new_mapping;
 128
 129 /*
 130  * The pool runs in 3 modes.  Ordered in degraded order for comparisons.
 131  */
 132 enum pool_mode {
 133         PM_WRITE,               /* metadata may be changed */
 134         PM_READ_ONLY,           /* metadata may not be changed */
 135         PM_FAIL,                /* all I/O fails */
 136 };
 137
 138 struct pool_features {
 139         enum pool_mode mode;
 140
 141         bool zero_new_blocks:1;
 142         bool discard_enabled:1;
 143         bool discard_passdown:1;
 144 };
 145
 146 struct thin_c;
 147 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
 148 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
 149
 150 struct pool {
 151         struct list_head list;
 152         struct dm_target *ti;   /* Only set if a pool target is bound */
 153
 154         struct mapped_device *pool_md;
 155         struct block_device *md_dev;
 156         struct dm_pool_metadata *pmd;
 157
 158         dm_block_t low_water_blocks;
 159         uint32_t sectors_per_block;
 160         int sectors_per_block_shift;
 161
 162         struct pool_features pf;
 163         unsigned low_water_triggered:1; /* A dm event has been sent */
 164         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
 165
 166         struct dm_bio_prison *prison;
 167         struct dm_kcopyd_client *copier;
 168
 169         struct workqueue_struct *wq;
 170         struct work_struct worker;
 171         struct delayed_work waker;
 172
 173         unsigned long last_commit_jiffies;
 174         unsigned ref_count;
 175
 176         spinlock_t lock;
 177         struct bio_list deferred_bios;
 178         struct bio_list deferred_flush_bios;
 179         struct list_head prepared_mappings;
 180         struct list_head prepared_discards;
 181
 182         struct bio_list retry_on_resume_list;
 183
 184         struct dm_deferred_set *shared_read_ds;
 185         struct dm_deferred_set *all_io_ds;
 186
 187         struct dm_thin_new_mapping *next_mapping;
 188         mempool_t *mapping_pool;
 189         mempool_t *endio_hook_pool;
 190
 191         process_bio_fn process_bio;
 192         process_bio_fn process_discard;
 193
 194         process_mapping_fn process_prepared_mapping;
 195         process_mapping_fn process_prepared_discard;
 196 };
 197
 198 static enum pool_mode get_pool_mode(struct pool *pool);
 199 static void set_pool_mode(struct pool *pool, enum pool_mode mode);
 200
 201 /*
 202  * Target context for a pool.
 203  */
 204 struct pool_c {
 205         struct dm_target *ti;
 206         struct pool *pool;
 207         struct dm_dev *data_dev;
 208         struct dm_dev *metadata_dev;
 209         struct dm_target_callbacks callbacks;
 210
 211         dm_block_t low_water_blocks;
 212         struct pool_features requested_pf; /* Features requested during table load */
 213         struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
 214 };
 215
 216 /*
 217  * Target context for a thin.
 218  */
 219 struct thin_c {
 220         struct dm_dev *pool_dev;
 221         struct dm_dev *origin_dev;
 222         dm_thin_id dev_id;
 223
 224         struct pool *pool;
 225         struct dm_thin_device *td;
 226 };
 227
 228 /*----------------------------------------------------------------*/
 229
 230 /*
 231  * A global list of pools that uses a struct mapped_device as a key.
 232  */
 233 static struct dm_thin_pool_table {
 234         struct mutex mutex;
 235         struct list_head pools;
 236 } dm_thin_pool_table;
 237
 238 static void pool_table_init(void)
 239 {
 240         mutex_init(&dm_thin_pool_table.mutex);
 241         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
 242 }
 243
 244 static void __pool_table_insert(struct pool *pool)
 245 {
 246         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 247         list_add(&pool->list, &dm_thin_pool_table.pools);
 248 }
 249
 250 static void __pool_table_remove(struct pool *pool)
 251 {
 252         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 253         list_del(&pool->list);
 254 }
 255
 256 static struct pool *__pool_table_lookup(struct mapped_device *md)
 257 {
 258         struct pool *pool = NULL, *tmp;
 259
 260         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 261
 262         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 263                 if (tmp->pool_md == md) {
 264                         pool = tmp;
 265                         break;
 266                 }
 267         }
 268
 269         return pool;
 270 }
 271
 272 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
 273 {
 274         struct pool *pool = NULL, *tmp;
 275
 276         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 277
 278         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 279                 if (tmp->md_dev == md_dev) {
 280                         pool = tmp;
 281                         break;
 282                 }
 283         }
 284
 285         return pool;
 286 }
 287
 288 /*----------------------------------------------------------------*/
 289
 290 struct dm_thin_endio_hook {
 291         struct thin_c *tc;
 292         struct dm_deferred_entry *shared_read_entry;
 293         struct dm_deferred_entry *all_io_entry;
 294         struct dm_thin_new_mapping *overwrite_mapping;
 295 };
 296
 297 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
 298 {
 299         struct bio *bio;
 300         struct bio_list bios;
 301
 302         bio_list_init(&bios);
 303         bio_list_merge(&bios, master);
 304         bio_list_init(master);
 305
 306         while ((bio = bio_list_pop(&bios))) {
 307                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 308
 309                 if (h->tc == tc)
 310                         bio_endio(bio, DM_ENDIO_REQUEUE);
 311                 else
 312                         bio_list_add(master, bio);
 313         }
 314 }
 315
 316 static void requeue_io(struct thin_c *tc)
 317 {
 318         struct pool *pool = tc->pool;
 319         unsigned long flags;
 320
 321         spin_lock_irqsave(&pool->lock, flags);
 322         __requeue_bio_list(tc, &pool->deferred_bios);
 323         __requeue_bio_list(tc, &pool->retry_on_resume_list);
 324         spin_unlock_irqrestore(&pool->lock, flags);
 325 }
 326
 327 /*
 328  * This section of code contains the logic for processing a thin device's IO.
 329  * Much of the code depends on pool object resources (lists, workqueues, etc)
 330  * but most is exclusively called from the thin target rather than the thin-pool
 331  * target.
 332  */
 333
 334 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
 335 {
 336         sector_t block_nr = bio->bi_sector;
 337
 338         if (tc->pool->sectors_per_block_shift < 0)
 339                 (void) sector_div(block_nr, tc->pool->sectors_per_block);
 340         else
 341                 block_nr >>= tc->pool->sectors_per_block_shift;
 342
 343         return block_nr;
 344 }
 345
 346 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
 347 {
 348         struct pool *pool = tc->pool;
 349         sector_t bi_sector = bio->bi_sector;
 350
 351         bio->bi_bdev = tc->pool_dev->bdev;
 352         if (tc->pool->sectors_per_block_shift < 0)
 353                 bio->bi_sector = (block * pool->sectors_per_block) +
 354                                  sector_div(bi_sector, pool->sectors_per_block);
 355         else
 356                 bio->bi_sector = (block << pool->sectors_per_block_shift) |
 357                                 (bi_sector & (pool->sectors_per_block - 1));
 358 }
 359
 360 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
 361 {
 362         bio->bi_bdev = tc->origin_dev->bdev;
 363 }
 364
 365 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
 366 {
 367         return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
 368                 dm_thin_changed_this_transaction(tc->td);
 369 }
 370
 371 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
 372 {
 373         struct dm_thin_endio_hook *h;
 374
 375         if (bio->bi_rw & REQ_DISCARD)
 376                 return;
 377
 378         h = dm_get_mapinfo(bio)->ptr;
 379         h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
 380 }
 381
 382 static void issue(struct thin_c *tc, struct bio *bio)
 383 {
 384         struct pool *pool = tc->pool;
 385         unsigned long flags;
 386
 387         if (!bio_triggers_commit(tc, bio)) {
 388                 generic_make_request(bio);
 389                 return;
 390         }
 391
 392         /*
 393          * Complete bio with an error if earlier I/O caused changes to
 394          * the metadata that can't be committed e.g, due to I/O errors
 395          * on the metadata device.
 396          */
 397         if (dm_thin_aborted_changes(tc->td)) {
 398                 bio_io_error(bio);
 399                 return;
 400         }
 401
 402         /*
 403          * Batch together any bios that trigger commits and then issue a
 404          * single commit for them in process_deferred_bios().
 405          */
 406         spin_lock_irqsave(&pool->lock, flags);
 407         bio_list_add(&pool->deferred_flush_bios, bio);
 408         spin_unlock_irqrestore(&pool->lock, flags);
 409 }
 410
 411 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
 412 {
 413         remap_to_origin(tc, bio);
 414         issue(tc, bio);
 415 }
 416
 417 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
 418                             dm_block_t block)
 419 {
 420         remap(tc, bio, block);
 421         issue(tc, bio);
 422 }
 423
 424 /*
 425  * wake_worker() is used when new work is queued and when pool_resume is
 426  * ready to continue deferred IO processing.
 427  */
 428 static void wake_worker(struct pool *pool)
 429 {
 430         queue_work(pool->wq, &pool->worker);
 431 }
 432
 433 /*----------------------------------------------------------------*/
 434
 435 /*
 436  * Bio endio functions.
 437  */
 438 struct dm_thin_new_mapping {
 439         struct list_head list;
 440
 441         unsigned quiesced:1;
 442         unsigned prepared:1;
 443         unsigned pass_discard:1;
 444
 445         struct thin_c *tc;
 446         dm_block_t virt_block;
 447         dm_block_t data_block;
 448         struct dm_bio_prison_cell *cell, *cell2;
 449         int err;
 450
 451         /*
 452          * If the bio covers the whole area of a block then we can avoid
 453          * zeroing or copying.  Instead this bio is hooked.  The bio will
 454          * still be in the cell, so care has to be taken to avoid issuing
 455          * the bio twice.
 456          */
 457         struct bio *bio;
 458         bio_end_io_t *saved_bi_end_io;
 459 };
 460
 461 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
 462 {
 463         struct pool *pool = m->tc->pool;
 464
 465         if (m->quiesced && m->prepared) {
 466                 list_add(&m->list, &pool->prepared_mappings);
 467                 wake_worker(pool);
 468         }
 469 }
 470
 471 static void copy_complete(int read_err, unsigned long write_err, void *context)
 472 {
 473         unsigned long flags;
 474         struct dm_thin_new_mapping *m = context;
 475         struct pool *pool = m->tc->pool;
 476
 477         m->err = read_err || write_err ? -EIO : 0;
 478
 479         spin_lock_irqsave(&pool->lock, flags);
 480         m->prepared = 1;
 481         __maybe_add_mapping(m);
 482         spin_unlock_irqrestore(&pool->lock, flags);
 483 }
 484
 485 static void overwrite_endio(struct bio *bio, int err)
 486 {
 487         unsigned long flags;
 488         struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 489         struct dm_thin_new_mapping *m = h->overwrite_mapping;
 490         struct pool *pool = m->tc->pool;
 491
 492         m->err = err;
 493
 494         spin_lock_irqsave(&pool->lock, flags);
 495         m->prepared = 1;
 496         __maybe_add_mapping(m);
 497         spin_unlock_irqrestore(&pool->lock, flags);
 498 }
 499
 500 /*----------------------------------------------------------------*/
 501
 502 /*
 503  * Workqueue.
 504  */
 505
 506 /*
 507  * Prepared mapping jobs.
 508  */
 509
 510 /*
 511  * This sends the bios in the cell back to the deferred_bios list.
 512  */
 513 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 514 {
 515         struct pool *pool = tc->pool;
 516         unsigned long flags;
 517
 518         spin_lock_irqsave(&pool->lock, flags);
 519         dm_cell_release(cell, &pool->deferred_bios);
 520         spin_unlock_irqrestore(&tc->pool->lock, flags);
 521
 522         wake_worker(pool);
 523 }
 524
 525 /*
 526  * Same as cell_defer except it omits the original holder of the cell.
 527  */
 528 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 529 {
 530         struct pool *pool = tc->pool;
 531         unsigned long flags;
 532
 533         spin_lock_irqsave(&pool->lock, flags);
 534         dm_cell_release_no_holder(cell, &pool->deferred_bios);
 535         spin_unlock_irqrestore(&pool->lock, flags);
 536
 537         wake_worker(pool);
 538 }
 539
 540 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
 541 {
 542         if (m->bio)
 543                 m->bio->bi_end_io = m->saved_bi_end_io;
 544         dm_cell_error(m->cell);
 545         list_del(&m->list);
 546         mempool_free(m, m->tc->pool->mapping_pool);
 547 }
 548 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
 549 {
 550         struct thin_c *tc = m->tc;
 551         struct bio *bio;
 552         int r;
 553
 554         bio = m->bio;
 555         if (bio)
 556                 bio->bi_end_io = m->saved_bi_end_io;
 557
 558         if (m->err) {
 559                 dm_cell_error(m->cell);
 560                 goto out;
 561         }
 562
 563         /*
 564          * Commit the prepared block into the mapping btree.
 565          * Any I/O for this block arriving after this point will get
 566          * remapped to it directly.
 567          */
 568         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
 569         if (r) {
 570                 DMERR("dm_thin_insert_block() failed");
 571                 dm_cell_error(m->cell);
 572                 goto out;
 573         }
 574
 575         /*
 576          * Release any bios held while the block was being provisioned.
 577          * If we are processing a write bio that completely covers the block,
 578          * we already processed it so can ignore it now when processing
 579          * the bios in the cell.
 580          */
 581         if (bio) {
 582                 cell_defer_no_holder(tc, m->cell);
 583                 bio_endio(bio, 0);
 584         } else
 585                 cell_defer(tc, m->cell);
 586
 587 out:
 588         list_del(&m->list);
 589         mempool_free(m, tc->pool->mapping_pool);
 590 }
 591
 592 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
 593 {
 594         struct thin_c *tc = m->tc;
 595
 596         bio_io_error(m->bio);
 597         cell_defer_no_holder(tc, m->cell);
 598         cell_defer_no_holder(tc, m->cell2);
 599         mempool_free(m, tc->pool->mapping_pool);
 600 }
 601
 602 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
 603 {
 604         struct thin_c *tc = m->tc;
 605
 606         inc_all_io_entry(tc->pool, m->bio);
 607         cell_defer_no_holder(tc, m->cell);
 608         cell_defer_no_holder(tc, m->cell2);
 609
 610         if (m->pass_discard)
 611                 remap_and_issue(tc, m->bio, m->data_block);
 612         else
 613                 bio_endio(m->bio, 0);
 614
 615         mempool_free(m, tc->pool->mapping_pool);
 616 }
 617
 618 static void process_prepared_discard(struct dm_thin_new_mapping *m)
 619 {
 620         int r;
 621         struct thin_c *tc = m->tc;
 622
 623         r = dm_thin_remove_block(tc->td, m->virt_block);
 624         if (r)
 625                 DMERR("dm_thin_remove_block() failed");
 626
 627         process_prepared_discard_passdown(m);
 628 }
 629
 630 static void process_prepared(struct pool *pool, struct list_head *head,
 631                              process_mapping_fn *fn)
 632 {
 633         unsigned long flags;
 634         struct list_head maps;
 635         struct dm_thin_new_mapping *m, *tmp;
 636
 637         INIT_LIST_HEAD(&maps);
 638         spin_lock_irqsave(&pool->lock, flags);
 639         list_splice_init(head, &maps);
 640         spin_unlock_irqrestore(&pool->lock, flags);
 641
 642         list_for_each_entry_safe(m, tmp, &maps, list)
 643                 (*fn)(m);
 644 }
 645
 646 /*
 647  * Deferred bio jobs.
 648  */
 649 static int io_overlaps_block(struct pool *pool, struct bio *bio)
 650 {
 651         return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
 652 }
 653
 654 static int io_overwrites_block(struct pool *pool, struct bio *bio)
 655 {
 656         return (bio_data_dir(bio) == WRITE) &&
 657                 io_overlaps_block(pool, bio);
 658 }
 659
 660 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
 661                                bio_end_io_t *fn)
 662 {
 663         *save = bio->bi_end_io;
 664         bio->bi_end_io = fn;
 665 }
 666
 667 static int ensure_next_mapping(struct pool *pool)
 668 {
 669         if (pool->next_mapping)
 670                 return 0;
 671
 672         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
 673
 674         return pool->next_mapping ? 0 : -ENOMEM;
 675 }
 676
 677 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
 678 {
 679         struct dm_thin_new_mapping *r = pool->next_mapping;
 680
 681         BUG_ON(!pool->next_mapping);
 682
 683         pool->next_mapping = NULL;
 684
 685         return r;
 686 }
 687
 688 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
 689                           struct dm_dev *origin, dm_block_t data_origin,
 690                           dm_block_t data_dest,
 691                           struct dm_bio_prison_cell *cell, struct bio *bio)
 692 {
 693         int r;
 694         struct pool *pool = tc->pool;
 695         struct dm_thin_new_mapping *m = get_next_mapping(pool);
 696
 697         INIT_LIST_HEAD(&m->list);
 698         m->quiesced = 0;
 699         m->prepared = 0;
 700         m->tc = tc;
 701         m->virt_block = virt_block;
 702         m->data_block = data_dest;
 703         m->cell = cell;
 704         m->err = 0;
 705         m->bio = NULL;
 706
 707         if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
 708                 m->quiesced = 1;
 709
 710         /*
 711          * IO to pool_dev remaps to the pool target's data_dev.
 712          *
 713          * If the whole block of data is being overwritten, we can issue the
 714          * bio immediately. Otherwise we use kcopyd to clone the data first.
 715          */
 716         if (io_overwrites_block(pool, bio)) {
 717                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 718
 719                 h->overwrite_mapping = m;
 720                 m->bio = bio;
 721                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 722                 inc_all_io_entry(pool, bio);
 723                 remap_and_issue(tc, bio, data_dest);
 724         } else {
 725                 struct dm_io_region from, to;
 726
 727                 from.bdev = origin->bdev;
 728                 from.sector = data_origin * pool->sectors_per_block;
 729                 from.count = pool->sectors_per_block;
 730
 731                 to.bdev = tc->pool_dev->bdev;
 732                 to.sector = data_dest * pool->sectors_per_block;
 733                 to.count = pool->sectors_per_block;
 734
 735                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
 736                                    0, copy_complete, m);
 737                 if (r < 0) {
 738                         mempool_free(m, pool->mapping_pool);
 739                         DMERR("dm_kcopyd_copy() failed");
 740                         dm_cell_error(cell);
 741                 }
 742         }
 743 }
 744
 745 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
 746                                    dm_block_t data_origin, dm_block_t data_dest,
 747                                    struct dm_bio_prison_cell *cell, struct bio *bio)
 748 {
 749         schedule_copy(tc, virt_block, tc->pool_dev,
 750                       data_origin, data_dest, cell, bio);
 751 }
 752
 753 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
 754                                    dm_block_t data_dest,
 755                                    struct dm_bio_prison_cell *cell, struct bio *bio)
 756 {
 757         schedule_copy(tc, virt_block, tc->origin_dev,
 758                       virt_block, data_dest, cell, bio);
 759 }
 760
 761 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
 762                           dm_block_t data_block, struct dm_bio_prison_cell *cell,
 763                           struct bio *bio)
 764 {
 765         struct pool *pool = tc->pool;
 766         struct dm_thin_new_mapping *m = get_next_mapping(pool);
 767
 768         INIT_LIST_HEAD(&m->list);
 769         m->quiesced = 1;
 770         m->prepared = 0;
 771         m->tc = tc;
 772         m->virt_block = virt_block;
 773         m->data_block = data_block;
 774         m->cell = cell;
 775         m->err = 0;
 776         m->bio = NULL;
 777
 778         /*
 779          * If the whole block of data is being overwritten or we are not
 780          * zeroing pre-existing data, we can issue the bio immediately.
 781          * Otherwise we use kcopyd to zero the data first.
 782          */
 783         if (!pool->pf.zero_new_blocks)
 784                 process_prepared_mapping(m);
 785
 786         else if (io_overwrites_block(pool, bio)) {
 787                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 788
 789                 h->overwrite_mapping = m;
 790                 m->bio = bio;
 791                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 792                 inc_all_io_entry(pool, bio);
 793                 remap_and_issue(tc, bio, data_block);
 794         } else {
 795                 int r;
 796                 struct dm_io_region to;
 797
 798                 to.bdev = tc->pool_dev->bdev;
 799                 to.sector = data_block * pool->sectors_per_block;
 800                 to.count = pool->sectors_per_block;
 801
 802                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
 803                 if (r < 0) {
 804                         mempool_free(m, pool->mapping_pool);
 805                         DMERR("dm_kcopyd_zero() failed");
 806                         dm_cell_error(cell);
 807                 }
 808         }
 809 }
 810
 811 static int commit(struct pool *pool)
 812 {
 813         int r;
 814
 815         r = dm_pool_commit_metadata(pool->pmd);
 816         if (r)
 817                 DMERR("commit failed, error = %d", r);
 818
 819         return r;
 820 }
 821
 822 /*
 823  * A non-zero return indicates read_only or fail_io mode.
 824  * Many callers don't care about the return value.
 825  */
 826 static int commit_or_fallback(struct pool *pool)
 827 {
 828         int r;
 829
 830         if (get_pool_mode(pool) != PM_WRITE)
 831                 return -EINVAL;
 832
 833         r = commit(pool);
 834         if (r)
 835                 set_pool_mode(pool, PM_READ_ONLY);
 836
 837         return r;
 838 }
 839
 840 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
 841 {
 842         int r;
 843         dm_block_t free_blocks;
 844         unsigned long flags;
 845         struct pool *pool = tc->pool;
 846
 847         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 848         if (r)
 849                 return r;
 850
 851         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
 852                 DMWARN("%s: reached low water mark, sending event.",
 853                        dm_device_name(pool->pool_md));
 854                 spin_lock_irqsave(&pool->lock, flags);
 855                 pool->low_water_triggered = 1;
 856                 spin_unlock_irqrestore(&pool->lock, flags);
 857                 dm_table_event(pool->ti->table);
 858         }
 859
 860         if (!free_blocks) {
 861                 if (pool->no_free_space)
 862                         return -ENOSPC;
 863                 else {
 864                         /*
 865                          * Try to commit to see if that will free up some
 866                          * more space.
 867                          */
 868                         (void) commit_or_fallback(pool);
 869
 870                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 871                         if (r)
 872                                 return r;
 873
 874                         /*
 875                          * If we still have no space we set a flag to avoid
 876                          * doing all this checking and return -ENOSPC.
 877                          */
 878                         if (!free_blocks) {
 879                                 DMWARN("%s: no free space available.",
 880                                        dm_device_name(pool->pool_md));
 881                                 spin_lock_irqsave(&pool->lock, flags);
 882                                 pool->no_free_space = 1;
 883                                 spin_unlock_irqrestore(&pool->lock, flags);
 884                                 return -ENOSPC;
 885                         }
 886                 }
 887         }
 888
 889         r = dm_pool_alloc_data_block(pool->pmd, result);
 890         if (r)
 891                 return r;
 892
 893         return 0;
 894 }
 895
 896 /*
 897  * If we have run out of space, queue bios until the device is
 898  * resumed, presumably after having been reloaded with more space.
 899  */
 900 static void retry_on_resume(struct bio *bio)
 901 {
 902         struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 903         struct thin_c *tc = h->tc;
 904         struct pool *pool = tc->pool;
 905         unsigned long flags;
 906
 907         spin_lock_irqsave(&pool->lock, flags);
 908         bio_list_add(&pool->retry_on_resume_list, bio);
 909         spin_unlock_irqrestore(&pool->lock, flags);
 910 }
 911
 912 static void no_space(struct dm_bio_prison_cell *cell)
 913 {
 914         struct bio *bio;
 915         struct bio_list bios;
 916
 917         bio_list_init(&bios);
 918         dm_cell_release(cell, &bios);
 919
 920         while ((bio = bio_list_pop(&bios)))
 921                 retry_on_resume(bio);
 922 }
 923
 924 static void process_discard(struct thin_c *tc, struct bio *bio)
 925 {
 926         int r;
 927         unsigned long flags;
 928         struct pool *pool = tc->pool;
 929         struct dm_bio_prison_cell *cell, *cell2;
 930         struct dm_cell_key key, key2;
 931         dm_block_t block = get_bio_block(tc, bio);
 932         struct dm_thin_lookup_result lookup_result;
 933         struct dm_thin_new_mapping *m;
 934
 935         build_virtual_key(tc->td, block, &key);
 936         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
 937                 return;
 938
 939         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
 940         switch (r) {
 941         case 0:
 942                 /*
 943                  * Check nobody is fiddling with this pool block.  This can
 944                  * happen if someone's in the process of breaking sharing
 945                  * on this block.
 946                  */
 947                 build_data_key(tc->td, lookup_result.block, &key2);
 948                 if (dm_bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
 949                         cell_defer_no_holder(tc, cell);
 950                         break;
 951                 }
 952
 953                 if (io_overlaps_block(pool, bio)) {
 954                         /*
 955                          * IO may still be going to the destination block.  We must
 956                          * quiesce before we can do the removal.
 957                          */
 958                         m = get_next_mapping(pool);
 959                         m->tc = tc;
 960                         m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
 961                         m->virt_block = block;
 962                         m->data_block = lookup_result.block;
 963                         m->cell = cell;
 964                         m->cell2 = cell2;
 965                         m->err = 0;
 966                         m->bio = bio;
 967
 968                         if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
 969                                 spin_lock_irqsave(&pool->lock, flags);
 970                                 list_add(&m->list, &pool->prepared_discards);
 971                                 spin_unlock_irqrestore(&pool->lock, flags);
 972                                 wake_worker(pool);
 973                         }
 974                 } else {
 975                         inc_all_io_entry(pool, bio);
 976                         cell_defer_no_holder(tc, cell);
 977                         cell_defer_no_holder(tc, cell2);
 978
 979                         /*
 980                          * The DM core makes sure that the discard doesn't span
 981                          * a block boundary.  So we submit the discard of a
 982                          * partial block appropriately.
 983                          */
 984                         if ((!lookup_result.shared) && pool->pf.discard_passdown)
 985                                 remap_and_issue(tc, bio, lookup_result.block);
 986                         else
 987                                 bio_endio(bio, 0);
 988                 }
 989                 break;
 990
 991         case -ENODATA:
 992                 /*
 993                  * It isn't provisioned, just forget it.
 994                  */
 995                 cell_defer_no_holder(tc, cell);
 996                 bio_endio(bio, 0);
 997                 break;
 998
 999         default:
1000                 DMERR("discard: find block unexpectedly returned %d", r);
1001                 cell_defer_no_holder(tc, cell);
1002                 bio_io_error(bio);
1003                 break;
1004         }
1005 }
1006
1007 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1008                           struct dm_cell_key *key,
1009                           struct dm_thin_lookup_result *lookup_result,
1010                           struct dm_bio_prison_cell *cell)
1011 {
1012         int r;
1013         dm_block_t data_block;
1014
1015         r = alloc_data_block(tc, &data_block);
1016         switch (r) {
1017         case 0:
1018                 schedule_internal_copy(tc, block, lookup_result->block,
1019                                        data_block, cell, bio);
1020                 break;
1021
1022         case -ENOSPC:
1023                 no_space(cell);
1024                 break;
1025
1026         default:
1027                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1028                 dm_cell_error(cell);
1029                 break;
1030         }
1031 }
1032
1033 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1034                                dm_block_t block,
1035                                struct dm_thin_lookup_result *lookup_result)
1036 {
1037         struct dm_bio_prison_cell *cell;
1038         struct pool *pool = tc->pool;
1039         struct dm_cell_key key;
1040
1041         /*
1042          * If cell is already occupied, then sharing is already in the process
1043          * of being broken so we have nothing further to do here.
1044          */
1045         build_data_key(tc->td, lookup_result->block, &key);
1046         if (dm_bio_detain(pool->prison, &key, bio, &cell))
1047                 return;
1048
1049         if (bio_data_dir(bio) == WRITE && bio->bi_size)
1050                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1051         else {
1052                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1053
1054                 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1055                 inc_all_io_entry(pool, bio);
1056                 cell_defer_no_holder(tc, cell);
1057
1058                 remap_and_issue(tc, bio, lookup_result->block);
1059         }
1060 }
1061
1062 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1063                             struct dm_bio_prison_cell *cell)
1064 {
1065         int r;
1066         dm_block_t data_block;
1067
1068         /*
1069          * Remap empty bios (flushes) immediately, without provisioning.
1070          */
1071         if (!bio->bi_size) {
1072                 inc_all_io_entry(tc->pool, bio);
1073                 cell_defer_no_holder(tc, cell);
1074
1075                 remap_and_issue(tc, bio, 0);
1076                 return;
1077         }
1078
1079         /*
1080          * Fill read bios with zeroes and complete them immediately.
1081          */
1082         if (bio_data_dir(bio) == READ) {
1083                 zero_fill_bio(bio);
1084                 cell_defer_no_holder(tc, cell);
1085                 bio_endio(bio, 0);
1086                 return;
1087         }
1088
1089         r = alloc_data_block(tc, &data_block);
1090         switch (r) {
1091         case 0:
1092                 if (tc->origin_dev)
1093                         schedule_external_copy(tc, block, data_block, cell, bio);
1094                 else
1095                         schedule_zero(tc, block, data_block, cell, bio);
1096                 break;
1097
1098         case -ENOSPC:
1099                 no_space(cell);
1100                 break;
1101
1102         default:
1103                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1104                 set_pool_mode(tc->pool, PM_READ_ONLY);
1105                 dm_cell_error(cell);
1106                 break;
1107         }
1108 }
1109
1110 static void process_bio(struct thin_c *tc, struct bio *bio)
1111 {
1112         int r;
1113         dm_block_t block = get_bio_block(tc, bio);
1114         struct dm_bio_prison_cell *cell;
1115         struct dm_cell_key key;
1116         struct dm_thin_lookup_result lookup_result;
1117
1118         /*
1119          * If cell is already occupied, then the block is already
1120          * being provisioned so we have nothing further to do here.
1121          */
1122         build_virtual_key(tc->td, block, &key);
1123         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
1124                 return;
1125
1126         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1127         switch (r) {
1128         case 0:
1129                 if (lookup_result.shared) {
1130                         process_shared_bio(tc, bio, block, &lookup_result);
1131                         cell_defer_no_holder(tc, cell);
1132                 } else {
1133                         inc_all_io_entry(tc->pool, bio);
1134                         cell_defer_no_holder(tc, cell);
1135
1136                         remap_and_issue(tc, bio, lookup_result.block);
1137                 }
1138                 break;
1139
1140         case -ENODATA:
1141                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1142                         inc_all_io_entry(tc->pool, bio);
1143                         cell_defer_no_holder(tc, cell);
1144
1145                         remap_to_origin_and_issue(tc, bio);
1146                 } else
1147                         provision_block(tc, bio, block, cell);
1148                 break;
1149
1150         default:
1151                 DMERR("dm_thin_find_block() failed, error = %d", r);
1152                 cell_defer_no_holder(tc, cell);
1153                 bio_io_error(bio);
1154                 break;
1155         }
1156 }
1157
1158 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1159 {
1160         int r;
1161         int rw = bio_data_dir(bio);
1162         dm_block_t block = get_bio_block(tc, bio);
1163         struct dm_thin_lookup_result lookup_result;
1164
1165         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1166         switch (r) {
1167         case 0:
1168                 if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1169                         bio_io_error(bio);
1170                 else {
1171                         inc_all_io_entry(tc->pool, bio);
1172                         remap_and_issue(tc, bio, lookup_result.block);
1173                 }
1174                 break;
1175
1176         case -ENODATA:
1177                 if (rw != READ) {
1178                         bio_io_error(bio);
1179                         break;
1180                 }
1181
1182                 if (tc->origin_dev) {
1183                         inc_all_io_entry(tc->pool, bio);
1184                         remap_to_origin_and_issue(tc, bio);
1185                         break;
1186                 }
1187
1188                 zero_fill_bio(bio);
1189                 bio_endio(bio, 0);
1190                 break;
1191
1192         default:
1193                 DMERR("dm_thin_find_block() failed, error = %d", r);
1194                 bio_io_error(bio);
1195                 break;
1196         }
1197 }
1198
1199 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1200 {
1201         bio_io_error(bio);
1202 }
1203
1204 static int need_commit_due_to_time(struct pool *pool)
1205 {
1206         return jiffies < pool->last_commit_jiffies ||
1207                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1208 }
1209
1210 static void process_deferred_bios(struct pool *pool)
1211 {
1212         unsigned long flags;
1213         struct bio *bio;
1214         struct bio_list bios;
1215
1216         bio_list_init(&bios);
1217
1218         spin_lock_irqsave(&pool->lock, flags);
1219         bio_list_merge(&bios, &pool->deferred_bios);
1220         bio_list_init(&pool->deferred_bios);
1221         spin_unlock_irqrestore(&pool->lock, flags);
1222
1223         while ((bio = bio_list_pop(&bios))) {
1224                 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1225                 struct thin_c *tc = h->tc;
1226
1227                 /*
1228                  * If we've got no free new_mapping structs, and processing
1229                  * this bio might require one, we pause until there are some
1230                  * prepared mappings to process.
1231                  */
1232                 if (ensure_next_mapping(pool)) {
1233                         spin_lock_irqsave(&pool->lock, flags);
1234                         bio_list_merge(&pool->deferred_bios, &bios);
1235                         spin_unlock_irqrestore(&pool->lock, flags);
1236
1237                         break;
1238                 }
1239
1240                 if (bio->bi_rw & REQ_DISCARD)
1241                         pool->process_discard(tc, bio);
1242                 else
1243                         pool->process_bio(tc, bio);
1244         }
1245
1246         /*
1247          * If there are any deferred flush bios, we must commit
1248          * the metadata before issuing them.
1249          */
1250         bio_list_init(&bios);
1251         spin_lock_irqsave(&pool->lock, flags);
1252         bio_list_merge(&bios, &pool->deferred_flush_bios);
1253         bio_list_init(&pool->deferred_flush_bios);
1254         spin_unlock_irqrestore(&pool->lock, flags);
1255
1256         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1257                 return;
1258
1259         if (commit_or_fallback(pool)) {
1260                 while ((bio = bio_list_pop(&bios)))
1261                         bio_io_error(bio);
1262                 return;
1263         }
1264         pool->last_commit_jiffies = jiffies;
1265
1266         while ((bio = bio_list_pop(&bios)))
1267                 generic_make_request(bio);
1268 }
1269
1270 static void do_worker(struct work_struct *ws)
1271 {
1272         struct pool *pool = container_of(ws, struct pool, worker);
1273
1274         process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1275         process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1276         process_deferred_bios(pool);
1277 }
1278
1279 /*
1280  * We want to commit periodically so that not too much
1281  * unwritten data builds up.
1282  */
1283 static void do_waker(struct work_struct *ws)
1284 {
1285         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1286         wake_worker(pool);
1287         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1288 }
1289
1290 /*----------------------------------------------------------------*/
1291
1292 static enum pool_mode get_pool_mode(struct pool *pool)
1293 {
1294         return pool->pf.mode;
1295 }
1296
1297 static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1298 {
1299         int r;
1300
1301         pool->pf.mode = mode;
1302
1303         switch (mode) {
1304         case PM_FAIL:
1305                 DMERR("switching pool to failure mode");
1306                 pool->process_bio = process_bio_fail;
1307                 pool->process_discard = process_bio_fail;
1308                 pool->process_prepared_mapping = process_prepared_mapping_fail;
1309                 pool->process_prepared_discard = process_prepared_discard_fail;
1310                 break;
1311
1312         case PM_READ_ONLY:
1313                 DMERR("switching pool to read-only mode");
1314                 r = dm_pool_abort_metadata(pool->pmd);
1315                 if (r) {
1316                         DMERR("aborting transaction failed");
1317                         set_pool_mode(pool, PM_FAIL);
1318                 } else {
1319                         dm_pool_metadata_read_only(pool->pmd);
1320                         pool->process_bio = process_bio_read_only;
1321                         pool->process_discard = process_discard;
1322                         pool->process_prepared_mapping = process_prepared_mapping_fail;
1323                         pool->process_prepared_discard = process_prepared_discard_passdown;
1324                 }
1325                 break;
1326
1327         case PM_WRITE:
1328                 pool->process_bio = process_bio;
1329                 pool->process_discard = process_discard;
1330                 pool->process_prepared_mapping = process_prepared_mapping;
1331                 pool->process_prepared_discard = process_prepared_discard;
1332                 break;
1333         }
1334 }
1335
1336 /*----------------------------------------------------------------*/
1337
1338 /*
1339  * Mapping functions.
1340  */
1341
1342 /*
1343  * Called only while mapping a thin bio to hand it over to the workqueue.
1344  */
1345 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1346 {
1347         unsigned long flags;
1348         struct pool *pool = tc->pool;
1349
1350         spin_lock_irqsave(&pool->lock, flags);
1351         bio_list_add(&pool->deferred_bios, bio);
1352         spin_unlock_irqrestore(&pool->lock, flags);
1353
1354         wake_worker(pool);
1355 }
1356
1357 static struct dm_thin_endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1358 {
1359         struct pool *pool = tc->pool;
1360         struct dm_thin_endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1361
1362         h->tc = tc;
1363         h->shared_read_entry = NULL;
1364         h->all_io_entry = NULL;
1365         h->overwrite_mapping = NULL;
1366
1367         return h;
1368 }
1369
1370 /*
1371  * Non-blocking function called from the thin target's map function.
1372  */
1373 static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1374                         union map_info *map_context)
1375 {
1376         int r;
1377         struct thin_c *tc = ti->private;
1378         dm_block_t block = get_bio_block(tc, bio);
1379         struct dm_thin_device *td = tc->td;
1380         struct dm_thin_lookup_result result;
1381         struct dm_bio_prison_cell *cell1, *cell2;
1382         struct dm_cell_key key;
1383
1384         map_context->ptr = thin_hook_bio(tc, bio);
1385
1386         if (get_pool_mode(tc->pool) == PM_FAIL) {
1387                 bio_io_error(bio);
1388                 return DM_MAPIO_SUBMITTED;
1389         }
1390
1391         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1392                 thin_defer_bio(tc, bio);
1393                 return DM_MAPIO_SUBMITTED;
1394         }
1395
1396         r = dm_thin_find_block(td, block, 0, &result);
1397
1398         /*
1399          * Note that we defer readahead too.
1400          */
1401         switch (r) {
1402         case 0:
1403                 if (unlikely(result.shared)) {
1404                         /*
1405                          * We have a race condition here between the
1406                          * result.shared value returned by the lookup and
1407                          * snapshot creation, which may cause new
1408                          * sharing.
1409                          *
1410                          * To avoid this always quiesce the origin before
1411                          * taking the snap.  You want to do this anyway to
1412                          * ensure a consistent application view
1413                          * (i.e. lockfs).
1414                          *
1415                          * More distant ancestors are irrelevant. The
1416                          * shared flag will be set in their case.
1417                          */
1418                         thin_defer_bio(tc, bio);
1419                         return DM_MAPIO_SUBMITTED;
1420                 }
1421
1422                 build_virtual_key(tc->td, block, &key);
1423                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1))
1424                         return DM_MAPIO_SUBMITTED;
1425
1426                 build_data_key(tc->td, result.block, &key);
1427                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2)) {
1428                         cell_defer_no_holder(tc, cell1);
1429                         return DM_MAPIO_SUBMITTED;
1430                 }
1431
1432                 inc_all_io_entry(tc->pool, bio);
1433                 cell_defer_no_holder(tc, cell2);
1434                 cell_defer_no_holder(tc, cell1);
1435
1436                 remap(tc, bio, result.block);
1437                 return DM_MAPIO_REMAPPED;
1438
1439         case -ENODATA:
1440                 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1441                         /*
1442                          * This block isn't provisioned, and we have no way
1443                          * of doing so.  Just error it.
1444                          */
1445                         bio_io_error(bio);
1446                         return DM_MAPIO_SUBMITTED;
1447                 }
1448                 /* fall through */
1449
1450         case -EWOULDBLOCK:
1451                 /*
1452                  * In future, the failed dm_thin_find_block above could
1453                  * provide the hint to load the metadata into cache.
1454                  */
1455                 thin_defer_bio(tc, bio);
1456                 return DM_MAPIO_SUBMITTED;
1457
1458         default:
1459                 /*
1460                  * Must always call bio_io_error on failure.
1461                  * dm_thin_find_block can fail with -EINVAL if the
1462                  * pool is switched to fail-io mode.
1463                  */
1464                 bio_io_error(bio);
1465                 return DM_MAPIO_SUBMITTED;
1466         }
1467 }
1468
1469 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1470 {
1471         int r;
1472         unsigned long flags;
1473         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1474
1475         spin_lock_irqsave(&pt->pool->lock, flags);
1476         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1477         spin_unlock_irqrestore(&pt->pool->lock, flags);
1478
1479         if (!r) {
1480                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1481                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1482         }
1483
1484         return r;
1485 }
1486
1487 static void __requeue_bios(struct pool *pool)
1488 {
1489         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1490         bio_list_init(&pool->retry_on_resume_list);
1491 }
1492
1493 /*----------------------------------------------------------------
1494  * Binding of control targets to a pool object
1495  *--------------------------------------------------------------*/
1496 static bool data_dev_supports_discard(struct pool_c *pt)
1497 {
1498         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1499
1500         return q && blk_queue_discard(q);
1501 }
1502
1503 /*
1504  * If discard_passdown was enabled verify that the data device
1505  * supports discards.  Disable discard_passdown if not.
1506  */
1507 static void disable_passdown_if_not_supported(struct pool_c *pt)
1508 {
1509         struct pool *pool = pt->pool;
1510         struct block_device *data_bdev = pt->data_dev->bdev;
1511         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1512         sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1513         const char *reason = NULL;
1514         char buf[BDEVNAME_SIZE];
1515
1516         if (!pt->adjusted_pf.discard_passdown)
1517                 return;
1518
1519         if (!data_dev_supports_discard(pt))
1520                 reason = "discard unsupported";
1521
1522         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1523                 reason = "max discard sectors smaller than a block";
1524
1525         else if (data_limits->discard_granularity > block_size)
1526                 reason = "discard granularity larger than a block";
1527
1528         else if (block_size & (data_limits->discard_granularity - 1))
1529                 reason = "discard granularity not a factor of block size";
1530
1531         if (reason) {
1532                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1533                 pt->adjusted_pf.discard_passdown = false;
1534         }
1535 }
1536
1537 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1538 {
1539         struct pool_c *pt = ti->private;
1540
1541         /*
1542          * We want to make sure that degraded pools are never upgraded.
1543          */
1544         enum pool_mode old_mode = pool->pf.mode;
1545         enum pool_mode new_mode = pt->adjusted_pf.mode;
1546
1547         if (old_mode > new_mode)
1548                 new_mode = old_mode;
1549
1550         pool->ti = ti;
1551         pool->low_water_blocks = pt->low_water_blocks;
1552         pool->pf = pt->adjusted_pf;
1553
1554         set_pool_mode(pool, new_mode);
1555
1556         return 0;
1557 }
1558
1559 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1560 {
1561         if (pool->ti == ti)
1562                 pool->ti = NULL;
1563 }
1564
1565 /*----------------------------------------------------------------
1566  * Pool creation
1567  *--------------------------------------------------------------*/
1568 /* Initialize pool features. */
1569 static void pool_features_init(struct pool_features *pf)
1570 {
1571         pf->mode = PM_WRITE;
1572         pf->zero_new_blocks = true;
1573         pf->discard_enabled = true;
1574         pf->discard_passdown = true;
1575 }
1576
1577 static void __pool_destroy(struct pool *pool)
1578 {
1579         __pool_table_remove(pool);
1580
1581         if (dm_pool_metadata_close(pool->pmd) < 0)
1582                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1583
1584         dm_bio_prison_destroy(pool->prison);
1585         dm_kcopyd_client_destroy(pool->copier);
1586
1587         if (pool->wq)
1588                 destroy_workqueue(pool->wq);
1589
1590         if (pool->next_mapping)
1591                 mempool_free(pool->next_mapping, pool->mapping_pool);
1592         mempool_destroy(pool->mapping_pool);
1593         mempool_destroy(pool->endio_hook_pool);
1594         dm_deferred_set_destroy(pool->shared_read_ds);
1595         dm_deferred_set_destroy(pool->all_io_ds);
1596         kfree(pool);
1597 }
1598
1599 static struct kmem_cache *_new_mapping_cache;
1600 static struct kmem_cache *_endio_hook_cache;
1601
1602 static struct pool *pool_create(struct mapped_device *pool_md,
1603                                 struct block_device *metadata_dev,
1604                                 unsigned long block_size,
1605                                 int read_only, char **error)
1606 {
1607         int r;
1608         void *err_p;
1609         struct pool *pool;
1610         struct dm_pool_metadata *pmd;
1611         bool format_device = read_only ? false : true;
1612
1613         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1614         if (IS_ERR(pmd)) {
1615                 *error = "Error creating metadata object";
1616                 return (struct pool *)pmd;
1617         }
1618
1619         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1620         if (!pool) {
1621                 *error = "Error allocating memory for pool";
1622                 err_p = ERR_PTR(-ENOMEM);
1623                 goto bad_pool;
1624         }
1625
1626         pool->pmd = pmd;
1627         pool->sectors_per_block = block_size;
1628         if (block_size & (block_size - 1))
1629                 pool->sectors_per_block_shift = -1;
1630         else
1631                 pool->sectors_per_block_shift = __ffs(block_size);
1632         pool->low_water_blocks = 0;
1633         pool_features_init(&pool->pf);
1634         pool->prison = dm_bio_prison_create(PRISON_CELLS);
1635         if (!pool->prison) {
1636                 *error = "Error creating pool's bio prison";
1637                 err_p = ERR_PTR(-ENOMEM);
1638                 goto bad_prison;
1639         }
1640
1641         pool->copier = dm_kcopyd_client_create();
1642         if (IS_ERR(pool->copier)) {
1643                 r = PTR_ERR(pool->copier);
1644                 *error = "Error creating pool's kcopyd client";
1645                 err_p = ERR_PTR(r);
1646                 goto bad_kcopyd_client;
1647         }
1648
1649         /*
1650          * Create singlethreaded workqueue that will service all devices
1651          * that use this metadata.
1652          */
1653         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1654         if (!pool->wq) {
1655                 *error = "Error creating pool's workqueue";
1656                 err_p = ERR_PTR(-ENOMEM);
1657                 goto bad_wq;
1658         }
1659
1660         INIT_WORK(&pool->worker, do_worker);
1661         INIT_DELAYED_WORK(&pool->waker, do_waker);
1662         spin_lock_init(&pool->lock);
1663         bio_list_init(&pool->deferred_bios);
1664         bio_list_init(&pool->deferred_flush_bios);
1665         INIT_LIST_HEAD(&pool->prepared_mappings);
1666         INIT_LIST_HEAD(&pool->prepared_discards);
1667         pool->low_water_triggered = 0;
1668         pool->no_free_space = 0;
1669         bio_list_init(&pool->retry_on_resume_list);
1670
1671         pool->shared_read_ds = dm_deferred_set_create();
1672         if (!pool->shared_read_ds) {
1673                 *error = "Error creating pool's shared read deferred set";
1674                 err_p = ERR_PTR(-ENOMEM);
1675                 goto bad_shared_read_ds;
1676         }
1677
1678         pool->all_io_ds = dm_deferred_set_create();
1679         if (!pool->all_io_ds) {
1680                 *error = "Error creating pool's all io deferred set";
1681                 err_p = ERR_PTR(-ENOMEM);
1682                 goto bad_all_io_ds;
1683         }
1684
1685         pool->next_mapping = NULL;
1686         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1687                                                       _new_mapping_cache);
1688         if (!pool->mapping_pool) {
1689                 *error = "Error creating pool's mapping mempool";
1690                 err_p = ERR_PTR(-ENOMEM);
1691                 goto bad_mapping_pool;
1692         }
1693
1694         pool->endio_hook_pool = mempool_create_slab_pool(ENDIO_HOOK_POOL_SIZE,
1695                                                          _endio_hook_cache);
1696         if (!pool->endio_hook_pool) {
1697                 *error = "Error creating pool's endio_hook mempool";
1698                 err_p = ERR_PTR(-ENOMEM);
1699                 goto bad_endio_hook_pool;
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_endio_hook_pool:
1710         mempool_destroy(pool->mapping_pool);
1711 bad_mapping_pool:
1712         dm_deferred_set_destroy(pool->all_io_ds);
1713 bad_all_io_ds:
1714         dm_deferred_set_destroy(pool->shared_read_ds);
1715 bad_shared_read_ds:
1716         destroy_workqueue(pool->wq);
1717 bad_wq:
1718         dm_kcopyd_client_destroy(pool->copier);
1719 bad_kcopyd_client:
1720         dm_bio_prison_destroy(pool->prison);
1721 bad_prison:
1722         kfree(pool);
1723 bad_pool:
1724         if (dm_pool_metadata_close(pmd))
1725                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1726
1727         return err_p;
1728 }
1729
1730 static void __pool_inc(struct pool *pool)
1731 {
1732         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1733         pool->ref_count++;
1734 }
1735
1736 static void __pool_dec(struct pool *pool)
1737 {
1738         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1739         BUG_ON(!pool->ref_count);
1740         if (!--pool->ref_count)
1741                 __pool_destroy(pool);
1742 }
1743
1744 static struct pool *__pool_find(struct mapped_device *pool_md,
1745                                 struct block_device *metadata_dev,
1746                                 unsigned long block_size, int read_only,
1747                                 char **error, int *created)
1748 {
1749         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1750
1751         if (pool) {
1752                 if (pool->pool_md != pool_md) {
1753                         *error = "metadata device already in use by a pool";
1754                         return ERR_PTR(-EBUSY);
1755                 }
1756                 __pool_inc(pool);
1757
1758         } else {
1759                 pool = __pool_table_lookup(pool_md);
1760                 if (pool) {
1761                         if (pool->md_dev != metadata_dev) {
1762                                 *error = "different pool cannot replace a pool";
1763                                 return ERR_PTR(-EINVAL);
1764                         }
1765                         __pool_inc(pool);
1766
1767                 } else {
1768                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1769                         *created = 1;
1770                 }
1771         }
1772
1773         return pool;
1774 }
1775
1776 /*----------------------------------------------------------------
1777  * Pool target methods
1778  *--------------------------------------------------------------*/
1779 static void pool_dtr(struct dm_target *ti)
1780 {
1781         struct pool_c *pt = ti->private;
1782
1783         mutex_lock(&dm_thin_pool_table.mutex);
1784
1785         unbind_control_target(pt->pool, ti);
1786         __pool_dec(pt->pool);
1787         dm_put_device(ti, pt->metadata_dev);
1788         dm_put_device(ti, pt->data_dev);
1789         kfree(pt);
1790
1791         mutex_unlock(&dm_thin_pool_table.mutex);
1792 }
1793
1794 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1795                                struct dm_target *ti)
1796 {
1797         int r;
1798         unsigned argc;
1799         const char *arg_name;
1800
1801         static struct dm_arg _args[] = {
1802                 {0, 3, "Invalid number of pool feature arguments"},
1803         };
1804
1805         /*
1806          * No feature arguments supplied.
1807          */
1808         if (!as->argc)
1809                 return 0;
1810
1811         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1812         if (r)
1813                 return -EINVAL;
1814
1815         while (argc && !r) {
1816                 arg_name = dm_shift_arg(as);
1817                 argc--;
1818
1819                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
1820                         pf->zero_new_blocks = false;
1821
1822                 else if (!strcasecmp(arg_name, "ignore_discard"))
1823                         pf->discard_enabled = false;
1824
1825                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
1826                         pf->discard_passdown = false;
1827
1828                 else if (!strcasecmp(arg_name, "read_only"))
1829                         pf->mode = PM_READ_ONLY;
1830
1831                 else {
1832                         ti->error = "Unrecognised pool feature requested";
1833                         r = -EINVAL;
1834                         break;
1835                 }
1836         }
1837
1838         return r;
1839 }
1840
1841 /*
1842  * thin-pool <metadata dev> <data dev>
1843  *           <data block size (sectors)>
1844  *           <low water mark (blocks)>
1845  *           [<#feature args> [<arg>]*]
1846  *
1847  * Optional feature arguments are:
1848  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1849  *           ignore_discard: disable discard
1850  *           no_discard_passdown: don't pass discards down to the data device
1851  */
1852 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1853 {
1854         int r, pool_created = 0;
1855         struct pool_c *pt;
1856         struct pool *pool;
1857         struct pool_features pf;
1858         struct dm_arg_set as;
1859         struct dm_dev *data_dev;
1860         unsigned long block_size;
1861         dm_block_t low_water_blocks;
1862         struct dm_dev *metadata_dev;
1863         sector_t metadata_dev_size;
1864         char b[BDEVNAME_SIZE];
1865
1866         /*
1867          * FIXME Remove validation from scope of lock.
1868          */
1869         mutex_lock(&dm_thin_pool_table.mutex);
1870
1871         if (argc < 4) {
1872                 ti->error = "Invalid argument count";
1873                 r = -EINVAL;
1874                 goto out_unlock;
1875         }
1876         as.argc = argc;
1877         as.argv = argv;
1878
1879         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1880         if (r) {
1881                 ti->error = "Error opening metadata block device";
1882                 goto out_unlock;
1883         }
1884
1885         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1886         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1887                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1888                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1889
1890         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1891         if (r) {
1892                 ti->error = "Error getting data device";
1893                 goto out_metadata;
1894         }
1895
1896         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1897             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1898             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1899             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1900                 ti->error = "Invalid block size";
1901                 r = -EINVAL;
1902                 goto out;
1903         }
1904
1905         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1906                 ti->error = "Invalid low water mark";
1907                 r = -EINVAL;
1908                 goto out;
1909         }
1910
1911         /*
1912          * Set default pool features.
1913          */
1914         pool_features_init(&pf);
1915
1916         dm_consume_args(&as, 4);
1917         r = parse_pool_features(&as, &pf, ti);
1918         if (r)
1919                 goto out;
1920
1921         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1922         if (!pt) {
1923                 r = -ENOMEM;
1924                 goto out;
1925         }
1926
1927         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1928                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
1929         if (IS_ERR(pool)) {
1930                 r = PTR_ERR(pool);
1931                 goto out_free_pt;
1932         }
1933
1934         /*
1935          * 'pool_created' reflects whether this is the first table load.
1936          * Top level discard support is not allowed to be changed after
1937          * initial load.  This would require a pool reload to trigger thin
1938          * device changes.
1939          */
1940         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1941                 ti->error = "Discard support cannot be disabled once enabled";
1942                 r = -EINVAL;
1943                 goto out_flags_changed;
1944         }
1945
1946         pt->pool = pool;
1947         pt->ti = ti;
1948         pt->metadata_dev = metadata_dev;
1949         pt->data_dev = data_dev;
1950         pt->low_water_blocks = low_water_blocks;
1951         pt->adjusted_pf = pt->requested_pf = pf;
1952         ti->num_flush_requests = 1;
1953
1954         /*
1955          * Only need to enable discards if the pool should pass
1956          * them down to the data device.  The thin device's discard
1957          * processing will cause mappings to be removed from the btree.
1958          */
1959         if (pf.discard_enabled && pf.discard_passdown) {
1960                 ti->num_discard_requests = 1;
1961
1962                 /*
1963                  * Setting 'discards_supported' circumvents the normal
1964                  * stacking of discard limits (this keeps the pool and
1965                  * thin devices' discard limits consistent).
1966                  */
1967                 ti->discards_supported = true;
1968                 ti->discard_zeroes_data_unsupported = true;
1969         }
1970         ti->private = pt;
1971
1972         pt->callbacks.congested_fn = pool_is_congested;
1973         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1974
1975         mutex_unlock(&dm_thin_pool_table.mutex);
1976
1977         return 0;
1978
1979 out_flags_changed:
1980         __pool_dec(pool);
1981 out_free_pt:
1982         kfree(pt);
1983 out:
1984         dm_put_device(ti, data_dev);
1985 out_metadata:
1986         dm_put_device(ti, metadata_dev);
1987 out_unlock:
1988         mutex_unlock(&dm_thin_pool_table.mutex);
1989
1990         return r;
1991 }
1992
1993 static int pool_map(struct dm_target *ti, struct bio *bio,
1994                     union map_info *map_context)
1995 {
1996         int r;
1997         struct pool_c *pt = ti->private;
1998         struct pool *pool = pt->pool;
1999         unsigned long flags;
2000
2001         /*
2002          * As this is a singleton target, ti->begin is always zero.
2003          */
2004         spin_lock_irqsave(&pool->lock, flags);
2005         bio->bi_bdev = pt->data_dev->bdev;
2006         r = DM_MAPIO_REMAPPED;
2007         spin_unlock_irqrestore(&pool->lock, flags);
2008
2009         return r;
2010 }
2011
2012 /*
2013  * Retrieves the number of blocks of the data device from
2014  * the superblock and compares it to the actual device size,
2015  * thus resizing the data device in case it has grown.
2016  *
2017  * This both copes with opening preallocated data devices in the ctr
2018  * being followed by a resume
2019  * -and-
2020  * calling the resume method individually after userspace has
2021  * grown the data device in reaction to a table event.
2022  */
2023 static int pool_preresume(struct dm_target *ti)
2024 {
2025         int r;
2026         struct pool_c *pt = ti->private;
2027         struct pool *pool = pt->pool;
2028         sector_t data_size = ti->len;
2029         dm_block_t sb_data_size;
2030
2031         /*
2032          * Take control of the pool object.
2033          */
2034         r = bind_control_target(pool, ti);
2035         if (r)
2036                 return r;
2037
2038         (void) sector_div(data_size, pool->sectors_per_block);
2039
2040         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2041         if (r) {
2042                 DMERR("failed to retrieve data device size");
2043                 return r;
2044         }
2045
2046         if (data_size < sb_data_size) {
2047                 DMERR("pool target too small, is %llu blocks (expected %llu)",
2048                       (unsigned long long)data_size, sb_data_size);
2049                 return -EINVAL;
2050
2051         } else if (data_size > sb_data_size) {
2052                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2053                 if (r) {
2054                         DMERR("failed to resize data device");
2055                         /* FIXME Stricter than necessary: Rollback transaction instead here */
2056                         set_pool_mode(pool, PM_READ_ONLY);
2057                         return r;
2058                 }
2059
2060                 (void) commit_or_fallback(pool);
2061         }
2062
2063         return 0;
2064 }
2065
2066 static void pool_resume(struct dm_target *ti)
2067 {
2068         struct pool_c *pt = ti->private;
2069         struct pool *pool = pt->pool;
2070         unsigned long flags;
2071
2072         spin_lock_irqsave(&pool->lock, flags);
2073         pool->low_water_triggered = 0;
2074         pool->no_free_space = 0;
2075         __requeue_bios(pool);
2076         spin_unlock_irqrestore(&pool->lock, flags);
2077
2078         do_waker(&pool->waker.work);
2079 }
2080
2081 static void pool_postsuspend(struct dm_target *ti)
2082 {
2083         struct pool_c *pt = ti->private;
2084         struct pool *pool = pt->pool;
2085
2086         cancel_delayed_work(&pool->waker);
2087         flush_workqueue(pool->wq);
2088         (void) commit_or_fallback(pool);
2089 }
2090
2091 static int check_arg_count(unsigned argc, unsigned args_required)
2092 {
2093         if (argc != args_required) {
2094                 DMWARN("Message received with %u arguments instead of %u.",
2095                        argc, args_required);
2096                 return -EINVAL;
2097         }
2098
2099         return 0;
2100 }
2101
2102 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2103 {
2104         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2105             *dev_id <= MAX_DEV_ID)
2106                 return 0;
2107
2108         if (warning)
2109                 DMWARN("Message received with invalid device id: %s", arg);
2110
2111         return -EINVAL;
2112 }
2113
2114 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2115 {
2116         dm_thin_id dev_id;
2117         int r;
2118
2119         r = check_arg_count(argc, 2);
2120         if (r)
2121                 return r;
2122
2123         r = read_dev_id(argv[1], &dev_id, 1);
2124         if (r)
2125                 return r;
2126
2127         r = dm_pool_create_thin(pool->pmd, dev_id);
2128         if (r) {
2129                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2130                        argv[1]);
2131                 return r;
2132         }
2133
2134         return 0;
2135 }
2136
2137 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2138 {
2139         dm_thin_id dev_id;
2140         dm_thin_id origin_dev_id;
2141         int r;
2142
2143         r = check_arg_count(argc, 3);
2144         if (r)
2145                 return r;
2146
2147         r = read_dev_id(argv[1], &dev_id, 1);
2148         if (r)
2149                 return r;
2150
2151         r = read_dev_id(argv[2], &origin_dev_id, 1);
2152         if (r)
2153                 return r;
2154
2155         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2156         if (r) {
2157                 DMWARN("Creation of new snapshot %s of device %s failed.",
2158                        argv[1], argv[2]);
2159                 return r;
2160         }
2161
2162         return 0;
2163 }
2164
2165 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2166 {
2167         dm_thin_id dev_id;
2168         int r;
2169
2170         r = check_arg_count(argc, 2);
2171         if (r)
2172                 return r;
2173
2174         r = read_dev_id(argv[1], &dev_id, 1);
2175         if (r)
2176                 return r;
2177
2178         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2179         if (r)
2180                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2181
2182         return r;
2183 }
2184
2185 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2186 {
2187         dm_thin_id old_id, new_id;
2188         int r;
2189
2190         r = check_arg_count(argc, 3);
2191         if (r)
2192                 return r;
2193
2194         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2195                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2196                 return -EINVAL;
2197         }
2198
2199         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2200                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2201                 return -EINVAL;
2202         }
2203
2204         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2205         if (r) {
2206                 DMWARN("Failed to change transaction id from %s to %s.",
2207                        argv[1], argv[2]);
2208                 return r;
2209         }
2210
2211         return 0;
2212 }
2213
2214 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2215 {
2216         int r;
2217
2218         r = check_arg_count(argc, 1);
2219         if (r)
2220                 return r;
2221
2222         (void) commit_or_fallback(pool);
2223
2224         r = dm_pool_reserve_metadata_snap(pool->pmd);
2225         if (r)
2226                 DMWARN("reserve_metadata_snap message failed.");
2227
2228         return r;
2229 }
2230
2231 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2232 {
2233         int r;
2234
2235         r = check_arg_count(argc, 1);
2236         if (r)
2237                 return r;
2238
2239         r = dm_pool_release_metadata_snap(pool->pmd);
2240         if (r)
2241                 DMWARN("release_metadata_snap message failed.");
2242
2243         return r;
2244 }
2245
2246 /*
2247  * Messages supported:
2248  *   create_thin        <dev_id>
2249  *   create_snap        <dev_id> <origin_id>
2250  *   delete             <dev_id>
2251  *   trim               <dev_id> <new_size_in_sectors>
2252  *   set_transaction_id <current_trans_id> <new_trans_id>
2253  *   reserve_metadata_snap
2254  *   release_metadata_snap
2255  */
2256 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2257 {
2258         int r = -EINVAL;
2259         struct pool_c *pt = ti->private;
2260         struct pool *pool = pt->pool;
2261
2262         if (!strcasecmp(argv[0], "create_thin"))
2263                 r = process_create_thin_mesg(argc, argv, pool);
2264
2265         else if (!strcasecmp(argv[0], "create_snap"))
2266                 r = process_create_snap_mesg(argc, argv, pool);
2267
2268         else if (!strcasecmp(argv[0], "delete"))
2269                 r = process_delete_mesg(argc, argv, pool);
2270
2271         else if (!strcasecmp(argv[0], "set_transaction_id"))
2272                 r = process_set_transaction_id_mesg(argc, argv, pool);
2273
2274         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2275                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2276
2277         else if (!strcasecmp(argv[0], "release_metadata_snap"))
2278                 r = process_release_metadata_snap_mesg(argc, argv, pool);
2279
2280         else
2281                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2282
2283         if (!r)
2284                 (void) commit_or_fallback(pool);
2285
2286         return r;
2287 }
2288
2289 static void emit_flags(struct pool_features *pf, char *result,
2290                        unsigned sz, unsigned maxlen)
2291 {
2292         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2293                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2294         DMEMIT("%u ", count);
2295
2296         if (!pf->zero_new_blocks)
2297                 DMEMIT("skip_block_zeroing ");
2298
2299         if (!pf->discard_enabled)
2300                 DMEMIT("ignore_discard ");
2301
2302         if (!pf->discard_passdown)
2303                 DMEMIT("no_discard_passdown ");
2304
2305         if (pf->mode == PM_READ_ONLY)
2306                 DMEMIT("read_only ");
2307 }
2308
2309 /*
2310  * Status line is:
2311  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2312  *    <used data sectors>/<total data sectors> <held metadata root>
2313  */
2314 static int pool_status(struct dm_target *ti, status_type_t type,
2315                        unsigned status_flags, char *result, unsigned maxlen)
2316 {
2317         int r;
2318         unsigned sz = 0;
2319         uint64_t transaction_id;
2320         dm_block_t nr_free_blocks_data;
2321         dm_block_t nr_free_blocks_metadata;
2322         dm_block_t nr_blocks_data;
2323         dm_block_t nr_blocks_metadata;
2324         dm_block_t held_root;
2325         char buf[BDEVNAME_SIZE];
2326         char buf2[BDEVNAME_SIZE];
2327         struct pool_c *pt = ti->private;
2328         struct pool *pool = pt->pool;
2329
2330         switch (type) {
2331         case STATUSTYPE_INFO:
2332                 if (get_pool_mode(pool) == PM_FAIL) {
2333                         DMEMIT("Fail");
2334                         break;
2335                 }
2336
2337                 /* Commit to ensure statistics aren't out-of-date */
2338                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2339                         (void) commit_or_fallback(pool);
2340
2341                 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2342                                                         &transaction_id);
2343                 if (r)
2344                         return r;
2345
2346                 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2347                                                           &nr_free_blocks_metadata);
2348                 if (r)
2349                         return r;
2350
2351                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2352                 if (r)
2353                         return r;
2354
2355                 r = dm_pool_get_free_block_count(pool->pmd,
2356                                                  &nr_free_blocks_data);
2357                 if (r)
2358                         return r;
2359
2360                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2361                 if (r)
2362                         return r;
2363
2364                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2365                 if (r)
2366                         return r;
2367
2368                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2369                        (unsigned long long)transaction_id,
2370                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2371                        (unsigned long long)nr_blocks_metadata,
2372                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2373                        (unsigned long long)nr_blocks_data);
2374
2375                 if (held_root)
2376                         DMEMIT("%llu ", held_root);
2377                 else
2378                         DMEMIT("- ");
2379
2380                 if (pool->pf.mode == PM_READ_ONLY)
2381                         DMEMIT("ro ");
2382                 else
2383                         DMEMIT("rw ");
2384
2385                 if (!pool->pf.discard_enabled)
2386                         DMEMIT("ignore_discard");
2387                 else if (pool->pf.discard_passdown)
2388                         DMEMIT("discard_passdown");
2389                 else
2390                         DMEMIT("no_discard_passdown");
2391
2392                 break;
2393
2394         case STATUSTYPE_TABLE:
2395                 DMEMIT("%s %s %lu %llu ",
2396                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2397                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2398                        (unsigned long)pool->sectors_per_block,
2399                        (unsigned long long)pt->low_water_blocks);
2400                 emit_flags(&pt->requested_pf, result, sz, maxlen);
2401                 break;
2402         }
2403
2404         return 0;
2405 }
2406
2407 static int pool_iterate_devices(struct dm_target *ti,
2408                                 iterate_devices_callout_fn fn, void *data)
2409 {
2410         struct pool_c *pt = ti->private;
2411
2412         return fn(ti, pt->data_dev, 0, ti->len, data);
2413 }
2414
2415 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2416                       struct bio_vec *biovec, int max_size)
2417 {
2418         struct pool_c *pt = ti->private;
2419         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2420
2421         if (!q->merge_bvec_fn)
2422                 return max_size;
2423
2424         bvm->bi_bdev = pt->data_dev->bdev;
2425
2426         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2427 }
2428
2429 static bool block_size_is_power_of_two(struct pool *pool)
2430 {
2431         return pool->sectors_per_block_shift >= 0;
2432 }
2433
2434 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2435 {
2436         struct pool *pool = pt->pool;
2437         struct queue_limits *data_limits;
2438
2439         limits->max_discard_sectors = pool->sectors_per_block;
2440
2441         /*
2442          * discard_granularity is just a hint, and not enforced.
2443          */
2444         if (pt->adjusted_pf.discard_passdown) {
2445                 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2446                 limits->discard_granularity = data_limits->discard_granularity;
2447         } else if (block_size_is_power_of_two(pool))
2448                 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2449         else
2450                 /*
2451                  * Use largest power of 2 that is a factor of sectors_per_block
2452                  * but at least DATA_DEV_BLOCK_SIZE_MIN_SECTORS.
2453                  */
2454                 limits->discard_granularity = max(1 << (ffs(pool->sectors_per_block) - 1),
2455                                                   DATA_DEV_BLOCK_SIZE_MIN_SECTORS) << SECTOR_SHIFT;
2456 }
2457
2458 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2459 {
2460         struct pool_c *pt = ti->private;
2461         struct pool *pool = pt->pool;
2462
2463         blk_limits_io_min(limits, 0);
2464         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2465
2466         /*
2467          * pt->adjusted_pf is a staging area for the actual features to use.
2468          * They get transferred to the live pool in bind_control_target()
2469          * called from pool_preresume().
2470          */
2471         if (!pt->adjusted_pf.discard_enabled)
2472                 return;
2473
2474         disable_passdown_if_not_supported(pt);
2475
2476         set_discard_limits(pt, limits);
2477 }
2478
2479 static struct target_type pool_target = {
2480         .name = "thin-pool",
2481         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2482                     DM_TARGET_IMMUTABLE,
2483         .version = {1, 6, 0},
2484         .module = THIS_MODULE,
2485         .ctr = pool_ctr,
2486         .dtr = pool_dtr,
2487         .map = pool_map,
2488         .postsuspend = pool_postsuspend,
2489         .preresume = pool_preresume,
2490         .resume = pool_resume,
2491         .message = pool_message,
2492         .status = pool_status,
2493         .merge = pool_merge,
2494         .iterate_devices = pool_iterate_devices,
2495         .io_hints = pool_io_hints,
2496 };
2497
2498 /*----------------------------------------------------------------
2499  * Thin target methods
2500  *--------------------------------------------------------------*/
2501 static void thin_dtr(struct dm_target *ti)
2502 {
2503         struct thin_c *tc = ti->private;
2504
2505         mutex_lock(&dm_thin_pool_table.mutex);
2506
2507         __pool_dec(tc->pool);
2508         dm_pool_close_thin_device(tc->td);
2509         dm_put_device(ti, tc->pool_dev);
2510         if (tc->origin_dev)
2511                 dm_put_device(ti, tc->origin_dev);
2512         kfree(tc);
2513
2514         mutex_unlock(&dm_thin_pool_table.mutex);
2515 }
2516
2517 /*
2518  * Thin target parameters:
2519  *
2520  * <pool_dev> <dev_id> [origin_dev]
2521  *
2522  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2523  * dev_id: the internal device identifier
2524  * origin_dev: a device external to the pool that should act as the origin
2525  *
2526  * If the pool device has discards disabled, they get disabled for the thin
2527  * device as well.
2528  */
2529 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2530 {
2531         int r;
2532         struct thin_c *tc;
2533         struct dm_dev *pool_dev, *origin_dev;
2534         struct mapped_device *pool_md;
2535
2536         mutex_lock(&dm_thin_pool_table.mutex);
2537
2538         if (argc != 2 && argc != 3) {
2539                 ti->error = "Invalid argument count";
2540                 r = -EINVAL;
2541                 goto out_unlock;
2542         }
2543
2544         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2545         if (!tc) {
2546                 ti->error = "Out of memory";
2547                 r = -ENOMEM;
2548                 goto out_unlock;
2549         }
2550
2551         if (argc == 3) {
2552                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2553                 if (r) {
2554                         ti->error = "Error opening origin device";
2555                         goto bad_origin_dev;
2556                 }
2557                 tc->origin_dev = origin_dev;
2558         }
2559
2560         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2561         if (r) {
2562                 ti->error = "Error opening pool device";
2563                 goto bad_pool_dev;
2564         }
2565         tc->pool_dev = pool_dev;
2566
2567         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2568                 ti->error = "Invalid device id";
2569                 r = -EINVAL;
2570                 goto bad_common;
2571         }
2572
2573         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2574         if (!pool_md) {
2575                 ti->error = "Couldn't get pool mapped device";
2576                 r = -EINVAL;
2577                 goto bad_common;
2578         }
2579
2580         tc->pool = __pool_table_lookup(pool_md);
2581         if (!tc->pool) {
2582                 ti->error = "Couldn't find pool object";
2583                 r = -EINVAL;
2584                 goto bad_pool_lookup;
2585         }
2586         __pool_inc(tc->pool);
2587
2588         if (get_pool_mode(tc->pool) == PM_FAIL) {
2589                 ti->error = "Couldn't open thin device, Pool is in fail mode";
2590                 goto bad_thin_open;
2591         }
2592
2593         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2594         if (r) {
2595                 ti->error = "Couldn't open thin internal device";
2596                 goto bad_thin_open;
2597         }
2598
2599         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2600         if (r)
2601                 goto bad_thin_open;
2602
2603         ti->num_flush_requests = 1;
2604         ti->flush_supported = true;
2605
2606         /* In case the pool supports discards, pass them on. */
2607         if (tc->pool->pf.discard_enabled) {
2608                 ti->discards_supported = true;
2609                 ti->num_discard_requests = 1;
2610                 ti->discard_zeroes_data_unsupported = true;
2611                 /* Discard requests must be split on a block boundary */
2612                 ti->split_discard_requests = true;
2613         }
2614
2615         dm_put(pool_md);
2616
2617         mutex_unlock(&dm_thin_pool_table.mutex);
2618
2619         return 0;
2620
2621 bad_thin_open:
2622         __pool_dec(tc->pool);
2623 bad_pool_lookup:
2624         dm_put(pool_md);
2625 bad_common:
2626         dm_put_device(ti, tc->pool_dev);
2627 bad_pool_dev:
2628         if (tc->origin_dev)
2629                 dm_put_device(ti, tc->origin_dev);
2630 bad_origin_dev:
2631         kfree(tc);
2632 out_unlock:
2633         mutex_unlock(&dm_thin_pool_table.mutex);
2634
2635         return r;
2636 }
2637
2638 static int thin_map(struct dm_target *ti, struct bio *bio,
2639                     union map_info *map_context)
2640 {
2641         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2642
2643         return thin_bio_map(ti, bio, map_context);
2644 }
2645
2646 static int thin_endio(struct dm_target *ti,
2647                       struct bio *bio, int err,
2648                       union map_info *map_context)
2649 {
2650         unsigned long flags;
2651         struct dm_thin_endio_hook *h = map_context->ptr;
2652         struct list_head work;
2653         struct dm_thin_new_mapping *m, *tmp;
2654         struct pool *pool = h->tc->pool;
2655
2656         if (h->shared_read_entry) {
2657                 INIT_LIST_HEAD(&work);
2658                 dm_deferred_entry_dec(h->shared_read_entry, &work);
2659
2660                 spin_lock_irqsave(&pool->lock, flags);
2661                 list_for_each_entry_safe(m, tmp, &work, list) {
2662                         list_del(&m->list);
2663                         m->quiesced = 1;
2664                         __maybe_add_mapping(m);
2665                 }
2666                 spin_unlock_irqrestore(&pool->lock, flags);
2667         }
2668
2669         if (h->all_io_entry) {
2670                 INIT_LIST_HEAD(&work);
2671                 dm_deferred_entry_dec(h->all_io_entry, &work);
2672                 if (!list_empty(&work)) {
2673                         spin_lock_irqsave(&pool->lock, flags);
2674                         list_for_each_entry_safe(m, tmp, &work, list)
2675                                 list_add(&m->list, &pool->prepared_discards);
2676                         spin_unlock_irqrestore(&pool->lock, flags);
2677                         wake_worker(pool);
2678                 }
2679         }
2680
2681         mempool_free(h, pool->endio_hook_pool);
2682
2683         return 0;
2684 }
2685
2686 static void thin_postsuspend(struct dm_target *ti)
2687 {
2688         if (dm_noflush_suspending(ti))
2689                 requeue_io((struct thin_c *)ti->private);
2690 }
2691
2692 /*
2693  * <nr mapped sectors> <highest mapped sector>
2694  */
2695 static int thin_status(struct dm_target *ti, status_type_t type,
2696                        unsigned status_flags, char *result, unsigned maxlen)
2697 {
2698         int r;
2699         ssize_t sz = 0;
2700         dm_block_t mapped, highest;
2701         char buf[BDEVNAME_SIZE];
2702         struct thin_c *tc = ti->private;
2703
2704         if (get_pool_mode(tc->pool) == PM_FAIL) {
2705                 DMEMIT("Fail");
2706                 return 0;
2707         }
2708
2709         if (!tc->td)
2710                 DMEMIT("-");
2711         else {
2712                 switch (type) {
2713                 case STATUSTYPE_INFO:
2714                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2715                         if (r)
2716                                 return r;
2717
2718                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2719                         if (r < 0)
2720                                 return r;
2721
2722                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2723                         if (r)
2724                                 DMEMIT("%llu", ((highest + 1) *
2725                                                 tc->pool->sectors_per_block) - 1);
2726                         else
2727                                 DMEMIT("-");
2728                         break;
2729
2730                 case STATUSTYPE_TABLE:
2731                         DMEMIT("%s %lu",
2732                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2733                                (unsigned long) tc->dev_id);
2734                         if (tc->origin_dev)
2735                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2736                         break;
2737                 }
2738         }
2739
2740         return 0;
2741 }
2742
2743 static int thin_iterate_devices(struct dm_target *ti,
2744                                 iterate_devices_callout_fn fn, void *data)
2745 {
2746         sector_t blocks;
2747         struct thin_c *tc = ti->private;
2748         struct pool *pool = tc->pool;
2749
2750         /*
2751          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2752          * we follow a more convoluted path through to the pool's target.
2753          */
2754         if (!pool->ti)
2755                 return 0;       /* nothing is bound */
2756
2757         blocks = pool->ti->len;
2758         (void) sector_div(blocks, pool->sectors_per_block);
2759         if (blocks)
2760                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2761
2762         return 0;
2763 }
2764
2765 /*
2766  * A thin device always inherits its queue limits from its pool.
2767  */
2768 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2769 {
2770         struct thin_c *tc = ti->private;
2771
2772         *limits = bdev_get_queue(tc->pool_dev->bdev)->limits;
2773 }
2774
2775 static struct target_type thin_target = {
2776         .name = "thin",
2777         .version = {1, 5, 0},
2778         .module = THIS_MODULE,
2779         .ctr = thin_ctr,
2780         .dtr = thin_dtr,
2781         .map = thin_map,
2782         .end_io = thin_endio,
2783         .postsuspend = thin_postsuspend,
2784         .status = thin_status,
2785         .iterate_devices = thin_iterate_devices,
2786         .io_hints = thin_io_hints,
2787 };
2788
2789 /*----------------------------------------------------------------*/
2790
2791 static int __init dm_thin_init(void)
2792 {
2793         int r;
2794
2795         pool_table_init();
2796
2797         r = dm_register_target(&thin_target);
2798         if (r)
2799                 return r;
2800
2801         r = dm_register_target(&pool_target);
2802         if (r)
2803                 goto bad_pool_target;
2804
2805         r = -ENOMEM;
2806
2807         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2808         if (!_new_mapping_cache)
2809                 goto bad_new_mapping_cache;
2810
2811         _endio_hook_cache = KMEM_CACHE(dm_thin_endio_hook, 0);
2812         if (!_endio_hook_cache)
2813                 goto bad_endio_hook_cache;
2814
2815         return 0;
2816
2817 bad_endio_hook_cache:
2818         kmem_cache_destroy(_new_mapping_cache);
2819 bad_new_mapping_cache:
2820         dm_unregister_target(&pool_target);
2821 bad_pool_target:
2822         dm_unregister_target(&thin_target);
2823
2824         return r;
2825 }
2826
2827 static void dm_thin_exit(void)
2828 {
2829         dm_unregister_target(&thin_target);
2830         dm_unregister_target(&pool_target);
2831
2832         kmem_cache_destroy(_new_mapping_cache);
2833         kmem_cache_destroy(_endio_hook_cache);
2834 }
2835
2836 module_init(dm_thin_init);
2837 module_exit(dm_thin_exit);
2838
2839 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2840 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2841 MODULE_LICENSE("GPL");