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/jiffies.h>
  15#include <linux/log2.h>
  16#include <linux/list.h>
  17#include <linux/rculist.h>
  18#include <linux/init.h>
  19#include <linux/module.h>
  20#include <linux/slab.h>
  21#include <linux/vmalloc.h>
  22#include <linux/sort.h>
  23#include <linux/rbtree.h>
  24
  25#define	DM_MSG_PREFIX	"thin"
  26
  27/*
  28 * Tunable constants
  29 */
  30#define ENDIO_HOOK_POOL_SIZE 1024
  31#define MAPPING_POOL_SIZE 1024
  32#define COMMIT_PERIOD HZ
  33#define NO_SPACE_TIMEOUT_SECS 60
  34
  35static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
  36
  37DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
  38		"A percentage of time allocated for copy on write");
  39
  40/*
  41 * The block size of the device holding pool data must be
  42 * between 64KB and 1GB.
  43 */
  44#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
  45#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
  46
  47/*
  48 * Device id is restricted to 24 bits.
  49 */
  50#define MAX_DEV_ID ((1 << 24) - 1)
  51
  52/*
  53 * How do we handle breaking sharing of data blocks?
  54 * =================================================
  55 *
  56 * We use a standard copy-on-write btree to store the mappings for the
  57 * devices (note I'm talking about copy-on-write of the metadata here, not
  58 * the data).  When you take an internal snapshot you clone the root node
  59 * of the origin btree.  After this there is no concept of an origin or a
  60 * snapshot.  They are just two device trees that happen to point to the
  61 * same data blocks.
  62 *
  63 * When we get a write in we decide if it's to a shared data block using
  64 * some timestamp magic.  If it is, we have to break sharing.
  65 *
  66 * Let's say we write to a shared block in what was the origin.  The
  67 * steps are:
  68 *
  69 * i) plug io further to this physical block. (see bio_prison code).
  70 *
  71 * ii) quiesce any read io to that shared data block.  Obviously
  72 * including all devices that share this block.  (see dm_deferred_set code)
  73 *
  74 * iii) copy the data block to a newly allocate block.  This step can be
  75 * missed out if the io covers the block. (schedule_copy).
  76 *
  77 * iv) insert the new mapping into the origin's btree
  78 * (process_prepared_mapping).  This act of inserting breaks some
  79 * sharing of btree nodes between the two devices.  Breaking sharing only
  80 * effects the btree of that specific device.  Btrees for the other
  81 * devices that share the block never change.  The btree for the origin
  82 * device as it was after the last commit is untouched, ie. we're using
  83 * persistent data structures in the functional programming sense.
  84 *
  85 * v) unplug io to this physical block, including the io that triggered
  86 * the breaking of sharing.
  87 *
  88 * Steps (ii) and (iii) occur in parallel.
  89 *
  90 * The metadata _doesn't_ need to be committed before the io continues.  We
  91 * get away with this because the io is always written to a _new_ block.
  92 * If there's a crash, then:
  93 *
  94 * - The origin mapping will point to the old origin block (the shared
  95 * one).  This will contain the data as it was before the io that triggered
  96 * the breaking of sharing came in.
  97 *
  98 * - The snap mapping still points to the old block.  As it would after
  99 * the commit.
 100 *
 101 * The downside of this scheme is the timestamp magic isn't perfect, and
 102 * will continue to think that data block in the snapshot device is shared
 103 * even after the write to the origin has broken sharing.  I suspect data
 104 * blocks will typically be shared by many different devices, so we're
 105 * breaking sharing n + 1 times, rather than n, where n is the number of
 106 * devices that reference this data block.  At the moment I think the
 107 * benefits far, far outweigh the disadvantages.
 108 */
 109
 110/*----------------------------------------------------------------*/
 111
 112/*
 113 * Key building.
 114 */
 115enum lock_space {
 116	VIRTUAL,
 117	PHYSICAL
 118};
 119
 120static void build_key(struct dm_thin_device *td, enum lock_space ls,
 121		      dm_block_t b, dm_block_t e, struct dm_cell_key *key)
 122{
 123	key->virtual = (ls == VIRTUAL);
 124	key->dev = dm_thin_dev_id(td);
 125	key->block_begin = b;
 126	key->block_end = e;
 127}
 128
 129static void build_data_key(struct dm_thin_device *td, dm_block_t b,
 130			   struct dm_cell_key *key)
 131{
 132	build_key(td, PHYSICAL, b, b + 1llu, key);
 133}
 134
 135static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
 136			      struct dm_cell_key *key)
 137{
 138	build_key(td, VIRTUAL, b, b + 1llu, key);
 139}
 140
 141/*----------------------------------------------------------------*/
 142
 143#define THROTTLE_THRESHOLD (1 * HZ)
 144
 145struct throttle {
 146	struct rw_semaphore lock;
 147	unsigned long threshold;
 148	bool throttle_applied;
 149};
 150
 151static void throttle_init(struct throttle *t)
 152{
 153	init_rwsem(&t->lock);
 154	t->throttle_applied = false;
 155}
 156
 157static void throttle_work_start(struct throttle *t)
 158{
 159	t->threshold = jiffies + THROTTLE_THRESHOLD;
 160}
 161
 162static void throttle_work_update(struct throttle *t)
 163{
 164	if (!t->throttle_applied && jiffies > t->threshold) {
 165		down_write(&t->lock);
 166		t->throttle_applied = true;
 167	}
 168}
 169
 170static void throttle_work_complete(struct throttle *t)
 171{
 172	if (t->throttle_applied) {
 173		t->throttle_applied = false;
 174		up_write(&t->lock);
 175	}
 176}
 177
 178static void throttle_lock(struct throttle *t)
 179{
 180	down_read(&t->lock);
 181}
 182
 183static void throttle_unlock(struct throttle *t)
 184{
 185	up_read(&t->lock);
 186}
 187
 188/*----------------------------------------------------------------*/
 189
 190/*
 191 * A pool device ties together a metadata device and a data device.  It
 192 * also provides the interface for creating and destroying internal
 193 * devices.
 194 */
 195struct dm_thin_new_mapping;
 196
 197/*
 198 * The pool runs in 4 modes.  Ordered in degraded order for comparisons.
 199 */
 200enum pool_mode {
 201	PM_WRITE,		/* metadata may be changed */
 202	PM_OUT_OF_DATA_SPACE,	/* metadata may be changed, though data may not be allocated */
 203	PM_READ_ONLY,		/* metadata may not be changed */
 204	PM_FAIL,		/* all I/O fails */
 205};
 206
 207struct pool_features {
 208	enum pool_mode mode;
 209
 210	bool zero_new_blocks:1;
 211	bool discard_enabled:1;
 212	bool discard_passdown:1;
 213	bool error_if_no_space:1;
 214};
 215
 216struct thin_c;
 217typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
 218typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
 219typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
 220
 221#define CELL_SORT_ARRAY_SIZE 8192
 222
 223struct pool {
 224	struct list_head list;
 225	struct dm_target *ti;	/* Only set if a pool target is bound */
 226
 227	struct mapped_device *pool_md;
 228	struct block_device *md_dev;
 229	struct dm_pool_metadata *pmd;
 230
 231	dm_block_t low_water_blocks;
 232	uint32_t sectors_per_block;
 233	int sectors_per_block_shift;
 234
 235	struct pool_features pf;
 236	bool low_water_triggered:1;	/* A dm event has been sent */
 237	bool suspended:1;
 238	bool out_of_data_space:1;
 239
 240	struct dm_bio_prison *prison;
 241	struct dm_kcopyd_client *copier;
 242
 243	struct workqueue_struct *wq;
 244	struct throttle throttle;
 245	struct work_struct worker;
 246	struct delayed_work waker;
 247	struct delayed_work no_space_timeout;
 248
 249	unsigned long last_commit_jiffies;
 250	unsigned ref_count;
 251
 252	spinlock_t lock;
 253	struct bio_list deferred_flush_bios;
 254	struct list_head prepared_mappings;
 255	struct list_head prepared_discards;
 256	struct list_head prepared_discards_pt2;
 257	struct list_head active_thins;
 258
 259	struct dm_deferred_set *shared_read_ds;
 260	struct dm_deferred_set *all_io_ds;
 261
 262	struct dm_thin_new_mapping *next_mapping;
 263	mempool_t *mapping_pool;
 264
 265	process_bio_fn process_bio;
 266	process_bio_fn process_discard;
 267
 268	process_cell_fn process_cell;
 269	process_cell_fn process_discard_cell;
 270
 271	process_mapping_fn process_prepared_mapping;
 272	process_mapping_fn process_prepared_discard;
 273	process_mapping_fn process_prepared_discard_pt2;
 274
 275	struct dm_bio_prison_cell **cell_sort_array;
 276};
 277
 278static enum pool_mode get_pool_mode(struct pool *pool);
 279static void metadata_operation_failed(struct pool *pool, const char *op, int r);
 280
 281/*
 282 * Target context for a pool.
 283 */
 284struct pool_c {
 285	struct dm_target *ti;
 286	struct pool *pool;
 287	struct dm_dev *data_dev;
 288	struct dm_dev *metadata_dev;
 289	struct dm_target_callbacks callbacks;
 290
 291	dm_block_t low_water_blocks;
 292	struct pool_features requested_pf; /* Features requested during table load */
 293	struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
 294};
 295
 296/*
 297 * Target context for a thin.
 298 */
 299struct thin_c {
 300	struct list_head list;
 301	struct dm_dev *pool_dev;
 302	struct dm_dev *origin_dev;
 303	sector_t origin_size;
 304	dm_thin_id dev_id;
 305
 306	struct pool *pool;
 307	struct dm_thin_device *td;
 308	struct mapped_device *thin_md;
 309
 310	bool requeue_mode:1;
 311	spinlock_t lock;
 312	struct list_head deferred_cells;
 313	struct bio_list deferred_bio_list;
 314	struct bio_list retry_on_resume_list;
 315	struct rb_root sort_bio_list; /* sorted list of deferred bios */
 316
 317	/*
 318	 * Ensures the thin is not destroyed until the worker has finished
 319	 * iterating the active_thins list.
 320	 */
 321	atomic_t refcount;
 322	struct completion can_destroy;
 323};
 324
 325/*----------------------------------------------------------------*/
 326
 327static bool block_size_is_power_of_two(struct pool *pool)
 328{
 329	return pool->sectors_per_block_shift >= 0;
 330}
 331
 332static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
 333{
 334	return block_size_is_power_of_two(pool) ?
 335		(b << pool->sectors_per_block_shift) :
 336		(b * pool->sectors_per_block);
 337}
 338
 339/*----------------------------------------------------------------*/
 340
 341struct discard_op {
 342	struct thin_c *tc;
 343	struct blk_plug plug;
 344	struct bio *parent_bio;
 345	struct bio *bio;
 346};
 347
 348static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
 349{
 350	BUG_ON(!parent);
 351
 352	op->tc = tc;
 353	blk_start_plug(&op->plug);
 354	op->parent_bio = parent;
 355	op->bio = NULL;
 356}
 357
 358static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
 359{
 360	struct thin_c *tc = op->tc;
 361	sector_t s = block_to_sectors(tc->pool, data_b);
 362	sector_t len = block_to_sectors(tc->pool, data_e - data_b);
 363
 364	return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
 365				      GFP_NOWAIT, 0, &op->bio);
 366}
 367
 368static void end_discard(struct discard_op *op, int r)
 369{
 370	if (op->bio) {
 371		/*
 372		 * Even if one of the calls to issue_discard failed, we
 373		 * need to wait for the chain to complete.
 374		 */
 375		bio_chain(op->bio, op->parent_bio);
 376		bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
 377		submit_bio(op->bio);
 378	}
 379
 380	blk_finish_plug(&op->plug);
 381
 382	/*
 383	 * Even if r is set, there could be sub discards in flight that we
 384	 * need to wait for.
 385	 */
 386	if (r && !op->parent_bio->bi_error)
 387		op->parent_bio->bi_error = r;
 388	bio_endio(op->parent_bio);
 389}
 390
 391/*----------------------------------------------------------------*/
 392
 393/*
 394 * wake_worker() is used when new work is queued and when pool_resume is
 395 * ready to continue deferred IO processing.
 396 */
 397static void wake_worker(struct pool *pool)
 398{
 399	queue_work(pool->wq, &pool->worker);
 400}
 401
 402/*----------------------------------------------------------------*/
 403
 404static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
 405		      struct dm_bio_prison_cell **cell_result)
 406{
 407	int r;
 408	struct dm_bio_prison_cell *cell_prealloc;
 409
 410	/*
 411	 * Allocate a cell from the prison's mempool.
 412	 * This might block but it can't fail.
 413	 */
 414	cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
 415
 416	r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
 417	if (r)
 418		/*
 419		 * We reused an old cell; we can get rid of
 420		 * the new one.
 421		 */
 422		dm_bio_prison_free_cell(pool->prison, cell_prealloc);
 423
 424	return r;
 425}
 426
 427static void cell_release(struct pool *pool,
 428			 struct dm_bio_prison_cell *cell,
 429			 struct bio_list *bios)
 430{
 431	dm_cell_release(pool->prison, cell, bios);
 432	dm_bio_prison_free_cell(pool->prison, cell);
 433}
 434
 435static void cell_visit_release(struct pool *pool,
 436			       void (*fn)(void *, struct dm_bio_prison_cell *),
 437			       void *context,
 438			       struct dm_bio_prison_cell *cell)
 439{
 440	dm_cell_visit_release(pool->prison, fn, context, cell);
 441	dm_bio_prison_free_cell(pool->prison, cell);
 442}
 443
 444static void cell_release_no_holder(struct pool *pool,
 445				   struct dm_bio_prison_cell *cell,
 446				   struct bio_list *bios)
 447{
 448	dm_cell_release_no_holder(pool->prison, cell, bios);
 449	dm_bio_prison_free_cell(pool->prison, cell);
 450}
 451
 452static void cell_error_with_code(struct pool *pool,
 453				 struct dm_bio_prison_cell *cell, int error_code)
 454{
 455	dm_cell_error(pool->prison, cell, error_code);
 456	dm_bio_prison_free_cell(pool->prison, cell);
 457}
 458
 459static int get_pool_io_error_code(struct pool *pool)
 460{
 461	return pool->out_of_data_space ? -ENOSPC : -EIO;
 462}
 463
 464static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
 465{
 466	int error = get_pool_io_error_code(pool);
 467
 468	cell_error_with_code(pool, cell, error);
 469}
 470
 471static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
 472{
 473	cell_error_with_code(pool, cell, 0);
 474}
 475
 476static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
 477{
 478	cell_error_with_code(pool, cell, DM_ENDIO_REQUEUE);
 479}
 480
 481/*----------------------------------------------------------------*/
 482
 483/*
 484 * A global list of pools that uses a struct mapped_device as a key.
 485 */
 486static struct dm_thin_pool_table {
 487	struct mutex mutex;
 488	struct list_head pools;
 489} dm_thin_pool_table;
 490
 491static void pool_table_init(void)
 492{
 493	mutex_init(&dm_thin_pool_table.mutex);
 494	INIT_LIST_HEAD(&dm_thin_pool_table.pools);
 495}
 496
 497static void __pool_table_insert(struct pool *pool)
 498{
 499	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 500	list_add(&pool->list, &dm_thin_pool_table.pools);
 501}
 502
 503static void __pool_table_remove(struct pool *pool)
 504{
 505	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 506	list_del(&pool->list);
 507}
 508
 509static struct pool *__pool_table_lookup(struct mapped_device *md)
 510{
 511	struct pool *pool = NULL, *tmp;
 512
 513	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 514
 515	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 516		if (tmp->pool_md == md) {
 517			pool = tmp;
 518			break;
 519		}
 520	}
 521
 522	return pool;
 523}
 524
 525static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
 526{
 527	struct pool *pool = NULL, *tmp;
 528
 529	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
 530
 531	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
 532		if (tmp->md_dev == md_dev) {
 533			pool = tmp;
 534			break;
 535		}
 536	}
 537
 538	return pool;
 539}
 540
 541/*----------------------------------------------------------------*/
 542
 543struct dm_thin_endio_hook {
 544	struct thin_c *tc;
 545	struct dm_deferred_entry *shared_read_entry;
 546	struct dm_deferred_entry *all_io_entry;
 547	struct dm_thin_new_mapping *overwrite_mapping;
 548	struct rb_node rb_node;
 549	struct dm_bio_prison_cell *cell;
 550};
 551
 552static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
 553{
 554	bio_list_merge(bios, master);
 555	bio_list_init(master);
 556}
 557
 558static void error_bio_list(struct bio_list *bios, int error)
 559{
 560	struct bio *bio;
 561
 562	while ((bio = bio_list_pop(bios))) {
 563		bio->bi_error = error;
 564		bio_endio(bio);
 565	}
 566}
 567
 568static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master, int error)
 569{
 570	struct bio_list bios;
 571	unsigned long flags;
 572
 573	bio_list_init(&bios);
 574
 575	spin_lock_irqsave(&tc->lock, flags);
 576	__merge_bio_list(&bios, master);
 577	spin_unlock_irqrestore(&tc->lock, flags);
 578
 579	error_bio_list(&bios, error);
 580}
 581
 582static void requeue_deferred_cells(struct thin_c *tc)
 583{
 584	struct pool *pool = tc->pool;
 585	unsigned long flags;
 586	struct list_head cells;
 587	struct dm_bio_prison_cell *cell, *tmp;
 588
 589	INIT_LIST_HEAD(&cells);
 590
 591	spin_lock_irqsave(&tc->lock, flags);
 592	list_splice_init(&tc->deferred_cells, &cells);
 593	spin_unlock_irqrestore(&tc->lock, flags);
 594
 595	list_for_each_entry_safe(cell, tmp, &cells, user_list)
 596		cell_requeue(pool, cell);
 597}
 598
 599static void requeue_io(struct thin_c *tc)
 600{
 601	struct bio_list bios;
 602	unsigned long flags;
 603
 604	bio_list_init(&bios);
 605
 606	spin_lock_irqsave(&tc->lock, flags);
 607	__merge_bio_list(&bios, &tc->deferred_bio_list);
 608	__merge_bio_list(&bios, &tc->retry_on_resume_list);
 609	spin_unlock_irqrestore(&tc->lock, flags);
 610
 611	error_bio_list(&bios, DM_ENDIO_REQUEUE);
 612	requeue_deferred_cells(tc);
 613}
 614
 615static void error_retry_list_with_code(struct pool *pool, int error)
 616{
 617	struct thin_c *tc;
 618
 619	rcu_read_lock();
 620	list_for_each_entry_rcu(tc, &pool->active_thins, list)
 621		error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
 622	rcu_read_unlock();
 623}
 624
 625static void error_retry_list(struct pool *pool)
 626{
 627	int error = get_pool_io_error_code(pool);
 628
 629	error_retry_list_with_code(pool, error);
 630}
 631
 632/*
 633 * This section of code contains the logic for processing a thin device's IO.
 634 * Much of the code depends on pool object resources (lists, workqueues, etc)
 635 * but most is exclusively called from the thin target rather than the thin-pool
 636 * target.
 637 */
 638
 639static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
 640{
 641	struct pool *pool = tc->pool;
 642	sector_t block_nr = bio->bi_iter.bi_sector;
 643
 644	if (block_size_is_power_of_two(pool))
 645		block_nr >>= pool->sectors_per_block_shift;
 646	else
 647		(void) sector_div(block_nr, pool->sectors_per_block);
 648
 649	return block_nr;
 650}
 651
 652/*
 653 * Returns the _complete_ blocks that this bio covers.
 654 */
 655static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
 656				dm_block_t *begin, dm_block_t *end)
 657{
 658	struct pool *pool = tc->pool;
 659	sector_t b = bio->bi_iter.bi_sector;
 660	sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
 661
 662	b += pool->sectors_per_block - 1ull; /* so we round up */
 663
 664	if (block_size_is_power_of_two(pool)) {
 665		b >>= pool->sectors_per_block_shift;
 666		e >>= pool->sectors_per_block_shift;
 667	} else {
 668		(void) sector_div(b, pool->sectors_per_block);
 669		(void) sector_div(e, pool->sectors_per_block);
 670	}
 671
 672	if (e < b)
 673		/* Can happen if the bio is within a single block. */
 674		e = b;
 675
 676	*begin = b;
 677	*end = e;
 678}
 679
 680static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
 681{
 682	struct pool *pool = tc->pool;
 683	sector_t bi_sector = bio->bi_iter.bi_sector;
 684
 685	bio->bi_bdev = tc->pool_dev->bdev;
 686	if (block_size_is_power_of_two(pool))
 687		bio->bi_iter.bi_sector =
 688			(block << pool->sectors_per_block_shift) |
 689			(bi_sector & (pool->sectors_per_block - 1));
 690	else
 691		bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
 692				 sector_div(bi_sector, pool->sectors_per_block);
 693}
 694
 695static void remap_to_origin(struct thin_c *tc, struct bio *bio)
 696{
 697	bio->bi_bdev = tc->origin_dev->bdev;
 698}
 699
 700static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
 701{
 702	return (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA)) &&
 703		dm_thin_changed_this_transaction(tc->td);
 704}
 705
 706static void inc_all_io_entry(struct pool *pool, struct bio *bio)
 707{
 708	struct dm_thin_endio_hook *h;
 709
 710	if (bio_op(bio) == REQ_OP_DISCARD)
 711		return;
 712
 713	h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 714	h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
 715}
 716
 717static void issue(struct thin_c *tc, struct bio *bio)
 718{
 719	struct pool *pool = tc->pool;
 720	unsigned long flags;
 721
 722	if (!bio_triggers_commit(tc, bio)) {
 723		generic_make_request(bio);
 724		return;
 725	}
 726
 727	/*
 728	 * Complete bio with an error if earlier I/O caused changes to
 729	 * the metadata that can't be committed e.g, due to I/O errors
 730	 * on the metadata device.
 731	 */
 732	if (dm_thin_aborted_changes(tc->td)) {
 733		bio_io_error(bio);
 734		return;
 735	}
 736
 737	/*
 738	 * Batch together any bios that trigger commits and then issue a
 739	 * single commit for them in process_deferred_bios().
 740	 */
 741	spin_lock_irqsave(&pool->lock, flags);
 742	bio_list_add(&pool->deferred_flush_bios, bio);
 743	spin_unlock_irqrestore(&pool->lock, flags);
 744}
 745
 746static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
 747{
 748	remap_to_origin(tc, bio);
 749	issue(tc, bio);
 750}
 751
 752static void remap_and_issue(struct thin_c *tc, struct bio *bio,
 753			    dm_block_t block)
 754{
 755	remap(tc, bio, block);
 756	issue(tc, bio);
 757}
 758
 759/*----------------------------------------------------------------*/
 760
 761/*
 762 * Bio endio functions.
 763 */
 764struct dm_thin_new_mapping {
 765	struct list_head list;
 766
 767	bool pass_discard:1;
 768	bool maybe_shared:1;
 769
 770	/*
 771	 * Track quiescing, copying and zeroing preparation actions.  When this
 772	 * counter hits zero the block is prepared and can be inserted into the
 773	 * btree.
 774	 */
 775	atomic_t prepare_actions;
 776
 777	int err;
 778	struct thin_c *tc;
 779	dm_block_t virt_begin, virt_end;
 780	dm_block_t data_block;
 781	struct dm_bio_prison_cell *cell;
 782
 783	/*
 784	 * If the bio covers the whole area of a block then we can avoid
 785	 * zeroing or copying.  Instead this bio is hooked.  The bio will
 786	 * still be in the cell, so care has to be taken to avoid issuing
 787	 * the bio twice.
 788	 */
 789	struct bio *bio;
 790	bio_end_io_t *saved_bi_end_io;
 791};
 792
 793static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
 794{
 795	struct pool *pool = m->tc->pool;
 796
 797	if (atomic_dec_and_test(&m->prepare_actions)) {
 798		list_add_tail(&m->list, &pool->prepared_mappings);
 799		wake_worker(pool);
 800	}
 801}
 802
 803static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
 804{
 805	unsigned long flags;
 806	struct pool *pool = m->tc->pool;
 807
 808	spin_lock_irqsave(&pool->lock, flags);
 809	__complete_mapping_preparation(m);
 810	spin_unlock_irqrestore(&pool->lock, flags);
 811}
 812
 813static void copy_complete(int read_err, unsigned long write_err, void *context)
 814{
 815	struct dm_thin_new_mapping *m = context;
 816
 817	m->err = read_err || write_err ? -EIO : 0;
 818	complete_mapping_preparation(m);
 819}
 820
 821static void overwrite_endio(struct bio *bio)
 822{
 823	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
 824	struct dm_thin_new_mapping *m = h->overwrite_mapping;
 825
 826	bio->bi_end_io = m->saved_bi_end_io;
 827
 828	m->err = bio->bi_error;
 829	complete_mapping_preparation(m);
 830}
 831
 832/*----------------------------------------------------------------*/
 833
 834/*
 835 * Workqueue.
 836 */
 837
 838/*
 839 * Prepared mapping jobs.
 840 */
 841
 842/*
 843 * This sends the bios in the cell, except the original holder, back
 844 * to the deferred_bios list.
 845 */
 846static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
 847{
 848	struct pool *pool = tc->pool;
 849	unsigned long flags;
 850
 851	spin_lock_irqsave(&tc->lock, flags);
 852	cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
 853	spin_unlock_irqrestore(&tc->lock, flags);
 854
 855	wake_worker(pool);
 856}
 857
 858static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
 859
 860struct remap_info {
 861	struct thin_c *tc;
 862	struct bio_list defer_bios;
 863	struct bio_list issue_bios;
 864};
 865
 866static void __inc_remap_and_issue_cell(void *context,
 867				       struct dm_bio_prison_cell *cell)
 868{
 869	struct remap_info *info = context;
 870	struct bio *bio;
 871
 872	while ((bio = bio_list_pop(&cell->bios))) {
 873		if (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA) ||
 874		    bio_op(bio) == REQ_OP_DISCARD)
 875			bio_list_add(&info->defer_bios, bio);
 876		else {
 877			inc_all_io_entry(info->tc->pool, bio);
 878
 879			/*
 880			 * We can't issue the bios with the bio prison lock
 881			 * held, so we add them to a list to issue on
 882			 * return from this function.
 883			 */
 884			bio_list_add(&info->issue_bios, bio);
 885		}
 886	}
 887}
 888
 889static void inc_remap_and_issue_cell(struct thin_c *tc,
 890				     struct dm_bio_prison_cell *cell,
 891				     dm_block_t block)
 892{
 893	struct bio *bio;
 894	struct remap_info info;
 895
 896	info.tc = tc;
 897	bio_list_init(&info.defer_bios);
 898	bio_list_init(&info.issue_bios);
 899
 900	/*
 901	 * We have to be careful to inc any bios we're about to issue
 902	 * before the cell is released, and avoid a race with new bios
 903	 * being added to the cell.
 904	 */
 905	cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
 906			   &info, cell);
 907
 908	while ((bio = bio_list_pop(&info.defer_bios)))
 909		thin_defer_bio(tc, bio);
 910
 911	while ((bio = bio_list_pop(&info.issue_bios)))
 912		remap_and_issue(info.tc, bio, block);
 913}
 914
 915static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
 916{
 917	cell_error(m->tc->pool, m->cell);
 918	list_del(&m->list);
 919	mempool_free(m, m->tc->pool->mapping_pool);
 920}
 921
 922static void process_prepared_mapping(struct dm_thin_new_mapping *m)
 923{
 924	struct thin_c *tc = m->tc;
 925	struct pool *pool = tc->pool;
 926	struct bio *bio = m->bio;
 927	int r;
 928
 929	if (m->err) {
 930		cell_error(pool, m->cell);
 931		goto out;
 932	}
 933
 934	/*
 935	 * Commit the prepared block into the mapping btree.
 936	 * Any I/O for this block arriving after this point will get
 937	 * remapped to it directly.
 938	 */
 939	r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
 940	if (r) {
 941		metadata_operation_failed(pool, "dm_thin_insert_block", r);
 942		cell_error(pool, m->cell);
 943		goto out;
 944	}
 945
 946	/*
 947	 * Release any bios held while the block was being provisioned.
 948	 * If we are processing a write bio that completely covers the block,
 949	 * we already processed it so can ignore it now when processing
 950	 * the bios in the cell.
 951	 */
 952	if (bio) {
 953		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
 954		bio_endio(bio);
 955	} else {
 956		inc_all_io_entry(tc->pool, m->cell->holder);
 957		remap_and_issue(tc, m->cell->holder, m->data_block);
 958		inc_remap_and_issue_cell(tc, m->cell, m->data_block);
 959	}
 960
 961out:
 962	list_del(&m->list);
 963	mempool_free(m, pool->mapping_pool);
 964}
 965
 966/*----------------------------------------------------------------*/
 967
 968static void free_discard_mapping(struct dm_thin_new_mapping *m)
 969{
 970	struct thin_c *tc = m->tc;
 971	if (m->cell)
 972		cell_defer_no_holder(tc, m->cell);
 973	mempool_free(m, tc->pool->mapping_pool);
 974}
 975
 976static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
 977{
 978	bio_io_error(m->bio);
 979	free_discard_mapping(m);
 980}
 981
 982static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
 983{
 984	bio_endio(m->bio);
 985	free_discard_mapping(m);
 986}
 987
 988static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
 989{
 990	int r;
 991	struct thin_c *tc = m->tc;
 992
 993	r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
 994	if (r) {
 995		metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
 996		bio_io_error(m->bio);
 997	} else
 998		bio_endio(m->bio);
 999
1000	cell_defer_no_holder(tc, m->cell);
1001	mempool_free(m, tc->pool->mapping_pool);
1002}
1003
1004/*----------------------------------------------------------------*/
1005
1006static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1007						   struct bio *discard_parent)
1008{
1009	/*
1010	 * We've already unmapped this range of blocks, but before we
1011	 * passdown we have to check that these blocks are now unused.
1012	 */
1013	int r = 0;
1014	bool used = true;
1015	struct thin_c *tc = m->tc;
1016	struct pool *pool = tc->pool;
1017	dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1018	struct discard_op op;
1019
1020	begin_discard(&op, tc, discard_parent);
1021	while (b != end) {
1022		/* find start of unmapped run */
1023		for (; b < end; b++) {
1024			r = dm_pool_block_is_used(pool->pmd, b, &used);
1025			if (r)
1026				goto out;
1027
1028			if (!used)
1029				break;
1030		}
1031
1032		if (b == end)
1033			break;
1034
1035		/* find end of run */
1036		for (e = b + 1; e != end; e++) {
1037			r = dm_pool_block_is_used(pool->pmd, e, &used);
1038			if (r)
1039				goto out;
1040
1041			if (used)
1042				break;
1043		}
1044
1045		r = issue_discard(&op, b, e);
1046		if (r)
1047			goto out;
1048
1049		b = e;
1050	}
1051out:
1052	end_discard(&op, r);
1053}
1054
1055static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1056{
1057	unsigned long flags;
1058	struct pool *pool = m->tc->pool;
1059
1060	spin_lock_irqsave(&pool->lock, flags);
1061	list_add_tail(&m->list, &pool->prepared_discards_pt2);
1062	spin_unlock_irqrestore(&pool->lock, flags);
1063	wake_worker(pool);
1064}
1065
1066static void passdown_endio(struct bio *bio)
1067{
1068	/*
1069	 * It doesn't matter if the passdown discard failed, we still want
1070	 * to unmap (we ignore err).
1071	 */
1072	queue_passdown_pt2(bio->bi_private);
1073}
1074
1075static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1076{
1077	int r;
1078	struct thin_c *tc = m->tc;
1079	struct pool *pool = tc->pool;
1080	struct bio *discard_parent;
1081	dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1082
1083	/*
1084	 * Only this thread allocates blocks, so we can be sure that the
1085	 * newly unmapped blocks will not be allocated before the end of
1086	 * the function.
1087	 */
1088	r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1089	if (r) {
1090		metadata_operation_failed(pool, "dm_thin_remove_range", r);
1091		bio_io_error(m->bio);
1092		cell_defer_no_holder(tc, m->cell);
1093		mempool_free(m, pool->mapping_pool);
1094		return;
1095	}
1096
1097	discard_parent = bio_alloc(GFP_NOIO, 1);
1098	if (!discard_parent) {
1099		DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1100		       dm_device_name(tc->pool->pool_md));
1101		queue_passdown_pt2(m);
1102
1103	} else {
1104		discard_parent->bi_end_io = passdown_endio;
1105		discard_parent->bi_private = m;
1106
1107		if (m->maybe_shared)
1108			passdown_double_checking_shared_status(m, discard_parent);
1109		else {
1110			struct discard_op op;
1111
1112			begin_discard(&op, tc, discard_parent);
1113			r = issue_discard(&op, m->data_block, data_end);
1114			end_discard(&op, r);
1115		}
1116	}
1117
1118	/*
1119	 * Increment the unmapped blocks.  This prevents a race between the
1120	 * passdown io and reallocation of freed blocks.
1121	 */
1122	r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1123	if (r) {
1124		metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1125		bio_io_error(m->bio);
1126		cell_defer_no_holder(tc, m->cell);
1127		mempool_free(m, pool->mapping_pool);
1128		return;
1129	}
1130}
1131
1132static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1133{
1134	int r;
1135	struct thin_c *tc = m->tc;
1136	struct pool *pool = tc->pool;
1137
1138	/*
1139	 * The passdown has completed, so now we can decrement all those
1140	 * unmapped blocks.
1141	 */
1142	r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1143				   m->data_block + (m->virt_end - m->virt_begin));
1144	if (r) {
1145		metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1146		bio_io_error(m->bio);
1147	} else
1148		bio_endio(m->bio);
1149
1150	cell_defer_no_holder(tc, m->cell);
1151	mempool_free(m, pool->mapping_pool);
1152}
1153
1154static void process_prepared(struct pool *pool, struct list_head *head,
1155			     process_mapping_fn *fn)
1156{
1157	unsigned long flags;
1158	struct list_head maps;
1159	struct dm_thin_new_mapping *m, *tmp;
1160
1161	INIT_LIST_HEAD(&maps);
1162	spin_lock_irqsave(&pool->lock, flags);
1163	list_splice_init(head, &maps);
1164	spin_unlock_irqrestore(&pool->lock, flags);
1165
1166	list_for_each_entry_safe(m, tmp, &maps, list)
1167		(*fn)(m);
1168}
1169
1170/*
1171 * Deferred bio jobs.
1172 */
1173static int io_overlaps_block(struct pool *pool, struct bio *bio)
1174{
1175	return bio->bi_iter.bi_size ==
1176		(pool->sectors_per_block << SECTOR_SHIFT);
1177}
1178
1179static int io_overwrites_block(struct pool *pool, struct bio *bio)
1180{
1181	return (bio_data_dir(bio) == WRITE) &&
1182		io_overlaps_block(pool, bio);
1183}
1184
1185static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1186			       bio_end_io_t *fn)
1187{
1188	*save = bio->bi_end_io;
1189	bio->bi_end_io = fn;
1190}
1191
1192static int ensure_next_mapping(struct pool *pool)
1193{
1194	if (pool->next_mapping)
1195		return 0;
1196
1197	pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
1198
1199	return pool->next_mapping ? 0 : -ENOMEM;
1200}
1201
1202static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1203{
1204	struct dm_thin_new_mapping *m = pool->next_mapping;
1205
1206	BUG_ON(!pool->next_mapping);
1207
1208	memset(m, 0, sizeof(struct dm_thin_new_mapping));
1209	INIT_LIST_HEAD(&m->list);
1210	m->bio = NULL;
1211
1212	pool->next_mapping = NULL;
1213
1214	return m;
1215}
1216
1217static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1218		    sector_t begin, sector_t end)
1219{
1220	int r;
1221	struct dm_io_region to;
1222
1223	to.bdev = tc->pool_dev->bdev;
1224	to.sector = begin;
1225	to.count = end - begin;
1226
1227	r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1228	if (r < 0) {
1229		DMERR_LIMIT("dm_kcopyd_zero() failed");
1230		copy_complete(1, 1, m);
1231	}
1232}
1233
1234static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1235				      dm_block_t data_begin,
1236				      struct dm_thin_new_mapping *m)
1237{
1238	struct pool *pool = tc->pool;
1239	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1240
1241	h->overwrite_mapping = m;
1242	m->bio = bio;
1243	save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1244	inc_all_io_entry(pool, bio);
1245	remap_and_issue(tc, bio, data_begin);
1246}
1247
1248/*
1249 * A partial copy also needs to zero the uncopied region.
1250 */
1251static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1252			  struct dm_dev *origin, dm_block_t data_origin,
1253			  dm_block_t data_dest,
1254			  struct dm_bio_prison_cell *cell, struct bio *bio,
1255			  sector_t len)
1256{
1257	int r;
1258	struct pool *pool = tc->pool;
1259	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1260
1261	m->tc = tc;
1262	m->virt_begin = virt_block;
1263	m->virt_end = virt_block + 1u;
1264	m->data_block = data_dest;
1265	m->cell = cell;
1266
1267	/*
1268	 * quiesce action + copy action + an extra reference held for the
1269	 * duration of this function (we may need to inc later for a
1270	 * partial zero).
1271	 */
1272	atomic_set(&m->prepare_actions, 3);
1273
1274	if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1275		complete_mapping_preparation(m); /* already quiesced */
1276
1277	/*
1278	 * IO to pool_dev remaps to the pool target's data_dev.
1279	 *
1280	 * If the whole block of data is being overwritten, we can issue the
1281	 * bio immediately. Otherwise we use kcopyd to clone the data first.
1282	 */
1283	if (io_overwrites_block(pool, bio))
1284		remap_and_issue_overwrite(tc, bio, data_dest, m);
1285	else {
1286		struct dm_io_region from, to;
1287
1288		from.bdev = origin->bdev;
1289		from.sector = data_origin * pool->sectors_per_block;
1290		from.count = len;
1291
1292		to.bdev = tc->pool_dev->bdev;
1293		to.sector = data_dest * pool->sectors_per_block;
1294		to.count = len;
1295
1296		r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1297				   0, copy_complete, m);
1298		if (r < 0) {
1299			DMERR_LIMIT("dm_kcopyd_copy() failed");
1300			copy_complete(1, 1, m);
1301
1302			/*
1303			 * We allow the zero to be issued, to simplify the
1304			 * error path.  Otherwise we'd need to start
1305			 * worrying about decrementing the prepare_actions
1306			 * counter.
1307			 */
1308		}
1309
1310		/*
1311		 * Do we need to zero a tail region?
1312		 */
1313		if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1314			atomic_inc(&m->prepare_actions);
1315			ll_zero(tc, m,
1316				data_dest * pool->sectors_per_block + len,
1317				(data_dest + 1) * pool->sectors_per_block);
1318		}
1319	}
1320
1321	complete_mapping_preparation(m); /* drop our ref */
1322}
1323
1324static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1325				   dm_block_t data_origin, dm_block_t data_dest,
1326				   struct dm_bio_prison_cell *cell, struct bio *bio)
1327{
1328	schedule_copy(tc, virt_block, tc->pool_dev,
1329		      data_origin, data_dest, cell, bio,
1330		      tc->pool->sectors_per_block);
1331}
1332
1333static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1334			  dm_block_t data_block, struct dm_bio_prison_cell *cell,
1335			  struct bio *bio)
1336{
1337	struct pool *pool = tc->pool;
1338	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1339
1340	atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1341	m->tc = tc;
1342	m->virt_begin = virt_block;
1343	m->virt_end = virt_block + 1u;
1344	m->data_block = data_block;
1345	m->cell = cell;
1346
1347	/*
1348	 * If the whole block of data is being overwritten or we are not
1349	 * zeroing pre-existing data, we can issue the bio immediately.
1350	 * Otherwise we use kcopyd to zero the data first.
1351	 */
1352	if (pool->pf.zero_new_blocks) {
1353		if (io_overwrites_block(pool, bio))
1354			remap_and_issue_overwrite(tc, bio, data_block, m);
1355		else
1356			ll_zero(tc, m, data_block * pool->sectors_per_block,
1357				(data_block + 1) * pool->sectors_per_block);
1358	} else
1359		process_prepared_mapping(m);
1360}
1361
1362static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1363				   dm_block_t data_dest,
1364				   struct dm_bio_prison_cell *cell, struct bio *bio)
1365{
1366	struct pool *pool = tc->pool;
1367	sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1368	sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1369
1370	if (virt_block_end <= tc->origin_size)
1371		schedule_copy(tc, virt_block, tc->origin_dev,
1372			      virt_block, data_dest, cell, bio,
1373			      pool->sectors_per_block);
1374
1375	else if (virt_block_begin < tc->origin_size)
1376		schedule_copy(tc, virt_block, tc->origin_dev,
1377			      virt_block, data_dest, cell, bio,
1378			      tc->origin_size - virt_block_begin);
1379
1380	else
1381		schedule_zero(tc, virt_block, data_dest, cell, bio);
1382}
1383
1384static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1385
1386static void check_for_space(struct pool *pool)
1387{
1388	int r;
1389	dm_block_t nr_free;
1390
1391	if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1392		return;
1393
1394	r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1395	if (r)
1396		return;
1397
1398	if (nr_free)
1399		set_pool_mode(pool, PM_WRITE);
1400}
1401
1402/*
1403 * A non-zero return indicates read_only or fail_io mode.
1404 * Many callers don't care about the return value.
1405 */
1406static int commit(struct pool *pool)
1407{
1408	int r;
1409
1410	if (get_pool_mode(pool) >= PM_READ_ONLY)
1411		return -EINVAL;
1412
1413	r = dm_pool_commit_metadata(pool->pmd);
1414	if (r)
1415		metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1416	else
1417		check_for_space(pool);
1418
1419	return r;
1420}
1421
1422static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1423{
1424	unsigned long flags;
1425
1426	if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1427		DMWARN("%s: reached low water mark for data device: sending event.",
1428		       dm_device_name(pool->pool_md));
1429		spin_lock_irqsave(&pool->lock, flags);
1430		pool->low_water_triggered = true;
1431		spin_unlock_irqrestore(&pool->lock, flags);
1432		dm_table_event(pool->ti->table);
1433	}
1434}
1435
1436static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1437{
1438	int r;
1439	dm_block_t free_blocks;
1440	struct pool *pool = tc->pool;
1441
1442	if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1443		return -EINVAL;
1444
1445	r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1446	if (r) {
1447		metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1448		return r;
1449	}
1450
1451	check_low_water_mark(pool, free_blocks);
1452
1453	if (!free_blocks) {
1454		/*
1455		 * Try to commit to see if that will free up some
1456		 * more space.
1457		 */
1458		r = commit(pool);
1459		if (r)
1460			return r;
1461
1462		r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1463		if (r) {
1464			metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1465			return r;
1466		}
1467
1468		if (!free_blocks) {
1469			set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1470			return -ENOSPC;
1471		}
1472	}
1473
1474	r = dm_pool_alloc_data_block(pool->pmd, result);
1475	if (r) {
1476		metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1477		return r;
1478	}
1479
1480	return 0;
1481}
1482
1483/*
1484 * If we have run out of space, queue bios until the device is
1485 * resumed, presumably after having been reloaded with more space.
1486 */
1487static void retry_on_resume(struct bio *bio)
1488{
1489	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1490	struct thin_c *tc = h->tc;
1491	unsigned long flags;
1492
1493	spin_lock_irqsave(&tc->lock, flags);
1494	bio_list_add(&tc->retry_on_resume_list, bio);
1495	spin_unlock_irqrestore(&tc->lock, flags);
1496}
1497
1498static int should_error_unserviceable_bio(struct pool *pool)
1499{
1500	enum pool_mode m = get_pool_mode(pool);
1501
1502	switch (m) {
1503	case PM_WRITE:
1504		/* Shouldn't get here */
1505		DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1506		return -EIO;
1507
1508	case PM_OUT_OF_DATA_SPACE:
1509		return pool->pf.error_if_no_space ? -ENOSPC : 0;
1510
1511	case PM_READ_ONLY:
1512	case PM_FAIL:
1513		return -EIO;
1514	default:
1515		/* Shouldn't get here */
1516		DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1517		return -EIO;
1518	}
1519}
1520
1521static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1522{
1523	int error = should_error_unserviceable_bio(pool);
1524
1525	if (error) {
1526		bio->bi_error = error;
1527		bio_endio(bio);
1528	} else
1529		retry_on_resume(bio);
1530}
1531
1532static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1533{
1534	struct bio *bio;
1535	struct bio_list bios;
1536	int error;
1537
1538	error = should_error_unserviceable_bio(pool);
1539	if (error) {
1540		cell_error_with_code(pool, cell, error);
1541		return;
1542	}
1543
1544	bio_list_init(&bios);
1545	cell_release(pool, cell, &bios);
1546
1547	while ((bio = bio_list_pop(&bios)))
1548		retry_on_resume(bio);
1549}
1550
1551static void process_discard_cell_no_passdown(struct thin_c *tc,
1552					     struct dm_bio_prison_cell *virt_cell)
1553{
1554	struct pool *pool = tc->pool;
1555	struct dm_thin_new_mapping *m = get_next_mapping(pool);
1556
1557	/*
1558	 * We don't need to lock the data blocks, since there's no
1559	 * passdown.  We only lock data blocks for allocation and breaking sharing.
1560	 */
1561	m->tc = tc;
1562	m->virt_begin = virt_cell->key.block_begin;
1563	m->virt_end = virt_cell->key.block_end;
1564	m->cell = virt_cell;
1565	m->bio = virt_cell->holder;
1566
1567	if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1568		pool->process_prepared_discard(m);
1569}
1570
1571static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1572				 struct bio *bio)
1573{
1574	struct pool *pool = tc->pool;
1575
1576	int r;
1577	bool maybe_shared;
1578	struct dm_cell_key data_key;
1579	struct dm_bio_prison_cell *data_cell;
1580	struct dm_thin_new_mapping *m;
1581	dm_block_t virt_begin, virt_end, data_begin;
1582
1583	while (begin != end) {
1584		r = ensure_next_mapping(pool);
1585		if (r)
1586			/* we did our best */
1587			return;
1588
1589		r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1590					      &data_begin, &maybe_shared);
1591		if (r)
1592			/*
1593			 * Silently fail, letting any mappings we've
1594			 * created complete.
1595			 */
1596			break;
1597
1598		build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1599		if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1600			/* contention, we'll give up with this range */
1601			begin = virt_end;
1602			continue;
1603		}
1604
1605		/*
1606		 * IO may still be going to the destination block.  We must
1607		 * quiesce before we can do the removal.
1608		 */
1609		m = get_next_mapping(pool);
1610		m->tc = tc;
1611		m->maybe_shared = maybe_shared;
1612		m->virt_begin = virt_begin;
1613		m->virt_end = virt_end;
1614		m->data_block = data_begin;
1615		m->cell = data_cell;
1616		m->bio = bio;
1617
1618		/*
1619		 * The parent bio must not complete before sub discard bios are
1620		 * chained to it (see end_discard's bio_chain)!
1621		 *
1622		 * This per-mapping bi_remaining increment is paired with
1623		 * the implicit decrement that occurs via bio_endio() in
1624		 * end_discard().
1625		 */
1626		bio_inc_remaining(bio);
1627		if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1628			pool->process_prepared_discard(m);
1629
1630		begin = virt_end;
1631	}
1632}
1633
1634static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1635{
1636	struct bio *bio = virt_cell->holder;
1637	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1638
1639	/*
1640	 * The virt_cell will only get freed once the origin bio completes.
1641	 * This means it will remain locked while all the individual
1642	 * passdown bios are in flight.
1643	 */
1644	h->cell = virt_cell;
1645	break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1646
1647	/*
1648	 * We complete the bio now, knowing that the bi_remaining field
1649	 * will prevent completion until the sub range discards have
1650	 * completed.
1651	 */
1652	bio_endio(bio);
1653}
1654
1655static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1656{
1657	dm_block_t begin, end;
1658	struct dm_cell_key virt_key;
1659	struct dm_bio_prison_cell *virt_cell;
1660
1661	get_bio_block_range(tc, bio, &begin, &end);
1662	if (begin == end) {
1663		/*
1664		 * The discard covers less than a block.
1665		 */
1666		bio_endio(bio);
1667		return;
1668	}
1669
1670	build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1671	if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1672		/*
1673		 * Potential starvation issue: We're relying on the
1674		 * fs/application being well behaved, and not trying to
1675		 * send IO to a region at the same time as discarding it.
1676		 * If they do this persistently then it's possible this
1677		 * cell will never be granted.
1678		 */
1679		return;
1680
1681	tc->pool->process_discard_cell(tc, virt_cell);
1682}
1683
1684static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1685			  struct dm_cell_key *key,
1686			  struct dm_thin_lookup_result *lookup_result,
1687			  struct dm_bio_prison_cell *cell)
1688{
1689	int r;
1690	dm_block_t data_block;
1691	struct pool *pool = tc->pool;
1692
1693	r = alloc_data_block(tc, &data_block);
1694	switch (r) {
1695	case 0:
1696		schedule_internal_copy(tc, block, lookup_result->block,
1697				       data_block, cell, bio);
1698		break;
1699
1700	case -ENOSPC:
1701		retry_bios_on_resume(pool, cell);
1702		break;
1703
1704	default:
1705		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1706			    __func__, r);
1707		cell_error(pool, cell);
1708		break;
1709	}
1710}
1711
1712static void __remap_and_issue_shared_cell(void *context,
1713					  struct dm_bio_prison_cell *cell)
1714{
1715	struct remap_info *info = context;
1716	struct bio *bio;
1717
1718	while ((bio = bio_list_pop(&cell->bios))) {
1719		if ((bio_data_dir(bio) == WRITE) ||
1720		    (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA) ||
1721		     bio_op(bio) == REQ_OP_DISCARD))
1722			bio_list_add(&info->defer_bios, bio);
1723		else {
1724			struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1725
1726			h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1727			inc_all_io_entry(info->tc->pool, bio);
1728			bio_list_add(&info->issue_bios, bio);
1729		}
1730	}
1731}
1732
1733static void remap_and_issue_shared_cell(struct thin_c *tc,
1734					struct dm_bio_prison_cell *cell,
1735					dm_block_t block)
1736{
1737	struct bio *bio;
1738	struct remap_info info;
1739
1740	info.tc = tc;
1741	bio_list_init(&info.defer_bios);
1742	bio_list_init(&info.issue_bios);
1743
1744	cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1745			   &info, cell);
1746
1747	while ((bio = bio_list_pop(&info.defer_bios)))
1748		thin_defer_bio(tc, bio);
1749
1750	while ((bio = bio_list_pop(&info.issue_bios)))
1751		remap_and_issue(tc, bio, block);
1752}
1753
1754static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1755			       dm_block_t block,
1756			       struct dm_thin_lookup_result *lookup_result,
1757			       struct dm_bio_prison_cell *virt_cell)
1758{
1759	struct dm_bio_prison_cell *data_cell;
1760	struct pool *pool = tc->pool;
1761	struct dm_cell_key key;
1762
1763	/*
1764	 * If cell is already occupied, then sharing is already in the process
1765	 * of being broken so we have nothing further to do here.
1766	 */
1767	build_data_key(tc->td, lookup_result->block, &key);
1768	if (bio_detain(pool, &key, bio, &data_cell)) {
1769		cell_defer_no_holder(tc, virt_cell);
1770		return;
1771	}
1772
1773	if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1774		break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1775		cell_defer_no_holder(tc, virt_cell);
1776	} else {
1777		struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1778
1779		h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1780		inc_all_io_entry(pool, bio);
1781		remap_and_issue(tc, bio, lookup_result->block);
1782
1783		remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1784		remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1785	}
1786}
1787
1788static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1789			    struct dm_bio_prison_cell *cell)
1790{
1791	int r;
1792	dm_block_t data_block;
1793	struct pool *pool = tc->pool;
1794
1795	/*
1796	 * Remap empty bios (flushes) immediately, without provisioning.
1797	 */
1798	if (!bio->bi_iter.bi_size) {
1799		inc_all_io_entry(pool, bio);
1800		cell_defer_no_holder(tc, cell);
1801
1802		remap_and_issue(tc, bio, 0);
1803		return;
1804	}
1805
1806	/*
1807	 * Fill read bios with zeroes and complete them immediately.
1808	 */
1809	if (bio_data_dir(bio) == READ) {
1810		zero_fill_bio(bio);
1811		cell_defer_no_holder(tc, cell);
1812		bio_endio(bio);
1813		return;
1814	}
1815
1816	r = alloc_data_block(tc, &data_block);
1817	switch (r) {
1818	case 0:
1819		if (tc->origin_dev)
1820			schedule_external_copy(tc, block, data_block, cell, bio);
1821		else
1822			schedule_zero(tc, block, data_block, cell, bio);
1823		break;
1824
1825	case -ENOSPC:
1826		retry_bios_on_resume(pool, cell);
1827		break;
1828
1829	default:
1830		DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1831			    __func__, r);
1832		cell_error(pool, cell);
1833		break;
1834	}
1835}
1836
1837static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1838{
1839	int r;
1840	struct pool *pool = tc->pool;
1841	struct bio *bio = cell->holder;
1842	dm_block_t block = get_bio_block(tc, bio);
1843	struct dm_thin_lookup_result lookup_result;
1844
1845	if (tc->requeue_mode) {
1846		cell_requeue(pool, cell);
1847		return;
1848	}
1849
1850	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1851	switch (r) {
1852	case 0:
1853		if (lookup_result.shared)
1854			process_shared_bio(tc, bio, block, &lookup_result, cell);
1855		else {
1856			inc_all_io_entry(pool, bio);
1857			remap_and_issue(tc, bio, lookup_result.block);
1858			inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1859		}
1860		break;
1861
1862	case -ENODATA:
1863		if (bio_data_dir(bio) == READ && tc->origin_dev) {
1864			inc_all_io_entry(pool, bio);
1865			cell_defer_no_holder(tc, cell);
1866
1867			if (bio_end_sector(bio) <= tc->origin_size)
1868				remap_to_origin_and_issue(tc, bio);
1869
1870			else if (bio->bi_iter.bi_sector < tc->origin_size) {
1871				zero_fill_bio(bio);
1872				bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1873				remap_to_origin_and_issue(tc, bio);
1874
1875			} else {
1876				zero_fill_bio(bio);
1877				bio_endio(bio);
1878			}
1879		} else
1880			provision_block(tc, bio, block, cell);
1881		break;
1882
1883	default:
1884		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1885			    __func__, r);
1886		cell_defer_no_holder(tc, cell);
1887		bio_io_error(bio);
1888		break;
1889	}
1890}
1891
1892static void process_bio(struct thin_c *tc, struct bio *bio)
1893{
1894	struct pool *pool = tc->pool;
1895	dm_block_t block = get_bio_block(tc, bio);
1896	struct dm_bio_prison_cell *cell;
1897	struct dm_cell_key key;
1898
1899	/*
1900	 * If cell is already occupied, then the block is already
1901	 * being provisioned so we have nothing further to do here.
1902	 */
1903	build_virtual_key(tc->td, block, &key);
1904	if (bio_detain(pool, &key, bio, &cell))
1905		return;
1906
1907	process_cell(tc, cell);
1908}
1909
1910static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1911				    struct dm_bio_prison_cell *cell)
1912{
1913	int r;
1914	int rw = bio_data_dir(bio);
1915	dm_block_t block = get_bio_block(tc, bio);
1916	struct dm_thin_lookup_result lookup_result;
1917
1918	r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1919	switch (r) {
1920	case 0:
1921		if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1922			handle_unserviceable_bio(tc->pool, bio);
1923			if (cell)
1924				cell_defer_no_holder(tc, cell);
1925		} else {
1926			inc_all_io_entry(tc->pool, bio);
1927			remap_and_issue(tc, bio, lookup_result.block);
1928			if (cell)
1929				inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1930		}
1931		break;
1932
1933	case -ENODATA:
1934		if (cell)
1935			cell_defer_no_holder(tc, cell);
1936		if (rw != READ) {
1937			handle_unserviceable_bio(tc->pool, bio);
1938			break;
1939		}
1940
1941		if (tc->origin_dev) {
1942			inc_all_io_entry(tc->pool, bio);
1943			remap_to_origin_and_issue(tc, bio);
1944			break;
1945		}
1946
1947		zero_fill_bio(bio);
1948		bio_endio(bio);
1949		break;
1950
1951	default:
1952		DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1953			    __func__, r);
1954		if (cell)
1955			cell_defer_no_holder(tc, cell);
1956		bio_io_error(bio);
1957		break;
1958	}
1959}
1960
1961static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1962{
1963	__process_bio_read_only(tc, bio, NULL);
1964}
1965
1966static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1967{
1968	__process_bio_read_only(tc, cell->holder, cell);
1969}
1970
1971static void process_bio_success(struct thin_c *tc, struct bio *bio)
1972{
1973	bio_endio(bio);
1974}
1975
1976static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1977{
1978	bio_io_error(bio);
1979}
1980
1981static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1982{
1983	cell_success(tc->pool, cell);
1984}
1985
1986static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1987{
1988	cell_error(tc->pool, cell);
1989}
1990
1991/*
1992 * FIXME: should we also commit due to size of transaction, measured in
1993 * metadata blocks?
1994 */
1995static int need_commit_due_to_time(struct pool *pool)
1996{
1997	return !time_in_range(jiffies, pool->last_commit_jiffies,
1998			      pool->last_commit_jiffies + COMMIT_PERIOD);
1999}
2000
2001#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2002#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2003
2004static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2005{
2006	struct rb_node **rbp, *parent;
2007	struct dm_thin_endio_hook *pbd;
2008	sector_t bi_sector = bio->bi_iter.bi_sector;
2009
2010	rbp = &tc->sort_bio_list.rb_node;
2011	parent = NULL;
2012	while (*rbp) {
2013		parent = *rbp;
2014		pbd = thin_pbd(parent);
2015
2016		if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2017			rbp = &(*rbp)->rb_left;
2018		else
2019			rbp = &(*rbp)->rb_right;
2020	}
2021
2022	pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2023	rb_link_node(&pbd->rb_node, parent, rbp);
2024	rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2025}
2026
2027static void __extract_sorted_bios(struct thin_c *tc)
2028{
2029	struct rb_node *node;
2030	struct dm_thin_endio_hook *pbd;
2031	struct bio *bio;
2032
2033	for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2034		pbd = thin_pbd(node);
2035		bio = thin_bio(pbd);
2036
2037		bio_list_add(&tc->deferred_bio_list, bio);
2038		rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2039	}
2040
2041	WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2042}
2043
2044static void __sort_thin_deferred_bios(struct thin_c *tc)
2045{
2046	struct bio *bio;
2047	struct bio_list bios;
2048
2049	bio_list_init(&bios);
2050	bio_list_merge(&bios, &tc->deferred_bio_list);
2051	bio_list_init(&tc->deferred_bio_list);
2052
2053	/* Sort deferred_bio_list using rb-tree */
2054	while ((bio = bio_list_pop(&bios)))
2055		__thin_bio_rb_add(tc, bio);
2056
2057	/*
2058	 * Transfer the sorted bios in sort_bio_list back to
2059	 * deferred_bio_list to allow lockless submission of
2060	 * all bios.
2061	 */
2062	__extract_sorted_bios(tc);
2063}
2064
2065static void process_thin_deferred_bios(struct thin_c *tc)
2066{
2067	struct pool *pool = tc->pool;
2068	unsigned long flags;
2069	struct bio *bio;
2070	struct bio_list bios;
2071	struct blk_plug plug;
2072	unsigned count = 0;
2073
2074	if (tc->requeue_mode) {
2075		error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
2076		return;
2077	}
2078
2079	bio_list_init(&bios);
2080
2081	spin_lock_irqsave(&tc->lock, flags);
2082
2083	if (bio_list_empty(&tc->deferred_bio_list)) {
2084		spin_unlock_irqrestore(&tc->lock, flags);
2085		return;
2086	}
2087
2088	__sort_thin_deferred_bios(tc);
2089
2090	bio_list_merge(&bios, &tc->deferred_bio_list);
2091	bio_list_init(&tc->deferred_bio_list);
2092
2093	spin_unlock_irqrestore(&tc->lock, flags);
2094
2095	blk_start_plug(&plug);
2096	while ((bio = bio_list_pop(&bios))) {
2097		/*
2098		 * If we've got no free new_mapping structs, and processing
2099		 * this bio might require one, we pause until there are some
2100		 * prepared mappings to process.
2101		 */
2102		if (ensure_next_mapping(pool)) {
2103			spin_lock_irqsave(&tc->lock, flags);
2104			bio_list_add(&tc->deferred_bio_list, bio);
2105			bio_list_merge(&tc->deferred_bio_list, &bios);
2106			spin_unlock_irqrestore(&tc->lock, flags);
2107			break;
2108		}
2109
2110		if (bio_op(bio) == REQ_OP_DISCARD)
2111			pool->process_discard(tc, bio);
2112		else
2113			pool->process_bio(tc, bio);
2114
2115		if ((count++ & 127) == 0) {
2116			throttle_work_update(&pool->throttle);
2117			dm_pool_issue_prefetches(pool->pmd);
2118		}
2119	}
2120	blk_finish_plug(&plug);
2121}
2122
2123static int cmp_cells(const void *lhs, const void *rhs)
2124{
2125	struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2126	struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2127
2128	BUG_ON(!lhs_cell->holder);
2129	BUG_ON(!rhs_cell->holder);
2130
2131	if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2132		return -1;
2133
2134	if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2135		return 1;
2136
2137	return 0;
2138}
2139
2140static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2141{
2142	unsigned count = 0;
2143	struct dm_bio_prison_cell *cell, *tmp;
2144
2145	list_for_each_entry_safe(cell, tmp, cells, user_list) {
2146		if (count >= CELL_SORT_ARRAY_SIZE)
2147			break;
2148
2149		pool->cell_sort_array[count++] = cell;
2150		list_del(&cell->user_list);
2151	}
2152
2153	sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2154
2155	return count;
2156}
2157
2158static void process_thin_deferred_cells(struct thin_c *tc)
2159{
2160	struct pool *pool = tc->pool;
2161	unsigned long flags;
2162	struct list_head cells;
2163	struct dm_bio_prison_cell *cell;
2164	unsigned i, j, count;
2165
2166	INIT_LIST_HEAD(&cells);
2167
2168	spin_lock_irqsave(&tc->lock, flags);
2169	list_splice_init(&tc->deferred_cells, &cells);
2170	spin_unlock_irqrestore(&tc->lock, flags);
2171
2172	if (list_empty(&cells))
2173		return;
2174
2175	do {
2176		count = sort_cells(tc->pool, &cells);
2177
2178		for (i = 0; i < count; i++) {
2179			cell = pool->cell_sort_array[i];
2180			BUG_ON(!cell->holder);
2181
2182			/*
2183			 * If we've got no free new_mapping structs, and processing
2184			 * this bio might require one, we pause until there are some
2185			 * prepared mappings to process.
2186			 */
2187			if (ensure_next_mapping(pool)) {
2188				for (j = i; j < count; j++)
2189					list_add(&pool->cell_sort_array[j]->user_list, &cells);
2190
2191				spin_lock_irqsave(&tc->lock, flags);
2192				list_splice(&cells, &tc->deferred_cells);
2193				spin_unlock_irqrestore(&tc->lock, flags);
2194				return;
2195			}
2196
2197			if (bio_op(cell->holder) == REQ_OP_DISCARD)
2198				pool->process_discard_cell(tc, cell);
2199			else
2200				pool->process_cell(tc, cell);
2201		}
2202	} while (!list_empty(&cells));
2203}
2204
2205static void thin_get(struct thin_c *tc);
2206static void thin_put(struct thin_c *tc);
2207
2208/*
2209 * We can't hold rcu_read_lock() around code that can block.  So we
2210 * find a thin with the rcu lock held; bump a refcount; then drop
2211 * the lock.
2212 */
2213static struct thin_c *get_first_thin(struct pool *pool)
2214{
2215	struct thin_c *tc = NULL;
2216
2217	rcu_read_lock();
2218	if (!list_empty(&pool->active_thins)) {
2219		tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2220		thin_get(tc);
2221	}
2222	rcu_read_unlock();
2223
2224	return tc;
2225}
2226
2227static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2228{
2229	struct thin_c *old_tc = tc;
2230
2231	rcu_read_lock();
2232	list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2233		thin_get(tc);
2234		thin_put(old_tc);
2235		rcu_read_unlock();
2236		return tc;
2237	}
2238	thin_put(old_tc);
2239	rcu_read_unlock();
2240
2241	return NULL;
2242}
2243
2244static void process_deferred_bios(struct pool *pool)
2245{
2246	unsigned long flags;
2247	struct bio *bio;
2248	struct bio_list bios;
2249	struct thin_c *tc;
2250
2251	tc = get_first_thin(pool);
2252	while (tc) {
2253		process_thin_deferred_cells(tc);
2254		process_thin_deferred_bios(tc);
2255		tc = get_next_thin(pool, tc);
2256	}
2257
2258	/*
2259	 * If there are any deferred flush bios, we must commit
2260	 * the metadata before issuing them.
2261	 */
2262	bio_list_init(&bios);
2263	spin_lock_irqsave(&pool->lock, flags);
2264	bio_list_merge(&bios, &pool->deferred_flush_bios);
2265	bio_list_init(&pool->deferred_flush_bios);
2266	spin_unlock_irqrestore(&pool->lock, flags);
2267
2268	if (bio_list_empty(&bios) &&
2269	    !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2270		return;
2271
2272	if (commit(pool)) {
2273		while ((bio = bio_list_pop(&bios)))
2274			bio_io_error(bio);
2275		return;
2276	}
2277	pool->last_commit_jiffies = jiffies;
2278
2279	while ((bio = bio_list_pop(&bios)))
2280		generic_make_request(bio);
2281}
2282
2283static void do_worker(struct work_struct *ws)
2284{
2285	struct pool *pool = container_of(ws, struct pool, worker);
2286
2287	throttle_work_start(&pool->throttle);
2288	dm_pool_issue_prefetches(pool->pmd);
2289	throttle_work_update(&pool->throttle);
2290	process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2291	throttle_work_update(&pool->throttle);
2292	process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2293	throttle_work_update(&pool->throttle);
2294	process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2295	throttle_work_update(&pool->throttle);
2296	process_deferred_bios(pool);
2297	throttle_work_complete(&pool->throttle);
2298}
2299
2300/*
2301 * We want to commit periodically so that not too much
2302 * unwritten data builds up.
2303 */
2304static void do_waker(struct work_struct *ws)
2305{
2306	struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2307	wake_worker(pool);
2308	queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2309}
2310
2311static void notify_of_pool_mode_change_to_oods(struct pool *pool);
2312
2313/*
2314 * We're holding onto IO to allow userland time to react.  After the
2315 * timeout either the pool will have been resized (and thus back in
2316 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2317 */
2318static void do_no_space_timeout(struct work_struct *ws)
2319{
2320	struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2321					 no_space_timeout);
2322
2323	if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2324		pool->pf.error_if_no_space = true;
2325		notify_of_pool_mode_change_to_oods(pool);
2326		error_retry_list_with_code(pool, -ENOSPC);
2327	}
2328}
2329
2330/*----------------------------------------------------------------*/
2331
2332struct pool_work {
2333	struct work_struct worker;
2334	struct completion complete;
2335};
2336
2337static struct pool_work *to_pool_work(struct work_struct *ws)
2338{
2339	return container_of(ws, struct pool_work, worker);
2340}
2341
2342static void pool_work_complete(struct pool_work *pw)
2343{
2344	complete(&pw->complete);
2345}
2346
2347static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2348			   void (*fn)(struct work_struct *))
2349{
2350	INIT_WORK_ONSTACK(&pw->worker, fn);
2351	init_completion(&pw->complete);
2352	queue_work(pool->wq, &pw->worker);
2353	wait_for_completion(&pw->complete);
2354}
2355
2356/*----------------------------------------------------------------*/
2357
2358struct noflush_work {
2359	struct pool_work pw;
2360	struct thin_c *tc;
2361};
2362
2363static struct noflush_work *to_noflush(struct work_struct *ws)
2364{
2365	return container_of(to_pool_work(ws), struct noflush_work, pw);
2366}
2367
2368static void do_noflush_start(struct work_struct *ws)
2369{
2370	struct noflush_work *w = to_noflush(ws);
2371	w->tc->requeue_mode = true;
2372	requeue_io(w->tc);
2373	pool_work_complete(&w->pw);
2374}
2375
2376static void do_noflush_stop(struct work_struct *ws)
2377{
2378	struct noflush_work *w = to_noflush(ws);
2379	w->tc->requeue_mode = false;
2380	pool_work_complete(&w->pw);
2381}
2382
2383static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2384{
2385	struct noflush_work w;
2386
2387	w.tc = tc;
2388	pool_work_wait(&w.pw, tc->pool, fn);
2389}
2390
2391/*----------------------------------------------------------------*/
2392
2393static enum pool_mode get_pool_mode(struct pool *pool)
2394{
2395	return pool->pf.mode;
2396}
2397
2398static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2399{
2400	dm_table_event(pool->ti->table);
2401	DMINFO("%s: switching pool to %s mode",
2402	       dm_device_name(pool->pool_md), new_mode);
2403}
2404
2405static void notify_of_pool_mode_change_to_oods(struct pool *pool)
2406{
2407	if (!pool->pf.error_if_no_space)
2408		notify_of_pool_mode_change(pool, "out-of-data-space (queue IO)");
2409	else
2410		notify_of_pool_mode_change(pool, "out-of-data-space (error IO)");
2411}
2412
2413static bool passdown_enabled(struct pool_c *pt)
2414{
2415	return pt->adjusted_pf.discard_passdown;
2416}
2417
2418static void set_discard_callbacks(struct pool *pool)
2419{
2420	struct pool_c *pt = pool->ti->private;
2421
2422	if (passdown_enabled(pt)) {
2423		pool->process_discard_cell = process_discard_cell_passdown;
2424		pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2425		pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2426	} else {
2427		pool->process_discard_cell = process_discard_cell_no_passdown;
2428		pool->process_prepared_discard = process_prepared_discard_no_passdown;
2429	}
2430}
2431
2432static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2433{
2434	struct pool_c *pt = pool->ti->private;
2435	bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2436	enum pool_mode old_mode = get_pool_mode(pool);
2437	unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2438
2439	/*
2440	 * Never allow the pool to transition to PM_WRITE mode if user
2441	 * intervention is required to verify metadata and data consistency.
2442	 */
2443	if (new_mode == PM_WRITE && needs_check) {
2444		DMERR("%s: unable to switch pool to write mode until repaired.",
2445		      dm_device_name(pool->pool_md));
2446		if (old_mode != new_mode)
2447			new_mode = old_mode;
2448		else
2449			new_mode = PM_READ_ONLY;
2450	}
2451	/*
2452	 * If we were in PM_FAIL mode, rollback of metadata failed.  We're
2453	 * not going to recover without a thin_repair.	So we never let the
2454	 * pool move out of the old mode.
2455	 */
2456	if (old_mode == PM_FAIL)
2457		new_mode = old_mode;
2458
2459	switch (new_mode) {
2460	case PM_FAIL:
2461		if (old_mode != new_mode)
2462			notify_of_pool_mode_change(pool, "failure");
2463		dm_pool_metadata_read_only(pool->pmd);
2464		pool->process_bio = process_bio_fail;
2465		pool->process_discard = process_bio_fail;
2466		pool->process_cell = process_cell_fail;
2467		pool->process_discard_cell = process_cell_fail;
2468		pool->process_prepared_mapping = process_prepared_mapping_fail;
2469		pool->process_prepared_discard = process_prepared_discard_fail;
2470
2471		error_retry_list(pool);
2472		break;
2473
2474	case PM_READ_ONLY:
2475		if (old_mode != new_mode)
2476			notify_of_pool_mode_change(pool, "read-only");
2477		dm_pool_metadata_read_only(pool->pmd);
2478		pool->process_bio = process_bio_read_only;
2479		pool->process_discard = process_bio_success;
2480		pool->process_cell = process_cell_read_only;
2481		pool->process_discard_cell = process_cell_success;
2482		pool->process_prepared_mapping = process_prepared_mapping_fail;
2483		pool->process_prepared_discard = process_prepared_discard_success;
2484
2485		error_retry_list(pool);
2486		break;
2487
2488	case PM_OUT_OF_DATA_SPACE:
2489		/*
2490		 * Ideally we'd never hit this state; the low water mark
2491		 * would trigger userland to extend the pool before we
2492		 * completely run out of data space.  However, many small
2493		 * IOs to unprovisioned space can consume data space at an
2494		 * alarming rate.  Adjust your low water mark if you're
2495		 * frequently seeing this mode.
2496		 */
2497		if (old_mode != new_mode)
2498			notify_of_pool_mode_change_to_oods(pool);
2499		pool->out_of_data_space = true;
2500		pool->process_bio = process_bio_read_only;
2501		pool->process_discard = process_discard_bio;
2502		pool->process_cell = process_cell_read_only;
2503		pool->process_prepared_mapping = process_prepared_mapping;
2504		set_discard_callbacks(pool);
2505
2506		if (!pool->pf.error_if_no_space && no_space_timeout)
2507			queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2508		break;
2509
2510	case PM_WRITE:
2511		if (old_mode != new_mode)
2512			notify_of_pool_mode_change(pool, "write");
2513		pool->out_of_data_space = false;
2514		pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2515		dm_pool_metadata_read_write(pool->pmd);
2516		pool->process_bio = process_bio;
2517		pool->process_discard = process_discard_bio;
2518		pool->process_cell = process_cell;
2519		pool->process_prepared_mapping = process_prepared_mapping;
2520		set_discard_callbacks(pool);
2521		break;
2522	}
2523
2524	pool->pf.mode = new_mode;
2525	/*
2526	 * The pool mode may have changed, sync it so bind_control_target()
2527	 * doesn't cause an unexpected mode transition on resume.
2528	 */
2529	pt->adjusted_pf.mode = new_mode;
2530}
2531
2532static void abort_transaction(struct pool *pool)
2533{
2534	const char *dev_name = dm_device_name(pool->pool_md);
2535
2536	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2537	if (dm_pool_abort_metadata(pool->pmd)) {
2538		DMERR("%s: failed to abort metadata transaction", dev_name);
2539		set_pool_mode(pool, PM_FAIL);
2540	}
2541
2542	if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2543		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2544		set_pool_mode(pool, PM_FAIL);
2545	}
2546}
2547
2548static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2549{
2550	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2551		    dm_device_name(pool->pool_md), op, r);
2552
2553	abort_transaction(pool);
2554	set_pool_mode(pool, PM_READ_ONLY);
2555}
2556
2557/*----------------------------------------------------------------*/
2558
2559/*
2560 * Mapping functions.
2561 */
2562
2563/*
2564 * Called only while mapping a thin bio to hand it over to the workqueue.
2565 */
2566static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2567{
2568	unsigned long flags;
2569	struct pool *pool = tc->pool;
2570
2571	spin_lock_irqsave(&tc->lock, flags);
2572	bio_list_add(&tc->deferred_bio_list, bio);
2573	spin_unlock_irqrestore(&tc->lock, flags);
2574
2575	wake_worker(pool);
2576}
2577
2578static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2579{
2580	struct pool *pool = tc->pool;
2581
2582	throttle_lock(&pool->throttle);
2583	thin_defer_bio(tc, bio);
2584	throttle_unlock(&pool->throttle);
2585}
2586
2587static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2588{
2589	unsigned long flags;
2590	struct pool *pool = tc->pool;
2591
2592	throttle_lock(&pool->throttle);
2593	spin_lock_irqsave(&tc->lock, flags);
2594	list_add_tail(&cell->user_list, &tc->deferred_cells);
2595	spin_unlock_irqrestore(&tc->lock, flags);
2596	throttle_unlock(&pool->throttle);
2597
2598	wake_worker(pool);
2599}
2600
2601static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2602{
2603	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2604
2605	h->tc = tc;
2606	h->shared_read_entry = NULL;
2607	h->all_io_entry = NULL;
2608	h->overwrite_mapping = NULL;
2609	h->cell = NULL;
2610}
2611
2612/*
2613 * Non-blocking function called from the thin target's map function.
2614 */
2615static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2616{
2617	int r;
2618	struct thin_c *tc = ti->private;
2619	dm_block_t block = get_bio_block(tc, bio);
2620	struct dm_thin_device *td = tc->td;
2621	struct dm_thin_lookup_result result;
2622	struct dm_bio_prison_cell *virt_cell, *data_cell;
2623	struct dm_cell_key key;
2624
2625	thin_hook_bio(tc, bio);
2626
2627	if (tc->requeue_mode) {
2628		bio->bi_error = DM_ENDIO_REQUEUE;
2629		bio_endio(bio);
2630		return DM_MAPIO_SUBMITTED;
2631	}
2632
2633	if (get_pool_mode(tc->pool) == PM_FAIL) {
2634		bio_io_error(bio);
2635		return DM_MAPIO_SUBMITTED;
2636	}
2637
2638	if (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA) ||
2639	    bio_op(bio) == REQ_OP_DISCARD) {
2640		thin_defer_bio_with_throttle(tc, bio);
2641		return DM_MAPIO_SUBMITTED;
2642	}
2643
2644	/*
2645	 * We must hold the virtual cell before doing the lookup, otherwise
2646	 * there's a race with discard.
2647	 */
2648	build_virtual_key(tc->td, block, &key);
2649	if (bio_detain(tc->pool, &key, bio, &virt_cell))
2650		return DM_MAPIO_SUBMITTED;
2651
2652	r = dm_thin_find_block(td, block, 0, &result);
2653
2654	/*
2655	 * Note that we defer readahead too.
2656	 */
2657	switch (r) {
2658	case 0:
2659		if (unlikely(result.shared)) {
2660			/*
2661			 * We have a race condition here between the
2662			 * result.shared value returned by the lookup and
2663			 * snapshot creation, which may cause new
2664			 * sharing.
2665			 *
2666			 * To avoid this always quiesce the origin before
2667			 * taking the snap.  You want to do this anyway to
2668			 * ensure a consistent application view
2669			 * (i.e. lockfs).
2670			 *
2671			 * More distant ancestors are irrelevant. The
2672			 * shared flag will be set in their case.
2673			 */
2674			thin_defer_cell(tc, virt_cell);
2675			return DM_MAPIO_SUBMITTED;
2676		}
2677
2678		build_data_key(tc->td, result.block, &key);
2679		if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2680			cell_defer_no_holder(tc, virt_cell);
2681			return DM_MAPIO_SUBMITTED;
2682		}
2683
2684		inc_all_io_entry(tc->pool, bio);
2685		cell_defer_no_holder(tc, data_cell);
2686		cell_defer_no_holder(tc, virt_cell);
2687
2688		remap(tc, bio, result.block);
2689		return DM_MAPIO_REMAPPED;
2690
2691	case -ENODATA:
2692	case -EWOULDBLOCK:
2693		thin_defer_cell(tc, virt_cell);
2694		return DM_MAPIO_SUBMITTED;
2695
2696	default:
2697		/*
2698		 * Must always call bio_io_error on failure.
2699		 * dm_thin_find_block can fail with -EINVAL if the
2700		 * pool is switched to fail-io mode.
2701		 */
2702		bio_io_error(bio);
2703		cell_defer_no_holder(tc, virt_cell);
2704		return DM_MAPIO_SUBMITTED;
2705	}
2706}
2707
2708static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2709{
2710	struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2711	struct request_queue *q;
2712
2713	if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2714		return 1;
2715
2716	q = bdev_get_queue(pt->data_dev->bdev);
2717	return bdi_congested(&q->backing_dev_info, bdi_bits);
2718}
2719
2720static void requeue_bios(struct pool *pool)
2721{
2722	unsigned long flags;
2723	struct thin_c *tc;
2724
2725	rcu_read_lock();
2726	list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2727		spin_lock_irqsave(&tc->lock, flags);
2728		bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2729		bio_list_init(&tc->retry_on_resume_list);
2730		spin_unlock_irqrestore(&tc->lock, flags);
2731	}
2732	rcu_read_unlock();
2733}
2734
2735/*----------------------------------------------------------------
2736 * Binding of control targets to a pool object
2737 *--------------------------------------------------------------*/
2738static bool data_dev_supports_discard(struct pool_c *pt)
2739{
2740	struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2741
2742	return q && blk_queue_discard(q);
2743}
2744
2745static bool is_factor(sector_t block_size, uint32_t n)
2746{
2747	return !sector_div(block_size, n);
2748}
2749
2750/*
2751 * If discard_passdown was enabled verify that the data device
2752 * supports discards.  Disable discard_passdown if not.
2753 */
2754static void disable_passdown_if_not_supported(struct pool_c *pt)
2755{
2756	struct pool *pool = pt->pool;
2757	struct block_device *data_bdev = pt->data_dev->bdev;
2758	struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2759	const char *reason = NULL;
2760	char buf[BDEVNAME_SIZE];
2761
2762	if (!pt->adjusted_pf.discard_passdown)
2763		return;
2764
2765	if (!data_dev_supports_discard(pt))
2766		reason = "discard unsupported";
2767
2768	else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2769		reason = "max discard sectors smaller than a block";
2770
2771	if (reason) {
2772		DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2773		pt->adjusted_pf.discard_passdown = false;
2774	}
2775}
2776
2777static int bind_control_target(struct pool *pool, struct dm_target *ti)
2778{
2779	struct pool_c *pt = ti->private;
2780
2781	/*
2782	 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2783	 */
2784	enum pool_mode old_mode = get_pool_mode(pool);
2785	enum pool_mode new_mode = pt->adjusted_pf.mode;
2786
2787	/*
2788	 * Don't change the pool's mode until set_pool_mode() below.
2789	 * Otherwise the pool's process_* function pointers may
2790	 * not match the desired pool mode.
2791	 */
2792	pt->adjusted_pf.mode = old_mode;
2793
2794	pool->ti = ti;
2795	pool->pf = pt->adjusted_pf;
2796	pool->low_water_blocks = pt->low_water_blocks;
2797
2798	set_pool_mode(pool, new_mode);
2799
2800	return 0;
2801}
2802
2803static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2804{
2805	if (pool->ti == ti)
2806		pool->ti = NULL;
2807}
2808
2809/*----------------------------------------------------------------
2810 * Pool creation
2811 *--------------------------------------------------------------*/
2812/* Initialize pool features. */
2813static void pool_features_init(struct pool_features *pf)
2814{
2815	pf->mode = PM_WRITE;
2816	pf->zero_new_blocks = true;
2817	pf->discard_enabled = true;
2818	pf->discard_passdown = true;
2819	pf->error_if_no_space = false;
2820}
2821
2822static void __pool_destroy(struct pool *pool)
2823{
2824	__pool_table_remove(pool);
2825
2826	vfree(pool->cell_sort_array);
2827	if (dm_pool_metadata_close(pool->pmd) < 0)
2828		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2829
2830	dm_bio_prison_destroy(pool->prison);
2831	dm_kcopyd_client_destroy(pool->copier);
2832
2833	if (pool->wq)
2834		destroy_workqueue(pool->wq);
2835
2836	if (pool->next_mapping)
2837		mempool_free(pool->next_mapping, pool->mapping_pool);
2838	mempool_destroy(pool->mapping_pool);
2839	dm_deferred_set_destroy(pool->shared_read_ds);
2840	dm_deferred_set_destroy(pool->all_io_ds);
2841	kfree(pool);
2842}
2843
2844static struct kmem_cache *_new_mapping_cache;
2845
2846static struct pool *pool_create(struct mapped_device *pool_md,
2847				struct block_device *metadata_dev,
2848				unsigned long block_size,
2849				int read_only, char **error)
2850{
2851	int r;
2852	void *err_p;
2853	struct pool *pool;
2854	struct dm_pool_metadata *pmd;
2855	bool format_device = read_only ? false : true;
2856
2857	pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2858	if (IS_ERR(pmd)) {
2859		*error = "Error creating metadata object";
2860		return (struct pool *)pmd;
2861	}
2862
2863	pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2864	if (!pool) {
2865		*error = "Error allocating memory for pool";
2866		err_p = ERR_PTR(-ENOMEM);
2867		goto bad_pool;
2868	}
2869
2870	pool->pmd = pmd;
2871	pool->sectors_per_block = block_size;
2872	if (block_size & (block_size - 1))
2873		pool->sectors_per_block_shift = -1;
2874	else
2875		pool->sectors_per_block_shift = __ffs(block_size);
2876	pool->low_water_blocks = 0;
2877	pool_features_init(&pool->pf);
2878	pool->prison = dm_bio_prison_create();
2879	if (!pool->prison) {
2880		*error = "Error creating pool's bio prison";
2881		err_p = ERR_PTR(-ENOMEM);
2882		goto bad_prison;
2883	}
2884
2885	pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2886	if (IS_ERR(pool->copier)) {
2887		r = PTR_ERR(pool->copier);
2888		*error = "Error creating pool's kcopyd client";
2889		err_p = ERR_PTR(r);
2890		goto bad_kcopyd_client;
2891	}
2892
2893	/*
2894	 * Create singlethreaded workqueue that will service all devices
2895	 * that use this metadata.
2896	 */
2897	pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2898	if (!pool->wq) {
2899		*error = "Error creating pool's workqueue";
2900		err_p = ERR_PTR(-ENOMEM);
2901		goto bad_wq;
2902	}
2903
2904	throttle_init(&pool->throttle);
2905	INIT_WORK(&pool->worker, do_worker);
2906	INIT_DELAYED_WORK(&pool->waker, do_waker);
2907	INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2908	spin_lock_init(&pool->lock);
2909	bio_list_init(&pool->deferred_flush_bios);
2910	INIT_LIST_HEAD(&pool->prepared_mappings);
2911	INIT_LIST_HEAD(&pool->prepared_discards);
2912	INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2913	INIT_LIST_HEAD(&pool->active_thins);
2914	pool->low_water_triggered = false;
2915	pool->suspended = true;
2916	pool->out_of_data_space = false;
2917
2918	pool->shared_read_ds = dm_deferred_set_create();
2919	if (!pool->shared_read_ds) {
2920		*error = "Error creating pool's shared read deferred set";
2921		err_p = ERR_PTR(-ENOMEM);
2922		goto bad_shared_read_ds;
2923	}
2924
2925	pool->all_io_ds = dm_deferred_set_create();
2926	if (!pool->all_io_ds) {
2927		*error = "Error creating pool's all io deferred set";
2928		err_p = ERR_PTR(-ENOMEM);
2929		goto bad_all_io_ds;
2930	}
2931
2932	pool->next_mapping = NULL;
2933	pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2934						      _new_mapping_cache);
2935	if (!pool->mapping_pool) {
2936		*error = "Error creating pool's mapping mempool";
2937		err_p = ERR_PTR(-ENOMEM);
2938		goto bad_mapping_pool;
2939	}
2940
2941	pool->cell_sort_array = vmalloc(sizeof(*pool->cell_sort_array) * CELL_SORT_ARRAY_SIZE);
2942	if (!pool->cell_sort_array) {
2943		*error = "Error allocating cell sort array";
2944		err_p = ERR_PTR(-ENOMEM);
2945		goto bad_sort_array;
2946	}
2947
2948	pool->ref_count = 1;
2949	pool->last_commit_jiffies = jiffies;
2950	pool->pool_md = pool_md;
2951	pool->md_dev = metadata_dev;
2952	__pool_table_insert(pool);
2953
2954	return pool;
2955
2956bad_sort_array:
2957	mempool_destroy(pool->mapping_pool);
2958bad_mapping_pool:
2959	dm_deferred_set_destroy(pool->all_io_ds);
2960bad_all_io_ds:
2961	dm_deferred_set_destroy(pool->shared_read_ds);
2962bad_shared_read_ds:
2963	destroy_workqueue(pool->wq);
2964bad_wq:
2965	dm_kcopyd_client_destroy(pool->copier);
2966bad_kcopyd_client:
2967	dm_bio_prison_destroy(pool->prison);
2968bad_prison:
2969	kfree(pool);
2970bad_pool:
2971	if (dm_pool_metadata_close(pmd))
2972		DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2973
2974	return err_p;
2975}
2976
2977static void __pool_inc(struct pool *pool)
2978{
2979	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2980	pool->ref_count++;
2981}
2982
2983static void __pool_dec(struct pool *pool)
2984{
2985	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2986	BUG_ON(!pool->ref_count);
2987	if (!--pool->ref_count)
2988		__pool_destroy(pool);
2989}
2990
2991static struct pool *__pool_find(struct mapped_device *pool_md,
2992				struct block_device *metadata_dev,
2993				unsigned long block_size, int read_only,
2994				char **error, int *created)
2995{
2996	struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2997
2998	if (pool) {
2999		if (pool->pool_md != pool_md) {
3000			*error = "metadata device already in use by a pool";
3001			return ERR_PTR(-EBUSY);
3002		}
3003		__pool_inc(pool);
3004
3005	} else {
3006		pool = __pool_table_lookup(pool_md);
3007		if (pool) {
3008			if (pool->md_dev != metadata_dev) {
3009				*error = "different pool cannot replace a pool";
3010				return ERR_PTR(-EINVAL);
3011			}
3012			__pool_inc(pool);
3013
3014		} else {
3015			pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
3016			*created = 1;
3017		}
3018	}
3019
3020	return pool;
3021}
3022
3023/*----------------------------------------------------------------
3024 * Pool target methods
3025 *--------------------------------------------------------------*/
3026static void pool_dtr(struct dm_target *ti)
3027{
3028	struct pool_c *pt = ti->private;
3029
3030	mutex_lock(&dm_thin_pool_table.mutex);
3031
3032	unbind_control_target(pt->pool, ti);
3033	__pool_dec(pt->pool);
3034	dm_put_device(ti, pt->metadata_dev);
3035	dm_put_device(ti, pt->data_dev);
3036	kfree(pt);
3037
3038	mutex_unlock(&dm_thin_pool_table.mutex);
3039}
3040
3041static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3042			       struct dm_target *ti)
3043{
3044	int r;
3045	unsigned argc;
3046	const char *arg_name;
3047
3048	static struct dm_arg _args[] = {
3049		{0, 4, "Invalid number of pool feature arguments"},
3050	};
3051
3052	/*
3053	 * No feature arguments supplied.
3054	 */
3055	if (!as->argc)
3056		return 0;
3057
3058	r = dm_read_arg_group(_args, as, &argc, &ti->error);
3059	if (r)
3060		return -EINVAL;
3061
3062	while (argc && !r) {
3063		arg_name = dm_shift_arg(as);
3064		argc--;
3065
3066		if (!strcasecmp(arg_name, "skip_block_zeroing"))
3067			pf->zero_new_blocks = false;
3068
3069		else if (!strcasecmp(arg_name, "ignore_discard"))
3070			pf->discard_enabled = false;
3071
3072		else if (!strcasecmp(arg_name, "no_discard_passdown"))
3073			pf->discard_passdown = false;
3074
3075		else if (!strcasecmp(arg_name, "read_only"))
3076			pf->mode = PM_READ_ONLY;
3077
3078		else if (!strcasecmp(arg_name, "error_if_no_space"))
3079			pf->error_if_no_space = true;
3080
3081		else {
3082			ti->error = "Unrecognised pool feature requested";
3083			r = -EINVAL;
3084			break;
3085		}
3086	}
3087
3088	return r;
3089}
3090
3091static void metadata_low_callback(void *context)
3092{
3093	struct pool *pool = context;
3094
3095	DMWARN("%s: reached low water mark for metadata device: sending event.",
3096	       dm_device_name(pool->pool_md));
3097
3098	dm_table_event(pool->ti->table);
3099}
3100
3101static sector_t get_dev_size(struct block_device *bdev)
3102{
3103	return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3104}
3105
3106static void warn_if_metadata_device_too_big(struct block_device *bdev)
3107{
3108	sector_t metadata_dev_size = get_dev_size(bdev);
3109	char buffer[BDEVNAME_SIZE];
3110
3111	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3112		DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3113		       bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3114}
3115
3116static sector_t get_metadata_dev_size(struct block_device *bdev)
3117{
3118	sector_t metadata_dev_size = get_dev_size(bdev);
3119
3120	if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3121		metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3122
3123	return metadata_dev_size;
3124}
3125
3126static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3127{
3128	sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3129
3130	sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3131
3132	return metadata_dev_size;
3133}
3134
3135/*
3136 * When a metadata threshold is crossed a dm event is triggered, and
3137 * userland should respond by growing the metadata device.  We could let
3138 * userland set the threshold, like we do with the data threshold, but I'm
3139 * not sure they know enough to do this well.
3140 */
3141static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3142{
3143	/*
3144	 * 4M is ample for all ops with the possible exception of thin
3145	 * device deletion which is harmless if it fails (just retry the
3146	 * delete after you've grown the device).
3147	 */
3148	dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3149	return min((dm_block_t)1024ULL /* 4M */, quarter);
3150}
3151
3152/*
3153 * thin-pool <metadata dev> <data dev>
3154 *	     <data block size (sectors)>
3155 *	     <low water mark (blocks)>
3156 *	     [<#feature args> [<arg>]*]
3157 *
3158 * Optional feature arguments are:
3159 *	     skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3160 *	     ignore_discard: disable discard
3161 *	     no_discard_passdown: don't pass discards down to the data device
3162 *	     read_only: Don't allow any changes to be made to the pool metadata.
3163 *	     error_if_no_space: error IOs, instead of queueing, if no space.
3164 */
3165static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3166{
3167	int r, pool_created = 0;
3168	struct pool_c *pt;
3169	struct pool *pool;
3170	struct pool_features pf;
3171	struct dm_arg_set as;
3172	struct dm_dev *data_dev;
3173	unsigned long block_size;
3174	dm_block_t low_water_blocks;
3175	struct dm_dev *metadata_dev;
3176	fmode_t metadata_mode;
3177
3178	/*
3179	 * FIXME Remove validation from scope of lock.
3180	 */
3181	mutex_lock(&dm_thin_pool_table.mutex);
3182
3183	if (argc < 4) {
3184		ti->error = "Invalid argument count";
3185		r = -EINVAL;
3186		goto out_unlock;
3187	}
3188
3189	as.argc = argc;
3190	as.argv = argv;
3191
3192	/*
3193	 * Set default pool features.
3194	 */
3195	pool_features_init(&pf);
3196
3197	dm_consume_args(&as, 4);
3198	r = parse_pool_features(&as, &pf, ti);
3199	if (r)
3200		goto out_unlock;
3201
3202	metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3203	r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3204	if (r) {
3205		ti->error = "Error opening metadata block device";
3206		goto out_unlock;
3207	}
3208	warn_if_metadata_device_too_big(metadata_dev->bdev);
3209
3210	r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3211	if (r) {
3212		ti->error = "Error getting data device";
3213		goto out_metadata;
3214	}
3215
3216	if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3217	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3218	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3219	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3220		ti->error = "Invalid block size";
3221		r = -EINVAL;
3222		goto out;
3223	}
3224
3225	if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3226		ti->error = "Invalid low water mark";
3227		r = -EINVAL;
3228		goto out;
3229	}
3230
3231	pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3232	if (!pt) {
3233		r = -ENOMEM;
3234		goto out;
3235	}
3236
3237	pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3238			   block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3239	if (IS_ERR(pool)) {
3240		r = PTR_ERR(pool);
3241		goto out_free_pt;
3242	}
3243
3244	/*
3245	 * 'pool_created' reflects whether this is the first table load.
3246	 * Top level discard support is not allowed to be changed after
3247	 * initial load.  This would require a pool reload to trigger thin
3248	 * device changes.
3249	 */
3250	if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3251		ti->error = "Discard support cannot be disabled once enabled";
3252		r = -EINVAL;
3253		goto out_flags_changed;
3254	}
3255
3256	pt->pool = pool;
3257	pt->ti = ti;
3258	pt->metadata_dev = metadata_dev;
3259	pt->data_dev = data_dev;
3260	pt->low_water_blocks = low_water_blocks;
3261	pt->adjusted_pf = pt->requested_pf = pf;
3262	ti->num_flush_bios = 1;
3263
3264	/*
3265	 * Only need to enable discards if the pool should pass
3266	 * them down to the data device.  The thin device's discard
3267	 * processing will cause mappings to be removed from the btree.
3268	 */
3269	ti->discard_zeroes_data_unsupported = true;
3270	if (pf.discard_enabled && pf.discard_passdown) {
3271		ti->num_discard_bios = 1;
3272
3273		/*
3274		 * Setting 'discards_supported' circumvents the normal
3275		 * stacking of discard limits (this keeps the pool and
3276		 * thin devices' discard limits consistent).
3277		 */
3278		ti->discards_supported = true;
3279	}
3280	ti->private = pt;
3281
3282	r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3283						calc_metadata_threshold(pt),
3284						metadata_low_callback,
3285						pool);
3286	if (r)
3287		goto out_flags_changed;
3288
3289	pt->callbacks.congested_fn = pool_is_congested;
3290	dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3291
3292	mutex_unlock(&dm_thin_pool_table.mutex);
3293
3294	return 0;
3295
3296out_flags_changed:
3297	__pool_dec(pool);
3298out_free_pt:
3299	kfree(pt);
3300out:
3301	dm_put_device(ti, data_dev);
3302out_metadata:
3303	dm_put_device(ti, metadata_dev);
3304out_unlock:
3305	mutex_unlock(&dm_thin_pool_table.mutex);
3306
3307	return r;
3308}
3309
3310static int pool_map(struct dm_target *ti, struct bio *bio)
3311{
3312	int r;
3313	struct pool_c *pt = ti->private;
3314	struct pool *pool = pt->pool;
3315	unsigned long flags;
3316
3317	/*
3318	 * As this is a singleton target, ti->begin is always zero.
3319	 */
3320	spin_lock_irqsave(&pool->lock, flags);
3321	bio->bi_bdev = pt->data_dev->bdev;
3322	r = DM_MAPIO_REMAPPED;
3323	spin_unlock_irqrestore(&pool->lock, flags);
3324
3325	return r;
3326}
3327
3328static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3329{
3330	int r;
3331	struct pool_c *pt = ti->private;
3332	struct pool *pool = pt->pool;
3333	sector_t data_size = ti->len;
3334	dm_block_t sb_data_size;
3335
3336	*need_commit = false;
3337
3338	(void) sector_div(data_size, pool->sectors_per_block);
3339
3340	r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3341	if (r) {
3342		DMERR("%s: failed to retrieve data device size",
3343		      dm_device_name(pool->pool_md));
3344		return r;
3345	}
3346
3347	if (data_size < sb_data_size) {
3348		DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3349		      dm_device_name(pool->pool_md),
3350		      (unsigned long long)data_size, sb_data_size);
3351		return -EINVAL;
3352
3353	} else if (data_size > sb_data_size) {
3354		if (dm_pool_metadata_needs_check(pool->pmd)) {
3355			DMERR("%s: unable to grow the data device until repaired.",
3356			      dm_device_name(pool->pool_md));
3357			return 0;
3358		}
3359
3360		if (sb_data_size)
3361			DMINFO("%s: growing the data device from %llu to %llu blocks",
3362			       dm_device_name(pool->pool_md),
3363			       sb_data_size, (unsigned long long)data_size);
3364		r = dm_pool_resize_data_dev(pool->pmd, data_size);
3365		if (r) {
3366			metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3367			return r;
3368		}
3369
3370		*need_commit = true;
3371	}
3372
3373	return 0;
3374}
3375
3376static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3377{
3378	int r;
3379	struct pool_c *pt = ti->private;
3380	struct pool *pool = pt->pool;
3381	dm_block_t metadata_dev_size, sb_metadata_dev_size;
3382
3383	*need_commit = false;
3384
3385	metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3386
3387	r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3388	if (r) {
3389		DMERR("%s: failed to retrieve metadata device size",
3390		      dm_device_name(pool->pool_md));
3391		return r;
3392	}
3393
3394	if (metadata_dev_size < sb_metadata_dev_size) {
3395		DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3396		      dm_device_name(pool->pool_md),
3397		      metadata_dev_size, sb_metadata_dev_size);
3398		return -EINVAL;
3399
3400	} else if (metadata_dev_size > sb_metadata_dev_size) {
3401		if (dm_pool_metadata_needs_check(pool->pmd)) {
3402			DMERR("%s: unable to grow the metadata device until repaired.",
3403			      dm_device_name(pool->pool_md));
3404			return 0;
3405		}
3406
3407		warn_if_metadata_device_too_big(pool->md_dev);
3408		DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3409		       dm_device_name(pool->pool_md),
3410		       sb_metadata_dev_size, metadata_dev_size);
3411		r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3412		if (r) {
3413			metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3414			return r;
3415		}
3416
3417		*need_commit = true;
3418	}
3419
3420	return 0;
3421}
3422
3423/*
3424 * Retrieves the number of blocks of the data device from
3425 * the superblock and compares it to the actual device size,
3426 * thus resizing the data device in case it has grown.
3427 *
3428 * This both copes with opening preallocated data devices in the ctr
3429 * being followed by a resume
3430 * -and-
3431 * calling the resume method individually after userspace has
3432 * grown the data device in reaction to a table event.
3433 */
3434static int pool_preresume(struct dm_target *ti)
3435{
3436	int r;
3437	bool need_commit1, need_commit2;
3438	struct pool_c *pt = ti->private;
3439	struct pool *pool = pt->pool;
3440
3441	/*
3442	 * Take control of the pool object.
3443	 */
3444	r = bind_control_target(pool, ti);
3445	if (r)
3446		return r;
3447
3448	r = maybe_resize_data_dev(ti, &need_commit1);
3449	if (r)
3450		return r;
3451
3452	r = maybe_resize_metadata_dev(ti, &need_commit2);
3453	if (r)
3454		return r;
3455
3456	if (need_commit1 || need_commit2)
3457		(void) commit(pool);
3458
3459	return 0;
3460}
3461
3462static void pool_suspend_active_thins(struct pool *pool)
3463{
3464	struct thin_c *tc;
3465
3466	/* Suspend all active thin devices */
3467	tc = get_first_thin(pool);
3468	while (tc) {
3469		dm_internal_suspend_noflush(tc->thin_md);
3470		tc = get_next_thin(pool, tc);
3471	}
3472}
3473
3474static void pool_resume_active_thins(struct pool *pool)
3475{
3476	struct thin_c *tc;
3477
3478	/* Resume all active thin devices */
3479	tc = get_first_thin(pool);
3480	while (tc) {
3481		dm_internal_resume(tc->thin_md);
3482		tc = get_next_thin(pool, tc);
3483	}
3484}
3485
3486static void pool_resume(struct dm_target *ti)
3487{
3488	struct pool_c *pt = ti->private;
3489	struct pool *pool = pt->pool;
3490	unsigned long flags;
3491
3492	/*
3493	 * Must requeue active_thins' bios and then resume
3494	 * active_thins _before_ clearing 'suspend' flag.
3495	 */
3496	requeue_bios(pool);
3497	pool_resume_active_thins(pool);
3498
3499	spin_lock_irqsave(&pool->lock, flags);
3500	pool->low_water_triggered = false;
3501	pool->suspended = false;
3502	spin_unlock_irqrestore(&pool->lock, flags);
3503
3504	do_waker(&pool->waker.work);
3505}
3506
3507static void pool_presuspend(struct dm_target *ti)
3508{
3509	struct pool_c *pt = ti->private;
3510	struct pool *pool = pt->pool;
3511	unsigned long flags;
3512
3513	spin_lock_irqsave(&pool->lock, flags);
3514	pool->suspended = true;
3515	spin_unlock_irqrestore(&pool->lock, flags);
3516
3517	pool_suspend_active_thins(pool);
3518}
3519
3520static void pool_presuspend_undo(struct dm_target *ti)
3521{
3522	struct pool_c *pt = ti->private;
3523	struct pool *pool = pt->pool;
3524	unsigned long flags;
3525
3526	pool_resume_active_thins(pool);
3527
3528	spin_lock_irqsave(&pool->lock, flags);
3529	pool->suspended = false;
3530	spin_unlock_irqrestore(&pool->lock, flags);
3531}
3532
3533static void pool_postsuspend(struct dm_target *ti)
3534{
3535	struct pool_c *pt = ti->private;
3536	struct pool *pool = pt->pool;
3537
3538	cancel_delayed_work_sync(&pool->waker);
3539	cancel_delayed_work_sync(&pool->no_space_timeout);
3540	flush_workqueue(pool->wq);
3541	(void) commit(pool);
3542}
3543
3544static int check_arg_count(unsigned argc, unsigned args_required)
3545{
3546	if (argc != args_required) {
3547		DMWARN("Message received with %u arguments instead of %u.",
3548		       argc, args_required);
3549		return -EINVAL;
3550	}
3551
3552	return 0;
3553}
3554
3555static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3556{
3557	if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3558	    *dev_id <= MAX_DEV_ID)
3559		return 0;
3560
3561	if (warning)
3562		DMWARN("Message received with invalid device id: %s", arg);
3563
3564	return -EINVAL;
3565}
3566
3567static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3568{
3569	dm_thin_id dev_id;
3570	int r;
3571
3572	r = check_arg_count(argc, 2);
3573	if (r)
3574		return r;
3575
3576	r = read_dev_id(argv[1], &dev_id, 1);
3577	if (r)
3578		return r;
3579
3580	r = dm_pool_create_thin(pool->pmd, dev_id);
3581	if (r) {
3582		DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3583		       argv[1]);
3584		return r;
3585	}
3586
3587	return 0;
3588}
3589
3590static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3591{
3592	dm_thin_id dev_id;
3593	dm_thin_id origin_dev_id;
3594	int r;
3595
3596	r = check_arg_count(argc, 3);
3597	if (r)
3598		return r;
3599
3600	r = read_dev_id(argv[1], &dev_id, 1);
3601	if (r)
3602		return r;
3603
3604	r = read_dev_id(argv[2], &origin_dev_id, 1);
3605	if (r)
3606		return r;
3607
3608	r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3609	if (r) {
3610		DMWARN("Creation of new snapshot %s of device %s failed.",
3611		       argv[1], argv[2]);
3612		return r;
3613	}
3614
3615	return 0;
3616}
3617
3618static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3619{
3620	dm_thin_id dev_id;
3621	int r;
3622
3623	r = check_arg_count(argc, 2);
3624	if (r)
3625		return r;
3626
3627	r = read_dev_id(argv[1], &dev_id, 1);
3628	if (r)
3629		return r;
3630
3631	r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3632	if (r)
3633		DMWARN("Deletion of thin device %s failed.", argv[1]);
3634
3635	return r;
3636}
3637
3638static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3639{
3640	dm_thin_id old_id, new_id;
3641	int r;
3642
3643	r = check_arg_count(argc, 3);
3644	if (r)
3645		return r;
3646
3647	if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3648		DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3649		return -EINVAL;
3650	}
3651
3652	if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3653		DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3654		return -EINVAL;
3655	}
3656
3657	r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3658	if (r) {
3659		DMWARN("Failed to change transaction id from %s to %s.",
3660		       argv[1], argv[2]);
3661		return r;
3662	}
3663
3664	return 0;
3665}
3666
3667static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3668{
3669	int r;
3670
3671	r = check_arg_count(argc, 1);
3672	if (r)
3673		return r;
3674
3675	(void) commit(pool);
3676
3677	r = dm_pool_reserve_metadata_snap(pool->pmd);
3678	if (r)
3679		DMWARN("reserve_metadata_snap message failed.");
3680
3681	return r;
3682}
3683
3684static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3685{
3686	int r;
3687
3688	r = check_arg_count(argc, 1);
3689	if (r)
3690		return r;
3691
3692	r = dm_pool_release_metadata_snap(pool->pmd);
3693	if (r)
3694		DMWARN("release_metadata_snap message failed.");
3695
3696	return r;
3697}
3698
3699/*
3700 * Messages supported:
3701 *   create_thin	<dev_id>
3702 *   create_snap	<dev_id> <origin_id>
3703 *   delete		<dev_id>
3704 *   set_transaction_id <current_trans_id> <new_trans_id>
3705 *   reserve_metadata_snap
3706 *   release_metadata_snap
3707 */
3708static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3709{
3710	int r = -EINVAL;
3711	struct pool_c *pt = ti->private;
3712	struct pool *pool = pt->pool;
3713
3714	if (get_pool_mode(pool) >= PM_READ_ONLY) {
3715		DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3716		      dm_device_name(pool->pool_md));
3717		return -EOPNOTSUPP;
3718	}
3719
3720	if (!strcasecmp(argv[0], "create_thin"))
3721		r = process_create_thin_mesg(argc, argv, pool);
3722
3723	else if (!strcasecmp(argv[0], "create_snap"))
3724		r = process_create_snap_mesg(argc, argv, pool);
3725
3726	else if (!strcasecmp(argv[0], "delete"))
3727		r = process_delete_mesg(argc, argv, pool);
3728
3729	else if (!strcasecmp(argv[0], "set_transaction_id"))
3730		r = process_set_transaction_id_mesg(argc, argv, pool);
3731
3732	else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3733		r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3734
3735	else if (!strcasecmp(argv[0], "release_metadata_snap"))
3736		r = process_release_metadata_snap_mesg(argc, argv, pool);
3737
3738	else
3739		DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3740
3741	if (!r)
3742		(void) commit(pool);
3743
3744	return r;
3745}
3746
3747static void emit_flags(struct pool_features *pf, char *result,
3748		       unsigned sz, unsigned maxlen)
3749{
3750	unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3751		!pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3752		pf->error_if_no_space;
3753	DMEMIT("%u ", count);
3754
3755	if (!pf->zero_new_blocks)
3756		DMEMIT("skip_block_zeroing ");
3757
3758	if (!pf->discard_enabled)
3759		DMEMIT("ignore_discard ");
3760
3761	if (!pf->discard_passdown)
3762		DMEMIT("no_discard_passdown ");
3763
3764	if (pf->mode == PM_READ_ONLY)
3765		DMEMIT("read_only ");
3766
3767	if (pf->error_if_no_space)
3768		DMEMIT("error_if_no_space ");
3769}
3770
3771/*
3772 * Status line is:
3773 *    <transaction id> <used metadata sectors>/<total metadata sectors>
3774 *    <used data sectors>/<total data sectors> <held metadata root>
3775 *    <pool mode> <discard config> <no space config> <needs_check>
3776 */
3777static void pool_status(struct dm_target *ti, status_type_t type,
3778			unsigned status_flags, char *result, unsigned maxlen)
3779{
3780	int r;
3781	unsigned sz = 0;
3782	uint64_t transaction_id;
3783	dm_block_t nr_free_blocks_data;
3784	dm_block_t nr_free_blocks_metadata;
3785	dm_block_t nr_blocks_data;
3786	dm_block_t nr_blocks_metadata;
3787	dm_block_t held_root;
3788	char buf[BDEVNAME_SIZE];
3789	char buf2[BDEVNAME_SIZE];
3790	struct pool_c *pt = ti->private;
3791	struct pool *pool = pt->pool;
3792
3793	switch (type) {
3794	case STATUSTYPE_INFO:
3795		if (get_pool_mode(pool) == PM_FAIL) {
3796			DMEMIT("Fail");
3797			break;
3798		}
3799
3800		/* Commit to ensure statistics aren't out-of-date */
3801		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3802			(void) commit(pool);
3803
3804		r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3805		if (r) {
3806			DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3807			      dm_device_name(pool->pool_md), r);
3808			goto err;
3809		}
3810
3811		r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3812		if (r) {
3813			DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3814			      dm_device_name(pool->pool_md), r);
3815			goto err;
3816		}
3817
3818		r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3819		if (r) {
3820			DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3821			      dm_device_name(pool->pool_md), r);
3822			goto err;
3823		}
3824
3825		r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3826		if (r) {
3827			DMERR("%s: dm_pool_get_free_block_count returned %d",
3828			      dm_device_name(pool->pool_md), r);
3829			goto err;
3830		}
3831
3832		r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3833		if (r) {
3834			DMERR("%s: dm_pool_get_data_dev_size returned %d",
3835			      dm_device_name(pool->pool_md), r);
3836			goto err;
3837		}
3838
3839		r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3840		if (r) {
3841			DMERR("%s: dm_pool_get_metadata_snap returned %d",
3842			      dm_device_name(pool->pool_md), r);
3843			goto err;
3844		}
3845
3846		DMEMIT("%llu %llu/%llu %llu/%llu ",
3847		       (unsigned long long)transaction_id,
3848		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3849		       (unsigned long long)nr_blocks_metadata,
3850		       (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3851		       (unsigned long long)nr_blocks_data);
3852
3853		if (held_root)
3854			DMEMIT("%llu ", held_root);
3855		else
3856			DMEMIT("- ");
3857
3858		if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3859			DMEMIT("out_of_data_space ");
3860		else if (pool->pf.mode == PM_READ_ONLY)
3861			DMEMIT("ro ");
3862		else
3863			DMEMIT("rw ");
3864
3865		if (!pool->pf.discard_enabled)
3866			DMEMIT("ignore_discard ");
3867		else if (pool->pf.discard_passdown)
3868			DMEMIT("discard_passdown ");
3869		else
3870			DMEMIT("no_discard_passdown ");
3871
3872		if (pool->pf.error_if_no_space)
3873			DMEMIT("error_if_no_space ");
3874		else
3875			DMEMIT("queue_if_no_space ");
3876
3877		if (dm_pool_metadata_needs_check(pool->pmd))
3878			DMEMIT("needs_check ");
3879		else
3880			DMEMIT("- ");
3881
3882		break;
3883
3884	case STATUSTYPE_TABLE:
3885		DMEMIT("%s %s %lu %llu ",
3886		       format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3887		       format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3888		       (unsigned long)pool->sectors_per_block,
3889		       (unsigned long long)pt->low_water_blocks);
3890		emit_flags(&pt->requested_pf, result, sz, maxlen);
3891		break;
3892	}
3893	return;
3894
3895err:
3896	DMEMIT("Error");
3897}
3898
3899static int pool_iterate_devices(struct dm_target *ti,
3900				iterate_devices_callout_fn fn, void *data)
3901{
3902	struct pool_c *pt = ti->private;
3903
3904	return fn(ti, pt->data_dev, 0, ti->len, data);
3905}
3906
3907static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3908{
3909	struct pool_c *pt = ti->private;
3910	struct pool *pool = pt->pool;
3911	sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3912
3913	/*
3914	 * If max_sectors is smaller than pool->sectors_per_block adjust it
3915	 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3916	 * This is especially beneficial when the pool's data device is a RAID
3917	 * device that has a full stripe width that matches pool->sectors_per_block
3918	 * -- because even though partial RAID stripe-sized IOs will be issued to a
3919	 *    single RAID stripe; when aggregated they will end on a full RAID stripe
3920	 *    boundary.. which avoids additional partial RAID stripe writes cascading
3921	 */
3922	if (limits->max_sectors < pool->sectors_per_block) {
3923		while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3924			if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3925				limits->max_sectors--;
3926			limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3927		}
3928	}
3929
3930	/*
3931	 * If the system-determined stacked limits are compatible with the
3932	 * pool's blocksize (io_opt is a factor) do not override them.
3933	 */
3934	if (io_opt_sectors < pool->sectors_per_block ||
3935	    !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3936		if (is_factor(pool->sectors_per_block, limits->max_sectors))
3937			blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3938		else
3939			blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3940		blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3941	}
3942
3943	/*
3944	 * pt->adjusted_pf is a staging area for the actual features to use.
3945	 * They get transferred to the live pool in bind_control_target()
3946	 * called from pool_preresume().
3947	 */
3948	if (!pt->adjusted_pf.discard_enabled) {
3949		/*
3950		 * Must explicitly disallow stacking discard limits otherwise the
3951		 * block layer will stack them if pool's data device has support.
3952		 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3953		 * user to see that, so make sure to set all discard limits to 0.
3954		 */
3955		limits->discard_granularity = 0;
3956		return;
3957	}
3958
3959	disable_passdown_if_not_supported(pt);
3960
3961	/*
3962	 * The pool uses the same discard limits as the underlying data
3963	 * device.  DM core has already set this up.
3964	 */
3965}
3966
3967static struct target_type pool_target = {
3968	.name = "thin-pool",
3969	.features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3970		    DM_TARGET_IMMUTABLE,
3971	.version = {1, 19, 0},
3972	.module = THIS_MODULE,
3973	.ctr = pool_ctr,
3974	.dtr = pool_dtr,
3975	.map = pool_map,
3976	.presuspend = pool_presuspend,
3977	.presuspend_undo = pool_presuspend_undo,
3978	.postsuspend = pool_postsuspend,
3979	.preresume = pool_preresume,
3980	.resume = pool_resume,
3981	.message = pool_message,
3982	.status = pool_status,
3983	.iterate_devices = pool_iterate_devices,
3984	.io_hints = pool_io_hints,
3985};
3986
3987/*----------------------------------------------------------------
3988 * Thin target methods
3989 *--------------------------------------------------------------*/
3990static void thin_get(struct thin_c *tc)
3991{
3992	atomic_inc(&tc->refcount);
3993}
3994
3995static void thin_put(struct thin_c *tc)
3996{
3997	if (atomic_dec_and_test(&tc->refcount))
3998		complete(&tc->can_destroy);
3999}
4000
4001static void thin_dtr(struct dm_target *ti)
4002{
4003	struct thin_c *tc = ti->private;
4004	unsigned long flags;
4005
4006	spin_lock_irqsave(&tc->pool->lock, flags);
4007	list_del_rcu(&tc->list);
4008	spin_unlock_irqrestore(&tc->pool->lock, flags);
4009	synchronize_rcu();
4010
4011	thin_put(tc);
4012	wait_for_completion(&tc->can_destroy);
4013
4014	mutex_lock(&dm_thin_pool_table.mutex);
4015
4016	__pool_dec(tc->pool);
4017	dm_pool_close_thin_device(tc->td);
4018	dm_put_device(ti, tc->pool_dev);
4019	if (tc->origin_dev)
4020		dm_put_device(ti, tc->origin_dev);
4021	kfree(tc);
4022
4023	mutex_unlock(&dm_thin_pool_table.mutex);
4024}
4025
4026/*
4027 * Thin target parameters:
4028 *
4029 * <pool_dev> <dev_id> [origin_dev]
4030 *
4031 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4032 * dev_id: the internal device identifier
4033 * origin_dev: a device external to the pool that should act as the origin
4034 *
4035 * If the pool device has discards disabled, they get disabled for the thin
4036 * device as well.
4037 */
4038static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4039{
4040	int r;
4041	struct thin_c *tc;
4042	struct dm_dev *pool_dev, *origin_dev;
4043	struct mapped_device *pool_md;
4044	unsigned long flags;
4045
4046	mutex_lock(&dm_thin_pool_table.mutex);
4047
4048	if (argc != 2 && argc != 3) {
4049		ti->error = "Invalid argument count";
4050		r = -EINVAL;
4051		goto out_unlock;
4052	}
4053
4054	tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4055	if (!tc) {
4056		ti->error = "Out of memory";
4057		r = -ENOMEM;
4058		goto out_unlock;
4059	}
4060	tc->thin_md = dm_table_get_md(ti->table);
4061	spin_lock_init(&tc->lock);
4062	INIT_LIST_HEAD(&tc->deferred_cells);
4063	bio_list_init(&tc->deferred_bio_list);
4064	bio_list_init(&tc->retry_on_resume_list);
4065	tc->sort_bio_list = RB_ROOT;
4066
4067	if (argc == 3) {
4068		r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4069		if (r) {
4070			ti->error = "Error opening origin device";
4071			goto bad_origin_dev;
4072		}
4073		tc->origin_dev = origin_dev;
4074	}
4075
4076	r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4077	if (r) {
4078		ti->error = "Error opening pool device";
4079		goto bad_pool_dev;
4080	}
4081	tc->pool_dev = pool_dev;
4082
4083	if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4084		ti->error = "Invalid device id";
4085		r = -EINVAL;
4086		goto bad_common;
4087	}
4088
4089	pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4090	if (!pool_md) {
4091		ti->error = "Couldn't get pool mapped device";
4092		r = -EINVAL;
4093		goto bad_common;
4094	}
4095
4096	tc->pool = __pool_table_lookup(pool_md);
4097	if (!tc->pool) {
4098		ti->error = "Couldn't find pool object";
4099		r = -EINVAL;
4100		goto bad_pool_lookup;
4101	}
4102	__pool_inc(tc->pool);
4103
4104	if (get_pool_mode(tc->pool) == PM_FAIL) {
4105		ti->error = "Couldn't open thin device, Pool is in fail mode";
4106		r = -EINVAL;
4107		goto bad_pool;
4108	}
4109
4110	r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4111	if (r) {
4112		ti->error = "Couldn't open thin internal device";
4113		goto bad_pool;
4114	}
4115
4116	r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4117	if (r)
4118		goto bad;
4119
4120	ti->num_flush_bios = 1;
4121	ti->flush_supported = true;
4122	ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4123
4124	/* In case the pool supports discards, pass them on. */
4125	ti->discard_zeroes_data_unsupported = true;
4126	if (tc->pool->pf.discard_enabled) {
4127		ti->discards_supported = true;
4128		ti->num_discard_bios = 1;
4129		ti->split_discard_bios = false;
4130	}
4131
4132	mutex_unlock(&dm_thin_pool_table.mutex);
4133
4134	spin_lock_irqsave(&tc->pool->lock, flags);
4135	if (tc->pool->suspended) {
4136		spin_unlock_irqrestore(&tc->pool->lock, flags);
4137		mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4138		ti->error = "Unable to activate thin device while pool is suspended";
4139		r = -EINVAL;
4140		goto bad;
4141	}
4142	atomic_set(&tc->refcount, 1);
4143	init_completion(&tc->can_destroy);
4144	list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4145	spin_unlock_irqrestore(&tc->pool->lock, flags);
4146	/*
4147	 * This synchronize_rcu() call is needed here otherwise we risk a
4148	 * wake_worker() call finding no bios to process (because the newly
4149	 * added tc isn't yet visible).  So this reduces latency since we
4150	 * aren't then dependent on the periodic commit to wake_worker().
4151	 */
4152	synchronize_rcu();
4153
4154	dm_put(pool_md);
4155
4156	return 0;
4157
4158bad:
4159	dm_pool_close_thin_device(tc->td);
4160bad_pool:
4161	__pool_dec(tc->pool);
4162bad_pool_lookup:
4163	dm_put(pool_md);
4164bad_common:
4165	dm_put_device(ti, tc->pool_dev);
4166bad_pool_dev:
4167	if (tc->origin_dev)
4168		dm_put_device(ti, tc->origin_dev);
4169bad_origin_dev:
4170	kfree(tc);
4171out_unlock:
4172	mutex_unlock(&dm_thin_pool_table.mutex);
4173
4174	return r;
4175}
4176
4177static int thin_map(struct dm_target *ti, struct bio *bio)
4178{
4179	bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4180
4181	return thin_bio_map(ti, bio);
4182}
4183
4184static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
4185{
4186	unsigned long flags;
4187	struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4188	struct list_head work;
4189	struct dm_thin_new_mapping *m, *tmp;
4190	struct pool *pool = h->tc->pool;
4191
4192	if (h->shared_read_entry) {
4193		INIT_LIST_HEAD(&work);
4194		dm_deferred_entry_dec(h->shared_read_entry, &work);
4195
4196		spin_lock_irqsave(&pool->lock, flags);
4197		list_for_each_entry_safe(m, tmp, &work, list) {
4198			list_del(&m->list);
4199			__complete_mapping_preparation(m);
4200		}
4201		spin_unlock_irqrestore(&pool->lock, flags);
4202	}
4203
4204	if (h->all_io_entry) {
4205		INIT_LIST_HEAD(&work);
4206		dm_deferred_entry_dec(h->all_io_entry, &work);
4207		if (!list_empty(&work)) {
4208			spin_lock_irqsave(&pool->lock, flags);
4209			list_for_each_entry_safe(m, tmp, &work, list)
4210				list_add_tail(&m->list, &pool->prepared_discards);
4211			spin_unlock_irqrestore(&pool->lock, flags);
4212			wake_worker(pool);
4213		}
4214	}
4215
4216	if (h->cell)
4217		cell_defer_no_holder(h->tc, h->cell);
4218
4219	return 0;
4220}
4221
4222static void thin_presuspend(struct dm_target *ti)
4223{
4224	struct thin_c *tc = ti->private;
4225
4226	if (dm_noflush_suspending(ti))
4227		noflush_work(tc, do_noflush_start);
4228}
4229
4230static void thin_postsuspend(struct dm_target *ti)
4231{
4232	struct thin_c *tc = ti->private;
4233
4234	/*
4235	 * The dm_noflush_suspending flag has been cleared by now, so
4236	 * unfortunately we must always run this.
4237	 */
4238	noflush_work(tc, do_noflush_stop);
4239}
4240
4241static int thin_preresume(struct dm_target *ti)
4242{
4243	struct thin_c *tc = ti->private;
4244
4245	if (tc->origin_dev)
4246		tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4247
4248	return 0;
4249}
4250
4251/*
4252 * <nr mapped sectors> <highest mapped sector>
4253 */
4254static void thin_status(struct dm_target *ti, status_type_t type,
4255			unsigned status_flags, char *result, unsigned maxlen)
4256{
4257	int r;
4258	ssize_t sz = 0;
4259	dm_block_t mapped, highest;
4260	char buf[BDEVNAME_SIZE];
4261	struct thin_c *tc = ti->private;
4262
4263	if (get_pool_mode(tc->pool) == PM_FAIL) {
4264		DMEMIT("Fail");
4265		return;
4266	}
4267
4268	if (!tc->td)
4269		DMEMIT("-");
4270	else {
4271		switch (type) {
4272		case STATUSTYPE_INFO:
4273			r = dm_thin_get_mapped_count(tc->td, &mapped);
4274			if (r) {
4275				DMERR("dm_thin_get_mapped_count returned %d", r);
4276				goto err;
4277			}
4278
4279			r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4280			if (r < 0) {
4281				DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4282				goto err;
4283			}
4284
4285			DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4286			if (r)
4287				DMEMIT("%llu", ((highest + 1) *
4288						tc->pool->sectors_per_block) - 1);
4289			else
4290				DMEMIT("-");
4291			break;
4292
4293		case STATUSTYPE_TABLE:
4294			DMEMIT("%s %lu",
4295			       format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4296			       (unsigned long) tc->dev_id);
4297			if (tc->origin_dev)
4298				DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4299			break;
4300		}
4301	}
4302
4303	return;
4304
4305err:
4306	DMEMIT("Error");
4307}
4308
4309static int thin_iterate_devices(struct dm_target *ti,
4310				iterate_devices_callout_fn fn, void *data)
4311{
4312	sector_t blocks;
4313	struct thin_c *tc = ti->private;
4314	struct pool *pool = tc->pool;
4315
4316	/*
4317	 * We can't call dm_pool_get_data_dev_size() since that blocks.  So
4318	 * we follow a more convoluted path through to the pool's target.
4319	 */
4320	if (!pool->ti)
4321		return 0;	/* nothing is bound */
4322
4323	blocks = pool->ti->len;
4324	(void) sector_div(blocks, pool->sectors_per_block);
4325	if (blocks)
4326		return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4327
4328	return 0;
4329}
4330
4331static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4332{
4333	struct thin_c *tc = ti->private;
4334	struct pool *pool = tc->pool;
4335
4336	if (!pool->pf.discard_enabled)
4337		return;
4338
4339	limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4340	limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4341}
4342
4343static struct target_type thin_target = {
4344	.name = "thin",
4345	.version = {1, 19, 0},
4346	.module	= THIS_MODULE,
4347	.ctr = thin_ctr,
4348	.dtr = thin_dtr,
4349	.map = thin_map,
4350	.end_io = thin_endio,
4351	.preresume = thin_preresume,
4352	.presuspend = thin_presuspend,
4353	.postsuspend = thin_postsuspend,
4354	.status = thin_status,
4355	.iterate_devices = thin_iterate_devices,
4356	.io_hints = thin_io_hints,
4357};
4358
4359/*----------------------------------------------------------------*/
4360
4361static int __init dm_thin_init(void)
4362{
4363	int r;
4364
4365	pool_table_init();
4366
4367	r = dm_register_target(&thin_target);
4368	if (r)
4369		return r;
4370
4371	r = dm_register_target(&pool_target);
4372	if (r)
4373		goto bad_pool_target;
4374
4375	r = -ENOMEM;
4376
4377	_new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4378	if (!_new_mapping_cache)
4379		goto bad_new_mapping_cache;
4380
4381	return 0;
4382
4383bad_new_mapping_cache:
4384	dm_unregister_target(&pool_target);
4385bad_pool_target:
4386	dm_unregister_target(&thin_target);
4387
4388	return r;
4389}
4390
4391static void dm_thin_exit(void)
4392{
4393	dm_unregister_target(&thin_target);
4394	dm_unregister_target(&pool_target);
4395
4396	kmem_cache_destroy(_new_mapping_cache);
4397}
4398
4399module_init(dm_thin_init);
4400module_exit(dm_thin_exit);
4401
4402module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4403MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4404
4405MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4406MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4407MODULE_LICENSE("GPL");