1// SPDX-License-Identifier: GPL-2.0
   2
   3#include <linux/err.h>
   4#include <linux/slab.h>
   5#include <linux/spinlock.h>
   6#include "messages.h"
   7#include "ctree.h"
   8#include "volumes.h"
   9#include "extent_map.h"
  10#include "compression.h"
  11#include "btrfs_inode.h"
  12
  13
  14static struct kmem_cache *extent_map_cache;
  15
  16int __init extent_map_init(void)
  17{
  18	extent_map_cache = kmem_cache_create("btrfs_extent_map",
  19			sizeof(struct extent_map), 0,
  20			SLAB_MEM_SPREAD, NULL);
  21	if (!extent_map_cache)
  22		return -ENOMEM;
  23	return 0;
  24}
  25
  26void __cold extent_map_exit(void)
  27{
  28	kmem_cache_destroy(extent_map_cache);
  29}
  30
  31/*
  32 * Initialize the extent tree @tree.  Should be called for each new inode or
  33 * other user of the extent_map interface.
 
 
 
  34 */
  35void extent_map_tree_init(struct extent_map_tree *tree)
  36{
  37	tree->map = RB_ROOT_CACHED;
  38	INIT_LIST_HEAD(&tree->modified_extents);
  39	rwlock_init(&tree->lock);
  40}
  41
  42/*
  43 * Allocate a new extent_map structure.  The new structure is returned with a
  44 * reference count of one and needs to be freed using free_extent_map()
 
 
 
  45 */
  46struct extent_map *alloc_extent_map(void)
  47{
  48	struct extent_map *em;
  49	em = kmem_cache_zalloc(extent_map_cache, GFP_NOFS);
  50	if (!em)
  51		return NULL;
  52	RB_CLEAR_NODE(&em->rb_node);
 
 
 
  53	refcount_set(&em->refs, 1);
  54	INIT_LIST_HEAD(&em->list);
  55	return em;
  56}
  57
  58/*
  59 * Drop the reference out on @em by one and free the structure if the reference
  60 * count hits zero.
 
 
 
  61 */
  62void free_extent_map(struct extent_map *em)
  63{
  64	if (!em)
  65		return;
 
  66	if (refcount_dec_and_test(&em->refs)) {
  67		WARN_ON(extent_map_in_tree(em));
  68		WARN_ON(!list_empty(&em->list));
 
 
  69		kmem_cache_free(extent_map_cache, em);
  70	}
  71}
  72
  73/* Do the math around the end of an extent, handling wrapping. */
  74static u64 range_end(u64 start, u64 len)
  75{
  76	if (start + len < start)
  77		return (u64)-1;
  78	return start + len;
  79}
  80
  81static int tree_insert(struct rb_root_cached *root, struct extent_map *em)
  82{
  83	struct rb_node **p = &root->rb_root.rb_node;
  84	struct rb_node *parent = NULL;
  85	struct extent_map *entry = NULL;
  86	struct rb_node *orig_parent = NULL;
  87	u64 end = range_end(em->start, em->len);
  88	bool leftmost = true;
  89
  90	while (*p) {
  91		parent = *p;
  92		entry = rb_entry(parent, struct extent_map, rb_node);
  93
  94		if (em->start < entry->start) {
  95			p = &(*p)->rb_left;
  96		} else if (em->start >= extent_map_end(entry)) {
  97			p = &(*p)->rb_right;
  98			leftmost = false;
  99		} else {
 100			return -EEXIST;
 101		}
 102	}
 103
 104	orig_parent = parent;
 105	while (parent && em->start >= extent_map_end(entry)) {
 106		parent = rb_next(parent);
 107		entry = rb_entry(parent, struct extent_map, rb_node);
 108	}
 109	if (parent)
 110		if (end > entry->start && em->start < extent_map_end(entry))
 111			return -EEXIST;
 112
 113	parent = orig_parent;
 114	entry = rb_entry(parent, struct extent_map, rb_node);
 115	while (parent && em->start < entry->start) {
 116		parent = rb_prev(parent);
 117		entry = rb_entry(parent, struct extent_map, rb_node);
 118	}
 119	if (parent)
 120		if (end > entry->start && em->start < extent_map_end(entry))
 121			return -EEXIST;
 122
 123	rb_link_node(&em->rb_node, orig_parent, p);
 124	rb_insert_color_cached(&em->rb_node, root, leftmost);
 125	return 0;
 126}
 127
 128/*
 129 * Search through the tree for an extent_map with a given offset.  If it can't
 130 * be found, try to find some neighboring extents
 131 */
 132static struct rb_node *__tree_search(struct rb_root *root, u64 offset,
 133				     struct rb_node **prev_or_next_ret)
 
 134{
 135	struct rb_node *n = root->rb_node;
 136	struct rb_node *prev = NULL;
 137	struct rb_node *orig_prev = NULL;
 138	struct extent_map *entry;
 139	struct extent_map *prev_entry = NULL;
 140
 141	ASSERT(prev_or_next_ret);
 142
 143	while (n) {
 144		entry = rb_entry(n, struct extent_map, rb_node);
 145		prev = n;
 146		prev_entry = entry;
 147
 148		if (offset < entry->start)
 149			n = n->rb_left;
 150		else if (offset >= extent_map_end(entry))
 151			n = n->rb_right;
 152		else
 153			return n;
 154	}
 155
 156	orig_prev = prev;
 157	while (prev && offset >= extent_map_end(prev_entry)) {
 158		prev = rb_next(prev);
 159		prev_entry = rb_entry(prev, struct extent_map, rb_node);
 160	}
 161
 162	/*
 163	 * Previous extent map found, return as in this case the caller does not
 164	 * care about the next one.
 165	 */
 166	if (prev) {
 167		*prev_or_next_ret = prev;
 168		return NULL;
 169	}
 170
 171	prev = orig_prev;
 172	prev_entry = rb_entry(prev, struct extent_map, rb_node);
 173	while (prev && offset < prev_entry->start) {
 174		prev = rb_prev(prev);
 175		prev_entry = rb_entry(prev, struct extent_map, rb_node);
 
 
 
 
 
 176	}
 177	*prev_or_next_ret = prev;
 178
 179	return NULL;
 180}
 181
 182static inline u64 extent_map_block_end(const struct extent_map *em)
 183{
 184	if (em->block_start + em->block_len < em->block_start)
 185		return (u64)-1;
 186	return em->block_start + em->block_len;
 187}
 188
 189static bool can_merge_extent_map(const struct extent_map *em)
 190{
 191	if (em->flags & EXTENT_FLAG_PINNED)
 192		return false;
 193
 194	/* Don't merge compressed extents, we need to know their actual size. */
 195	if (extent_map_is_compressed(em))
 196		return false;
 
 
 
 197
 198	if (em->flags & EXTENT_FLAG_LOGGING)
 199		return false;
 
 200
 201	/*
 202	 * We don't want to merge stuff that hasn't been written to the log yet
 203	 * since it may not reflect exactly what is on disk, and that would be
 204	 * bad.
 205	 */
 206	if (!list_empty(&em->list))
 207		return false;
 208
 209	return true;
 210}
 211
 212/* Check to see if two extent_map structs are adjacent and safe to merge. */
 213static bool mergeable_maps(const struct extent_map *prev, const struct extent_map *next)
 214{
 215	if (extent_map_end(prev) != next->start)
 216		return false;
 217
 218	if (prev->flags != next->flags)
 219		return false;
 220
 221	if (next->block_start < EXTENT_MAP_LAST_BYTE - 1)
 222		return next->block_start == extent_map_block_end(prev);
 223
 224	/* HOLES and INLINE extents. */
 225	return next->block_start == prev->block_start;
 
 
 
 
 
 
 
 
 
 226}
 227
 228static void try_merge_map(struct extent_map_tree *tree, struct extent_map *em)
 229{
 230	struct extent_map *merge = NULL;
 231	struct rb_node *rb;
 232
 233	/*
 234	 * We can't modify an extent map that is in the tree and that is being
 235	 * used by another task, as it can cause that other task to see it in
 236	 * inconsistent state during the merging. We always have 1 reference for
 237	 * the tree and 1 for this task (which is unpinning the extent map or
 238	 * clearing the logging flag), so anything > 2 means it's being used by
 239	 * other tasks too.
 240	 */
 241	if (refcount_read(&em->refs) > 2)
 242		return;
 243
 244	if (!can_merge_extent_map(em))
 245		return;
 246
 247	if (em->start != 0) {
 248		rb = rb_prev(&em->rb_node);
 249		if (rb)
 250			merge = rb_entry(rb, struct extent_map, rb_node);
 251		if (rb && can_merge_extent_map(merge) && mergeable_maps(merge, em)) {
 252			em->start = merge->start;
 253			em->orig_start = merge->orig_start;
 254			em->len += merge->len;
 255			em->block_len += merge->block_len;
 256			em->block_start = merge->block_start;
 257			em->mod_len = (em->mod_len + em->mod_start) - merge->mod_start;
 258			em->mod_start = merge->mod_start;
 259			em->generation = max(em->generation, merge->generation);
 260			em->flags |= EXTENT_FLAG_MERGED;
 261
 262			rb_erase_cached(&merge->rb_node, &tree->map);
 263			RB_CLEAR_NODE(&merge->rb_node);
 264			free_extent_map(merge);
 265		}
 266	}
 267
 268	rb = rb_next(&em->rb_node);
 269	if (rb)
 270		merge = rb_entry(rb, struct extent_map, rb_node);
 271	if (rb && can_merge_extent_map(merge) && mergeable_maps(em, merge)) {
 272		em->len += merge->len;
 273		em->block_len += merge->block_len;
 274		rb_erase_cached(&merge->rb_node, &tree->map);
 275		RB_CLEAR_NODE(&merge->rb_node);
 276		em->mod_len = (merge->mod_start + merge->mod_len) - em->mod_start;
 277		em->generation = max(em->generation, merge->generation);
 278		em->flags |= EXTENT_FLAG_MERGED;
 279		free_extent_map(merge);
 280	}
 281}
 282
 283/*
 284 * Unpin an extent from the cache.
 285 *
 286 * @inode:	the inode from which we are unpinning an extent range
 287 * @start:	logical offset in the file
 288 * @len:	length of the extent
 289 * @gen:	generation that this extent has been modified in
 290 *
 291 * Called after an extent has been written to disk properly.  Set the generation
 292 * to the generation that actually added the file item to the inode so we know
 293 * we need to sync this extent when we call fsync().
 294 */
 295int unpin_extent_cache(struct btrfs_inode *inode, u64 start, u64 len, u64 gen)
 
 296{
 297	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 298	struct extent_map_tree *tree = &inode->extent_tree;
 299	int ret = 0;
 300	struct extent_map *em;
 301	bool prealloc = false;
 302
 303	write_lock(&tree->lock);
 304	em = lookup_extent_mapping(tree, start, len);
 305
 306	if (WARN_ON(!em)) {
 307		btrfs_warn(fs_info,
 308"no extent map found for inode %llu (root %lld) when unpinning extent range [%llu, %llu), generation %llu",
 309			   btrfs_ino(inode), btrfs_root_id(inode->root),
 310			   start, len, gen);
 311		goto out;
 312	}
 313
 314	if (WARN_ON(em->start != start))
 315		btrfs_warn(fs_info,
 316"found extent map for inode %llu (root %lld) with unexpected start offset %llu when unpinning extent range [%llu, %llu), generation %llu",
 317			   btrfs_ino(inode), btrfs_root_id(inode->root),
 318			   em->start, start, len, gen);
 319
 320	em->generation = gen;
 321	em->flags &= ~EXTENT_FLAG_PINNED;
 322	em->mod_start = em->start;
 323	em->mod_len = em->len;
 324
 325	if (em->flags & EXTENT_FLAG_FILLING) {
 326		prealloc = true;
 327		em->flags &= ~EXTENT_FLAG_FILLING;
 328	}
 329
 330	try_merge_map(tree, em);
 331
 332	if (prealloc) {
 333		em->mod_start = em->start;
 334		em->mod_len = em->len;
 335	}
 336
 337	free_extent_map(em);
 338out:
 339	write_unlock(&tree->lock);
 340	return ret;
 341
 342}
 343
 344void clear_em_logging(struct extent_map_tree *tree, struct extent_map *em)
 345{
 346	lockdep_assert_held_write(&tree->lock);
 347
 348	em->flags &= ~EXTENT_FLAG_LOGGING;
 349	if (extent_map_in_tree(em))
 350		try_merge_map(tree, em);
 351}
 352
 353static inline void setup_extent_mapping(struct extent_map_tree *tree,
 354					struct extent_map *em,
 355					int modified)
 356{
 357	refcount_inc(&em->refs);
 358	em->mod_start = em->start;
 359	em->mod_len = em->len;
 360
 361	ASSERT(list_empty(&em->list));
 362
 363	if (modified)
 364		list_add(&em->list, &tree->modified_extents);
 365	else
 366		try_merge_map(tree, em);
 367}
 368
 369/*
 370 * Add new extent map to the extent tree
 371 *
 372 * @tree:	tree to insert new map in
 373 * @em:		map to insert
 374 * @modified:	indicate whether the given @em should be added to the
 375 *	        modified list, which indicates the extent needs to be logged
 376 *
 377 * Insert @em into @tree or perform a simple forward/backward merge with
 378 * existing mappings.  The extent_map struct passed in will be inserted
 379 * into the tree directly, with an additional reference taken, or a
 380 * reference dropped if the merge attempt was successful.
 381 */
 382static int add_extent_mapping(struct extent_map_tree *tree,
 383			      struct extent_map *em, int modified)
 384{
 385	int ret = 0;
 386
 387	lockdep_assert_held_write(&tree->lock);
 388
 389	ret = tree_insert(&tree->map, em);
 390	if (ret)
 391		goto out;
 392
 393	setup_extent_mapping(tree, em, modified);
 394out:
 395	return ret;
 396}
 397
 398static struct extent_map *
 399__lookup_extent_mapping(struct extent_map_tree *tree,
 400			u64 start, u64 len, int strict)
 401{
 402	struct extent_map *em;
 403	struct rb_node *rb_node;
 404	struct rb_node *prev_or_next = NULL;
 
 405	u64 end = range_end(start, len);
 406
 407	rb_node = __tree_search(&tree->map.rb_root, start, &prev_or_next);
 408	if (!rb_node) {
 409		if (prev_or_next)
 410			rb_node = prev_or_next;
 
 
 411		else
 412			return NULL;
 413	}
 414
 415	em = rb_entry(rb_node, struct extent_map, rb_node);
 416
 417	if (strict && !(end > em->start && start < extent_map_end(em)))
 418		return NULL;
 419
 420	refcount_inc(&em->refs);
 421	return em;
 422}
 423
 424/*
 425 * Lookup extent_map that intersects @start + @len range.
 426 *
 427 * @tree:	tree to lookup in
 428 * @start:	byte offset to start the search
 429 * @len:	length of the lookup range
 430 *
 431 * Find and return the first extent_map struct in @tree that intersects the
 432 * [start, len] range.  There may be additional objects in the tree that
 433 * intersect, so check the object returned carefully to make sure that no
 434 * additional lookups are needed.
 435 */
 436struct extent_map *lookup_extent_mapping(struct extent_map_tree *tree,
 437					 u64 start, u64 len)
 438{
 439	return __lookup_extent_mapping(tree, start, len, 1);
 440}
 441
 442/*
 443 * Find a nearby extent map intersecting @start + @len (not an exact search).
 444 *
 445 * @tree:	tree to lookup in
 446 * @start:	byte offset to start the search
 447 * @len:	length of the lookup range
 448 *
 449 * Find and return the first extent_map struct in @tree that intersects the
 450 * [start, len] range.
 451 *
 452 * If one can't be found, any nearby extent may be returned
 453 */
 454struct extent_map *search_extent_mapping(struct extent_map_tree *tree,
 455					 u64 start, u64 len)
 456{
 457	return __lookup_extent_mapping(tree, start, len, 0);
 458}
 459
 460/*
 461 * Remove an extent_map from the extent tree.
 462 *
 463 * @tree:	extent tree to remove from
 464 * @em:		extent map being removed
 465 *
 466 * Remove @em from @tree.  No reference counts are dropped, and no checks
 467 * are done to see if the range is in use.
 468 */
 469void remove_extent_mapping(struct extent_map_tree *tree, struct extent_map *em)
 470{
 471	lockdep_assert_held_write(&tree->lock);
 472
 473	WARN_ON(em->flags & EXTENT_FLAG_PINNED);
 474	rb_erase_cached(&em->rb_node, &tree->map);
 475	if (!(em->flags & EXTENT_FLAG_LOGGING))
 476		list_del_init(&em->list);
 477	RB_CLEAR_NODE(&em->rb_node);
 
 478}
 479
 480static void replace_extent_mapping(struct extent_map_tree *tree,
 481				   struct extent_map *cur,
 482				   struct extent_map *new,
 483				   int modified)
 484{
 485	lockdep_assert_held_write(&tree->lock);
 486
 487	WARN_ON(cur->flags & EXTENT_FLAG_PINNED);
 488	ASSERT(extent_map_in_tree(cur));
 489	if (!(cur->flags & EXTENT_FLAG_LOGGING))
 490		list_del_init(&cur->list);
 491	rb_replace_node_cached(&cur->rb_node, &new->rb_node, &tree->map);
 492	RB_CLEAR_NODE(&cur->rb_node);
 493
 494	setup_extent_mapping(tree, new, modified);
 495}
 496
 497static struct extent_map *next_extent_map(const struct extent_map *em)
 498{
 499	struct rb_node *next;
 500
 501	next = rb_next(&em->rb_node);
 502	if (!next)
 503		return NULL;
 504	return container_of(next, struct extent_map, rb_node);
 505}
 506
 507static struct extent_map *prev_extent_map(struct extent_map *em)
 508{
 509	struct rb_node *prev;
 510
 511	prev = rb_prev(&em->rb_node);
 512	if (!prev)
 513		return NULL;
 514	return container_of(prev, struct extent_map, rb_node);
 515}
 516
 517/*
 518 * Helper for btrfs_get_extent.  Given an existing extent in the tree,
 519 * the existing extent is the nearest extent to map_start,
 520 * and an extent that you want to insert, deal with overlap and insert
 521 * the best fitted new extent into the tree.
 522 */
 523static noinline int merge_extent_mapping(struct extent_map_tree *em_tree,
 524					 struct extent_map *existing,
 525					 struct extent_map *em,
 526					 u64 map_start)
 527{
 528	struct extent_map *prev;
 529	struct extent_map *next;
 530	u64 start;
 531	u64 end;
 532	u64 start_diff;
 533
 534	BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
 535
 536	if (existing->start > map_start) {
 537		next = existing;
 538		prev = prev_extent_map(next);
 539	} else {
 540		prev = existing;
 541		next = next_extent_map(prev);
 542	}
 543
 544	start = prev ? extent_map_end(prev) : em->start;
 545	start = max_t(u64, start, em->start);
 546	end = next ? next->start : extent_map_end(em);
 547	end = min_t(u64, end, extent_map_end(em));
 548	start_diff = start - em->start;
 549	em->start = start;
 550	em->len = end - start;
 551	if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 552	    !extent_map_is_compressed(em)) {
 553		em->block_start += start_diff;
 554		em->block_len = em->len;
 555	}
 556	return add_extent_mapping(em_tree, em, 0);
 557}
 558
 559/*
 560 * Add extent mapping into em_tree.
 561 *
 562 * @fs_info:  the filesystem
 563 * @em_tree:  extent tree into which we want to insert the extent mapping
 564 * @em_in:    extent we are inserting
 565 * @start:    start of the logical range btrfs_get_extent() is requesting
 566 * @len:      length of the logical range btrfs_get_extent() is requesting
 567 *
 568 * Note that @em_in's range may be different from [start, start+len),
 569 * but they must be overlapped.
 570 *
 571 * Insert @em_in into @em_tree. In case there is an overlapping range, handle
 572 * the -EEXIST by either:
 573 * a) Returning the existing extent in @em_in if @start is within the
 574 *    existing em.
 575 * b) Merge the existing extent with @em_in passed in.
 576 *
 577 * Return 0 on success, otherwise -EEXIST.
 578 *
 579 */
 580int btrfs_add_extent_mapping(struct btrfs_fs_info *fs_info,
 581			     struct extent_map_tree *em_tree,
 582			     struct extent_map **em_in, u64 start, u64 len)
 583{
 584	int ret;
 585	struct extent_map *em = *em_in;
 586
 587	/*
 588	 * Tree-checker should have rejected any inline extent with non-zero
 589	 * file offset. Here just do a sanity check.
 590	 */
 591	if (em->block_start == EXTENT_MAP_INLINE)
 592		ASSERT(em->start == 0);
 593
 594	ret = add_extent_mapping(em_tree, em, 0);
 595	/* it is possible that someone inserted the extent into the tree
 596	 * while we had the lock dropped.  It is also possible that
 597	 * an overlapping map exists in the tree
 598	 */
 599	if (ret == -EEXIST) {
 600		struct extent_map *existing;
 601
 
 
 602		existing = search_extent_mapping(em_tree, start, len);
 603
 604		trace_btrfs_handle_em_exist(fs_info, existing, em, start, len);
 605
 606		/*
 607		 * existing will always be non-NULL, since there must be
 608		 * extent causing the -EEXIST.
 609		 */
 610		if (start >= existing->start &&
 611		    start < extent_map_end(existing)) {
 612			free_extent_map(em);
 613			*em_in = existing;
 614			ret = 0;
 615		} else {
 616			u64 orig_start = em->start;
 617			u64 orig_len = em->len;
 618
 619			/*
 620			 * The existing extent map is the one nearest to
 621			 * the [start, start + len) range which overlaps
 622			 */
 623			ret = merge_extent_mapping(em_tree, existing,
 624						   em, start);
 625			if (ret) {
 626				free_extent_map(em);
 627				*em_in = NULL;
 628				WARN_ONCE(ret,
 629"unexpected error %d: merge existing(start %llu len %llu) with em(start %llu len %llu)\n",
 630					  ret, existing->start, existing->len,
 631					  orig_start, orig_len);
 632			}
 633			free_extent_map(existing);
 634		}
 635	}
 636
 637	ASSERT(ret == 0 || ret == -EEXIST);
 638	return ret;
 639}
 640
 641/*
 642 * Drop all extent maps from a tree in the fastest possible way, rescheduling
 643 * if needed. This avoids searching the tree, from the root down to the first
 644 * extent map, before each deletion.
 645 */
 646static void drop_all_extent_maps_fast(struct extent_map_tree *tree)
 647{
 648	write_lock(&tree->lock);
 649	while (!RB_EMPTY_ROOT(&tree->map.rb_root)) {
 650		struct extent_map *em;
 651		struct rb_node *node;
 652
 653		node = rb_first_cached(&tree->map);
 654		em = rb_entry(node, struct extent_map, rb_node);
 655		em->flags &= ~(EXTENT_FLAG_PINNED | EXTENT_FLAG_LOGGING);
 656		remove_extent_mapping(tree, em);
 657		free_extent_map(em);
 658		cond_resched_rwlock_write(&tree->lock);
 659	}
 660	write_unlock(&tree->lock);
 661}
 662
 663/*
 664 * Drop all extent maps in a given range.
 665 *
 666 * @inode:       The target inode.
 667 * @start:       Start offset of the range.
 668 * @end:         End offset of the range (inclusive value).
 669 * @skip_pinned: Indicate if pinned extent maps should be ignored or not.
 670 *
 671 * This drops all the extent maps that intersect the given range [@start, @end].
 672 * Extent maps that partially overlap the range and extend behind or beyond it,
 673 * are split.
 674 * The caller should have locked an appropriate file range in the inode's io
 675 * tree before calling this function.
 676 */
 677void btrfs_drop_extent_map_range(struct btrfs_inode *inode, u64 start, u64 end,
 678				 bool skip_pinned)
 679{
 680	struct extent_map *split;
 681	struct extent_map *split2;
 682	struct extent_map *em;
 683	struct extent_map_tree *em_tree = &inode->extent_tree;
 684	u64 len = end - start + 1;
 685
 686	WARN_ON(end < start);
 687	if (end == (u64)-1) {
 688		if (start == 0 && !skip_pinned) {
 689			drop_all_extent_maps_fast(em_tree);
 690			return;
 691		}
 692		len = (u64)-1;
 693	} else {
 694		/* Make end offset exclusive for use in the loop below. */
 695		end++;
 696	}
 697
 698	/*
 699	 * It's ok if we fail to allocate the extent maps, see the comment near
 700	 * the bottom of the loop below. We only need two spare extent maps in
 701	 * the worst case, where the first extent map that intersects our range
 702	 * starts before the range and the last extent map that intersects our
 703	 * range ends after our range (and they might be the same extent map),
 704	 * because we need to split those two extent maps at the boundaries.
 705	 */
 706	split = alloc_extent_map();
 707	split2 = alloc_extent_map();
 708
 709	write_lock(&em_tree->lock);
 710	em = lookup_extent_mapping(em_tree, start, len);
 711
 712	while (em) {
 713		/* extent_map_end() returns exclusive value (last byte + 1). */
 714		const u64 em_end = extent_map_end(em);
 715		struct extent_map *next_em = NULL;
 716		u64 gen;
 717		unsigned long flags;
 718		bool modified;
 719		bool compressed;
 720
 721		if (em_end < end) {
 722			next_em = next_extent_map(em);
 723			if (next_em) {
 724				if (next_em->start < end)
 725					refcount_inc(&next_em->refs);
 726				else
 727					next_em = NULL;
 728			}
 729		}
 730
 731		if (skip_pinned && (em->flags & EXTENT_FLAG_PINNED)) {
 732			start = em_end;
 733			goto next;
 734		}
 735
 736		flags = em->flags;
 737		/*
 738		 * In case we split the extent map, we want to preserve the
 739		 * EXTENT_FLAG_LOGGING flag on our extent map, but we don't want
 740		 * it on the new extent maps.
 741		 */
 742		em->flags &= ~(EXTENT_FLAG_PINNED | EXTENT_FLAG_LOGGING);
 743		modified = !list_empty(&em->list);
 744
 745		/*
 746		 * The extent map does not cross our target range, so no need to
 747		 * split it, we can remove it directly.
 748		 */
 749		if (em->start >= start && em_end <= end)
 750			goto remove_em;
 751
 752		gen = em->generation;
 753		compressed = extent_map_is_compressed(em);
 754
 755		if (em->start < start) {
 756			if (!split) {
 757				split = split2;
 758				split2 = NULL;
 759				if (!split)
 760					goto remove_em;
 761			}
 762			split->start = em->start;
 763			split->len = start - em->start;
 764
 765			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 766				split->orig_start = em->orig_start;
 767				split->block_start = em->block_start;
 768
 769				if (compressed)
 770					split->block_len = em->block_len;
 771				else
 772					split->block_len = split->len;
 773				split->orig_block_len = max(split->block_len,
 774						em->orig_block_len);
 775				split->ram_bytes = em->ram_bytes;
 776			} else {
 777				split->orig_start = split->start;
 778				split->block_len = 0;
 779				split->block_start = em->block_start;
 780				split->orig_block_len = 0;
 781				split->ram_bytes = split->len;
 782			}
 783
 784			split->generation = gen;
 785			split->flags = flags;
 786			replace_extent_mapping(em_tree, em, split, modified);
 787			free_extent_map(split);
 788			split = split2;
 789			split2 = NULL;
 790		}
 791		if (em_end > end) {
 792			if (!split) {
 793				split = split2;
 794				split2 = NULL;
 795				if (!split)
 796					goto remove_em;
 797			}
 798			split->start = end;
 799			split->len = em_end - end;
 800			split->block_start = em->block_start;
 801			split->flags = flags;
 802			split->generation = gen;
 803
 804			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 805				split->orig_block_len = max(em->block_len,
 806						    em->orig_block_len);
 807
 808				split->ram_bytes = em->ram_bytes;
 809				if (compressed) {
 810					split->block_len = em->block_len;
 811					split->orig_start = em->orig_start;
 812				} else {
 813					const u64 diff = start + len - em->start;
 814
 815					split->block_len = split->len;
 816					split->block_start += diff;
 817					split->orig_start = em->orig_start;
 818				}
 819			} else {
 820				split->ram_bytes = split->len;
 821				split->orig_start = split->start;
 822				split->block_len = 0;
 823				split->orig_block_len = 0;
 824			}
 825
 826			if (extent_map_in_tree(em)) {
 827				replace_extent_mapping(em_tree, em, split,
 828						       modified);
 829			} else {
 830				int ret;
 831
 832				ret = add_extent_mapping(em_tree, split,
 833							 modified);
 834				/* Logic error, shouldn't happen. */
 835				ASSERT(ret == 0);
 836				if (WARN_ON(ret != 0) && modified)
 837					btrfs_set_inode_full_sync(inode);
 838			}
 839			free_extent_map(split);
 840			split = NULL;
 841		}
 842remove_em:
 843		if (extent_map_in_tree(em)) {
 844			/*
 845			 * If the extent map is still in the tree it means that
 846			 * either of the following is true:
 847			 *
 848			 * 1) It fits entirely in our range (doesn't end beyond
 849			 *    it or starts before it);
 850			 *
 851			 * 2) It starts before our range and/or ends after our
 852			 *    range, and we were not able to allocate the extent
 853			 *    maps for split operations, @split and @split2.
 854			 *
 855			 * If we are at case 2) then we just remove the entire
 856			 * extent map - this is fine since if anyone needs it to
 857			 * access the subranges outside our range, will just
 858			 * load it again from the subvolume tree's file extent
 859			 * item. However if the extent map was in the list of
 860			 * modified extents, then we must mark the inode for a
 861			 * full fsync, otherwise a fast fsync will miss this
 862			 * extent if it's new and needs to be logged.
 863			 */
 864			if ((em->start < start || em_end > end) && modified) {
 865				ASSERT(!split);
 866				btrfs_set_inode_full_sync(inode);
 867			}
 868			remove_extent_mapping(em_tree, em);
 869		}
 870
 871		/*
 872		 * Once for the tree reference (we replaced or removed the
 873		 * extent map from the tree).
 874		 */
 875		free_extent_map(em);
 876next:
 877		/* Once for us (for our lookup reference). */
 878		free_extent_map(em);
 879
 880		em = next_em;
 881	}
 882
 883	write_unlock(&em_tree->lock);
 884
 885	free_extent_map(split);
 886	free_extent_map(split2);
 887}
 888
 889/*
 890 * Replace a range in the inode's extent map tree with a new extent map.
 891 *
 892 * @inode:      The target inode.
 893 * @new_em:     The new extent map to add to the inode's extent map tree.
 894 * @modified:   Indicate if the new extent map should be added to the list of
 895 *              modified extents (for fast fsync tracking).
 896 *
 897 * Drops all the extent maps in the inode's extent map tree that intersect the
 898 * range of the new extent map and adds the new extent map to the tree.
 899 * The caller should have locked an appropriate file range in the inode's io
 900 * tree before calling this function.
 901 */
 902int btrfs_replace_extent_map_range(struct btrfs_inode *inode,
 903				   struct extent_map *new_em,
 904				   bool modified)
 905{
 906	const u64 end = new_em->start + new_em->len - 1;
 907	struct extent_map_tree *tree = &inode->extent_tree;
 908	int ret;
 909
 910	ASSERT(!extent_map_in_tree(new_em));
 911
 912	/*
 913	 * The caller has locked an appropriate file range in the inode's io
 914	 * tree, but getting -EEXIST when adding the new extent map can still
 915	 * happen in case there are extents that partially cover the range, and
 916	 * this is due to two tasks operating on different parts of the extent.
 917	 * See commit 18e83ac75bfe67 ("Btrfs: fix unexpected EEXIST from
 918	 * btrfs_get_extent") for an example and details.
 919	 */
 920	do {
 921		btrfs_drop_extent_map_range(inode, new_em->start, end, false);
 922		write_lock(&tree->lock);
 923		ret = add_extent_mapping(tree, new_em, modified);
 924		write_unlock(&tree->lock);
 925	} while (ret == -EEXIST);
 926
 927	return ret;
 928}
 929
 930/*
 931 * Split off the first pre bytes from the extent_map at [start, start + len],
 932 * and set the block_start for it to new_logical.
 933 *
 934 * This function is used when an ordered_extent needs to be split.
 935 */
 936int split_extent_map(struct btrfs_inode *inode, u64 start, u64 len, u64 pre,
 937		     u64 new_logical)
 938{
 939	struct extent_map_tree *em_tree = &inode->extent_tree;
 940	struct extent_map *em;
 941	struct extent_map *split_pre = NULL;
 942	struct extent_map *split_mid = NULL;
 943	int ret = 0;
 944	unsigned long flags;
 945
 946	ASSERT(pre != 0);
 947	ASSERT(pre < len);
 948
 949	split_pre = alloc_extent_map();
 950	if (!split_pre)
 951		return -ENOMEM;
 952	split_mid = alloc_extent_map();
 953	if (!split_mid) {
 954		ret = -ENOMEM;
 955		goto out_free_pre;
 956	}
 957
 958	lock_extent(&inode->io_tree, start, start + len - 1, NULL);
 959	write_lock(&em_tree->lock);
 960	em = lookup_extent_mapping(em_tree, start, len);
 961	if (!em) {
 962		ret = -EIO;
 963		goto out_unlock;
 964	}
 965
 966	ASSERT(em->len == len);
 967	ASSERT(!extent_map_is_compressed(em));
 968	ASSERT(em->block_start < EXTENT_MAP_LAST_BYTE);
 969	ASSERT(em->flags & EXTENT_FLAG_PINNED);
 970	ASSERT(!(em->flags & EXTENT_FLAG_LOGGING));
 971	ASSERT(!list_empty(&em->list));
 972
 973	flags = em->flags;
 974	em->flags &= ~EXTENT_FLAG_PINNED;
 975
 976	/* First, replace the em with a new extent_map starting from * em->start */
 977	split_pre->start = em->start;
 978	split_pre->len = pre;
 979	split_pre->orig_start = split_pre->start;
 980	split_pre->block_start = new_logical;
 981	split_pre->block_len = split_pre->len;
 982	split_pre->orig_block_len = split_pre->block_len;
 983	split_pre->ram_bytes = split_pre->len;
 984	split_pre->flags = flags;
 985	split_pre->generation = em->generation;
 986
 987	replace_extent_mapping(em_tree, em, split_pre, 1);
 988
 989	/*
 990	 * Now we only have an extent_map at:
 991	 *     [em->start, em->start + pre]
 992	 */
 993
 994	/* Insert the middle extent_map. */
 995	split_mid->start = em->start + pre;
 996	split_mid->len = em->len - pre;
 997	split_mid->orig_start = split_mid->start;
 998	split_mid->block_start = em->block_start + pre;
 999	split_mid->block_len = split_mid->len;
1000	split_mid->orig_block_len = split_mid->block_len;
1001	split_mid->ram_bytes = split_mid->len;
1002	split_mid->flags = flags;
1003	split_mid->generation = em->generation;
1004	add_extent_mapping(em_tree, split_mid, 1);
1005
1006	/* Once for us */
1007	free_extent_map(em);
1008	/* Once for the tree */
1009	free_extent_map(em);
1010
1011out_unlock:
1012	write_unlock(&em_tree->lock);
1013	unlock_extent(&inode->io_tree, start, start + len - 1, NULL);
1014	free_extent_map(split_mid);
1015out_free_pre:
1016	free_extent_map(split_pre);
1017	return ret;
1018}