drivers/md/dm-thin.c

   1 /*
   2  * Copyright (C) 2011 Red Hat UK.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-thin-metadata.h"
   8
   9 #include <linux/device-mapper.h>
  10 #include <linux/dm-io.h>
  11 #include <linux/dm-kcopyd.h>
  12 #include <linux/list.h>
  13 #include <linux/init.h>
  14 #include <linux/module.h>
  15 #include <linux/slab.h>
  16
  17 #define DM_MSG_PREFIX   "thin"
  18
  19 /*
  20  * Tunable constants
  21  */
  22 #define ENDIO_HOOK_POOL_SIZE 10240
  23 #define DEFERRED_SET_SIZE 64
  24 #define MAPPING_POOL_SIZE 1024
  25 #define PRISON_CELLS 1024
  26 #define COMMIT_PERIOD HZ
  27
  28 /*
  29  * The block size of the device holding pool data must be
  30  * between 64KB and 1GB.
  31  */
  32 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
  33 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
  34
  35 /*
  36  * Device id is restricted to 24 bits.
  37  */
  38 #define MAX_DEV_ID ((1 << 24) - 1)
  39
  40 /*
  41  * How do we handle breaking sharing of data blocks?
  42  * =================================================
  43  *
  44  * We use a standard copy-on-write btree to store the mappings for the
  45  * devices (note I'm talking about copy-on-write of the metadata here, not
  46  * the data).  When you take an internal snapshot you clone the root node
  47  * of the origin btree.  After this there is no concept of an origin or a
  48  * snapshot.  They are just two device trees that happen to point to the
  49  * same data blocks.
  50  *
  51  * When we get a write in we decide if it's to a shared data block using
  52  * some timestamp magic.  If it is, we have to break sharing.
  53  *
  54  * Let's say we write to a shared block in what was the origin.  The
  55  * steps are:
  56  *
  57  * i) plug io further to this physical block. (see bio_prison code).
  58  *
  59  * ii) quiesce any read io to that shared data block.  Obviously
  60  * including all devices that share this block.  (see deferred_set code)
  61  *
  62  * iii) copy the data block to a newly allocate block.  This step can be
  63  * missed out if the io covers the block. (schedule_copy).
  64  *
  65  * iv) insert the new mapping into the origin's btree
  66  * (process_prepared_mapping).  This act of inserting breaks some
  67  * sharing of btree nodes between the two devices.  Breaking sharing only
  68  * effects the btree of that specific device.  Btrees for the other
  69  * devices that share the block never change.  The btree for the origin
  70  * device as it was after the last commit is untouched, ie. we're using
  71  * persistent data structures in the functional programming sense.
  72  *
  73  * v) unplug io to this physical block, including the io that triggered
  74  * the breaking of sharing.
  75  *
  76  * Steps (ii) and (iii) occur in parallel.
  77  *
  78  * The metadata _doesn't_ need to be committed before the io continues.  We
  79  * get away with this because the io is always written to a _new_ block.
  80  * If there's a crash, then:
  81  *
  82  * - The origin mapping will point to the old origin block (the shared
  83  * one).  This will contain the data as it was before the io that triggered
  84  * the breaking of sharing came in.
  85  *
  86  * - The snap mapping still points to the old block.  As it would after
  87  * the commit.
  88  *
  89  * The downside of this scheme is the timestamp magic isn't perfect, and
  90  * will continue to think that data block in the snapshot device is shared
  91  * even after the write to the origin has broken sharing.  I suspect data
  92  * blocks will typically be shared by many different devices, so we're
  93  * breaking sharing n + 1 times, rather than n, where n is the number of
  94  * devices that reference this data block.  At the moment I think the
  95  * benefits far, far outweigh the disadvantages.
  96  */
  97
  98 /*----------------------------------------------------------------*/
  99
 100 /*
 101  * Sometimes we can't deal with a bio straight away.  We put them in prison
 102  * where they can't cause any mischief.  Bios are put in a cell identified
 103  * by a key, multiple bios can be in the same cell.  When the cell is
 104  * subsequently unlocked the bios become available.
 105  */
 106 struct bio_prison;
 107
 108 struct cell_key {
 109         int virtual;
 110         dm_thin_id dev;
 111         dm_block_t block;
 112 };
 113
 114 struct cell {
 115         struct hlist_node list;
 116         struct bio_prison *prison;
 117         struct cell_key key;
 118         struct bio *holder;
 119         struct bio_list bios;
 120 };
 121
 122 struct bio_prison {
 123         spinlock_t lock;
 124         mempool_t *cell_pool;
 125
 126         unsigned nr_buckets;
 127         unsigned hash_mask;
 128         struct hlist_head *cells;
 129 };
 130
 131 static uint32_t calc_nr_buckets(unsigned nr_cells)
 132 {
 133         uint32_t n = 128;
 134
 135         nr_cells /= 4;
 136         nr_cells = min(nr_cells, 8192u);
 137
 138         while (n < nr_cells)
 139                 n <<= 1;
 140
 141         return n;
 142 }
 143
 144 /*
 145  * @nr_cells should be the number of cells you want in use _concurrently_.
 146  * Don't confuse it with the number of distinct keys.
 147  */
 148 static struct bio_prison *prison_create(unsigned nr_cells)
 149 {
 150         unsigned i;
 151         uint32_t nr_buckets = calc_nr_buckets(nr_cells);
 152         size_t len = sizeof(struct bio_prison) +
 153                 (sizeof(struct hlist_head) * nr_buckets);
 154         struct bio_prison *prison = kmalloc(len, GFP_KERNEL);
 155
 156         if (!prison)
 157                 return NULL;
 158
 159         spin_lock_init(&prison->lock);
 160         prison->cell_pool = mempool_create_kmalloc_pool(nr_cells,
 161                                                         sizeof(struct cell));
 162         if (!prison->cell_pool) {
 163                 kfree(prison);
 164                 return NULL;
 165         }
 166
 167         prison->nr_buckets = nr_buckets;
 168         prison->hash_mask = nr_buckets - 1;
 169         prison->cells = (struct hlist_head *) (prison + 1);
 170         for (i = 0; i < nr_buckets; i++)
 171                 INIT_HLIST_HEAD(prison->cells + i);
 172
 173         return prison;
 174 }
 175
 176 static void prison_destroy(struct bio_prison *prison)
 177 {
 178         mempool_destroy(prison->cell_pool);
 179         kfree(prison);
 180 }
 181
 182 static uint32_t hash_key(struct bio_prison *prison, struct cell_key *key)
 183 {
 184         const unsigned long BIG_PRIME = 4294967291UL;
 185         uint64_t hash = key->block * BIG_PRIME;
 186
 187         return (uint32_t) (hash & prison->hash_mask);
 188 }
 189
 190 static int keys_equal(struct cell_key *lhs, struct cell_key *rhs)
 191 {
 192                return (lhs->virtual == rhs->virtual) &&
 193                        (lhs->dev == rhs->dev) &&
 194                        (lhs->block == rhs->block);
 195 }
 196
 197 static struct cell *__search_bucket(struct hlist_head *bucket,
 198                                     struct cell_key *key)
 199 {
 200         struct cell *cell;
 201         struct hlist_node *tmp;
 202
 203         hlist_for_each_entry(cell, tmp, bucket, list)
 204                 if (keys_equal(&cell->key, key))
 205                         return cell;
 206
 207         return NULL;
 208 }
 209
 210 /*
 211  * This may block if a new cell needs allocating.  You must ensure that
 212  * cells will be unlocked even if the calling thread is blocked.
 213  *
 214  * Returns 1 if the cell was already held, 0 if @inmate is the new holder.
 215  */
 216 static int bio_detain(struct bio_prison *prison, struct cell_key *key,
 217                       struct bio *inmate, struct cell **ref)
 218 {
 219         int r = 1;
 220         unsigned long flags;
 221         uint32_t hash = hash_key(prison, key);
 222         struct cell *cell, *cell2;
 223
 224         BUG_ON(hash > prison->nr_buckets);
 225
 226         spin_lock_irqsave(&prison->lock, flags);
 227
 228         cell = __search_bucket(prison->cells + hash, key);
 229         if (cell) {
 230                 bio_list_add(&cell->bios, inmate);
 231                 goto out;
 232         }
 233
 234         /*
 235          * Allocate a new cell
 236          */
 237         spin_unlock_irqrestore(&prison->lock, flags);
 238         cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
 239         spin_lock_irqsave(&prison->lock, flags);
 240
 241         /*
 242          * We've been unlocked, so we have to double check that
 243          * nobody else has inserted this cell in the meantime.
 244          */
 245         cell = __search_bucket(prison->cells + hash, key);
 246         if (cell) {
 247                 mempool_free(cell2, prison->cell_pool);
 248                 bio_list_add(&cell->bios, inmate);
 249                 goto out;
 250         }
 251
 252         /*
 253          * Use new cell.
 254          */
 255         cell = cell2;
 256
 257         cell->prison = prison;
 258         memcpy(&cell->key, key, sizeof(cell->key));
 259         cell->holder = inmate;
 260         bio_list_init(&cell->bios);
 261         hlist_add_head(&cell->list, prison->cells + hash);
 262
 263         r = 0;
 264
 265 out:
 266         spin_unlock_irqrestore(&prison->lock, flags);
 267
 268         *ref = cell;
 269
 270         return r;
 271 }
 272
 273 /*
 274  * @inmates must have been initialised prior to this call
 275  */
 276 static void __cell_release(struct cell *cell, struct bio_list *inmates)
 277 {
 278         struct bio_prison *prison = cell->prison;
 279
 280         hlist_del(&cell->list);
 281
 282         bio_list_add(inmates, cell->holder);
 283         bio_list_merge(inmates, &cell->bios);
 284
 285         mempool_free(cell, prison->cell_pool);
 286 }
 287
 288 static void cell_release(struct cell *cell, struct bio_list *bios)
 289 {
 290         unsigned long flags;
 291         struct bio_prison *prison = cell->prison;
 292
 293         spin_lock_irqsave(&prison->lock, flags);
 294         __cell_release(cell, bios);
 295         spin_unlock_irqrestore(&prison->lock, flags);
 296 }
 297
 298 /*
 299  * There are a couple of places where we put a bio into a cell briefly
 300  * before taking it out again.  In these situations we know that no other
 301  * bio may be in the cell.  This function releases the cell, and also does
 302  * a sanity check.
 303  */
 304 static void __cell_release_singleton(struct cell *cell, struct bio *bio)
 305 {
 306         hlist_del(&cell->list);
 307         BUG_ON(cell->holder != bio);
 308         BUG_ON(!bio_list_empty(&cell->bios));
 309 }
 310
 311 static void cell_release_singleton(struct cell *cell, struct bio *bio)
 312 {
 313         unsigned long flags;
 314         struct bio_prison *prison = cell->prison;
 315
 316         spin_lock_irqsave(&prison->lock, flags);
 317         __cell_release_singleton(cell, bio);
 318         spin_unlock_irqrestore(&prison->lock, flags);
 319 }
 320
 321 /*
 322  * Sometimes we don't want the holder, just the additional bios.
 323  */
 324 static void __cell_release_no_holder(struct cell *cell, struct bio_list *inmates)
 325 {
 326         struct bio_prison *prison = cell->prison;
 327
 328         hlist_del(&cell->list);
 329         bio_list_merge(inmates, &cell->bios);
 330
 331         mempool_free(cell, prison->cell_pool);
 332 }
 333
 334 static void cell_release_no_holder(struct cell *cell, struct bio_list *inmates)
 335 {
 336         unsigned long flags;
 337         struct bio_prison *prison = cell->prison;
 338
 339         spin_lock_irqsave(&prison->lock, flags);
 340         __cell_release_no_holder(cell, inmates);
 341         spin_unlock_irqrestore(&prison->lock, flags);
 342 }
 343
 344 static void cell_error(struct cell *cell)
 345 {
 346         struct bio_prison *prison = cell->prison;
 347         struct bio_list bios;
 348         struct bio *bio;
 349         unsigned long flags;
 350
 351         bio_list_init(&bios);
 352
 353         spin_lock_irqsave(&prison->lock, flags);
 354         __cell_release(cell, &bios);
 355         spin_unlock_irqrestore(&prison->lock, flags);
 356
 357         while ((bio = bio_list_pop(&bios)))
 358                 bio_io_error(bio);
 359 }
 360
 361 /*----------------------------------------------------------------*/
 362
 363 /*
 364  * We use the deferred set to keep track of pending reads to shared blocks.
 365  * We do this to ensure the new mapping caused by a write isn't performed
 366  * until these prior reads have completed.  Otherwise the insertion of the
 367  * new mapping could free the old block that the read bios are mapped to.
 368  */
 369
 370 struct deferred_set;
 371 struct deferred_entry {
 372         struct deferred_set *ds;
 373         unsigned count;
 374         struct list_head work_items;
 375 };
 376
 377 struct deferred_set {
 378         spinlock_t lock;
 379         unsigned current_entry;
 380         unsigned sweeper;
 381         struct deferred_entry entries[DEFERRED_SET_SIZE];
 382 };
 383
 384 static void ds_init(struct deferred_set *ds)
 385 {
 386         int i;
 387
 388         spin_lock_init(&ds->lock);
 389         ds->current_entry = 0;
 390         ds->sweeper = 0;
 391         for (i = 0; i < DEFERRED_SET_SIZE; i++) {
 392                 ds->entries[i].ds = ds;
 393                 ds->entries[i].count = 0;
 394                 INIT_LIST_HEAD(&ds->entries[i].work_items);
 395         }
 396 }
 397
 398 static struct deferred_entry *ds_inc(struct deferred_set *ds)
 399 {
 400         unsigned long flags;
 401         struct deferred_entry *entry;
 402
 403         spin_lock_irqsave(&ds->lock, flags);
 404         entry = ds->entries + ds->current_entry;
 405         entry->count++;
 406         spin_unlock_irqrestore(&ds->lock, flags);
 407
 408         return entry;
 409 }
 410
 411 static unsigned ds_next(unsigned index)
 412 {
 413         return (index + 1) % DEFERRED_SET_SIZE;
 414 }
 415
 416 static void __sweep(struct deferred_set *ds, struct list_head *head)
 417 {
 418         while ((ds->sweeper != ds->current_entry) &&
 419                !ds->entries[ds->sweeper].count) {
 420                 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
 421                 ds->sweeper = ds_next(ds->sweeper);
 422         }
 423
 424         if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
 425                 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
 426 }
 427
 428 static void ds_dec(struct deferred_entry *entry, struct list_head *head)
 429 {
 430         unsigned long flags;
 431
 432         spin_lock_irqsave(&entry->ds->lock, flags);
 433         BUG_ON(!entry->count);
 434         --entry->count;
 435         __sweep(entry->ds, head);
 436         spin_unlock_irqrestore(&entry->ds->lock, flags);
 437 }
 438
 439 /*
 440  * Returns 1 if deferred or 0 if no pending items to delay job.
 441  */
 442 static int ds_add_work(struct deferred_set *ds, struct list_head *work)
 443 {
 444         int r = 1;
 445         unsigned long flags;
 446         unsigned next_entry;
 447
 448         spin_lock_irqsave(&ds->lock, flags);
 449         if ((ds->sweeper == ds->current_entry) &&
 450             !ds->entries[ds->current_entry].count)
 451                 r = 0;
 452         else {
 453                 list_add(work, &ds->entries[ds->current_entry].work_items);
 454                 next_entry = ds_next(ds->current_entry);
 455                 if (!ds->entries[next_entry].count)
 456                         ds->current_entry = next_entry;
 457         }
 458         spin_unlock_irqrestore(&ds->lock, flags);
 459
 460         return r;
 461 }
 462
 463 /*----------------------------------------------------------------*/
 464
 465 /*
 466  * Key building.
 467  */
 468 static void build_data_key(struct dm_thin_device *td,
 469                            dm_block_t b, struct cell_key *key)
 470 {
 471         key->virtual = 0;
 472         key->dev = dm_thin_dev_id(td);
 473         key->block = b;
 474 }
 475
 476 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
 477                               struct cell_key *key)
 478 {
 479         key->virtual = 1;
 480         key->dev = dm_thin_dev_id(td);
 481         key->block = b;
 482 }
 483
 484 /*----------------------------------------------------------------*/
 485
 486 /*
 487  * A pool device ties together a metadata device and a data device.  It
 488  * also provides the interface for creating and destroying internal
 489  * devices.
 490  */
 491 struct new_mapping;
 492 struct pool {
 493         struct list_head list;
 494         struct dm_target *ti;   /* Only set if a pool target is bound */
 495
 496         struct mapped_device *pool_md;
 497         struct block_device *md_dev;
 498         struct dm_pool_metadata *pmd;
 499
 500         uint32_t sectors_per_block;
 501         unsigned block_shift;
 502         dm_block_t offset_mask;
 503         dm_block_t low_water_blocks;
 504
 505         unsigned zero_new_blocks:1;
 506         unsigned low_water_triggered:1; /* A dm event has been sent */
 507         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
 508
 509         struct bio_prison *prison;
 510         struct dm_kcopyd_client *copier;
 511
 512         struct workqueue_struct *wq;
 513         struct work_struct worker;
 514         struct delayed_work waker;
 515
 516         unsigned ref_count;
 517         unsigned long last_commit_jiffies;
 518
 519         spinlock_t lock;
 520         struct bio_list deferred_bios;
 521         struct bio_list deferred_flush_bios;
 522         struct list_head prepared_mappings;
 523
 524         struct bio_list retry_on_resume_list;
 525
 526         struct deferred_set ds; /* FIXME: move to thin_c */
 527
 528         struct new_mapping *next_mapping;
 529         mempool_t *mapping_pool;
 530         mempool_t *endio_hook_pool;
 531 };
 532
 533 /*
 534  * Target context for a pool.
 535  */
 536 struct pool_c {
 537         struct dm_target *ti;
 538         struct pool *pool;
 539         struct dm_dev *data_dev;
 540         struct dm_dev *metadata_dev;
 541         struct dm_target_callbacks callbacks;
 542
 543         dm_block_t low_water_blocks;
 544         unsigned zero_new_blocks:1;
 545 };
 546
 547 /*
 548  * Target context for a thin.
 549  */
 550 struct thin_c {
 551         struct dm_dev *pool_dev;
 552         struct dm_dev *origin_dev;
 553         dm_thin_id dev_id;
 554
 555         struct pool *pool;
 556         struct dm_thin_device *td;
 557 };
 558
 559 /*----------------------------------------------------------------*/
 560
 561 /*
 562  * A global list of pools that uses a struct mapped_device as a key.
 563  */
 564 static struct dm_thin_pool_table {
 565         struct mutex mutex;
 566         struct list_head pools;
 567 } dm_thin_pool_table;
 568
 569 static void pool_table_init(void)
 570 {
 571         mutex_init(&dm_thin_pool_table.mutex);
 572         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
 573 }
 574
 575 static void __pool_table_insert(struct pool *pool)
 576 {
 577         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 578         list_add(&pool->list, &dm_thin_pool_table.pools);
 579 }
 580
 581 static void __pool_table_remove(struct pool *pool)
 582 {
 583         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 584         list_del(&pool->list);
 585 }
 586
 587 static struct pool *__pool_table_lookup(struct mapped_device *md)
 588 {
 589         struct pool *pool = NULL, *tmp;
 590
 591         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 592
 593         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 594                 if (tmp->pool_md == md) {
 595                         pool = tmp;
 596                         break;
 597                 }
 598         }
 599
 600         return pool;
 601 }
 602
 603 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
 604 {
 605         struct pool *pool = NULL, *tmp;
 606
 607         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 608
 609         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 610                 if (tmp->md_dev == md_dev) {
 611                         pool = tmp;
 612                         break;
 613                 }
 614         }
 615
 616         return pool;
 617 }
 618
 619 /*----------------------------------------------------------------*/
 620
 621 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
 622 {
 623         struct bio *bio;
 624         struct bio_list bios;
 625
 626         bio_list_init(&bios);
 627         bio_list_merge(&bios, master);
 628         bio_list_init(master);
 629
 630         while ((bio = bio_list_pop(&bios))) {
 631                 if (dm_get_mapinfo(bio)->ptr == tc)
 632                         bio_endio(bio, DM_ENDIO_REQUEUE);
 633                 else
 634                         bio_list_add(master, bio);
 635         }
 636 }
 637
 638 static void requeue_io(struct thin_c *tc)
 639 {
 640         struct pool *pool = tc->pool;
 641         unsigned long flags;
 642
 643         spin_lock_irqsave(&pool->lock, flags);
 644         __requeue_bio_list(tc, &pool->deferred_bios);
 645         __requeue_bio_list(tc, &pool->retry_on_resume_list);
 646         spin_unlock_irqrestore(&pool->lock, flags);
 647 }
 648
 649 /*
 650  * This section of code contains the logic for processing a thin device's IO.
 651  * Much of the code depends on pool object resources (lists, workqueues, etc)
 652  * but most is exclusively called from the thin target rather than the thin-pool
 653  * target.
 654  */
 655
 656 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
 657 {
 658         return bio->bi_sector >> tc->pool->block_shift;
 659 }
 660
 661 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
 662 {
 663         struct pool *pool = tc->pool;
 664
 665         bio->bi_bdev = tc->pool_dev->bdev;
 666         bio->bi_sector = (block << pool->block_shift) +
 667                 (bio->bi_sector & pool->offset_mask);
 668 }
 669
 670 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
 671 {
 672         bio->bi_bdev = tc->origin_dev->bdev;
 673 }
 674
 675 static void issue(struct thin_c *tc, struct bio *bio)
 676 {
 677         struct pool *pool = tc->pool;
 678         unsigned long flags;
 679
 680         /*
 681          * Batch together any FUA/FLUSH bios we find and then issue
 682          * a single commit for them in process_deferred_bios().
 683          */
 684         if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
 685                 spin_lock_irqsave(&pool->lock, flags);
 686                 bio_list_add(&pool->deferred_flush_bios, bio);
 687                 spin_unlock_irqrestore(&pool->lock, flags);
 688         } else
 689                 generic_make_request(bio);
 690 }
 691
 692 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
 693 {
 694         remap_to_origin(tc, bio);
 695         issue(tc, bio);
 696 }
 697
 698 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
 699                             dm_block_t block)
 700 {
 701         remap(tc, bio, block);
 702         issue(tc, bio);
 703 }
 704
 705 /*
 706  * wake_worker() is used when new work is queued and when pool_resume is
 707  * ready to continue deferred IO processing.
 708  */
 709 static void wake_worker(struct pool *pool)
 710 {
 711         queue_work(pool->wq, &pool->worker);
 712 }
 713
 714 /*----------------------------------------------------------------*/
 715
 716 /*
 717  * Bio endio functions.
 718  */
 719 struct endio_hook {
 720         struct thin_c *tc;
 721         bio_end_io_t *saved_bi_end_io;
 722         struct deferred_entry *entry;
 723 };
 724
 725 struct new_mapping {
 726         struct list_head list;
 727
 728         int prepared;
 729
 730         struct thin_c *tc;
 731         dm_block_t virt_block;
 732         dm_block_t data_block;
 733         struct cell *cell;
 734         int err;
 735
 736         /*
 737          * If the bio covers the whole area of a block then we can avoid
 738          * zeroing or copying.  Instead this bio is hooked.  The bio will
 739          * still be in the cell, so care has to be taken to avoid issuing
 740          * the bio twice.
 741          */
 742         struct bio *bio;
 743         bio_end_io_t *saved_bi_end_io;
 744 };
 745
 746 static void __maybe_add_mapping(struct new_mapping *m)
 747 {
 748         struct pool *pool = m->tc->pool;
 749
 750         if (list_empty(&m->list) && m->prepared) {
 751                 list_add(&m->list, &pool->prepared_mappings);
 752                 wake_worker(pool);
 753         }
 754 }
 755
 756 static void copy_complete(int read_err, unsigned long write_err, void *context)
 757 {
 758         unsigned long flags;
 759         struct new_mapping *m = context;
 760         struct pool *pool = m->tc->pool;
 761
 762         m->err = read_err || write_err ? -EIO : 0;
 763
 764         spin_lock_irqsave(&pool->lock, flags);
 765         m->prepared = 1;
 766         __maybe_add_mapping(m);
 767         spin_unlock_irqrestore(&pool->lock, flags);
 768 }
 769
 770 static void overwrite_endio(struct bio *bio, int err)
 771 {
 772         unsigned long flags;
 773         struct new_mapping *m = dm_get_mapinfo(bio)->ptr;
 774         struct pool *pool = m->tc->pool;
 775
 776         m->err = err;
 777
 778         spin_lock_irqsave(&pool->lock, flags);
 779         m->prepared = 1;
 780         __maybe_add_mapping(m);
 781         spin_unlock_irqrestore(&pool->lock, flags);
 782 }
 783
 784 static void shared_read_endio(struct bio *bio, int err)
 785 {
 786         struct list_head mappings;
 787         struct new_mapping *m, *tmp;
 788         struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
 789         unsigned long flags;
 790         struct pool *pool = h->tc->pool;
 791
 792         bio->bi_end_io = h->saved_bi_end_io;
 793         bio_endio(bio, err);
 794
 795         INIT_LIST_HEAD(&mappings);
 796         ds_dec(h->entry, &mappings);
 797
 798         spin_lock_irqsave(&pool->lock, flags);
 799         list_for_each_entry_safe(m, tmp, &mappings, list) {
 800                 list_del(&m->list);
 801                 INIT_LIST_HEAD(&m->list);
 802                 __maybe_add_mapping(m);
 803         }
 804         spin_unlock_irqrestore(&pool->lock, flags);
 805
 806         mempool_free(h, pool->endio_hook_pool);
 807 }
 808
 809 /*----------------------------------------------------------------*/
 810
 811 /*
 812  * Workqueue.
 813  */
 814
 815 /*
 816  * Prepared mapping jobs.
 817  */
 818
 819 /*
 820  * This sends the bios in the cell back to the deferred_bios list.
 821  */
 822 static void cell_defer(struct thin_c *tc, struct cell *cell,
 823                        dm_block_t data_block)
 824 {
 825         struct pool *pool = tc->pool;
 826         unsigned long flags;
 827
 828         spin_lock_irqsave(&pool->lock, flags);
 829         cell_release(cell, &pool->deferred_bios);
 830         spin_unlock_irqrestore(&tc->pool->lock, flags);
 831
 832         wake_worker(pool);
 833 }
 834
 835 /*
 836  * Same as cell_defer above, except it omits one particular detainee,
 837  * a write bio that covers the block and has already been processed.
 838  */
 839 static void cell_defer_except(struct thin_c *tc, struct cell *cell)
 840 {
 841         struct bio_list bios;
 842         struct pool *pool = tc->pool;
 843         unsigned long flags;
 844
 845         bio_list_init(&bios);
 846
 847         spin_lock_irqsave(&pool->lock, flags);
 848         cell_release_no_holder(cell, &pool->deferred_bios);
 849         spin_unlock_irqrestore(&pool->lock, flags);
 850
 851         wake_worker(pool);
 852 }
 853
 854 static void process_prepared_mapping(struct new_mapping *m)
 855 {
 856         struct thin_c *tc = m->tc;
 857         struct bio *bio;
 858         int r;
 859
 860         bio = m->bio;
 861         if (bio)
 862                 bio->bi_end_io = m->saved_bi_end_io;
 863
 864         if (m->err) {
 865                 cell_error(m->cell);
 866                 return;
 867         }
 868
 869         /*
 870          * Commit the prepared block into the mapping btree.
 871          * Any I/O for this block arriving after this point will get
 872          * remapped to it directly.
 873          */
 874         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
 875         if (r) {
 876                 DMERR("dm_thin_insert_block() failed");
 877                 cell_error(m->cell);
 878                 return;
 879         }
 880
 881         /*
 882          * Release any bios held while the block was being provisioned.
 883          * If we are processing a write bio that completely covers the block,
 884          * we already processed it so can ignore it now when processing
 885          * the bios in the cell.
 886          */
 887         if (bio) {
 888                 cell_defer_except(tc, m->cell);
 889                 bio_endio(bio, 0);
 890         } else
 891                 cell_defer(tc, m->cell, m->data_block);
 892
 893         list_del(&m->list);
 894         mempool_free(m, tc->pool->mapping_pool);
 895 }
 896
 897 static void process_prepared_mappings(struct pool *pool)
 898 {
 899         unsigned long flags;
 900         struct list_head maps;
 901         struct new_mapping *m, *tmp;
 902
 903         INIT_LIST_HEAD(&maps);
 904         spin_lock_irqsave(&pool->lock, flags);
 905         list_splice_init(&pool->prepared_mappings, &maps);
 906         spin_unlock_irqrestore(&pool->lock, flags);
 907
 908         list_for_each_entry_safe(m, tmp, &maps, list)
 909                 process_prepared_mapping(m);
 910 }
 911
 912 /*
 913  * Deferred bio jobs.
 914  */
 915 static int io_overwrites_block(struct pool *pool, struct bio *bio)
 916 {
 917         return ((bio_data_dir(bio) == WRITE) &&
 918                 !(bio->bi_sector & pool->offset_mask)) &&
 919                 (bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT));
 920 }
 921
 922 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
 923                                bio_end_io_t *fn)
 924 {
 925         *save = bio->bi_end_io;
 926         bio->bi_end_io = fn;
 927 }
 928
 929 static int ensure_next_mapping(struct pool *pool)
 930 {
 931         if (pool->next_mapping)
 932                 return 0;
 933
 934         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
 935
 936         return pool->next_mapping ? 0 : -ENOMEM;
 937 }
 938
 939 static struct new_mapping *get_next_mapping(struct pool *pool)
 940 {
 941         struct new_mapping *r = pool->next_mapping;
 942
 943         BUG_ON(!pool->next_mapping);
 944
 945         pool->next_mapping = NULL;
 946
 947         return r;
 948 }
 949
 950 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
 951                           struct dm_dev *origin, dm_block_t data_origin,
 952                           dm_block_t data_dest,
 953                           struct cell *cell, struct bio *bio)
 954 {
 955         int r;
 956         struct pool *pool = tc->pool;
 957         struct new_mapping *m = get_next_mapping(pool);
 958
 959         INIT_LIST_HEAD(&m->list);
 960         m->prepared = 0;
 961         m->tc = tc;
 962         m->virt_block = virt_block;
 963         m->data_block = data_dest;
 964         m->cell = cell;
 965         m->err = 0;
 966         m->bio = NULL;
 967
 968         ds_add_work(&pool->ds, &m->list);
 969
 970         /*
 971          * IO to pool_dev remaps to the pool target's data_dev.
 972          *
 973          * If the whole block of data is being overwritten, we can issue the
 974          * bio immediately. Otherwise we use kcopyd to clone the data first.
 975          */
 976         if (io_overwrites_block(pool, bio)) {
 977                 m->bio = bio;
 978                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 979                 dm_get_mapinfo(bio)->ptr = m;
 980                 remap_and_issue(tc, bio, data_dest);
 981         } else {
 982                 struct dm_io_region from, to;
 983
 984                 from.bdev = origin->bdev;
 985                 from.sector = data_origin * pool->sectors_per_block;
 986                 from.count = pool->sectors_per_block;
 987
 988                 to.bdev = tc->pool_dev->bdev;
 989                 to.sector = data_dest * pool->sectors_per_block;
 990                 to.count = pool->sectors_per_block;
 991
 992                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
 993                                    0, copy_complete, m);
 994                 if (r < 0) {
 995                         mempool_free(m, pool->mapping_pool);
 996                         DMERR("dm_kcopyd_copy() failed");
 997                         cell_error(cell);
 998                 }
 999         }
1000 }
1001
1002 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1003                                    dm_block_t data_origin, dm_block_t data_dest,
1004                                    struct cell *cell, struct bio *bio)
1005 {
1006         schedule_copy(tc, virt_block, tc->pool_dev,
1007                       data_origin, data_dest, cell, bio);
1008 }
1009
1010 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1011                                    dm_block_t data_dest,
1012                                    struct cell *cell, struct bio *bio)
1013 {
1014         schedule_copy(tc, virt_block, tc->origin_dev,
1015                       virt_block, data_dest, cell, bio);
1016 }
1017
1018 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1019                           dm_block_t data_block, struct cell *cell,
1020                           struct bio *bio)
1021 {
1022         struct pool *pool = tc->pool;
1023         struct new_mapping *m = get_next_mapping(pool);
1024
1025         INIT_LIST_HEAD(&m->list);
1026         m->prepared = 0;
1027         m->tc = tc;
1028         m->virt_block = virt_block;
1029         m->data_block = data_block;
1030         m->cell = cell;
1031         m->err = 0;
1032         m->bio = NULL;
1033
1034         /*
1035          * If the whole block of data is being overwritten or we are not
1036          * zeroing pre-existing data, we can issue the bio immediately.
1037          * Otherwise we use kcopyd to zero the data first.
1038          */
1039         if (!pool->zero_new_blocks)
1040                 process_prepared_mapping(m);
1041
1042         else if (io_overwrites_block(pool, bio)) {
1043                 m->bio = bio;
1044                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1045                 dm_get_mapinfo(bio)->ptr = m;
1046                 remap_and_issue(tc, bio, data_block);
1047
1048         } else {
1049                 int r;
1050                 struct dm_io_region to;
1051
1052                 to.bdev = tc->pool_dev->bdev;
1053                 to.sector = data_block * pool->sectors_per_block;
1054                 to.count = pool->sectors_per_block;
1055
1056                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
1057                 if (r < 0) {
1058                         mempool_free(m, pool->mapping_pool);
1059                         DMERR("dm_kcopyd_zero() failed");
1060                         cell_error(cell);
1061                 }
1062         }
1063 }
1064
1065 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1066 {
1067         int r;
1068         dm_block_t free_blocks;
1069         unsigned long flags;
1070         struct pool *pool = tc->pool;
1071
1072         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1073         if (r)
1074                 return r;
1075
1076         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1077                 DMWARN("%s: reached low water mark, sending event.",
1078                        dm_device_name(pool->pool_md));
1079                 spin_lock_irqsave(&pool->lock, flags);
1080                 pool->low_water_triggered = 1;
1081                 spin_unlock_irqrestore(&pool->lock, flags);
1082                 dm_table_event(pool->ti->table);
1083         }
1084
1085         if (!free_blocks) {
1086                 if (pool->no_free_space)
1087                         return -ENOSPC;
1088                 else {
1089                         /*
1090                          * Try to commit to see if that will free up some
1091                          * more space.
1092                          */
1093                         r = dm_pool_commit_metadata(pool->pmd);
1094                         if (r) {
1095                                 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1096                                       __func__, r);
1097                                 return r;
1098                         }
1099
1100                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1101                         if (r)
1102                                 return r;
1103
1104                         /*
1105                          * If we still have no space we set a flag to avoid
1106                          * doing all this checking and return -ENOSPC.
1107                          */
1108                         if (!free_blocks) {
1109                                 DMWARN("%s: no free space available.",
1110                                        dm_device_name(pool->pool_md));
1111                                 spin_lock_irqsave(&pool->lock, flags);
1112                                 pool->no_free_space = 1;
1113                                 spin_unlock_irqrestore(&pool->lock, flags);
1114                                 return -ENOSPC;
1115                         }
1116                 }
1117         }
1118
1119         r = dm_pool_alloc_data_block(pool->pmd, result);
1120         if (r)
1121                 return r;
1122
1123         return 0;
1124 }
1125
1126 /*
1127  * If we have run out of space, queue bios until the device is
1128  * resumed, presumably after having been reloaded with more space.
1129  */
1130 static void retry_on_resume(struct bio *bio)
1131 {
1132         struct thin_c *tc = dm_get_mapinfo(bio)->ptr;
1133         struct pool *pool = tc->pool;
1134         unsigned long flags;
1135
1136         spin_lock_irqsave(&pool->lock, flags);
1137         bio_list_add(&pool->retry_on_resume_list, bio);
1138         spin_unlock_irqrestore(&pool->lock, flags);
1139 }
1140
1141 static void no_space(struct cell *cell)
1142 {
1143         struct bio *bio;
1144         struct bio_list bios;
1145
1146         bio_list_init(&bios);
1147         cell_release(cell, &bios);
1148
1149         while ((bio = bio_list_pop(&bios)))
1150                 retry_on_resume(bio);
1151 }
1152
1153 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1154                           struct cell_key *key,
1155                           struct dm_thin_lookup_result *lookup_result,
1156                           struct cell *cell)
1157 {
1158         int r;
1159         dm_block_t data_block;
1160
1161         r = alloc_data_block(tc, &data_block);
1162         switch (r) {
1163         case 0:
1164                 schedule_internal_copy(tc, block, lookup_result->block,
1165                                        data_block, cell, bio);
1166                 break;
1167
1168         case -ENOSPC:
1169                 no_space(cell);
1170                 break;
1171
1172         default:
1173                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1174                 cell_error(cell);
1175                 break;
1176         }
1177 }
1178
1179 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1180                                dm_block_t block,
1181                                struct dm_thin_lookup_result *lookup_result)
1182 {
1183         struct cell *cell;
1184         struct pool *pool = tc->pool;
1185         struct cell_key key;
1186
1187         /*
1188          * If cell is already occupied, then sharing is already in the process
1189          * of being broken so we have nothing further to do here.
1190          */
1191         build_data_key(tc->td, lookup_result->block, &key);
1192         if (bio_detain(pool->prison, &key, bio, &cell))
1193                 return;
1194
1195         if (bio_data_dir(bio) == WRITE)
1196                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1197         else {
1198                 struct endio_hook *h;
1199                 h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1200
1201                 h->tc = tc;
1202                 h->entry = ds_inc(&pool->ds);
1203                 save_and_set_endio(bio, &h->saved_bi_end_io, shared_read_endio);
1204                 dm_get_mapinfo(bio)->ptr = h;
1205
1206                 cell_release_singleton(cell, bio);
1207                 remap_and_issue(tc, bio, lookup_result->block);
1208         }
1209 }
1210
1211 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1212                             struct cell *cell)
1213 {
1214         int r;
1215         dm_block_t data_block;
1216
1217         /*
1218          * Remap empty bios (flushes) immediately, without provisioning.
1219          */
1220         if (!bio->bi_size) {
1221                 cell_release_singleton(cell, bio);
1222                 remap_and_issue(tc, bio, 0);
1223                 return;
1224         }
1225
1226         /*
1227          * Fill read bios with zeroes and complete them immediately.
1228          */
1229         if (bio_data_dir(bio) == READ) {
1230                 zero_fill_bio(bio);
1231                 cell_release_singleton(cell, bio);
1232                 bio_endio(bio, 0);
1233                 return;
1234         }
1235
1236         r = alloc_data_block(tc, &data_block);
1237         switch (r) {
1238         case 0:
1239                 if (tc->origin_dev)
1240                         schedule_external_copy(tc, block, data_block, cell, bio);
1241                 else
1242                         schedule_zero(tc, block, data_block, cell, bio);
1243                 break;
1244
1245         case -ENOSPC:
1246                 no_space(cell);
1247                 break;
1248
1249         default:
1250                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1251                 cell_error(cell);
1252                 break;
1253         }
1254 }
1255
1256 static void process_bio(struct thin_c *tc, struct bio *bio)
1257 {
1258         int r;
1259         dm_block_t block = get_bio_block(tc, bio);
1260         struct cell *cell;
1261         struct cell_key key;
1262         struct dm_thin_lookup_result lookup_result;
1263
1264         /*
1265          * If cell is already occupied, then the block is already
1266          * being provisioned so we have nothing further to do here.
1267          */
1268         build_virtual_key(tc->td, block, &key);
1269         if (bio_detain(tc->pool->prison, &key, bio, &cell))
1270                 return;
1271
1272         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1273         switch (r) {
1274         case 0:
1275                 /*
1276                  * We can release this cell now.  This thread is the only
1277                  * one that puts bios into a cell, and we know there were
1278                  * no preceding bios.
1279                  */
1280                 /*
1281                  * TODO: this will probably have to change when discard goes
1282                  * back in.
1283                  */
1284                 cell_release_singleton(cell, bio);
1285
1286                 if (lookup_result.shared)
1287                         process_shared_bio(tc, bio, block, &lookup_result);
1288                 else
1289                         remap_and_issue(tc, bio, lookup_result.block);
1290                 break;
1291
1292         case -ENODATA:
1293                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1294                         cell_release_singleton(cell, bio);
1295                         remap_to_origin_and_issue(tc, bio);
1296                 } else
1297                         provision_block(tc, bio, block, cell);
1298                 break;
1299
1300         default:
1301                 DMERR("dm_thin_find_block() failed, error = %d", r);
1302                 bio_io_error(bio);
1303                 break;
1304         }
1305 }
1306
1307 static int need_commit_due_to_time(struct pool *pool)
1308 {
1309         return jiffies < pool->last_commit_jiffies ||
1310                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1311 }
1312
1313 static void process_deferred_bios(struct pool *pool)
1314 {
1315         unsigned long flags;
1316         struct bio *bio;
1317         struct bio_list bios;
1318         int r;
1319
1320         bio_list_init(&bios);
1321
1322         spin_lock_irqsave(&pool->lock, flags);
1323         bio_list_merge(&bios, &pool->deferred_bios);
1324         bio_list_init(&pool->deferred_bios);
1325         spin_unlock_irqrestore(&pool->lock, flags);
1326
1327         while ((bio = bio_list_pop(&bios))) {
1328                 struct thin_c *tc = dm_get_mapinfo(bio)->ptr;
1329                 /*
1330                  * If we've got no free new_mapping structs, and processing
1331                  * this bio might require one, we pause until there are some
1332                  * prepared mappings to process.
1333                  */
1334                 if (ensure_next_mapping(pool)) {
1335                         spin_lock_irqsave(&pool->lock, flags);
1336                         bio_list_merge(&pool->deferred_bios, &bios);
1337                         spin_unlock_irqrestore(&pool->lock, flags);
1338
1339                         break;
1340                 }
1341                 process_bio(tc, bio);
1342         }
1343
1344         /*
1345          * If there are any deferred flush bios, we must commit
1346          * the metadata before issuing them.
1347          */
1348         bio_list_init(&bios);
1349         spin_lock_irqsave(&pool->lock, flags);
1350         bio_list_merge(&bios, &pool->deferred_flush_bios);
1351         bio_list_init(&pool->deferred_flush_bios);
1352         spin_unlock_irqrestore(&pool->lock, flags);
1353
1354         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1355                 return;
1356
1357         r = dm_pool_commit_metadata(pool->pmd);
1358         if (r) {
1359                 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1360                       __func__, r);
1361                 while ((bio = bio_list_pop(&bios)))
1362                         bio_io_error(bio);
1363                 return;
1364         }
1365         pool->last_commit_jiffies = jiffies;
1366
1367         while ((bio = bio_list_pop(&bios)))
1368                 generic_make_request(bio);
1369 }
1370
1371 static void do_worker(struct work_struct *ws)
1372 {
1373         struct pool *pool = container_of(ws, struct pool, worker);
1374
1375         process_prepared_mappings(pool);
1376         process_deferred_bios(pool);
1377 }
1378
1379 /*
1380  * We want to commit periodically so that not too much
1381  * unwritten data builds up.
1382  */
1383 static void do_waker(struct work_struct *ws)
1384 {
1385         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1386         wake_worker(pool);
1387         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1388 }
1389
1390 /*----------------------------------------------------------------*/
1391
1392 /*
1393  * Mapping functions.
1394  */
1395
1396 /*
1397  * Called only while mapping a thin bio to hand it over to the workqueue.
1398  */
1399 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1400 {
1401         unsigned long flags;
1402         struct pool *pool = tc->pool;
1403
1404         spin_lock_irqsave(&pool->lock, flags);
1405         bio_list_add(&pool->deferred_bios, bio);
1406         spin_unlock_irqrestore(&pool->lock, flags);
1407
1408         wake_worker(pool);
1409 }
1410
1411 /*
1412  * Non-blocking function called from the thin target's map function.
1413  */
1414 static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1415                         union map_info *map_context)
1416 {
1417         int r;
1418         struct thin_c *tc = ti->private;
1419         dm_block_t block = get_bio_block(tc, bio);
1420         struct dm_thin_device *td = tc->td;
1421         struct dm_thin_lookup_result result;
1422
1423         /*
1424          * Save the thin context for easy access from the deferred bio later.
1425          */
1426         map_context->ptr = tc;
1427
1428         if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
1429                 thin_defer_bio(tc, bio);
1430                 return DM_MAPIO_SUBMITTED;
1431         }
1432
1433         r = dm_thin_find_block(td, block, 0, &result);
1434
1435         /*
1436          * Note that we defer readahead too.
1437          */
1438         switch (r) {
1439         case 0:
1440                 if (unlikely(result.shared)) {
1441                         /*
1442                          * We have a race condition here between the
1443                          * result.shared value returned by the lookup and
1444                          * snapshot creation, which may cause new
1445                          * sharing.
1446                          *
1447                          * To avoid this always quiesce the origin before
1448                          * taking the snap.  You want to do this anyway to
1449                          * ensure a consistent application view
1450                          * (i.e. lockfs).
1451                          *
1452                          * More distant ancestors are irrelevant. The
1453                          * shared flag will be set in their case.
1454                          */
1455                         thin_defer_bio(tc, bio);
1456                         r = DM_MAPIO_SUBMITTED;
1457                 } else {
1458                         remap(tc, bio, result.block);
1459                         r = DM_MAPIO_REMAPPED;
1460                 }
1461                 break;
1462
1463         case -ENODATA:
1464                 /*
1465                  * In future, the failed dm_thin_find_block above could
1466                  * provide the hint to load the metadata into cache.
1467                  */
1468         case -EWOULDBLOCK:
1469                 thin_defer_bio(tc, bio);
1470                 r = DM_MAPIO_SUBMITTED;
1471                 break;
1472         }
1473
1474         return r;
1475 }
1476
1477 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1478 {
1479         int r;
1480         unsigned long flags;
1481         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1482
1483         spin_lock_irqsave(&pt->pool->lock, flags);
1484         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1485         spin_unlock_irqrestore(&pt->pool->lock, flags);
1486
1487         if (!r) {
1488                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1489                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1490         }
1491
1492         return r;
1493 }
1494
1495 static void __requeue_bios(struct pool *pool)
1496 {
1497         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1498         bio_list_init(&pool->retry_on_resume_list);
1499 }
1500
1501 /*----------------------------------------------------------------
1502  * Binding of control targets to a pool object
1503  *--------------------------------------------------------------*/
1504 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1505 {
1506         struct pool_c *pt = ti->private;
1507
1508         pool->ti = ti;
1509         pool->low_water_blocks = pt->low_water_blocks;
1510         pool->zero_new_blocks = pt->zero_new_blocks;
1511
1512         return 0;
1513 }
1514
1515 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1516 {
1517         if (pool->ti == ti)
1518                 pool->ti = NULL;
1519 }
1520
1521 /*----------------------------------------------------------------
1522  * Pool creation
1523  *--------------------------------------------------------------*/
1524 static void __pool_destroy(struct pool *pool)
1525 {
1526         __pool_table_remove(pool);
1527
1528         if (dm_pool_metadata_close(pool->pmd) < 0)
1529                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1530
1531         prison_destroy(pool->prison);
1532         dm_kcopyd_client_destroy(pool->copier);
1533
1534         if (pool->wq)
1535                 destroy_workqueue(pool->wq);
1536
1537         if (pool->next_mapping)
1538                 mempool_free(pool->next_mapping, pool->mapping_pool);
1539         mempool_destroy(pool->mapping_pool);
1540         mempool_destroy(pool->endio_hook_pool);
1541         kfree(pool);
1542 }
1543
1544 static struct pool *pool_create(struct mapped_device *pool_md,
1545                                 struct block_device *metadata_dev,
1546                                 unsigned long block_size, char **error)
1547 {
1548         int r;
1549         void *err_p;
1550         struct pool *pool;
1551         struct dm_pool_metadata *pmd;
1552
1553         pmd = dm_pool_metadata_open(metadata_dev, block_size);
1554         if (IS_ERR(pmd)) {
1555                 *error = "Error creating metadata object";
1556                 return (struct pool *)pmd;
1557         }
1558
1559         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1560         if (!pool) {
1561                 *error = "Error allocating memory for pool";
1562                 err_p = ERR_PTR(-ENOMEM);
1563                 goto bad_pool;
1564         }
1565
1566         pool->pmd = pmd;
1567         pool->sectors_per_block = block_size;
1568         pool->block_shift = ffs(block_size) - 1;
1569         pool->offset_mask = block_size - 1;
1570         pool->low_water_blocks = 0;
1571         pool->zero_new_blocks = 1;
1572         pool->prison = prison_create(PRISON_CELLS);
1573         if (!pool->prison) {
1574                 *error = "Error creating pool's bio prison";
1575                 err_p = ERR_PTR(-ENOMEM);
1576                 goto bad_prison;
1577         }
1578
1579         pool->copier = dm_kcopyd_client_create();
1580         if (IS_ERR(pool->copier)) {
1581                 r = PTR_ERR(pool->copier);
1582                 *error = "Error creating pool's kcopyd client";
1583                 err_p = ERR_PTR(r);
1584                 goto bad_kcopyd_client;
1585         }
1586
1587         /*
1588          * Create singlethreaded workqueue that will service all devices
1589          * that use this metadata.
1590          */
1591         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1592         if (!pool->wq) {
1593                 *error = "Error creating pool's workqueue";
1594                 err_p = ERR_PTR(-ENOMEM);
1595                 goto bad_wq;
1596         }
1597
1598         INIT_WORK(&pool->worker, do_worker);
1599         INIT_DELAYED_WORK(&pool->waker, do_waker);
1600         spin_lock_init(&pool->lock);
1601         bio_list_init(&pool->deferred_bios);
1602         bio_list_init(&pool->deferred_flush_bios);
1603         INIT_LIST_HEAD(&pool->prepared_mappings);
1604         pool->low_water_triggered = 0;
1605         pool->no_free_space = 0;
1606         bio_list_init(&pool->retry_on_resume_list);
1607         ds_init(&pool->ds);
1608
1609         pool->next_mapping = NULL;
1610         pool->mapping_pool =
1611                 mempool_create_kmalloc_pool(MAPPING_POOL_SIZE, sizeof(struct new_mapping));
1612         if (!pool->mapping_pool) {
1613                 *error = "Error creating pool's mapping mempool";
1614                 err_p = ERR_PTR(-ENOMEM);
1615                 goto bad_mapping_pool;
1616         }
1617
1618         pool->endio_hook_pool =
1619                 mempool_create_kmalloc_pool(ENDIO_HOOK_POOL_SIZE, sizeof(struct endio_hook));
1620         if (!pool->endio_hook_pool) {
1621                 *error = "Error creating pool's endio_hook mempool";
1622                 err_p = ERR_PTR(-ENOMEM);
1623                 goto bad_endio_hook_pool;
1624         }
1625         pool->ref_count = 1;
1626         pool->last_commit_jiffies = jiffies;
1627         pool->pool_md = pool_md;
1628         pool->md_dev = metadata_dev;
1629         __pool_table_insert(pool);
1630
1631         return pool;
1632
1633 bad_endio_hook_pool:
1634         mempool_destroy(pool->mapping_pool);
1635 bad_mapping_pool:
1636         destroy_workqueue(pool->wq);
1637 bad_wq:
1638         dm_kcopyd_client_destroy(pool->copier);
1639 bad_kcopyd_client:
1640         prison_destroy(pool->prison);
1641 bad_prison:
1642         kfree(pool);
1643 bad_pool:
1644         if (dm_pool_metadata_close(pmd))
1645                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1646
1647         return err_p;
1648 }
1649
1650 static void __pool_inc(struct pool *pool)
1651 {
1652         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1653         pool->ref_count++;
1654 }
1655
1656 static void __pool_dec(struct pool *pool)
1657 {
1658         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1659         BUG_ON(!pool->ref_count);
1660         if (!--pool->ref_count)
1661                 __pool_destroy(pool);
1662 }
1663
1664 static struct pool *__pool_find(struct mapped_device *pool_md,
1665                                 struct block_device *metadata_dev,
1666                                 unsigned long block_size, char **error)
1667 {
1668         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1669
1670         if (pool) {
1671                 if (pool->pool_md != pool_md)
1672                         return ERR_PTR(-EBUSY);
1673                 __pool_inc(pool);
1674
1675         } else {
1676                 pool = __pool_table_lookup(pool_md);
1677                 if (pool) {
1678                         if (pool->md_dev != metadata_dev)
1679                                 return ERR_PTR(-EINVAL);
1680                         __pool_inc(pool);
1681
1682                 } else
1683                         pool = pool_create(pool_md, metadata_dev, block_size, error);
1684         }
1685
1686         return pool;
1687 }
1688
1689 /*----------------------------------------------------------------
1690  * Pool target methods
1691  *--------------------------------------------------------------*/
1692 static void pool_dtr(struct dm_target *ti)
1693 {
1694         struct pool_c *pt = ti->private;
1695
1696         mutex_lock(&dm_thin_pool_table.mutex);
1697
1698         unbind_control_target(pt->pool, ti);
1699         __pool_dec(pt->pool);
1700         dm_put_device(ti, pt->metadata_dev);
1701         dm_put_device(ti, pt->data_dev);
1702         kfree(pt);
1703
1704         mutex_unlock(&dm_thin_pool_table.mutex);
1705 }
1706
1707 struct pool_features {
1708         unsigned zero_new_blocks:1;
1709 };
1710
1711 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1712                                struct dm_target *ti)
1713 {
1714         int r;
1715         unsigned argc;
1716         const char *arg_name;
1717
1718         static struct dm_arg _args[] = {
1719                 {0, 1, "Invalid number of pool feature arguments"},
1720         };
1721
1722         /*
1723          * No feature arguments supplied.
1724          */
1725         if (!as->argc)
1726                 return 0;
1727
1728         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1729         if (r)
1730                 return -EINVAL;
1731
1732         while (argc && !r) {
1733                 arg_name = dm_shift_arg(as);
1734                 argc--;
1735
1736                 if (!strcasecmp(arg_name, "skip_block_zeroing")) {
1737                         pf->zero_new_blocks = 0;
1738                         continue;
1739                 }
1740
1741                 ti->error = "Unrecognised pool feature requested";
1742                 r = -EINVAL;
1743         }
1744
1745         return r;
1746 }
1747
1748 /*
1749  * thin-pool <metadata dev> <data dev>
1750  *           <data block size (sectors)>
1751  *           <low water mark (blocks)>
1752  *           [<#feature args> [<arg>]*]
1753  *
1754  * Optional feature arguments are:
1755  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1756  */
1757 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1758 {
1759         int r;
1760         struct pool_c *pt;
1761         struct pool *pool;
1762         struct pool_features pf;
1763         struct dm_arg_set as;
1764         struct dm_dev *data_dev;
1765         unsigned long block_size;
1766         dm_block_t low_water_blocks;
1767         struct dm_dev *metadata_dev;
1768         sector_t metadata_dev_size;
1769         char b[BDEVNAME_SIZE];
1770
1771         /*
1772          * FIXME Remove validation from scope of lock.
1773          */
1774         mutex_lock(&dm_thin_pool_table.mutex);
1775
1776         if (argc < 4) {
1777                 ti->error = "Invalid argument count";
1778                 r = -EINVAL;
1779                 goto out_unlock;
1780         }
1781         as.argc = argc;
1782         as.argv = argv;
1783
1784         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1785         if (r) {
1786                 ti->error = "Error opening metadata block device";
1787                 goto out_unlock;
1788         }
1789
1790         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1791         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1792                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1793                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1794
1795         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1796         if (r) {
1797                 ti->error = "Error getting data device";
1798                 goto out_metadata;
1799         }
1800
1801         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1802             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1803             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1804             !is_power_of_2(block_size)) {
1805                 ti->error = "Invalid block size";
1806                 r = -EINVAL;
1807                 goto out;
1808         }
1809
1810         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1811                 ti->error = "Invalid low water mark";
1812                 r = -EINVAL;
1813                 goto out;
1814         }
1815
1816         /*
1817          * Set default pool features.
1818          */
1819         memset(&pf, 0, sizeof(pf));
1820         pf.zero_new_blocks = 1;
1821
1822         dm_consume_args(&as, 4);
1823         r = parse_pool_features(&as, &pf, ti);
1824         if (r)
1825                 goto out;
1826
1827         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1828         if (!pt) {
1829                 r = -ENOMEM;
1830                 goto out;
1831         }
1832
1833         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1834                            block_size, &ti->error);
1835         if (IS_ERR(pool)) {
1836                 r = PTR_ERR(pool);
1837                 goto out_free_pt;
1838         }
1839
1840         pt->pool = pool;
1841         pt->ti = ti;
1842         pt->metadata_dev = metadata_dev;
1843         pt->data_dev = data_dev;
1844         pt->low_water_blocks = low_water_blocks;
1845         pt->zero_new_blocks = pf.zero_new_blocks;
1846         ti->num_flush_requests = 1;
1847         ti->num_discard_requests = 0;
1848         ti->private = pt;
1849
1850         pt->callbacks.congested_fn = pool_is_congested;
1851         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1852
1853         mutex_unlock(&dm_thin_pool_table.mutex);
1854
1855         return 0;
1856
1857 out_free_pt:
1858         kfree(pt);
1859 out:
1860         dm_put_device(ti, data_dev);
1861 out_metadata:
1862         dm_put_device(ti, metadata_dev);
1863 out_unlock:
1864         mutex_unlock(&dm_thin_pool_table.mutex);
1865
1866         return r;
1867 }
1868
1869 static int pool_map(struct dm_target *ti, struct bio *bio,
1870                     union map_info *map_context)
1871 {
1872         int r;
1873         struct pool_c *pt = ti->private;
1874         struct pool *pool = pt->pool;
1875         unsigned long flags;
1876
1877         /*
1878          * As this is a singleton target, ti->begin is always zero.
1879          */
1880         spin_lock_irqsave(&pool->lock, flags);
1881         bio->bi_bdev = pt->data_dev->bdev;
1882         r = DM_MAPIO_REMAPPED;
1883         spin_unlock_irqrestore(&pool->lock, flags);
1884
1885         return r;
1886 }
1887
1888 /*
1889  * Retrieves the number of blocks of the data device from
1890  * the superblock and compares it to the actual device size,
1891  * thus resizing the data device in case it has grown.
1892  *
1893  * This both copes with opening preallocated data devices in the ctr
1894  * being followed by a resume
1895  * -and-
1896  * calling the resume method individually after userspace has
1897  * grown the data device in reaction to a table event.
1898  */
1899 static int pool_preresume(struct dm_target *ti)
1900 {
1901         int r;
1902         struct pool_c *pt = ti->private;
1903         struct pool *pool = pt->pool;
1904         dm_block_t data_size, sb_data_size;
1905
1906         /*
1907          * Take control of the pool object.
1908          */
1909         r = bind_control_target(pool, ti);
1910         if (r)
1911                 return r;
1912
1913         data_size = ti->len >> pool->block_shift;
1914         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
1915         if (r) {
1916                 DMERR("failed to retrieve data device size");
1917                 return r;
1918         }
1919
1920         if (data_size < sb_data_size) {
1921                 DMERR("pool target too small, is %llu blocks (expected %llu)",
1922                       data_size, sb_data_size);
1923                 return -EINVAL;
1924
1925         } else if (data_size > sb_data_size) {
1926                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
1927                 if (r) {
1928                         DMERR("failed to resize data device");
1929                         return r;
1930                 }
1931
1932                 r = dm_pool_commit_metadata(pool->pmd);
1933                 if (r) {
1934                         DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1935                               __func__, r);
1936                         return r;
1937                 }
1938         }
1939
1940         return 0;
1941 }
1942
1943 static void pool_resume(struct dm_target *ti)
1944 {
1945         struct pool_c *pt = ti->private;
1946         struct pool *pool = pt->pool;
1947         unsigned long flags;
1948
1949         spin_lock_irqsave(&pool->lock, flags);
1950         pool->low_water_triggered = 0;
1951         pool->no_free_space = 0;
1952         __requeue_bios(pool);
1953         spin_unlock_irqrestore(&pool->lock, flags);
1954
1955         do_waker(&pool->waker.work);
1956 }
1957
1958 static void pool_postsuspend(struct dm_target *ti)
1959 {
1960         int r;
1961         struct pool_c *pt = ti->private;
1962         struct pool *pool = pt->pool;
1963
1964         cancel_delayed_work(&pool->waker);
1965         flush_workqueue(pool->wq);
1966
1967         r = dm_pool_commit_metadata(pool->pmd);
1968         if (r < 0) {
1969                 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1970                       __func__, r);
1971                 /* FIXME: invalidate device? error the next FUA or FLUSH bio ?*/
1972         }
1973 }
1974
1975 static int check_arg_count(unsigned argc, unsigned args_required)
1976 {
1977         if (argc != args_required) {
1978                 DMWARN("Message received with %u arguments instead of %u.",
1979                        argc, args_required);
1980                 return -EINVAL;
1981         }
1982
1983         return 0;
1984 }
1985
1986 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
1987 {
1988         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
1989             *dev_id <= MAX_DEV_ID)
1990                 return 0;
1991
1992         if (warning)
1993                 DMWARN("Message received with invalid device id: %s", arg);
1994
1995         return -EINVAL;
1996 }
1997
1998 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
1999 {
2000         dm_thin_id dev_id;
2001         int r;
2002
2003         r = check_arg_count(argc, 2);
2004         if (r)
2005                 return r;
2006
2007         r = read_dev_id(argv[1], &dev_id, 1);
2008         if (r)
2009                 return r;
2010
2011         r = dm_pool_create_thin(pool->pmd, dev_id);
2012         if (r) {
2013                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2014                        argv[1]);
2015                 return r;
2016         }
2017
2018         return 0;
2019 }
2020
2021 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2022 {
2023         dm_thin_id dev_id;
2024         dm_thin_id origin_dev_id;
2025         int r;
2026
2027         r = check_arg_count(argc, 3);
2028         if (r)
2029                 return r;
2030
2031         r = read_dev_id(argv[1], &dev_id, 1);
2032         if (r)
2033                 return r;
2034
2035         r = read_dev_id(argv[2], &origin_dev_id, 1);
2036         if (r)
2037                 return r;
2038
2039         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2040         if (r) {
2041                 DMWARN("Creation of new snapshot %s of device %s failed.",
2042                        argv[1], argv[2]);
2043                 return r;
2044         }
2045
2046         return 0;
2047 }
2048
2049 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2050 {
2051         dm_thin_id dev_id;
2052         int r;
2053
2054         r = check_arg_count(argc, 2);
2055         if (r)
2056                 return r;
2057
2058         r = read_dev_id(argv[1], &dev_id, 1);
2059         if (r)
2060                 return r;
2061
2062         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2063         if (r)
2064                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2065
2066         return r;
2067 }
2068
2069 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2070 {
2071         dm_thin_id old_id, new_id;
2072         int r;
2073
2074         r = check_arg_count(argc, 3);
2075         if (r)
2076                 return r;
2077
2078         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2079                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2080                 return -EINVAL;
2081         }
2082
2083         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2084                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2085                 return -EINVAL;
2086         }
2087
2088         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2089         if (r) {
2090                 DMWARN("Failed to change transaction id from %s to %s.",
2091                        argv[1], argv[2]);
2092                 return r;
2093         }
2094
2095         return 0;
2096 }
2097
2098 /*
2099  * Messages supported:
2100  *   create_thin        <dev_id>
2101  *   create_snap        <dev_id> <origin_id>
2102  *   delete             <dev_id>
2103  *   trim               <dev_id> <new_size_in_sectors>
2104  *   set_transaction_id <current_trans_id> <new_trans_id>
2105  */
2106 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2107 {
2108         int r = -EINVAL;
2109         struct pool_c *pt = ti->private;
2110         struct pool *pool = pt->pool;
2111
2112         if (!strcasecmp(argv[0], "create_thin"))
2113                 r = process_create_thin_mesg(argc, argv, pool);
2114
2115         else if (!strcasecmp(argv[0], "create_snap"))
2116                 r = process_create_snap_mesg(argc, argv, pool);
2117
2118         else if (!strcasecmp(argv[0], "delete"))
2119                 r = process_delete_mesg(argc, argv, pool);
2120
2121         else if (!strcasecmp(argv[0], "set_transaction_id"))
2122                 r = process_set_transaction_id_mesg(argc, argv, pool);
2123
2124         else
2125                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2126
2127         if (!r) {
2128                 r = dm_pool_commit_metadata(pool->pmd);
2129                 if (r)
2130                         DMERR("%s message: dm_pool_commit_metadata() failed, error = %d",
2131                               argv[0], r);
2132         }
2133
2134         return r;
2135 }
2136
2137 /*
2138  * Status line is:
2139  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2140  *    <used data sectors>/<total data sectors> <held metadata root>
2141  */
2142 static int pool_status(struct dm_target *ti, status_type_t type,
2143                        char *result, unsigned maxlen)
2144 {
2145         int r;
2146         unsigned sz = 0;
2147         uint64_t transaction_id;
2148         dm_block_t nr_free_blocks_data;
2149         dm_block_t nr_free_blocks_metadata;
2150         dm_block_t nr_blocks_data;
2151         dm_block_t nr_blocks_metadata;
2152         dm_block_t held_root;
2153         char buf[BDEVNAME_SIZE];
2154         char buf2[BDEVNAME_SIZE];
2155         struct pool_c *pt = ti->private;
2156         struct pool *pool = pt->pool;
2157
2158         switch (type) {
2159         case STATUSTYPE_INFO:
2160                 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2161                                                         &transaction_id);
2162                 if (r)
2163                         return r;
2164
2165                 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2166                                                           &nr_free_blocks_metadata);
2167                 if (r)
2168                         return r;
2169
2170                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2171                 if (r)
2172                         return r;
2173
2174                 r = dm_pool_get_free_block_count(pool->pmd,
2175                                                  &nr_free_blocks_data);
2176                 if (r)
2177                         return r;
2178
2179                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2180                 if (r)
2181                         return r;
2182
2183                 r = dm_pool_get_held_metadata_root(pool->pmd, &held_root);
2184                 if (r)
2185                         return r;
2186
2187                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2188                        (unsigned long long)transaction_id,
2189                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2190                        (unsigned long long)nr_blocks_metadata,
2191                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2192                        (unsigned long long)nr_blocks_data);
2193
2194                 if (held_root)
2195                         DMEMIT("%llu", held_root);
2196                 else
2197                         DMEMIT("-");
2198
2199                 break;
2200
2201         case STATUSTYPE_TABLE:
2202                 DMEMIT("%s %s %lu %llu ",
2203                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2204                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2205                        (unsigned long)pool->sectors_per_block,
2206                        (unsigned long long)pt->low_water_blocks);
2207
2208                 DMEMIT("%u ", !pool->zero_new_blocks);
2209
2210                 if (!pool->zero_new_blocks)
2211                         DMEMIT("skip_block_zeroing ");
2212                 break;
2213         }
2214
2215         return 0;
2216 }
2217
2218 static int pool_iterate_devices(struct dm_target *ti,
2219                                 iterate_devices_callout_fn fn, void *data)
2220 {
2221         struct pool_c *pt = ti->private;
2222
2223         return fn(ti, pt->data_dev, 0, ti->len, data);
2224 }
2225
2226 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2227                       struct bio_vec *biovec, int max_size)
2228 {
2229         struct pool_c *pt = ti->private;
2230         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2231
2232         if (!q->merge_bvec_fn)
2233                 return max_size;
2234
2235         bvm->bi_bdev = pt->data_dev->bdev;
2236
2237         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2238 }
2239
2240 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2241 {
2242         struct pool_c *pt = ti->private;
2243         struct pool *pool = pt->pool;
2244
2245         blk_limits_io_min(limits, 0);
2246         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2247 }
2248
2249 static struct target_type pool_target = {
2250         .name = "thin-pool",
2251         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2252                     DM_TARGET_IMMUTABLE,
2253         .version = {1, 0, 0},
2254         .module = THIS_MODULE,
2255         .ctr = pool_ctr,
2256         .dtr = pool_dtr,
2257         .map = pool_map,
2258         .postsuspend = pool_postsuspend,
2259         .preresume = pool_preresume,
2260         .resume = pool_resume,
2261         .message = pool_message,
2262         .status = pool_status,
2263         .merge = pool_merge,
2264         .iterate_devices = pool_iterate_devices,
2265         .io_hints = pool_io_hints,
2266 };
2267
2268 /*----------------------------------------------------------------
2269  * Thin target methods
2270  *--------------------------------------------------------------*/
2271 static void thin_dtr(struct dm_target *ti)
2272 {
2273         struct thin_c *tc = ti->private;
2274
2275         mutex_lock(&dm_thin_pool_table.mutex);
2276
2277         __pool_dec(tc->pool);
2278         dm_pool_close_thin_device(tc->td);
2279         dm_put_device(ti, tc->pool_dev);
2280         if (tc->origin_dev)
2281                 dm_put_device(ti, tc->origin_dev);
2282         kfree(tc);
2283
2284         mutex_unlock(&dm_thin_pool_table.mutex);
2285 }
2286
2287 /*
2288  * Thin target parameters:
2289  *
2290  * <pool_dev> <dev_id> [origin_dev]
2291  *
2292  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2293  * dev_id: the internal device identifier
2294  * origin_dev: a device external to the pool that should act as the origin
2295  */
2296 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2297 {
2298         int r;
2299         struct thin_c *tc;
2300         struct dm_dev *pool_dev, *origin_dev;
2301         struct mapped_device *pool_md;
2302
2303         mutex_lock(&dm_thin_pool_table.mutex);
2304
2305         if (argc != 2 && argc != 3) {
2306                 ti->error = "Invalid argument count";
2307                 r = -EINVAL;
2308                 goto out_unlock;
2309         }
2310
2311         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2312         if (!tc) {
2313                 ti->error = "Out of memory";
2314                 r = -ENOMEM;
2315                 goto out_unlock;
2316         }
2317
2318         if (argc == 3) {
2319                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2320                 if (r) {
2321                         ti->error = "Error opening origin device";
2322                         goto bad_origin_dev;
2323                 }
2324                 tc->origin_dev = origin_dev;
2325         }
2326
2327         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2328         if (r) {
2329                 ti->error = "Error opening pool device";
2330                 goto bad_pool_dev;
2331         }
2332         tc->pool_dev = pool_dev;
2333
2334         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2335                 ti->error = "Invalid device id";
2336                 r = -EINVAL;
2337                 goto bad_common;
2338         }
2339
2340         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2341         if (!pool_md) {
2342                 ti->error = "Couldn't get pool mapped device";
2343                 r = -EINVAL;
2344                 goto bad_common;
2345         }
2346
2347         tc->pool = __pool_table_lookup(pool_md);
2348         if (!tc->pool) {
2349                 ti->error = "Couldn't find pool object";
2350                 r = -EINVAL;
2351                 goto bad_pool_lookup;
2352         }
2353         __pool_inc(tc->pool);
2354
2355         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2356         if (r) {
2357                 ti->error = "Couldn't open thin internal device";
2358                 goto bad_thin_open;
2359         }
2360
2361         ti->split_io = tc->pool->sectors_per_block;
2362         ti->num_flush_requests = 1;
2363         ti->num_discard_requests = 0;
2364         ti->discards_supported = 0;
2365
2366         dm_put(pool_md);
2367
2368         mutex_unlock(&dm_thin_pool_table.mutex);
2369
2370         return 0;
2371
2372 bad_thin_open:
2373         __pool_dec(tc->pool);
2374 bad_pool_lookup:
2375         dm_put(pool_md);
2376 bad_common:
2377         dm_put_device(ti, tc->pool_dev);
2378 bad_pool_dev:
2379         if (tc->origin_dev)
2380                 dm_put_device(ti, tc->origin_dev);
2381 bad_origin_dev:
2382         kfree(tc);
2383 out_unlock:
2384         mutex_unlock(&dm_thin_pool_table.mutex);
2385
2386         return r;
2387 }
2388
2389 static int thin_map(struct dm_target *ti, struct bio *bio,
2390                     union map_info *map_context)
2391 {
2392         bio->bi_sector -= ti->begin;
2393
2394         return thin_bio_map(ti, bio, map_context);
2395 }
2396
2397 static void thin_postsuspend(struct dm_target *ti)
2398 {
2399         if (dm_noflush_suspending(ti))
2400                 requeue_io((struct thin_c *)ti->private);
2401 }
2402
2403 /*
2404  * <nr mapped sectors> <highest mapped sector>
2405  */
2406 static int thin_status(struct dm_target *ti, status_type_t type,
2407                        char *result, unsigned maxlen)
2408 {
2409         int r;
2410         ssize_t sz = 0;
2411         dm_block_t mapped, highest;
2412         char buf[BDEVNAME_SIZE];
2413         struct thin_c *tc = ti->private;
2414
2415         if (!tc->td)
2416                 DMEMIT("-");
2417         else {
2418                 switch (type) {
2419                 case STATUSTYPE_INFO:
2420                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2421                         if (r)
2422                                 return r;
2423
2424                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2425                         if (r < 0)
2426                                 return r;
2427
2428                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2429                         if (r)
2430                                 DMEMIT("%llu", ((highest + 1) *
2431                                                 tc->pool->sectors_per_block) - 1);
2432                         else
2433                                 DMEMIT("-");
2434                         break;
2435
2436                 case STATUSTYPE_TABLE:
2437                         DMEMIT("%s %lu",
2438                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2439                                (unsigned long) tc->dev_id);
2440                         if (tc->origin_dev)
2441                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2442                         break;
2443                 }
2444         }
2445
2446         return 0;
2447 }
2448
2449 static int thin_iterate_devices(struct dm_target *ti,
2450                                 iterate_devices_callout_fn fn, void *data)
2451 {
2452         dm_block_t blocks;
2453         struct thin_c *tc = ti->private;
2454
2455         /*
2456          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2457          * we follow a more convoluted path through to the pool's target.
2458          */
2459         if (!tc->pool->ti)
2460                 return 0;       /* nothing is bound */
2461
2462         blocks = tc->pool->ti->len >> tc->pool->block_shift;
2463         if (blocks)
2464                 return fn(ti, tc->pool_dev, 0, tc->pool->sectors_per_block * blocks, data);
2465
2466         return 0;
2467 }
2468
2469 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2470 {
2471         struct thin_c *tc = ti->private;
2472
2473         blk_limits_io_min(limits, 0);
2474         blk_limits_io_opt(limits, tc->pool->sectors_per_block << SECTOR_SHIFT);
2475 }
2476
2477 static struct target_type thin_target = {
2478         .name = "thin",
2479         .version = {1, 1, 0},
2480         .module = THIS_MODULE,
2481         .ctr = thin_ctr,
2482         .dtr = thin_dtr,
2483         .map = thin_map,
2484         .postsuspend = thin_postsuspend,
2485         .status = thin_status,
2486         .iterate_devices = thin_iterate_devices,
2487         .io_hints = thin_io_hints,
2488 };
2489
2490 /*----------------------------------------------------------------*/
2491
2492 static int __init dm_thin_init(void)
2493 {
2494         int r;
2495
2496         pool_table_init();
2497
2498         r = dm_register_target(&thin_target);
2499         if (r)
2500                 return r;
2501
2502         r = dm_register_target(&pool_target);
2503         if (r)
2504                 dm_unregister_target(&thin_target);
2505
2506         return r;
2507 }
2508
2509 static void dm_thin_exit(void)
2510 {
2511         dm_unregister_target(&thin_target);
2512         dm_unregister_target(&pool_target);
2513 }
2514
2515 module_init(dm_thin_init);
2516 module_exit(dm_thin_exit);
2517
2518 MODULE_DESCRIPTION(DM_NAME "device-mapper thin provisioning target");
2519 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2520 MODULE_LICENSE("GPL");