linux/drivers/md/dm-thin.c
<<
>>
Prefs
   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 */
 104static 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
 112static 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 */
 127struct dm_thin_new_mapping;
 128
 129/*
 130 * The pool runs in 3 modes.  Ordered in degraded order for comparisons.
 131 */
 132enum 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
 138struct 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
 146struct thin_c;
 147typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
 148typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
 149
 150struct 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
 198static enum pool_mode get_pool_mode(struct pool *pool);
 199static void set_pool_mode(struct pool *pool, enum pool_mode mode);
 200
 201/*
 202 * Target context for a pool.
 203 */
 204struct 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 */
 219struct 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 */
 233static struct dm_thin_pool_table {
 234        struct mutex mutex;
 235        struct list_head pools;
 236} dm_thin_pool_table;
 237
 238static 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
 244static 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
 250static 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
 256static 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
 272static 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
 290struct 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
 297static 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
 316static 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
 334static 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
 346static 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
 360static void remap_to_origin(struct thin_c *tc, struct bio *bio)
 361{
 362        bio->bi_bdev = tc->origin_dev->bdev;
 363}
 364
 365static 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
 371static 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
 382static 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
 411static 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
 417static 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 */
 428static void wake_worker(struct pool *pool)
 429{
 430        queue_work(pool->wq, &pool->worker);
 431}
 432
 433/*----------------------------------------------------------------*/
 434
 435/*
 436 * Bio endio functions.
 437 */
 438struct 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
 461static 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
 471static 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
 485static 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 */
 513static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell,
 514                       dm_block_t data_block)
 515{
 516        struct pool *pool = tc->pool;
 517        unsigned long flags;
 518
 519        spin_lock_irqsave(&pool->lock, flags);
 520        dm_cell_release(cell, &pool->deferred_bios);
 521        spin_unlock_irqrestore(&tc->pool->lock, flags);
 522
 523        wake_worker(pool);
 524}
 525
 526/*
 527 * Same as cell_defer except it omits the original holder of the cell.
 528 */
 529static void cell_defer_except(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 530{
 531        struct bio_list bios;
 532        struct pool *pool = tc->pool;
 533        unsigned long flags;
 534
 535        bio_list_init(&bios);
 536
 537        spin_lock_irqsave(&pool->lock, flags);
 538        dm_cell_release_no_holder(cell, &pool->deferred_bios);
 539        spin_unlock_irqrestore(&pool->lock, flags);
 540
 541        wake_worker(pool);
 542}
 543
 544static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
 545{
 546        if (m->bio)
 547                m->bio->bi_end_io = m->saved_bi_end_io;
 548        dm_cell_error(m->cell);
 549        list_del(&m->list);
 550        mempool_free(m, m->tc->pool->mapping_pool);
 551}
 552static void process_prepared_mapping(struct dm_thin_new_mapping *m)
 553{
 554        struct thin_c *tc = m->tc;
 555        struct bio *bio;
 556        int r;
 557
 558        bio = m->bio;
 559        if (bio)
 560                bio->bi_end_io = m->saved_bi_end_io;
 561
 562        if (m->err) {
 563                dm_cell_error(m->cell);
 564                goto out;
 565        }
 566
 567        /*
 568         * Commit the prepared block into the mapping btree.
 569         * Any I/O for this block arriving after this point will get
 570         * remapped to it directly.
 571         */
 572        r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
 573        if (r) {
 574                DMERR("dm_thin_insert_block() failed");
 575                dm_cell_error(m->cell);
 576                goto out;
 577        }
 578
 579        /*
 580         * Release any bios held while the block was being provisioned.
 581         * If we are processing a write bio that completely covers the block,
 582         * we already processed it so can ignore it now when processing
 583         * the bios in the cell.
 584         */
 585        if (bio) {
 586                cell_defer_except(tc, m->cell);
 587                bio_endio(bio, 0);
 588        } else
 589                cell_defer(tc, m->cell, m->data_block);
 590
 591out:
 592        list_del(&m->list);
 593        mempool_free(m, tc->pool->mapping_pool);
 594}
 595
 596static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
 597{
 598        struct thin_c *tc = m->tc;
 599
 600        bio_io_error(m->bio);
 601        cell_defer_except(tc, m->cell);
 602        cell_defer_except(tc, m->cell2);
 603        mempool_free(m, tc->pool->mapping_pool);
 604}
 605
 606static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
 607{
 608        struct thin_c *tc = m->tc;
 609
 610        inc_all_io_entry(tc->pool, m->bio);
 611        cell_defer_except(tc, m->cell);
 612        cell_defer_except(tc, m->cell2);
 613
 614        if (m->pass_discard)
 615                remap_and_issue(tc, m->bio, m->data_block);
 616        else
 617                bio_endio(m->bio, 0);
 618
 619        mempool_free(m, tc->pool->mapping_pool);
 620}
 621
 622static void process_prepared_discard(struct dm_thin_new_mapping *m)
 623{
 624        int r;
 625        struct thin_c *tc = m->tc;
 626
 627        r = dm_thin_remove_block(tc->td, m->virt_block);
 628        if (r)
 629                DMERR("dm_thin_remove_block() failed");
 630
 631        process_prepared_discard_passdown(m);
 632}
 633
 634static void process_prepared(struct pool *pool, struct list_head *head,
 635                             process_mapping_fn *fn)
 636{
 637        unsigned long flags;
 638        struct list_head maps;
 639        struct dm_thin_new_mapping *m, *tmp;
 640
 641        INIT_LIST_HEAD(&maps);
 642        spin_lock_irqsave(&pool->lock, flags);
 643        list_splice_init(head, &maps);
 644        spin_unlock_irqrestore(&pool->lock, flags);
 645
 646        list_for_each_entry_safe(m, tmp, &maps, list)
 647                (*fn)(m);
 648}
 649
 650/*
 651 * Deferred bio jobs.
 652 */
 653static int io_overlaps_block(struct pool *pool, struct bio *bio)
 654{
 655        return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
 656}
 657
 658static int io_overwrites_block(struct pool *pool, struct bio *bio)
 659{
 660        return (bio_data_dir(bio) == WRITE) &&
 661                io_overlaps_block(pool, bio);
 662}
 663
 664static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
 665                               bio_end_io_t *fn)
 666{
 667        *save = bio->bi_end_io;
 668        bio->bi_end_io = fn;
 669}
 670
 671static int ensure_next_mapping(struct pool *pool)
 672{
 673        if (pool->next_mapping)
 674                return 0;
 675
 676        pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
 677
 678        return pool->next_mapping ? 0 : -ENOMEM;
 679}
 680
 681static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
 682{
 683        struct dm_thin_new_mapping *r = pool->next_mapping;
 684
 685        BUG_ON(!pool->next_mapping);
 686
 687        pool->next_mapping = NULL;
 688
 689        return r;
 690}
 691
 692static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
 693                          struct dm_dev *origin, dm_block_t data_origin,
 694                          dm_block_t data_dest,
 695                          struct dm_bio_prison_cell *cell, struct bio *bio)
 696{
 697        int r;
 698        struct pool *pool = tc->pool;
 699        struct dm_thin_new_mapping *m = get_next_mapping(pool);
 700
 701        INIT_LIST_HEAD(&m->list);
 702        m->quiesced = 0;
 703        m->prepared = 0;
 704        m->tc = tc;
 705        m->virt_block = virt_block;
 706        m->data_block = data_dest;
 707        m->cell = cell;
 708        m->err = 0;
 709        m->bio = NULL;
 710
 711        if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
 712                m->quiesced = 1;
 713
 714        /*
 715         * IO to pool_dev remaps to the pool target's data_dev.
 716         *
 717         * If the whole block of data is being overwritten, we can issue the
 718         * bio immediately. Otherwise we use kcopyd to clone the data first.
 719         */
 720        if (io_overwrites_block(pool, bio)) {
 721                struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 722
 723                h->overwrite_mapping = m;
 724                m->bio = bio;
 725                save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 726                inc_all_io_entry(pool, bio);
 727                remap_and_issue(tc, bio, data_dest);
 728        } else {
 729                struct dm_io_region from, to;
 730
 731                from.bdev = origin->bdev;
 732                from.sector = data_origin * pool->sectors_per_block;
 733                from.count = pool->sectors_per_block;
 734
 735                to.bdev = tc->pool_dev->bdev;
 736                to.sector = data_dest * pool->sectors_per_block;
 737                to.count = pool->sectors_per_block;
 738
 739                r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
 740                                   0, copy_complete, m);
 741                if (r < 0) {
 742                        mempool_free(m, pool->mapping_pool);
 743                        DMERR("dm_kcopyd_copy() failed");
 744                        dm_cell_error(cell);
 745                }
 746        }
 747}
 748
 749static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
 750                                   dm_block_t data_origin, dm_block_t data_dest,
 751                                   struct dm_bio_prison_cell *cell, struct bio *bio)
 752{
 753        schedule_copy(tc, virt_block, tc->pool_dev,
 754                      data_origin, data_dest, cell, bio);
 755}
 756
 757static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
 758                                   dm_block_t data_dest,
 759                                   struct dm_bio_prison_cell *cell, struct bio *bio)
 760{
 761        schedule_copy(tc, virt_block, tc->origin_dev,
 762                      virt_block, data_dest, cell, bio);
 763}
 764
 765static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
 766                          dm_block_t data_block, struct dm_bio_prison_cell *cell,
 767                          struct bio *bio)
 768{
 769        struct pool *pool = tc->pool;
 770        struct dm_thin_new_mapping *m = get_next_mapping(pool);
 771
 772        INIT_LIST_HEAD(&m->list);
 773        m->quiesced = 1;
 774        m->prepared = 0;
 775        m->tc = tc;
 776        m->virt_block = virt_block;
 777        m->data_block = data_block;
 778        m->cell = cell;
 779        m->err = 0;
 780        m->bio = NULL;
 781
 782        /*
 783         * If the whole block of data is being overwritten or we are not
 784         * zeroing pre-existing data, we can issue the bio immediately.
 785         * Otherwise we use kcopyd to zero the data first.
 786         */
 787        if (!pool->pf.zero_new_blocks)
 788                process_prepared_mapping(m);
 789
 790        else if (io_overwrites_block(pool, bio)) {
 791                struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 792
 793                h->overwrite_mapping = m;
 794                m->bio = bio;
 795                save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
 796                inc_all_io_entry(pool, bio);
 797                remap_and_issue(tc, bio, data_block);
 798        } else {
 799                int r;
 800                struct dm_io_region to;
 801
 802                to.bdev = tc->pool_dev->bdev;
 803                to.sector = data_block * pool->sectors_per_block;
 804                to.count = pool->sectors_per_block;
 805
 806                r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
 807                if (r < 0) {
 808                        mempool_free(m, pool->mapping_pool);
 809                        DMERR("dm_kcopyd_zero() failed");
 810                        dm_cell_error(cell);
 811                }
 812        }
 813}
 814
 815static int commit(struct pool *pool)
 816{
 817        int r;
 818
 819        r = dm_pool_commit_metadata(pool->pmd);
 820        if (r)
 821                DMERR("commit failed, error = %d", r);
 822
 823        return r;
 824}
 825
 826/*
 827 * A non-zero return indicates read_only or fail_io mode.
 828 * Many callers don't care about the return value.
 829 */
 830static int commit_or_fallback(struct pool *pool)
 831{
 832        int r;
 833
 834        if (get_pool_mode(pool) != PM_WRITE)
 835                return -EINVAL;
 836
 837        r = commit(pool);
 838        if (r)
 839                set_pool_mode(pool, PM_READ_ONLY);
 840
 841        return r;
 842}
 843
 844static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
 845{
 846        int r;
 847        dm_block_t free_blocks;
 848        unsigned long flags;
 849        struct pool *pool = tc->pool;
 850
 851        r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 852        if (r)
 853                return r;
 854
 855        if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
 856                DMWARN("%s: reached low water mark, sending event.",
 857                       dm_device_name(pool->pool_md));
 858                spin_lock_irqsave(&pool->lock, flags);
 859                pool->low_water_triggered = 1;
 860                spin_unlock_irqrestore(&pool->lock, flags);
 861                dm_table_event(pool->ti->table);
 862        }
 863
 864        if (!free_blocks) {
 865                if (pool->no_free_space)
 866                        return -ENOSPC;
 867                else {
 868                        /*
 869                         * Try to commit to see if that will free up some
 870                         * more space.
 871                         */
 872                        (void) commit_or_fallback(pool);
 873
 874                        r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
 875                        if (r)
 876                                return r;
 877
 878                        /*
 879                         * If we still have no space we set a flag to avoid
 880                         * doing all this checking and return -ENOSPC.
 881                         */
 882                        if (!free_blocks) {
 883                                DMWARN("%s: no free space available.",
 884                                       dm_device_name(pool->pool_md));
 885                                spin_lock_irqsave(&pool->lock, flags);
 886                                pool->no_free_space = 1;
 887                                spin_unlock_irqrestore(&pool->lock, flags);
 888                                return -ENOSPC;
 889                        }
 890                }
 891        }
 892
 893        r = dm_pool_alloc_data_block(pool->pmd, result);
 894        if (r)
 895                return r;
 896
 897        return 0;
 898}
 899
 900/*
 901 * If we have run out of space, queue bios until the device is
 902 * resumed, presumably after having been reloaded with more space.
 903 */
 904static void retry_on_resume(struct bio *bio)
 905{
 906        struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
 907        struct thin_c *tc = h->tc;
 908        struct pool *pool = tc->pool;
 909        unsigned long flags;
 910
 911        spin_lock_irqsave(&pool->lock, flags);
 912        bio_list_add(&pool->retry_on_resume_list, bio);
 913        spin_unlock_irqrestore(&pool->lock, flags);
 914}
 915
 916static void no_space(struct dm_bio_prison_cell *cell)
 917{
 918        struct bio *bio;
 919        struct bio_list bios;
 920
 921        bio_list_init(&bios);
 922        dm_cell_release(cell, &bios);
 923
 924        while ((bio = bio_list_pop(&bios)))
 925                retry_on_resume(bio);
 926}
 927
 928static void process_discard(struct thin_c *tc, struct bio *bio)
 929{
 930        int r;
 931        unsigned long flags;
 932        struct pool *pool = tc->pool;
 933        struct dm_bio_prison_cell *cell, *cell2;
 934        struct dm_cell_key key, key2;
 935        dm_block_t block = get_bio_block(tc, bio);
 936        struct dm_thin_lookup_result lookup_result;
 937        struct dm_thin_new_mapping *m;
 938
 939        build_virtual_key(tc->td, block, &key);
 940        if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
 941                return;
 942
 943        r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
 944        switch (r) {
 945        case 0:
 946                /*
 947                 * Check nobody is fiddling with this pool block.  This can
 948                 * happen if someone's in the process of breaking sharing
 949                 * on this block.
 950                 */
 951                build_data_key(tc->td, lookup_result.block, &key2);
 952                if (dm_bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
 953                        cell_defer_except(tc, cell);
 954                        break;
 955                }
 956
 957                if (io_overlaps_block(pool, bio)) {
 958                        /*
 959                         * IO may still be going to the destination block.  We must
 960                         * quiesce before we can do the removal.
 961                         */
 962                        m = get_next_mapping(pool);
 963                        m->tc = tc;
 964                        m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
 965                        m->virt_block = block;
 966                        m->data_block = lookup_result.block;
 967                        m->cell = cell;
 968                        m->cell2 = cell2;
 969                        m->err = 0;
 970                        m->bio = bio;
 971
 972                        if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
 973                                spin_lock_irqsave(&pool->lock, flags);
 974                                list_add(&m->list, &pool->prepared_discards);
 975                                spin_unlock_irqrestore(&pool->lock, flags);
 976                                wake_worker(pool);
 977                        }
 978                } else {
 979                        inc_all_io_entry(pool, bio);
 980                        cell_defer_except(tc, cell);
 981                        cell_defer_except(tc, cell2);
 982
 983                        /*
 984                         * The DM core makes sure that the discard doesn't span
 985                         * a block boundary.  So we submit the discard of a
 986                         * partial block appropriately.
 987                         */
 988                        if ((!lookup_result.shared) && pool->pf.discard_passdown)
 989                                remap_and_issue(tc, bio, lookup_result.block);
 990                        else
 991                                bio_endio(bio, 0);
 992                }
 993                break;
 994
 995        case -ENODATA:
 996                /*
 997                 * It isn't provisioned, just forget it.
 998                 */
 999                cell_defer_except(tc, cell);
1000                bio_endio(bio, 0);
1001                break;
1002
1003        default:
1004                DMERR("discard: find block unexpectedly returned %d", r);
1005                cell_defer_except(tc, cell);
1006                bio_io_error(bio);
1007                break;
1008        }
1009}
1010
1011static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1012                          struct dm_cell_key *key,
1013                          struct dm_thin_lookup_result *lookup_result,
1014                          struct dm_bio_prison_cell *cell)
1015{
1016        int r;
1017        dm_block_t data_block;
1018
1019        r = alloc_data_block(tc, &data_block);
1020        switch (r) {
1021        case 0:
1022                schedule_internal_copy(tc, block, lookup_result->block,
1023                                       data_block, cell, bio);
1024                break;
1025
1026        case -ENOSPC:
1027                no_space(cell);
1028                break;
1029
1030        default:
1031                DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1032                dm_cell_error(cell);
1033                break;
1034        }
1035}
1036
1037static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1038                               dm_block_t block,
1039                               struct dm_thin_lookup_result *lookup_result)
1040{
1041        struct dm_bio_prison_cell *cell;
1042        struct pool *pool = tc->pool;
1043        struct dm_cell_key key;
1044
1045        /*
1046         * If cell is already occupied, then sharing is already in the process
1047         * of being broken so we have nothing further to do here.
1048         */
1049        build_data_key(tc->td, lookup_result->block, &key);
1050        if (dm_bio_detain(pool->prison, &key, bio, &cell))
1051                return;
1052
1053        if (bio_data_dir(bio) == WRITE && bio->bi_size)
1054                break_sharing(tc, bio, block, &key, lookup_result, cell);
1055        else {
1056                struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1057
1058                h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1059                inc_all_io_entry(pool, bio);
1060                cell_defer_except(tc, cell);
1061
1062                remap_and_issue(tc, bio, lookup_result->block);
1063        }
1064}
1065
1066static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1067                            struct dm_bio_prison_cell *cell)
1068{
1069        int r;
1070        dm_block_t data_block;
1071
1072        /*
1073         * Remap empty bios (flushes) immediately, without provisioning.
1074         */
1075        if (!bio->bi_size) {
1076                inc_all_io_entry(tc->pool, bio);
1077                cell_defer_except(tc, cell);
1078
1079                remap_and_issue(tc, bio, 0);
1080                return;
1081        }
1082
1083        /*
1084         * Fill read bios with zeroes and complete them immediately.
1085         */
1086        if (bio_data_dir(bio) == READ) {
1087                zero_fill_bio(bio);
1088                cell_defer_except(tc, cell);
1089                bio_endio(bio, 0);
1090                return;
1091        }
1092
1093        r = alloc_data_block(tc, &data_block);
1094        switch (r) {
1095        case 0:
1096                if (tc->origin_dev)
1097                        schedule_external_copy(tc, block, data_block, cell, bio);
1098                else
1099                        schedule_zero(tc, block, data_block, cell, bio);
1100                break;
1101
1102        case -ENOSPC:
1103                no_space(cell);
1104                break;
1105
1106        default:
1107                DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1108                set_pool_mode(tc->pool, PM_READ_ONLY);
1109                dm_cell_error(cell);
1110                break;
1111        }
1112}
1113
1114static void process_bio(struct thin_c *tc, struct bio *bio)
1115{
1116        int r;
1117        dm_block_t block = get_bio_block(tc, bio);
1118        struct dm_bio_prison_cell *cell;
1119        struct dm_cell_key key;
1120        struct dm_thin_lookup_result lookup_result;
1121
1122        /*
1123         * If cell is already occupied, then the block is already
1124         * being provisioned so we have nothing further to do here.
1125         */
1126        build_virtual_key(tc->td, block, &key);
1127        if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
1128                return;
1129
1130        r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1131        switch (r) {
1132        case 0:
1133                if (lookup_result.shared) {
1134                        process_shared_bio(tc, bio, block, &lookup_result);
1135                        cell_defer_except(tc, cell);
1136                } else {
1137                        inc_all_io_entry(tc->pool, bio);
1138                        cell_defer_except(tc, cell);
1139
1140                        remap_and_issue(tc, bio, lookup_result.block);
1141                }
1142                break;
1143
1144        case -ENODATA:
1145                if (bio_data_dir(bio) == READ && tc->origin_dev) {
1146                        inc_all_io_entry(tc->pool, bio);
1147                        cell_defer_except(tc, cell);
1148
1149                        remap_to_origin_and_issue(tc, bio);
1150                } else
1151                        provision_block(tc, bio, block, cell);
1152                break;
1153
1154        default:
1155                DMERR("dm_thin_find_block() failed, error = %d", r);
1156                cell_defer_except(tc, cell);
1157                bio_io_error(bio);
1158                break;
1159        }
1160}
1161
1162static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1163{
1164        int r;
1165        int rw = bio_data_dir(bio);
1166        dm_block_t block = get_bio_block(tc, bio);
1167        struct dm_thin_lookup_result lookup_result;
1168
1169        r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1170        switch (r) {
1171        case 0:
1172                if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1173                        bio_io_error(bio);
1174                else {
1175                        inc_all_io_entry(tc->pool, bio);
1176                        remap_and_issue(tc, bio, lookup_result.block);
1177                }
1178                break;
1179
1180        case -ENODATA:
1181                if (rw != READ) {
1182                        bio_io_error(bio);
1183                        break;
1184                }
1185
1186                if (tc->origin_dev) {
1187                        inc_all_io_entry(tc->pool, bio);
1188                        remap_to_origin_and_issue(tc, bio);
1189                        break;
1190                }
1191
1192                zero_fill_bio(bio);
1193                bio_endio(bio, 0);
1194                break;
1195
1196        default:
1197                DMERR("dm_thin_find_block() failed, error = %d", r);
1198                bio_io_error(bio);
1199                break;
1200        }
1201}
1202
1203static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1204{
1205        bio_io_error(bio);
1206}
1207
1208static int need_commit_due_to_time(struct pool *pool)
1209{
1210        return jiffies < pool->last_commit_jiffies ||
1211               jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1212}
1213
1214static void process_deferred_bios(struct pool *pool)
1215{
1216        unsigned long flags;
1217        struct bio *bio;
1218        struct bio_list bios;
1219
1220        bio_list_init(&bios);
1221
1222        spin_lock_irqsave(&pool->lock, flags);
1223        bio_list_merge(&bios, &pool->deferred_bios);
1224        bio_list_init(&pool->deferred_bios);
1225        spin_unlock_irqrestore(&pool->lock, flags);
1226
1227        while ((bio = bio_list_pop(&bios))) {
1228                struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1229                struct thin_c *tc = h->tc;
1230
1231                /*
1232                 * If we've got no free new_mapping structs, and processing
1233                 * this bio might require one, we pause until there are some
1234                 * prepared mappings to process.
1235                 */
1236                if (ensure_next_mapping(pool)) {
1237                        spin_lock_irqsave(&pool->lock, flags);
1238                        bio_list_merge(&pool->deferred_bios, &bios);
1239                        spin_unlock_irqrestore(&pool->lock, flags);
1240
1241                        break;
1242                }
1243
1244                if (bio->bi_rw & REQ_DISCARD)
1245                        pool->process_discard(tc, bio);
1246                else
1247                        pool->process_bio(tc, bio);
1248        }
1249
1250        /*
1251         * If there are any deferred flush bios, we must commit
1252         * the metadata before issuing them.
1253         */
1254        bio_list_init(&bios);
1255        spin_lock_irqsave(&pool->lock, flags);
1256        bio_list_merge(&bios, &pool->deferred_flush_bios);
1257        bio_list_init(&pool->deferred_flush_bios);
1258        spin_unlock_irqrestore(&pool->lock, flags);
1259
1260        if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1261                return;
1262
1263        if (commit_or_fallback(pool)) {
1264                while ((bio = bio_list_pop(&bios)))
1265                        bio_io_error(bio);
1266                return;
1267        }
1268        pool->last_commit_jiffies = jiffies;
1269
1270        while ((bio = bio_list_pop(&bios)))
1271                generic_make_request(bio);
1272}
1273
1274static void do_worker(struct work_struct *ws)
1275{
1276        struct pool *pool = container_of(ws, struct pool, worker);
1277
1278        process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1279        process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1280        process_deferred_bios(pool);
1281}
1282
1283/*
1284 * We want to commit periodically so that not too much
1285 * unwritten data builds up.
1286 */
1287static void do_waker(struct work_struct *ws)
1288{
1289        struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1290        wake_worker(pool);
1291        queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1292}
1293
1294/*----------------------------------------------------------------*/
1295
1296static enum pool_mode get_pool_mode(struct pool *pool)
1297{
1298        return pool->pf.mode;
1299}
1300
1301static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1302{
1303        int r;
1304
1305        pool->pf.mode = mode;
1306
1307        switch (mode) {
1308        case PM_FAIL:
1309                DMERR("switching pool to failure mode");
1310                pool->process_bio = process_bio_fail;
1311                pool->process_discard = process_bio_fail;
1312                pool->process_prepared_mapping = process_prepared_mapping_fail;
1313                pool->process_prepared_discard = process_prepared_discard_fail;
1314                break;
1315
1316        case PM_READ_ONLY:
1317                DMERR("switching pool to read-only mode");
1318                r = dm_pool_abort_metadata(pool->pmd);
1319                if (r) {
1320                        DMERR("aborting transaction failed");
1321                        set_pool_mode(pool, PM_FAIL);
1322                } else {
1323                        dm_pool_metadata_read_only(pool->pmd);
1324                        pool->process_bio = process_bio_read_only;
1325                        pool->process_discard = process_discard;
1326                        pool->process_prepared_mapping = process_prepared_mapping_fail;
1327                        pool->process_prepared_discard = process_prepared_discard_passdown;
1328                }
1329                break;
1330
1331        case PM_WRITE:
1332                pool->process_bio = process_bio;
1333                pool->process_discard = process_discard;
1334                pool->process_prepared_mapping = process_prepared_mapping;
1335                pool->process_prepared_discard = process_prepared_discard;
1336                break;
1337        }
1338}
1339
1340/*----------------------------------------------------------------*/
1341
1342/*
1343 * Mapping functions.
1344 */
1345
1346/*
1347 * Called only while mapping a thin bio to hand it over to the workqueue.
1348 */
1349static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1350{
1351        unsigned long flags;
1352        struct pool *pool = tc->pool;
1353
1354        spin_lock_irqsave(&pool->lock, flags);
1355        bio_list_add(&pool->deferred_bios, bio);
1356        spin_unlock_irqrestore(&pool->lock, flags);
1357
1358        wake_worker(pool);
1359}
1360
1361static struct dm_thin_endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1362{
1363        struct pool *pool = tc->pool;
1364        struct dm_thin_endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1365
1366        h->tc = tc;
1367        h->shared_read_entry = NULL;
1368        h->all_io_entry = NULL;
1369        h->overwrite_mapping = NULL;
1370
1371        return h;
1372}
1373
1374/*
1375 * Non-blocking function called from the thin target's map function.
1376 */
1377static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1378                        union map_info *map_context)
1379{
1380        int r;
1381        struct thin_c *tc = ti->private;
1382        dm_block_t block = get_bio_block(tc, bio);
1383        struct dm_thin_device *td = tc->td;
1384        struct dm_thin_lookup_result result;
1385        struct dm_bio_prison_cell *cell1, *cell2;
1386        struct dm_cell_key key;
1387
1388        map_context->ptr = thin_hook_bio(tc, bio);
1389
1390        if (get_pool_mode(tc->pool) == PM_FAIL) {
1391                bio_io_error(bio);
1392                return DM_MAPIO_SUBMITTED;
1393        }
1394
1395        if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1396                thin_defer_bio(tc, bio);
1397                return DM_MAPIO_SUBMITTED;
1398        }
1399
1400        r = dm_thin_find_block(td, block, 0, &result);
1401
1402        /*
1403         * Note that we defer readahead too.
1404         */
1405        switch (r) {
1406        case 0:
1407                if (unlikely(result.shared)) {
1408                        /*
1409                         * We have a race condition here between the
1410                         * result.shared value returned by the lookup and
1411                         * snapshot creation, which may cause new
1412                         * sharing.
1413                         *
1414                         * To avoid this always quiesce the origin before
1415                         * taking the snap.  You want to do this anyway to
1416                         * ensure a consistent application view
1417                         * (i.e. lockfs).
1418                         *
1419                         * More distant ancestors are irrelevant. The
1420                         * shared flag will be set in their case.
1421                         */
1422                        thin_defer_bio(tc, bio);
1423                        return DM_MAPIO_SUBMITTED;
1424                }
1425
1426                build_virtual_key(tc->td, block, &key);
1427                if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1))
1428                        return DM_MAPIO_SUBMITTED;
1429
1430                build_data_key(tc->td, result.block, &key);
1431                if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2)) {
1432                        cell_defer_except(tc, cell1);
1433                        return DM_MAPIO_SUBMITTED;
1434                }
1435
1436                inc_all_io_entry(tc->pool, bio);
1437                cell_defer_except(tc, cell2);
1438                cell_defer_except(tc, cell1);
1439
1440                remap(tc, bio, result.block);
1441                return DM_MAPIO_REMAPPED;
1442
1443        case -ENODATA:
1444                if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1445                        /*
1446                         * This block isn't provisioned, and we have no way
1447                         * of doing so.  Just error it.
1448                         */
1449                        bio_io_error(bio);
1450                        r = DM_MAPIO_SUBMITTED;
1451                        break;
1452                }
1453                /* fall through */
1454
1455        case -EWOULDBLOCK:
1456                /*
1457                 * In future, the failed dm_thin_find_block above could
1458                 * provide the hint to load the metadata into cache.
1459                 */
1460                thin_defer_bio(tc, bio);
1461                r = DM_MAPIO_SUBMITTED;
1462                break;
1463
1464        default:
1465                /*
1466                 * Must always call bio_io_error on failure.
1467                 * dm_thin_find_block can fail with -EINVAL if the
1468                 * pool is switched to fail-io mode.
1469                 */
1470                bio_io_error(bio);
1471                r = DM_MAPIO_SUBMITTED;
1472                break;
1473        }
1474
1475        return r;
1476}
1477
1478static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1479{
1480        int r;
1481        unsigned long flags;
1482        struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1483
1484        spin_lock_irqsave(&pt->pool->lock, flags);
1485        r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1486        spin_unlock_irqrestore(&pt->pool->lock, flags);
1487
1488        if (!r) {
1489                struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1490                r = bdi_congested(&q->backing_dev_info, bdi_bits);
1491        }
1492
1493        return r;
1494}
1495
1496static void __requeue_bios(struct pool *pool)
1497{
1498        bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1499        bio_list_init(&pool->retry_on_resume_list);
1500}
1501
1502/*----------------------------------------------------------------
1503 * Binding of control targets to a pool object
1504 *--------------------------------------------------------------*/
1505static bool data_dev_supports_discard(struct pool_c *pt)
1506{
1507        struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1508
1509        return q && blk_queue_discard(q);
1510}
1511
1512/*
1513 * If discard_passdown was enabled verify that the data device
1514 * supports discards.  Disable discard_passdown if not.
1515 */
1516static void disable_passdown_if_not_supported(struct pool_c *pt)
1517{
1518        struct pool *pool = pt->pool;
1519        struct block_device *data_bdev = pt->data_dev->bdev;
1520        struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1521        sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1522        const char *reason = NULL;
1523        char buf[BDEVNAME_SIZE];
1524
1525        if (!pt->adjusted_pf.discard_passdown)
1526                return;
1527
1528        if (!data_dev_supports_discard(pt))
1529                reason = "discard unsupported";
1530
1531        else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1532                reason = "max discard sectors smaller than a block";
1533
1534        else if (data_limits->discard_granularity > block_size)
1535                reason = "discard granularity larger than a block";
1536
1537        else if (block_size & (data_limits->discard_granularity - 1))
1538                reason = "discard granularity not a factor of block size";
1539
1540        if (reason) {
1541                DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1542                pt->adjusted_pf.discard_passdown = false;
1543        }
1544}
1545
1546static int bind_control_target(struct pool *pool, struct dm_target *ti)
1547{
1548        struct pool_c *pt = ti->private;
1549
1550        /*
1551         * We want to make sure that degraded pools are never upgraded.
1552         */
1553        enum pool_mode old_mode = pool->pf.mode;
1554        enum pool_mode new_mode = pt->adjusted_pf.mode;
1555
1556        if (old_mode > new_mode)
1557                new_mode = old_mode;
1558
1559        pool->ti = ti;
1560        pool->low_water_blocks = pt->low_water_blocks;
1561        pool->pf = pt->adjusted_pf;
1562
1563        set_pool_mode(pool, new_mode);
1564
1565        return 0;
1566}
1567
1568static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1569{
1570        if (pool->ti == ti)
1571                pool->ti = NULL;
1572}
1573
1574/*----------------------------------------------------------------
1575 * Pool creation
1576 *--------------------------------------------------------------*/
1577/* Initialize pool features. */
1578static void pool_features_init(struct pool_features *pf)
1579{
1580        pf->mode = PM_WRITE;
1581        pf->zero_new_blocks = true;
1582        pf->discard_enabled = true;
1583        pf->discard_passdown = true;
1584}
1585
1586static void __pool_destroy(struct pool *pool)
1587{
1588        __pool_table_remove(pool);
1589
1590        if (dm_pool_metadata_close(pool->pmd) < 0)
1591                DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1592
1593        dm_bio_prison_destroy(pool->prison);
1594        dm_kcopyd_client_destroy(pool->copier);
1595
1596        if (pool->wq)
1597                destroy_workqueue(pool->wq);
1598
1599        if (pool->next_mapping)
1600                mempool_free(pool->next_mapping, pool->mapping_pool);
1601        mempool_destroy(pool->mapping_pool);
1602        mempool_destroy(pool->endio_hook_pool);
1603        dm_deferred_set_destroy(pool->shared_read_ds);
1604        dm_deferred_set_destroy(pool->all_io_ds);
1605        kfree(pool);
1606}
1607
1608static struct kmem_cache *_new_mapping_cache;
1609static struct kmem_cache *_endio_hook_cache;
1610
1611static struct pool *pool_create(struct mapped_device *pool_md,
1612                                struct block_device *metadata_dev,
1613                                unsigned long block_size,
1614                                int read_only, char **error)
1615{
1616        int r;
1617        void *err_p;
1618        struct pool *pool;
1619        struct dm_pool_metadata *pmd;
1620        bool format_device = read_only ? false : true;
1621
1622        pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1623        if (IS_ERR(pmd)) {
1624                *error = "Error creating metadata object";
1625                return (struct pool *)pmd;
1626        }
1627
1628        pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1629        if (!pool) {
1630                *error = "Error allocating memory for pool";
1631                err_p = ERR_PTR(-ENOMEM);
1632                goto bad_pool;
1633        }
1634
1635        pool->pmd = pmd;
1636        pool->sectors_per_block = block_size;
1637        if (block_size & (block_size - 1))
1638                pool->sectors_per_block_shift = -1;
1639        else
1640                pool->sectors_per_block_shift = __ffs(block_size);
1641        pool->low_water_blocks = 0;
1642        pool_features_init(&pool->pf);
1643        pool->prison = dm_bio_prison_create(PRISON_CELLS);
1644        if (!pool->prison) {
1645                *error = "Error creating pool's bio prison";
1646                err_p = ERR_PTR(-ENOMEM);
1647                goto bad_prison;
1648        }
1649
1650        pool->copier = dm_kcopyd_client_create();
1651        if (IS_ERR(pool->copier)) {
1652                r = PTR_ERR(pool->copier);
1653                *error = "Error creating pool's kcopyd client";
1654                err_p = ERR_PTR(r);
1655                goto bad_kcopyd_client;
1656        }
1657
1658        /*
1659         * Create singlethreaded workqueue that will service all devices
1660         * that use this metadata.
1661         */
1662        pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1663        if (!pool->wq) {
1664                *error = "Error creating pool's workqueue";
1665                err_p = ERR_PTR(-ENOMEM);
1666                goto bad_wq;
1667        }
1668
1669        INIT_WORK(&pool->worker, do_worker);
1670        INIT_DELAYED_WORK(&pool->waker, do_waker);
1671        spin_lock_init(&pool->lock);
1672        bio_list_init(&pool->deferred_bios);
1673        bio_list_init(&pool->deferred_flush_bios);
1674        INIT_LIST_HEAD(&pool->prepared_mappings);
1675        INIT_LIST_HEAD(&pool->prepared_discards);
1676        pool->low_water_triggered = 0;
1677        pool->no_free_space = 0;
1678        bio_list_init(&pool->retry_on_resume_list);
1679
1680        pool->shared_read_ds = dm_deferred_set_create();
1681        if (!pool->shared_read_ds) {
1682                *error = "Error creating pool's shared read deferred set";
1683                err_p = ERR_PTR(-ENOMEM);
1684                goto bad_shared_read_ds;
1685        }
1686
1687        pool->all_io_ds = dm_deferred_set_create();
1688        if (!pool->all_io_ds) {
1689                *error = "Error creating pool's all io deferred set";
1690                err_p = ERR_PTR(-ENOMEM);
1691                goto bad_all_io_ds;
1692        }
1693
1694        pool->next_mapping = NULL;
1695        pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1696                                                      _new_mapping_cache);
1697        if (!pool->mapping_pool) {
1698                *error = "Error creating pool's mapping mempool";
1699                err_p = ERR_PTR(-ENOMEM);
1700                goto bad_mapping_pool;
1701        }
1702
1703        pool->endio_hook_pool = mempool_create_slab_pool(ENDIO_HOOK_POOL_SIZE,
1704                                                         _endio_hook_cache);
1705        if (!pool->endio_hook_pool) {
1706                *error = "Error creating pool's endio_hook mempool";
1707                err_p = ERR_PTR(-ENOMEM);
1708                goto bad_endio_hook_pool;
1709        }
1710        pool->ref_count = 1;
1711        pool->last_commit_jiffies = jiffies;
1712        pool->pool_md = pool_md;
1713        pool->md_dev = metadata_dev;
1714        __pool_table_insert(pool);
1715
1716        return pool;
1717
1718bad_endio_hook_pool:
1719        mempool_destroy(pool->mapping_pool);
1720bad_mapping_pool:
1721        dm_deferred_set_destroy(pool->all_io_ds);
1722bad_all_io_ds:
1723        dm_deferred_set_destroy(pool->shared_read_ds);
1724bad_shared_read_ds:
1725        destroy_workqueue(pool->wq);
1726bad_wq:
1727        dm_kcopyd_client_destroy(pool->copier);
1728bad_kcopyd_client:
1729        dm_bio_prison_destroy(pool->prison);
1730bad_prison:
1731        kfree(pool);
1732bad_pool:
1733        if (dm_pool_metadata_close(pmd))
1734                DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1735
1736        return err_p;
1737}
1738
1739static void __pool_inc(struct pool *pool)
1740{
1741        BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1742        pool->ref_count++;
1743}
1744
1745static void __pool_dec(struct pool *pool)
1746{
1747        BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1748        BUG_ON(!pool->ref_count);
1749        if (!--pool->ref_count)
1750                __pool_destroy(pool);
1751}
1752
1753static struct pool *__pool_find(struct mapped_device *pool_md,
1754                                struct block_device *metadata_dev,
1755                                unsigned long block_size, int read_only,
1756                                char **error, int *created)
1757{
1758        struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1759
1760        if (pool) {
1761                if (pool->pool_md != pool_md) {
1762                        *error = "metadata device already in use by a pool";
1763                        return ERR_PTR(-EBUSY);
1764                }
1765                __pool_inc(pool);
1766
1767        } else {
1768                pool = __pool_table_lookup(pool_md);
1769                if (pool) {
1770                        if (pool->md_dev != metadata_dev) {
1771                                *error = "different pool cannot replace a pool";
1772                                return ERR_PTR(-EINVAL);
1773                        }
1774                        __pool_inc(pool);
1775
1776                } else {
1777                        pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1778                        *created = 1;
1779                }
1780        }
1781
1782        return pool;
1783}
1784
1785/*----------------------------------------------------------------
1786 * Pool target methods
1787 *--------------------------------------------------------------*/
1788static void pool_dtr(struct dm_target *ti)
1789{
1790        struct pool_c *pt = ti->private;
1791
1792        mutex_lock(&dm_thin_pool_table.mutex);
1793
1794        unbind_control_target(pt->pool, ti);
1795        __pool_dec(pt->pool);
1796        dm_put_device(ti, pt->metadata_dev);
1797        dm_put_device(ti, pt->data_dev);
1798        kfree(pt);
1799
1800        mutex_unlock(&dm_thin_pool_table.mutex);
1801}
1802
1803static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1804                               struct dm_target *ti)
1805{
1806        int r;
1807        unsigned argc;
1808        const char *arg_name;
1809
1810        static struct dm_arg _args[] = {
1811                {0, 3, "Invalid number of pool feature arguments"},
1812        };
1813
1814        /*
1815         * No feature arguments supplied.
1816         */
1817        if (!as->argc)
1818                return 0;
1819
1820        r = dm_read_arg_group(_args, as, &argc, &ti->error);
1821        if (r)
1822                return -EINVAL;
1823
1824        while (argc && !r) {
1825                arg_name = dm_shift_arg(as);
1826                argc--;
1827
1828                if (!strcasecmp(arg_name, "skip_block_zeroing"))
1829                        pf->zero_new_blocks = false;
1830
1831                else if (!strcasecmp(arg_name, "ignore_discard"))
1832                        pf->discard_enabled = false;
1833
1834                else if (!strcasecmp(arg_name, "no_discard_passdown"))
1835                        pf->discard_passdown = false;
1836
1837                else if (!strcasecmp(arg_name, "read_only"))
1838                        pf->mode = PM_READ_ONLY;
1839
1840                else {
1841                        ti->error = "Unrecognised pool feature requested";
1842                        r = -EINVAL;
1843                        break;
1844                }
1845        }
1846
1847        return r;
1848}
1849
1850/*
1851 * thin-pool <metadata dev> <data dev>
1852 *           <data block size (sectors)>
1853 *           <low water mark (blocks)>
1854 *           [<#feature args> [<arg>]*]
1855 *
1856 * Optional feature arguments are:
1857 *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1858 *           ignore_discard: disable discard
1859 *           no_discard_passdown: don't pass discards down to the data device
1860 */
1861static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1862{
1863        int r, pool_created = 0;
1864        struct pool_c *pt;
1865        struct pool *pool;
1866        struct pool_features pf;
1867        struct dm_arg_set as;
1868        struct dm_dev *data_dev;
1869        unsigned long block_size;
1870        dm_block_t low_water_blocks;
1871        struct dm_dev *metadata_dev;
1872        sector_t metadata_dev_size;
1873        char b[BDEVNAME_SIZE];
1874
1875        /*
1876         * FIXME Remove validation from scope of lock.
1877         */
1878        mutex_lock(&dm_thin_pool_table.mutex);
1879
1880        if (argc < 4) {
1881                ti->error = "Invalid argument count";
1882                r = -EINVAL;
1883                goto out_unlock;
1884        }
1885        as.argc = argc;
1886        as.argv = argv;
1887
1888        r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1889        if (r) {
1890                ti->error = "Error opening metadata block device";
1891                goto out_unlock;
1892        }
1893
1894        metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1895        if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1896                DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1897                       bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1898
1899        r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1900        if (r) {
1901                ti->error = "Error getting data device";
1902                goto out_metadata;
1903        }
1904
1905        if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1906            block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1907            block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1908            block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1909                ti->error = "Invalid block size";
1910                r = -EINVAL;
1911                goto out;
1912        }
1913
1914        if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1915                ti->error = "Invalid low water mark";
1916                r = -EINVAL;
1917                goto out;
1918        }
1919
1920        /*
1921         * Set default pool features.
1922         */
1923        pool_features_init(&pf);
1924
1925        dm_consume_args(&as, 4);
1926        r = parse_pool_features(&as, &pf, ti);
1927        if (r)
1928                goto out;
1929
1930        pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1931        if (!pt) {
1932                r = -ENOMEM;
1933                goto out;
1934        }
1935
1936        pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1937                           block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
1938        if (IS_ERR(pool)) {
1939                r = PTR_ERR(pool);
1940                goto out_free_pt;
1941        }
1942
1943        /*
1944         * 'pool_created' reflects whether this is the first table load.
1945         * Top level discard support is not allowed to be changed after
1946         * initial load.  This would require a pool reload to trigger thin
1947         * device changes.
1948         */
1949        if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1950                ti->error = "Discard support cannot be disabled once enabled";
1951                r = -EINVAL;
1952                goto out_flags_changed;
1953        }
1954
1955        pt->pool = pool;
1956        pt->ti = ti;
1957        pt->metadata_dev = metadata_dev;
1958        pt->data_dev = data_dev;
1959        pt->low_water_blocks = low_water_blocks;
1960        pt->adjusted_pf = pt->requested_pf = pf;
1961        ti->num_flush_requests = 1;
1962
1963        /*
1964         * Only need to enable discards if the pool should pass
1965         * them down to the data device.  The thin device's discard
1966         * processing will cause mappings to be removed from the btree.
1967         */
1968        if (pf.discard_enabled && pf.discard_passdown) {
1969                ti->num_discard_requests = 1;
1970
1971                /*
1972                 * Setting 'discards_supported' circumvents the normal
1973                 * stacking of discard limits (this keeps the pool and
1974                 * thin devices' discard limits consistent).
1975                 */
1976                ti->discards_supported = true;
1977                ti->discard_zeroes_data_unsupported = true;
1978        }
1979        ti->private = pt;
1980
1981        pt->callbacks.congested_fn = pool_is_congested;
1982        dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1983
1984        mutex_unlock(&dm_thin_pool_table.mutex);
1985
1986        return 0;
1987
1988out_flags_changed:
1989        __pool_dec(pool);
1990out_free_pt:
1991        kfree(pt);
1992out:
1993        dm_put_device(ti, data_dev);
1994out_metadata:
1995        dm_put_device(ti, metadata_dev);
1996out_unlock:
1997        mutex_unlock(&dm_thin_pool_table.mutex);
1998
1999        return r;
2000}
2001
2002static int pool_map(struct dm_target *ti, struct bio *bio,
2003                    union map_info *map_context)
2004{
2005        int r;
2006        struct pool_c *pt = ti->private;
2007        struct pool *pool = pt->pool;
2008        unsigned long flags;
2009
2010        /*
2011         * As this is a singleton target, ti->begin is always zero.
2012         */
2013        spin_lock_irqsave(&pool->lock, flags);
2014        bio->bi_bdev = pt->data_dev->bdev;
2015        r = DM_MAPIO_REMAPPED;
2016        spin_unlock_irqrestore(&pool->lock, flags);
2017
2018        return r;
2019}
2020
2021/*
2022 * Retrieves the number of blocks of the data device from
2023 * the superblock and compares it to the actual device size,
2024 * thus resizing the data device in case it has grown.
2025 *
2026 * This both copes with opening preallocated data devices in the ctr
2027 * being followed by a resume
2028 * -and-
2029 * calling the resume method individually after userspace has
2030 * grown the data device in reaction to a table event.
2031 */
2032static int pool_preresume(struct dm_target *ti)
2033{
2034        int r;
2035        struct pool_c *pt = ti->private;
2036        struct pool *pool = pt->pool;
2037        sector_t data_size = ti->len;
2038        dm_block_t sb_data_size;
2039
2040        /*
2041         * Take control of the pool object.
2042         */
2043        r = bind_control_target(pool, ti);
2044        if (r)
2045                return r;
2046
2047        (void) sector_div(data_size, pool->sectors_per_block);
2048
2049        r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2050        if (r) {
2051                DMERR("failed to retrieve data device size");
2052                return r;
2053        }
2054
2055        if (data_size < sb_data_size) {
2056                DMERR("pool target too small, is %llu blocks (expected %llu)",
2057                      (unsigned long long)data_size, sb_data_size);
2058                return -EINVAL;
2059
2060        } else if (data_size > sb_data_size) {
2061                r = dm_pool_resize_data_dev(pool->pmd, data_size);
2062                if (r) {
2063                        DMERR("failed to resize data device");
2064                        /* FIXME Stricter than necessary: Rollback transaction instead here */
2065                        set_pool_mode(pool, PM_READ_ONLY);
2066                        return r;
2067                }
2068
2069                (void) commit_or_fallback(pool);
2070        }
2071
2072        return 0;
2073}
2074
2075static void pool_resume(struct dm_target *ti)
2076{
2077        struct pool_c *pt = ti->private;
2078        struct pool *pool = pt->pool;
2079        unsigned long flags;
2080
2081        spin_lock_irqsave(&pool->lock, flags);
2082        pool->low_water_triggered = 0;
2083        pool->no_free_space = 0;
2084        __requeue_bios(pool);
2085        spin_unlock_irqrestore(&pool->lock, flags);
2086
2087        do_waker(&pool->waker.work);
2088}
2089
2090static void pool_postsuspend(struct dm_target *ti)
2091{
2092        struct pool_c *pt = ti->private;
2093        struct pool *pool = pt->pool;
2094
2095        cancel_delayed_work(&pool->waker);
2096        flush_workqueue(pool->wq);
2097        (void) commit_or_fallback(pool);
2098}
2099
2100static int check_arg_count(unsigned argc, unsigned args_required)
2101{
2102        if (argc != args_required) {
2103                DMWARN("Message received with %u arguments instead of %u.",
2104                       argc, args_required);
2105                return -EINVAL;
2106        }
2107
2108        return 0;
2109}
2110
2111static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2112{
2113        if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2114            *dev_id <= MAX_DEV_ID)
2115                return 0;
2116
2117        if (warning)
2118                DMWARN("Message received with invalid device id: %s", arg);
2119
2120        return -EINVAL;
2121}
2122
2123static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2124{
2125        dm_thin_id dev_id;
2126        int r;
2127
2128        r = check_arg_count(argc, 2);
2129        if (r)
2130                return r;
2131
2132        r = read_dev_id(argv[1], &dev_id, 1);
2133        if (r)
2134                return r;
2135
2136        r = dm_pool_create_thin(pool->pmd, dev_id);
2137        if (r) {
2138                DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2139                       argv[1]);
2140                return r;
2141        }
2142
2143        return 0;
2144}
2145
2146static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2147{
2148        dm_thin_id dev_id;
2149        dm_thin_id origin_dev_id;
2150        int r;
2151
2152        r = check_arg_count(argc, 3);
2153        if (r)
2154                return r;
2155
2156        r = read_dev_id(argv[1], &dev_id, 1);
2157        if (r)
2158                return r;
2159
2160        r = read_dev_id(argv[2], &origin_dev_id, 1);
2161        if (r)
2162                return r;
2163
2164        r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2165        if (r) {
2166                DMWARN("Creation of new snapshot %s of device %s failed.",
2167                       argv[1], argv[2]);
2168                return r;
2169        }
2170
2171        return 0;
2172}
2173
2174static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2175{
2176        dm_thin_id dev_id;
2177        int r;
2178
2179        r = check_arg_count(argc, 2);
2180        if (r)
2181                return r;
2182
2183        r = read_dev_id(argv[1], &dev_id, 1);
2184        if (r)
2185                return r;
2186
2187        r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2188        if (r)
2189                DMWARN("Deletion of thin device %s failed.", argv[1]);
2190
2191        return r;
2192}
2193
2194static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2195{
2196        dm_thin_id old_id, new_id;
2197        int r;
2198
2199        r = check_arg_count(argc, 3);
2200        if (r)
2201                return r;
2202
2203        if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2204                DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2205                return -EINVAL;
2206        }
2207
2208        if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2209                DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2210                return -EINVAL;
2211        }
2212
2213        r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2214        if (r) {
2215                DMWARN("Failed to change transaction id from %s to %s.",
2216                       argv[1], argv[2]);
2217                return r;
2218        }
2219
2220        return 0;
2221}
2222
2223static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2224{
2225        int r;
2226
2227        r = check_arg_count(argc, 1);
2228        if (r)
2229                return r;
2230
2231        (void) commit_or_fallback(pool);
2232
2233        r = dm_pool_reserve_metadata_snap(pool->pmd);
2234        if (r)
2235                DMWARN("reserve_metadata_snap message failed.");
2236
2237        return r;
2238}
2239
2240static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2241{
2242        int r;
2243
2244        r = check_arg_count(argc, 1);
2245        if (r)
2246                return r;
2247
2248        r = dm_pool_release_metadata_snap(pool->pmd);
2249        if (r)
2250                DMWARN("release_metadata_snap message failed.");
2251
2252        return r;
2253}
2254
2255/*
2256 * Messages supported:
2257 *   create_thin        <dev_id>
2258 *   create_snap        <dev_id> <origin_id>
2259 *   delete             <dev_id>
2260 *   trim               <dev_id> <new_size_in_sectors>
2261 *   set_transaction_id <current_trans_id> <new_trans_id>
2262 *   reserve_metadata_snap
2263 *   release_metadata_snap
2264 */
2265static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2266{
2267        int r = -EINVAL;
2268        struct pool_c *pt = ti->private;
2269        struct pool *pool = pt->pool;
2270
2271        if (!strcasecmp(argv[0], "create_thin"))
2272                r = process_create_thin_mesg(argc, argv, pool);
2273
2274        else if (!strcasecmp(argv[0], "create_snap"))
2275                r = process_create_snap_mesg(argc, argv, pool);
2276
2277        else if (!strcasecmp(argv[0], "delete"))
2278                r = process_delete_mesg(argc, argv, pool);
2279
2280        else if (!strcasecmp(argv[0], "set_transaction_id"))
2281                r = process_set_transaction_id_mesg(argc, argv, pool);
2282
2283        else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2284                r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2285
2286        else if (!strcasecmp(argv[0], "release_metadata_snap"))
2287                r = process_release_metadata_snap_mesg(argc, argv, pool);
2288
2289        else
2290                DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2291
2292        if (!r)
2293                (void) commit_or_fallback(pool);
2294
2295        return r;
2296}
2297
2298static void emit_flags(struct pool_features *pf, char *result,
2299                       unsigned sz, unsigned maxlen)
2300{
2301        unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2302                !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2303        DMEMIT("%u ", count);
2304
2305        if (!pf->zero_new_blocks)
2306                DMEMIT("skip_block_zeroing ");
2307
2308        if (!pf->discard_enabled)
2309                DMEMIT("ignore_discard ");
2310
2311        if (!pf->discard_passdown)
2312                DMEMIT("no_discard_passdown ");
2313
2314        if (pf->mode == PM_READ_ONLY)
2315                DMEMIT("read_only ");
2316}
2317
2318/*
2319 * Status line is:
2320 *    <transaction id> <used metadata sectors>/<total metadata sectors>
2321 *    <used data sectors>/<total data sectors> <held metadata root>
2322 */
2323static int pool_status(struct dm_target *ti, status_type_t type,
2324                       unsigned status_flags, char *result, unsigned maxlen)
2325{
2326        int r;
2327        unsigned sz = 0;
2328        uint64_t transaction_id;
2329        dm_block_t nr_free_blocks_data;
2330        dm_block_t nr_free_blocks_metadata;
2331        dm_block_t nr_blocks_data;
2332        dm_block_t nr_blocks_metadata;
2333        dm_block_t held_root;
2334        char buf[BDEVNAME_SIZE];
2335        char buf2[BDEVNAME_SIZE];
2336        struct pool_c *pt = ti->private;
2337        struct pool *pool = pt->pool;
2338
2339        switch (type) {
2340        case STATUSTYPE_INFO:
2341                if (get_pool_mode(pool) == PM_FAIL) {
2342                        DMEMIT("Fail");
2343                        break;
2344                }
2345
2346                /* Commit to ensure statistics aren't out-of-date */
2347                if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2348                        (void) commit_or_fallback(pool);
2349
2350                r = dm_pool_get_metadata_transaction_id(pool->pmd,
2351                                                        &transaction_id);
2352                if (r)
2353                        return r;
2354
2355                r = dm_pool_get_free_metadata_block_count(pool->pmd,
2356                                                          &nr_free_blocks_metadata);
2357                if (r)
2358                        return r;
2359
2360                r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2361                if (r)
2362                        return r;
2363
2364                r = dm_pool_get_free_block_count(pool->pmd,
2365                                                 &nr_free_blocks_data);
2366                if (r)
2367                        return r;
2368
2369                r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2370                if (r)
2371                        return r;
2372
2373                r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2374                if (r)
2375                        return r;
2376
2377                DMEMIT("%llu %llu/%llu %llu/%llu ",
2378                       (unsigned long long)transaction_id,
2379                       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2380                       (unsigned long long)nr_blocks_metadata,
2381                       (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2382                       (unsigned long long)nr_blocks_data);
2383
2384                if (held_root)
2385                        DMEMIT("%llu ", held_root);
2386                else
2387                        DMEMIT("- ");
2388
2389                if (pool->pf.mode == PM_READ_ONLY)
2390                        DMEMIT("ro ");
2391                else
2392                        DMEMIT("rw ");
2393
2394                if (pool->pf.discard_enabled && pool->pf.discard_passdown)
2395                        DMEMIT("discard_passdown");
2396                else
2397                        DMEMIT("no_discard_passdown");
2398
2399                break;
2400
2401        case STATUSTYPE_TABLE:
2402                DMEMIT("%s %s %lu %llu ",
2403                       format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2404                       format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2405                       (unsigned long)pool->sectors_per_block,
2406                       (unsigned long long)pt->low_water_blocks);
2407                emit_flags(&pt->requested_pf, result, sz, maxlen);
2408                break;
2409        }
2410
2411        return 0;
2412}
2413
2414static int pool_iterate_devices(struct dm_target *ti,
2415                                iterate_devices_callout_fn fn, void *data)
2416{
2417        struct pool_c *pt = ti->private;
2418
2419        return fn(ti, pt->data_dev, 0, ti->len, data);
2420}
2421
2422static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2423                      struct bio_vec *biovec, int max_size)
2424{
2425        struct pool_c *pt = ti->private;
2426        struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2427
2428        if (!q->merge_bvec_fn)
2429                return max_size;
2430
2431        bvm->bi_bdev = pt->data_dev->bdev;
2432
2433        return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2434}
2435
2436static bool block_size_is_power_of_two(struct pool *pool)
2437{
2438        return pool->sectors_per_block_shift >= 0;
2439}
2440
2441static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2442{
2443        struct pool *pool = pt->pool;
2444        struct queue_limits *data_limits;
2445
2446        limits->max_discard_sectors = pool->sectors_per_block;
2447
2448        /*
2449         * discard_granularity is just a hint, and not enforced.
2450         */
2451        if (pt->adjusted_pf.discard_passdown) {
2452                data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2453                limits->discard_granularity = data_limits->discard_granularity;
2454        } else if (block_size_is_power_of_two(pool))
2455                limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2456        else
2457                /*
2458                 * Use largest power of 2 that is a factor of sectors_per_block
2459                 * but at least DATA_DEV_BLOCK_SIZE_MIN_SECTORS.
2460                 */
2461                limits->discard_granularity = max(1 << (ffs(pool->sectors_per_block) - 1),
2462                                                  DATA_DEV_BLOCK_SIZE_MIN_SECTORS) << SECTOR_SHIFT;
2463}
2464
2465static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2466{
2467        struct pool_c *pt = ti->private;
2468        struct pool *pool = pt->pool;
2469
2470        blk_limits_io_min(limits, 0);
2471        blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2472
2473        /*
2474         * pt->adjusted_pf is a staging area for the actual features to use.
2475         * They get transferred to the live pool in bind_control_target()
2476         * called from pool_preresume().
2477         */
2478        if (!pt->adjusted_pf.discard_enabled)
2479                return;
2480
2481        disable_passdown_if_not_supported(pt);
2482
2483        set_discard_limits(pt, limits);
2484}
2485
2486static struct target_type pool_target = {
2487        .name = "thin-pool",
2488        .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2489                    DM_TARGET_IMMUTABLE,
2490        .version = {1, 5, 0},
2491        .module = THIS_MODULE,
2492        .ctr = pool_ctr,
2493        .dtr = pool_dtr,
2494        .map = pool_map,
2495        .postsuspend = pool_postsuspend,
2496        .preresume = pool_preresume,
2497        .resume = pool_resume,
2498        .message = pool_message,
2499        .status = pool_status,
2500        .merge = pool_merge,
2501        .iterate_devices = pool_iterate_devices,
2502        .io_hints = pool_io_hints,
2503};
2504
2505/*----------------------------------------------------------------
2506 * Thin target methods
2507 *--------------------------------------------------------------*/
2508static void thin_dtr(struct dm_target *ti)
2509{
2510        struct thin_c *tc = ti->private;
2511
2512        mutex_lock(&dm_thin_pool_table.mutex);
2513
2514        __pool_dec(tc->pool);
2515        dm_pool_close_thin_device(tc->td);
2516        dm_put_device(ti, tc->pool_dev);
2517        if (tc->origin_dev)
2518                dm_put_device(ti, tc->origin_dev);
2519        kfree(tc);
2520
2521        mutex_unlock(&dm_thin_pool_table.mutex);
2522}
2523
2524/*
2525 * Thin target parameters:
2526 *
2527 * <pool_dev> <dev_id> [origin_dev]
2528 *
2529 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2530 * dev_id: the internal device identifier
2531 * origin_dev: a device external to the pool that should act as the origin
2532 *
2533 * If the pool device has discards disabled, they get disabled for the thin
2534 * device as well.
2535 */
2536static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2537{
2538        int r;
2539        struct thin_c *tc;
2540        struct dm_dev *pool_dev, *origin_dev;
2541        struct mapped_device *pool_md;
2542
2543        mutex_lock(&dm_thin_pool_table.mutex);
2544
2545        if (argc != 2 && argc != 3) {
2546                ti->error = "Invalid argument count";
2547                r = -EINVAL;
2548                goto out_unlock;
2549        }
2550
2551        tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2552        if (!tc) {
2553                ti->error = "Out of memory";
2554                r = -ENOMEM;
2555                goto out_unlock;
2556        }
2557
2558        if (argc == 3) {
2559                r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2560                if (r) {
2561                        ti->error = "Error opening origin device";
2562                        goto bad_origin_dev;
2563                }
2564                tc->origin_dev = origin_dev;
2565        }
2566
2567        r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2568        if (r) {
2569                ti->error = "Error opening pool device";
2570                goto bad_pool_dev;
2571        }
2572        tc->pool_dev = pool_dev;
2573
2574        if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2575                ti->error = "Invalid device id";
2576                r = -EINVAL;
2577                goto bad_common;
2578        }
2579
2580        pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2581        if (!pool_md) {
2582                ti->error = "Couldn't get pool mapped device";
2583                r = -EINVAL;
2584                goto bad_common;
2585        }
2586
2587        tc->pool = __pool_table_lookup(pool_md);
2588        if (!tc->pool) {
2589                ti->error = "Couldn't find pool object";
2590                r = -EINVAL;
2591                goto bad_pool_lookup;
2592        }
2593        __pool_inc(tc->pool);
2594
2595        if (get_pool_mode(tc->pool) == PM_FAIL) {
2596                ti->error = "Couldn't open thin device, Pool is in fail mode";
2597                goto bad_thin_open;
2598        }
2599
2600        r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2601        if (r) {
2602                ti->error = "Couldn't open thin internal device";
2603                goto bad_thin_open;
2604        }
2605
2606        r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2607        if (r)
2608                goto bad_thin_open;
2609
2610        ti->num_flush_requests = 1;
2611        ti->flush_supported = true;
2612
2613        /* In case the pool supports discards, pass them on. */
2614        if (tc->pool->pf.discard_enabled) {
2615                ti->discards_supported = true;
2616                ti->num_discard_requests = 1;
2617                ti->discard_zeroes_data_unsupported = true;
2618                /* Discard requests must be split on a block boundary */
2619                ti->split_discard_requests = true;
2620        }
2621
2622        dm_put(pool_md);
2623
2624        mutex_unlock(&dm_thin_pool_table.mutex);
2625
2626        return 0;
2627
2628bad_thin_open:
2629        __pool_dec(tc->pool);
2630bad_pool_lookup:
2631        dm_put(pool_md);
2632bad_common:
2633        dm_put_device(ti, tc->pool_dev);
2634bad_pool_dev:
2635        if (tc->origin_dev)
2636                dm_put_device(ti, tc->origin_dev);
2637bad_origin_dev:
2638        kfree(tc);
2639out_unlock:
2640        mutex_unlock(&dm_thin_pool_table.mutex);
2641
2642        return r;
2643}
2644
2645static int thin_map(struct dm_target *ti, struct bio *bio,
2646                    union map_info *map_context)
2647{
2648        bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2649
2650        return thin_bio_map(ti, bio, map_context);
2651}
2652
2653static int thin_endio(struct dm_target *ti,
2654                      struct bio *bio, int err,
2655                      union map_info *map_context)
2656{
2657        unsigned long flags;
2658        struct dm_thin_endio_hook *h = map_context->ptr;
2659        struct list_head work;
2660        struct dm_thin_new_mapping *m, *tmp;
2661        struct pool *pool = h->tc->pool;
2662
2663        if (h->shared_read_entry) {
2664                INIT_LIST_HEAD(&work);
2665                dm_deferred_entry_dec(h->shared_read_entry, &work);
2666
2667                spin_lock_irqsave(&pool->lock, flags);
2668                list_for_each_entry_safe(m, tmp, &work, list) {
2669                        list_del(&m->list);
2670                        m->quiesced = 1;
2671                        __maybe_add_mapping(m);
2672                }
2673                spin_unlock_irqrestore(&pool->lock, flags);
2674        }
2675
2676        if (h->all_io_entry) {
2677                INIT_LIST_HEAD(&work);
2678                dm_deferred_entry_dec(h->all_io_entry, &work);
2679                spin_lock_irqsave(&pool->lock, flags);
2680                list_for_each_entry_safe(m, tmp, &work, list)
2681                        list_add(&m->list, &pool->prepared_discards);
2682                spin_unlock_irqrestore(&pool->lock, flags);
2683        }
2684
2685        mempool_free(h, pool->endio_hook_pool);
2686
2687        return 0;
2688}
2689
2690static void thin_postsuspend(struct dm_target *ti)
2691{
2692        if (dm_noflush_suspending(ti))
2693                requeue_io((struct thin_c *)ti->private);
2694}
2695
2696/*
2697 * <nr mapped sectors> <highest mapped sector>
2698 */
2699static int thin_status(struct dm_target *ti, status_type_t type,
2700                       unsigned status_flags, char *result, unsigned maxlen)
2701{
2702        int r;
2703        ssize_t sz = 0;
2704        dm_block_t mapped, highest;
2705        char buf[BDEVNAME_SIZE];
2706        struct thin_c *tc = ti->private;
2707
2708        if (get_pool_mode(tc->pool) == PM_FAIL) {
2709                DMEMIT("Fail");
2710                return 0;
2711        }
2712
2713        if (!tc->td)
2714                DMEMIT("-");
2715        else {
2716                switch (type) {
2717                case STATUSTYPE_INFO:
2718                        r = dm_thin_get_mapped_count(tc->td, &mapped);
2719                        if (r)
2720                                return r;
2721
2722                        r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2723                        if (r < 0)
2724                                return r;
2725
2726                        DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2727                        if (r)
2728                                DMEMIT("%llu", ((highest + 1) *
2729                                                tc->pool->sectors_per_block) - 1);
2730                        else
2731                                DMEMIT("-");
2732                        break;
2733
2734                case STATUSTYPE_TABLE:
2735                        DMEMIT("%s %lu",
2736                               format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2737                               (unsigned long) tc->dev_id);
2738                        if (tc->origin_dev)
2739                                DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2740                        break;
2741                }
2742        }
2743
2744        return 0;
2745}
2746
2747static int thin_iterate_devices(struct dm_target *ti,
2748                                iterate_devices_callout_fn fn, void *data)
2749{
2750        sector_t blocks;
2751        struct thin_c *tc = ti->private;
2752        struct pool *pool = tc->pool;
2753
2754        /*
2755         * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2756         * we follow a more convoluted path through to the pool's target.
2757         */
2758        if (!pool->ti)
2759                return 0;       /* nothing is bound */
2760
2761        blocks = pool->ti->len;
2762        (void) sector_div(blocks, pool->sectors_per_block);
2763        if (blocks)
2764                return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2765
2766        return 0;
2767}
2768
2769/*
2770 * A thin device always inherits its queue limits from its pool.
2771 */
2772static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2773{
2774        struct thin_c *tc = ti->private;
2775
2776        *limits = bdev_get_queue(tc->pool_dev->bdev)->limits;
2777}
2778
2779static struct target_type thin_target = {
2780        .name = "thin",
2781        .version = {1, 5, 0},
2782        .module = THIS_MODULE,
2783        .ctr = thin_ctr,
2784        .dtr = thin_dtr,
2785        .map = thin_map,
2786        .end_io = thin_endio,
2787        .postsuspend = thin_postsuspend,
2788        .status = thin_status,
2789        .iterate_devices = thin_iterate_devices,
2790        .io_hints = thin_io_hints,
2791};
2792
2793/*----------------------------------------------------------------*/
2794
2795static int __init dm_thin_init(void)
2796{
2797        int r;
2798
2799        pool_table_init();
2800
2801        r = dm_register_target(&thin_target);
2802        if (r)
2803                return r;
2804
2805        r = dm_register_target(&pool_target);
2806        if (r)
2807                goto bad_pool_target;
2808
2809        r = -ENOMEM;
2810
2811        _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2812        if (!_new_mapping_cache)
2813                goto bad_new_mapping_cache;
2814
2815        _endio_hook_cache = KMEM_CACHE(dm_thin_endio_hook, 0);
2816        if (!_endio_hook_cache)
2817                goto bad_endio_hook_cache;
2818
2819        return 0;
2820
2821bad_endio_hook_cache:
2822        kmem_cache_destroy(_new_mapping_cache);
2823bad_new_mapping_cache:
2824        dm_unregister_target(&pool_target);
2825bad_pool_target:
2826        dm_unregister_target(&thin_target);
2827
2828        return r;
2829}
2830
2831static void dm_thin_exit(void)
2832{
2833        dm_unregister_target(&thin_target);
2834        dm_unregister_target(&pool_target);
2835
2836        kmem_cache_destroy(_new_mapping_cache);
2837        kmem_cache_destroy(_endio_hook_cache);
2838}
2839
2840module_init(dm_thin_init);
2841module_exit(dm_thin_exit);
2842
2843MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2844MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2845MODULE_LICENSE("GPL");
2846
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.