block-group.c

   1// SPDX-License-Identifier: GPL-2.0
   2
   3#include <linux/sizes.h>
   4#include <linux/list_sort.h>
   5#include "misc.h"
   6#include "ctree.h"
   7#include "block-group.h"
   8#include "space-info.h"
   9#include "disk-io.h"
  10#include "free-space-cache.h"
  11#include "free-space-tree.h"
  12#include "volumes.h"
  13#include "transaction.h"
  14#include "ref-verify.h"
  15#include "sysfs.h"
  16#include "tree-log.h"
  17#include "delalloc-space.h"
  18#include "discard.h"
  19#include "raid56.h"
  20#include "zoned.h"
  21#include "fs.h"
  22#include "accessors.h"
  23#include "extent-tree.h"
  24
  25#ifdef CONFIG_BTRFS_DEBUG
  26int btrfs_should_fragment_free_space(const struct btrfs_block_group *block_group)
  27{
  28	struct btrfs_fs_info *fs_info = block_group->fs_info;
  29
  30	return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
  31		block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
  32	       (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
  33		block_group->flags &  BTRFS_BLOCK_GROUP_DATA);
  34}
  35#endif
  36
  37static inline bool has_unwritten_metadata(struct btrfs_block_group *block_group)
  38{
  39	/* The meta_write_pointer is available only on the zoned setup. */
  40	if (!btrfs_is_zoned(block_group->fs_info))
  41		return false;
  42
  43	if (block_group->flags & BTRFS_BLOCK_GROUP_DATA)
  44		return false;
  45
  46	return block_group->start + block_group->alloc_offset >
  47		block_group->meta_write_pointer;
  48}
  49
  50/*
  51 * Return target flags in extended format or 0 if restripe for this chunk_type
  52 * is not in progress
  53 *
  54 * Should be called with balance_lock held
  55 */
  56static u64 get_restripe_target(const struct btrfs_fs_info *fs_info, u64 flags)
  57{
  58	const struct btrfs_balance_control *bctl = fs_info->balance_ctl;
  59	u64 target = 0;
  60
  61	if (!bctl)
  62		return 0;
  63
  64	if (flags & BTRFS_BLOCK_GROUP_DATA &&
  65	    bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  66		target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
  67	} else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
  68		   bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  69		target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
  70	} else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
  71		   bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  72		target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
  73	}
  74
  75	return target;
  76}
  77
  78/*
  79 * @flags: available profiles in extended format (see ctree.h)
  80 *
  81 * Return reduced profile in chunk format.  If profile changing is in progress
  82 * (either running or paused) picks the target profile (if it's already
  83 * available), otherwise falls back to plain reducing.
  84 */
  85static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
  86{
  87	u64 num_devices = fs_info->fs_devices->rw_devices;
  88	u64 target;
  89	u64 raid_type;
  90	u64 allowed = 0;
  91
  92	/*
  93	 * See if restripe for this chunk_type is in progress, if so try to
  94	 * reduce to the target profile
  95	 */
  96	spin_lock(&fs_info->balance_lock);
  97	target = get_restripe_target(fs_info, flags);
  98	if (target) {
  99		spin_unlock(&fs_info->balance_lock);
 100		return extended_to_chunk(target);
 101	}
 102	spin_unlock(&fs_info->balance_lock);
 103
 104	/* First, mask out the RAID levels which aren't possible */
 105	for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
 106		if (num_devices >= btrfs_raid_array[raid_type].devs_min)
 107			allowed |= btrfs_raid_array[raid_type].bg_flag;
 108	}
 109	allowed &= flags;
 110
 111	/* Select the highest-redundancy RAID level. */
 112	if (allowed & BTRFS_BLOCK_GROUP_RAID1C4)
 113		allowed = BTRFS_BLOCK_GROUP_RAID1C4;
 114	else if (allowed & BTRFS_BLOCK_GROUP_RAID6)
 115		allowed = BTRFS_BLOCK_GROUP_RAID6;
 116	else if (allowed & BTRFS_BLOCK_GROUP_RAID1C3)
 117		allowed = BTRFS_BLOCK_GROUP_RAID1C3;
 118	else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
 119		allowed = BTRFS_BLOCK_GROUP_RAID5;
 120	else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
 121		allowed = BTRFS_BLOCK_GROUP_RAID10;
 122	else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
 123		allowed = BTRFS_BLOCK_GROUP_RAID1;
 124	else if (allowed & BTRFS_BLOCK_GROUP_DUP)
 125		allowed = BTRFS_BLOCK_GROUP_DUP;
 126	else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
 127		allowed = BTRFS_BLOCK_GROUP_RAID0;
 128
 129	flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
 130
 131	return extended_to_chunk(flags | allowed);
 132}
 133
 134u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
 135{
 136	unsigned seq;
 137	u64 flags;
 138
 139	do {
 140		flags = orig_flags;
 141		seq = read_seqbegin(&fs_info->profiles_lock);
 142
 143		if (flags & BTRFS_BLOCK_GROUP_DATA)
 144			flags |= fs_info->avail_data_alloc_bits;
 145		else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 146			flags |= fs_info->avail_system_alloc_bits;
 147		else if (flags & BTRFS_BLOCK_GROUP_METADATA)
 148			flags |= fs_info->avail_metadata_alloc_bits;
 149	} while (read_seqretry(&fs_info->profiles_lock, seq));
 150
 151	return btrfs_reduce_alloc_profile(fs_info, flags);
 152}
 153
 154void btrfs_get_block_group(struct btrfs_block_group *cache)
 155{
 156	refcount_inc(&cache->refs);
 157}
 158
 159void btrfs_put_block_group(struct btrfs_block_group *cache)
 160{
 161	if (refcount_dec_and_test(&cache->refs)) {
 162		WARN_ON(cache->pinned > 0);
 163		/*
 164		 * If there was a failure to cleanup a log tree, very likely due
 165		 * to an IO failure on a writeback attempt of one or more of its
 166		 * extent buffers, we could not do proper (and cheap) unaccounting
 167		 * of their reserved space, so don't warn on reserved > 0 in that
 168		 * case.
 169		 */
 170		if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
 171		    !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
 172			WARN_ON(cache->reserved > 0);
 173
 174		/*
 175		 * A block_group shouldn't be on the discard_list anymore.
 176		 * Remove the block_group from the discard_list to prevent us
 177		 * from causing a panic due to NULL pointer dereference.
 178		 */
 179		if (WARN_ON(!list_empty(&cache->discard_list)))
 180			btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
 181						  cache);
 182
 183		kfree(cache->free_space_ctl);
 184		btrfs_free_chunk_map(cache->physical_map);
 185		kfree(cache);
 186	}
 187}
 188
 189static int btrfs_bg_start_cmp(const struct rb_node *new,
 190			      const struct rb_node *exist)
 191{
 192	const struct btrfs_block_group *new_bg =
 193		rb_entry(new, struct btrfs_block_group, cache_node);
 194	const struct btrfs_block_group *exist_bg =
 195		rb_entry(exist, struct btrfs_block_group, cache_node);
 196
 197	if (new_bg->start < exist_bg->start)
 198		return -1;
 199	if (new_bg->start > exist_bg->start)
 200		return 1;
 201	return 0;
 202}
 203
 204/*
 205 * This adds the block group to the fs_info rb tree for the block group cache
 206 */
 207static int btrfs_add_block_group_cache(struct btrfs_block_group *block_group)
 208{
 209	struct btrfs_fs_info *fs_info = block_group->fs_info;
 210	struct rb_node *exist;
 211	int ret = 0;
 212
 213	ASSERT(block_group->length != 0);
 214
 215	write_lock(&fs_info->block_group_cache_lock);
 216
 217	exist = rb_find_add_cached(&block_group->cache_node,
 218			&fs_info->block_group_cache_tree, btrfs_bg_start_cmp);
 219	if (exist)
 220		ret = -EEXIST;
 221	write_unlock(&fs_info->block_group_cache_lock);
 222
 223	return ret;
 224}
 225
 226/*
 227 * This will return the block group at or after bytenr if contains is 0, else
 228 * it will return the block group that contains the bytenr
 229 */
 230static struct btrfs_block_group *block_group_cache_tree_search(
 231		struct btrfs_fs_info *info, u64 bytenr, int contains)
 232{
 233	struct btrfs_block_group *cache, *ret = NULL;
 234	struct rb_node *n;
 235	u64 end, start;
 236
 237	read_lock(&info->block_group_cache_lock);
 238	n = info->block_group_cache_tree.rb_root.rb_node;
 239
 240	while (n) {
 241		cache = rb_entry(n, struct btrfs_block_group, cache_node);
 242		end = cache->start + cache->length - 1;
 243		start = cache->start;
 244
 245		if (bytenr < start) {
 246			if (!contains && (!ret || start < ret->start))
 247				ret = cache;
 248			n = n->rb_left;
 249		} else if (bytenr > start) {
 250			if (contains && bytenr <= end) {
 251				ret = cache;
 252				break;
 253			}
 254			n = n->rb_right;
 255		} else {
 256			ret = cache;
 257			break;
 258		}
 259	}
 260	if (ret)
 261		btrfs_get_block_group(ret);
 262	read_unlock(&info->block_group_cache_lock);
 263
 264	return ret;
 265}
 266
 267/*
 268 * Return the block group that starts at or after bytenr
 269 */
 270struct btrfs_block_group *btrfs_lookup_first_block_group(
 271		struct btrfs_fs_info *info, u64 bytenr)
 272{
 273	return block_group_cache_tree_search(info, bytenr, 0);
 274}
 275
 276/*
 277 * Return the block group that contains the given bytenr
 278 */
 279struct btrfs_block_group *btrfs_lookup_block_group(
 280		struct btrfs_fs_info *info, u64 bytenr)
 281{
 282	return block_group_cache_tree_search(info, bytenr, 1);
 283}
 284
 285struct btrfs_block_group *btrfs_next_block_group(
 286		struct btrfs_block_group *cache)
 287{
 288	struct btrfs_fs_info *fs_info = cache->fs_info;
 289	struct rb_node *node;
 290
 291	read_lock(&fs_info->block_group_cache_lock);
 292
 293	/* If our block group was removed, we need a full search. */
 294	if (RB_EMPTY_NODE(&cache->cache_node)) {
 295		const u64 next_bytenr = cache->start + cache->length;
 296
 297		read_unlock(&fs_info->block_group_cache_lock);
 298		btrfs_put_block_group(cache);
 299		return btrfs_lookup_first_block_group(fs_info, next_bytenr);
 300	}
 301	node = rb_next(&cache->cache_node);
 302	btrfs_put_block_group(cache);
 303	if (node) {
 304		cache = rb_entry(node, struct btrfs_block_group, cache_node);
 305		btrfs_get_block_group(cache);
 306	} else
 307		cache = NULL;
 308	read_unlock(&fs_info->block_group_cache_lock);
 309	return cache;
 310}
 311
 312/*
 313 * Check if we can do a NOCOW write for a given extent.
 314 *
 315 * @fs_info:       The filesystem information object.
 316 * @bytenr:        Logical start address of the extent.
 317 *
 318 * Check if we can do a NOCOW write for the given extent, and increments the
 319 * number of NOCOW writers in the block group that contains the extent, as long
 320 * as the block group exists and it's currently not in read-only mode.
 321 *
 322 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
 323 *          is responsible for calling btrfs_dec_nocow_writers() later.
 324 *
 325 *          Or NULL if we can not do a NOCOW write
 326 */
 327struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
 328						  u64 bytenr)
 329{
 330	struct btrfs_block_group *bg;
 331	bool can_nocow = true;
 332
 333	bg = btrfs_lookup_block_group(fs_info, bytenr);
 334	if (!bg)
 335		return NULL;
 336
 337	spin_lock(&bg->lock);
 338	if (bg->ro)
 339		can_nocow = false;
 340	else
 341		atomic_inc(&bg->nocow_writers);
 342	spin_unlock(&bg->lock);
 343
 344	if (!can_nocow) {
 345		btrfs_put_block_group(bg);
 346		return NULL;
 347	}
 348
 349	/* No put on block group, done by btrfs_dec_nocow_writers(). */
 350	return bg;
 351}
 352
 353/*
 354 * Decrement the number of NOCOW writers in a block group.
 355 *
 356 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
 357 * and on the block group returned by that call. Typically this is called after
 358 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
 359 * relocation.
 360 *
 361 * After this call, the caller should not use the block group anymore. It it wants
 362 * to use it, then it should get a reference on it before calling this function.
 363 */
 364void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
 365{
 366	if (atomic_dec_and_test(&bg->nocow_writers))
 367		wake_up_var(&bg->nocow_writers);
 368
 369	/* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
 370	btrfs_put_block_group(bg);
 371}
 372
 373void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
 374{
 375	wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
 376}
 377
 378void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
 379					const u64 start)
 380{
 381	struct btrfs_block_group *bg;
 382
 383	bg = btrfs_lookup_block_group(fs_info, start);
 384	ASSERT(bg);
 385	if (atomic_dec_and_test(&bg->reservations))
 386		wake_up_var(&bg->reservations);
 387	btrfs_put_block_group(bg);
 388}
 389
 390void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
 391{
 392	struct btrfs_space_info *space_info = bg->space_info;
 393
 394	ASSERT(bg->ro);
 395
 396	if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
 397		return;
 398
 399	/*
 400	 * Our block group is read only but before we set it to read only,
 401	 * some task might have had allocated an extent from it already, but it
 402	 * has not yet created a respective ordered extent (and added it to a
 403	 * root's list of ordered extents).
 404	 * Therefore wait for any task currently allocating extents, since the
 405	 * block group's reservations counter is incremented while a read lock
 406	 * on the groups' semaphore is held and decremented after releasing
 407	 * the read access on that semaphore and creating the ordered extent.
 408	 */
 409	down_write(&space_info->groups_sem);
 410	up_write(&space_info->groups_sem);
 411
 412	wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
 413}
 414
 415struct btrfs_caching_control *btrfs_get_caching_control(
 416		struct btrfs_block_group *cache)
 417{
 418	struct btrfs_caching_control *ctl;
 419
 420	spin_lock(&cache->lock);
 421	if (!cache->caching_ctl) {
 422		spin_unlock(&cache->lock);
 423		return NULL;
 424	}
 425
 426	ctl = cache->caching_ctl;
 427	refcount_inc(&ctl->count);
 428	spin_unlock(&cache->lock);
 429	return ctl;
 430}
 431
 432static void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
 433{
 434	if (refcount_dec_and_test(&ctl->count))
 435		kfree(ctl);
 436}
 437
 438/*
 439 * When we wait for progress in the block group caching, its because our
 440 * allocation attempt failed at least once.  So, we must sleep and let some
 441 * progress happen before we try again.
 442 *
 443 * This function will sleep at least once waiting for new free space to show
 444 * up, and then it will check the block group free space numbers for our min
 445 * num_bytes.  Another option is to have it go ahead and look in the rbtree for
 446 * a free extent of a given size, but this is a good start.
 447 *
 448 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
 449 * any of the information in this block group.
 450 */
 451void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
 452					   u64 num_bytes)
 453{
 454	struct btrfs_caching_control *caching_ctl;
 455	int progress;
 456
 457	caching_ctl = btrfs_get_caching_control(cache);
 458	if (!caching_ctl)
 459		return;
 460
 461	/*
 462	 * We've already failed to allocate from this block group, so even if
 463	 * there's enough space in the block group it isn't contiguous enough to
 464	 * allow for an allocation, so wait for at least the next wakeup tick,
 465	 * or for the thing to be done.
 466	 */
 467	progress = atomic_read(&caching_ctl->progress);
 468
 469	wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
 470		   (progress != atomic_read(&caching_ctl->progress) &&
 471		    (cache->free_space_ctl->free_space >= num_bytes)));
 472
 473	btrfs_put_caching_control(caching_ctl);
 474}
 475
 476static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
 477				       struct btrfs_caching_control *caching_ctl)
 478{
 479	wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
 480	return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
 481}
 482
 483static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
 484{
 485	struct btrfs_caching_control *caching_ctl;
 486	int ret;
 487
 488	caching_ctl = btrfs_get_caching_control(cache);
 489	if (!caching_ctl)
 490		return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
 491	ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
 492	btrfs_put_caching_control(caching_ctl);
 493	return ret;
 494}
 495
 496#ifdef CONFIG_BTRFS_DEBUG
 497static void fragment_free_space(struct btrfs_block_group *block_group)
 498{
 499	struct btrfs_fs_info *fs_info = block_group->fs_info;
 500	u64 start = block_group->start;
 501	u64 len = block_group->length;
 502	u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 503		fs_info->nodesize : fs_info->sectorsize;
 504	u64 step = chunk << 1;
 505
 506	while (len > chunk) {
 507		btrfs_remove_free_space(block_group, start, chunk);
 508		start += step;
 509		if (len < step)
 510			len = 0;
 511		else
 512			len -= step;
 513	}
 514}
 515#endif
 516
 517/*
 518 * Add a free space range to the in memory free space cache of a block group.
 519 * This checks if the range contains super block locations and any such
 520 * locations are not added to the free space cache.
 521 *
 522 * @block_group:      The target block group.
 523 * @start:            Start offset of the range.
 524 * @end:              End offset of the range (exclusive).
 525 * @total_added_ret:  Optional pointer to return the total amount of space
 526 *                    added to the block group's free space cache.
 527 *
 528 * Returns 0 on success or < 0 on error.
 529 */
 530int btrfs_add_new_free_space(struct btrfs_block_group *block_group, u64 start,
 531			     u64 end, u64 *total_added_ret)
 532{
 533	struct btrfs_fs_info *info = block_group->fs_info;
 534	u64 extent_start, extent_end, size;
 535	int ret;
 536
 537	if (total_added_ret)
 538		*total_added_ret = 0;
 539
 540	while (start < end) {
 541		if (!btrfs_find_first_extent_bit(&info->excluded_extents, start,
 542						 &extent_start, &extent_end,
 543						 EXTENT_DIRTY, NULL))
 544			break;
 545
 546		if (extent_start <= start) {
 547			start = extent_end + 1;
 548		} else if (extent_start > start && extent_start < end) {
 549			size = extent_start - start;
 550			ret = btrfs_add_free_space_async_trimmed(block_group,
 551								 start, size);
 552			if (ret)
 553				return ret;
 554			if (total_added_ret)
 555				*total_added_ret += size;
 556			start = extent_end + 1;
 557		} else {
 558			break;
 559		}
 560	}
 561
 562	if (start < end) {
 563		size = end - start;
 564		ret = btrfs_add_free_space_async_trimmed(block_group, start,
 565							 size);
 566		if (ret)
 567			return ret;
 568		if (total_added_ret)
 569			*total_added_ret += size;
 570	}
 571
 572	return 0;
 573}
 574
 575/*
 576 * Get an arbitrary extent item index / max_index through the block group
 577 *
 578 * @block_group   the block group to sample from
 579 * @index:        the integral step through the block group to grab from
 580 * @max_index:    the granularity of the sampling
 581 * @key:          return value parameter for the item we find
 582 *
 583 * Pre-conditions on indices:
 584 * 0 <= index <= max_index
 585 * 0 < max_index
 586 *
 587 * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
 588 * error code on error.
 589 */
 590static int sample_block_group_extent_item(struct btrfs_caching_control *caching_ctl,
 591					  struct btrfs_block_group *block_group,
 592					  int index, int max_index,
 593					  struct btrfs_key *found_key)
 594{
 595	struct btrfs_fs_info *fs_info = block_group->fs_info;
 596	struct btrfs_root *extent_root;
 597	u64 search_offset;
 598	u64 search_end = block_group->start + block_group->length;
 599	BTRFS_PATH_AUTO_FREE(path);
 600	struct btrfs_key search_key;
 601	int ret = 0;
 602
 603	ASSERT(index >= 0);
 604	ASSERT(index <= max_index);
 605	ASSERT(max_index > 0);
 606	lockdep_assert_held(&caching_ctl->mutex);
 607	lockdep_assert_held_read(&fs_info->commit_root_sem);
 608
 609	path = btrfs_alloc_path();
 610	if (!path)
 611		return -ENOMEM;
 612
 613	extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
 614						       BTRFS_SUPER_INFO_OFFSET));
 615
 616	path->skip_locking = true;
 617	path->search_commit_root = true;
 618	path->reada = READA_FORWARD;
 619
 620	search_offset = index * div_u64(block_group->length, max_index);
 621	search_key.objectid = block_group->start + search_offset;
 622	search_key.type = BTRFS_EXTENT_ITEM_KEY;
 623	search_key.offset = 0;
 624
 625	btrfs_for_each_slot(extent_root, &search_key, found_key, path, ret) {
 626		/* Success; sampled an extent item in the block group */
 627		if (found_key->type == BTRFS_EXTENT_ITEM_KEY &&
 628		    found_key->objectid >= block_group->start &&
 629		    found_key->objectid + found_key->offset <= search_end)
 630			break;
 631
 632		/* We can't possibly find a valid extent item anymore */
 633		if (found_key->objectid >= search_end) {
 634			ret = 1;
 635			break;
 636		}
 637	}
 638
 639	lockdep_assert_held(&caching_ctl->mutex);
 640	lockdep_assert_held_read(&fs_info->commit_root_sem);
 641	return ret;
 642}
 643
 644/*
 645 * Best effort attempt to compute a block group's size class while caching it.
 646 *
 647 * @block_group: the block group we are caching
 648 *
 649 * We cannot infer the size class while adding free space extents, because that
 650 * logic doesn't care about contiguous file extents (it doesn't differentiate
 651 * between a 100M extent and 100 contiguous 1M extents). So we need to read the
 652 * file extent items. Reading all of them is quite wasteful, because usually
 653 * only a handful are enough to give a good answer. Therefore, we just grab 5 of
 654 * them at even steps through the block group and pick the smallest size class
 655 * we see. Since size class is best effort, and not guaranteed in general,
 656 * inaccuracy is acceptable.
 657 *
 658 * To be more explicit about why this algorithm makes sense:
 659 *
 660 * If we are caching in a block group from disk, then there are three major cases
 661 * to consider:
 662 * 1. the block group is well behaved and all extents in it are the same size
 663 *    class.
 664 * 2. the block group is mostly one size class with rare exceptions for last
 665 *    ditch allocations
 666 * 3. the block group was populated before size classes and can have a totally
 667 *    arbitrary mix of size classes.
 668 *
 669 * In case 1, looking at any extent in the block group will yield the correct
 670 * result. For the mixed cases, taking the minimum size class seems like a good
 671 * approximation, since gaps from frees will be usable to the size class. For
 672 * 2., a small handful of file extents is likely to yield the right answer. For
 673 * 3, we can either read every file extent, or admit that this is best effort
 674 * anyway and try to stay fast.
 675 *
 676 * Returns: 0 on success, negative error code on error.
 677 */
 678static int load_block_group_size_class(struct btrfs_caching_control *caching_ctl,
 679				       struct btrfs_block_group *block_group)
 680{
 681	struct btrfs_fs_info *fs_info = block_group->fs_info;
 682	struct btrfs_key key;
 683	int i;
 684	u64 min_size = block_group->length;
 685	enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
 686	int ret;
 687
 688	if (!btrfs_block_group_should_use_size_class(block_group))
 689		return 0;
 690
 691	lockdep_assert_held(&caching_ctl->mutex);
 692	lockdep_assert_held_read(&fs_info->commit_root_sem);
 693	for (i = 0; i < 5; ++i) {
 694		ret = sample_block_group_extent_item(caching_ctl, block_group, i, 5, &key);
 695		if (ret < 0)
 696			goto out;
 697		if (ret > 0)
 698			continue;
 699		min_size = min_t(u64, min_size, key.offset);
 700		size_class = btrfs_calc_block_group_size_class(min_size);
 701	}
 702	if (size_class != BTRFS_BG_SZ_NONE) {
 703		spin_lock(&block_group->lock);
 704		block_group->size_class = size_class;
 705		spin_unlock(&block_group->lock);
 706	}
 707out:
 708	return ret;
 709}
 710
 711static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 712{
 713	struct btrfs_block_group *block_group = caching_ctl->block_group;
 714	struct btrfs_fs_info *fs_info = block_group->fs_info;
 715	struct btrfs_root *extent_root;
 716	BTRFS_PATH_AUTO_FREE(path);
 717	struct extent_buffer *leaf;
 718	struct btrfs_key key;
 719	u64 total_found = 0;
 720	u64 last = 0;
 721	u32 nritems;
 722	int ret;
 723	bool wakeup = true;
 724
 725	path = btrfs_alloc_path();
 726	if (!path)
 727		return -ENOMEM;
 728
 729	last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
 730	extent_root = btrfs_extent_root(fs_info, last);
 731
 732#ifdef CONFIG_BTRFS_DEBUG
 733	/*
 734	 * If we're fragmenting we don't want to make anybody think we can
 735	 * allocate from this block group until we've had a chance to fragment
 736	 * the free space.
 737	 */
 738	if (btrfs_should_fragment_free_space(block_group))
 739		wakeup = false;
 740#endif
 741	/*
 742	 * We don't want to deadlock with somebody trying to allocate a new
 743	 * extent for the extent root while also trying to search the extent
 744	 * root to add free space.  So we skip locking and search the commit
 745	 * root, since its read-only
 746	 */
 747	path->skip_locking = true;
 748	path->search_commit_root = true;
 749	path->reada = READA_FORWARD;
 750
 751	key.objectid = last;
 752	key.type = BTRFS_EXTENT_ITEM_KEY;
 753	key.offset = 0;
 754
 755next:
 756	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 757	if (ret < 0)
 758		goto out;
 759
 760	leaf = path->nodes[0];
 761	nritems = btrfs_header_nritems(leaf);
 762
 763	while (1) {
 764		if (btrfs_fs_closing(fs_info) > 1) {
 765			last = (u64)-1;
 766			break;
 767		}
 768
 769		if (path->slots[0] < nritems) {
 770			btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 771		} else {
 772			ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
 773			if (ret)
 774				break;
 775
 776			if (need_resched() ||
 777			    rwsem_is_contended(&fs_info->commit_root_sem)) {
 778				btrfs_release_path(path);
 779				up_read(&fs_info->commit_root_sem);
 780				mutex_unlock(&caching_ctl->mutex);
 781				cond_resched();
 782				mutex_lock(&caching_ctl->mutex);
 783				down_read(&fs_info->commit_root_sem);
 784				goto next;
 785			}
 786
 787			ret = btrfs_next_leaf(extent_root, path);
 788			if (ret < 0)
 789				goto out;
 790			if (ret)
 791				break;
 792			leaf = path->nodes[0];
 793			nritems = btrfs_header_nritems(leaf);
 794			continue;
 795		}
 796
 797		if (key.objectid < last) {
 798			key.objectid = last;
 799			key.type = BTRFS_EXTENT_ITEM_KEY;
 800			key.offset = 0;
 801			btrfs_release_path(path);
 802			goto next;
 803		}
 804
 805		if (key.objectid < block_group->start) {
 806			path->slots[0]++;
 807			continue;
 808		}
 809
 810		if (key.objectid >= block_group->start + block_group->length)
 811			break;
 812
 813		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 814		    key.type == BTRFS_METADATA_ITEM_KEY) {
 815			u64 space_added;
 816
 817			ret = btrfs_add_new_free_space(block_group, last,
 818						       key.objectid, &space_added);
 819			if (ret)
 820				goto out;
 821			total_found += space_added;
 822			if (key.type == BTRFS_METADATA_ITEM_KEY)
 823				last = key.objectid +
 824					fs_info->nodesize;
 825			else
 826				last = key.objectid + key.offset;
 827
 828			if (total_found > CACHING_CTL_WAKE_UP) {
 829				total_found = 0;
 830				if (wakeup) {
 831					atomic_inc(&caching_ctl->progress);
 832					wake_up(&caching_ctl->wait);
 833				}
 834			}
 835		}
 836		path->slots[0]++;
 837	}
 838
 839	ret = btrfs_add_new_free_space(block_group, last,
 840				       block_group->start + block_group->length,
 841				       NULL);
 842out:
 843	return ret;
 844}
 845
 846static inline void btrfs_free_excluded_extents(const struct btrfs_block_group *bg)
 847{
 848	btrfs_clear_extent_bit(&bg->fs_info->excluded_extents, bg->start,
 849			       bg->start + bg->length - 1, EXTENT_DIRTY, NULL);
 850}
 851
 852static noinline void caching_thread(struct btrfs_work *work)
 853{
 854	struct btrfs_block_group *block_group;
 855	struct btrfs_fs_info *fs_info;
 856	struct btrfs_caching_control *caching_ctl;
 857	int ret;
 858
 859	caching_ctl = container_of(work, struct btrfs_caching_control, work);
 860	block_group = caching_ctl->block_group;
 861	fs_info = block_group->fs_info;
 862
 863	mutex_lock(&caching_ctl->mutex);
 864	down_read(&fs_info->commit_root_sem);
 865
 866	load_block_group_size_class(caching_ctl, block_group);
 867	if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 868		ret = load_free_space_cache(block_group);
 869		if (ret == 1) {
 870			ret = 0;
 871			goto done;
 872		}
 873
 874		/*
 875		 * We failed to load the space cache, set ourselves to
 876		 * CACHE_STARTED and carry on.
 877		 */
 878		spin_lock(&block_group->lock);
 879		block_group->cached = BTRFS_CACHE_STARTED;
 880		spin_unlock(&block_group->lock);
 881		wake_up(&caching_ctl->wait);
 882	}
 883
 884	/*
 885	 * If we are in the transaction that populated the free space tree we
 886	 * can't actually cache from the free space tree as our commit root and
 887	 * real root are the same, so we could change the contents of the blocks
 888	 * while caching.  Instead do the slow caching in this case, and after
 889	 * the transaction has committed we will be safe.
 890	 */
 891	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
 892	    !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
 893		ret = btrfs_load_free_space_tree(caching_ctl);
 894	else
 895		ret = load_extent_tree_free(caching_ctl);
 896done:
 897	spin_lock(&block_group->lock);
 898	block_group->caching_ctl = NULL;
 899	block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 900	spin_unlock(&block_group->lock);
 901
 902#ifdef CONFIG_BTRFS_DEBUG
 903	if (btrfs_should_fragment_free_space(block_group)) {
 904		u64 bytes_used;
 905
 906		spin_lock(&block_group->space_info->lock);
 907		spin_lock(&block_group->lock);
 908		bytes_used = block_group->length - block_group->used;
 909		block_group->space_info->bytes_used += bytes_used >> 1;
 910		spin_unlock(&block_group->lock);
 911		spin_unlock(&block_group->space_info->lock);
 912		fragment_free_space(block_group);
 913	}
 914#endif
 915
 916	up_read(&fs_info->commit_root_sem);
 917	btrfs_free_excluded_extents(block_group);
 918	mutex_unlock(&caching_ctl->mutex);
 919
 920	wake_up(&caching_ctl->wait);
 921
 922	btrfs_put_caching_control(caching_ctl);
 923	btrfs_put_block_group(block_group);
 924}
 925
 926int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
 927{
 928	struct btrfs_fs_info *fs_info = cache->fs_info;
 929	struct btrfs_caching_control *caching_ctl = NULL;
 930	int ret = 0;
 931
 932	/* Allocator for zoned filesystems does not use the cache at all */
 933	if (btrfs_is_zoned(fs_info))
 934		return 0;
 935
 936	caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 937	if (!caching_ctl)
 938		return -ENOMEM;
 939
 940	INIT_LIST_HEAD(&caching_ctl->list);
 941	mutex_init(&caching_ctl->mutex);
 942	init_waitqueue_head(&caching_ctl->wait);
 943	caching_ctl->block_group = cache;
 944	refcount_set(&caching_ctl->count, 2);
 945	atomic_set(&caching_ctl->progress, 0);
 946	btrfs_init_work(&caching_ctl->work, caching_thread, NULL);
 947
 948	spin_lock(&cache->lock);
 949	if (cache->cached != BTRFS_CACHE_NO) {
 950		kfree(caching_ctl);
 951
 952		caching_ctl = cache->caching_ctl;
 953		if (caching_ctl)
 954			refcount_inc(&caching_ctl->count);
 955		spin_unlock(&cache->lock);
 956		goto out;
 957	}
 958	WARN_ON(cache->caching_ctl);
 959	cache->caching_ctl = caching_ctl;
 960	cache->cached = BTRFS_CACHE_STARTED;
 961	spin_unlock(&cache->lock);
 962
 963	write_lock(&fs_info->block_group_cache_lock);
 964	refcount_inc(&caching_ctl->count);
 965	list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 966	write_unlock(&fs_info->block_group_cache_lock);
 967
 968	btrfs_get_block_group(cache);
 969
 970	btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 971out:
 972	if (wait && caching_ctl)
 973		ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
 974	if (caching_ctl)
 975		btrfs_put_caching_control(caching_ctl);
 976
 977	return ret;
 978}
 979
 980static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
 981{
 982	u64 extra_flags = chunk_to_extended(flags) &
 983				BTRFS_EXTENDED_PROFILE_MASK;
 984
 985	write_seqlock(&fs_info->profiles_lock);
 986	if (flags & BTRFS_BLOCK_GROUP_DATA)
 987		fs_info->avail_data_alloc_bits &= ~extra_flags;
 988	if (flags & BTRFS_BLOCK_GROUP_METADATA)
 989		fs_info->avail_metadata_alloc_bits &= ~extra_flags;
 990	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 991		fs_info->avail_system_alloc_bits &= ~extra_flags;
 992	write_sequnlock(&fs_info->profiles_lock);
 993}
 994
 995/*
 996 * Clear incompat bits for the following feature(s):
 997 *
 998 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
 999 *            in the whole filesystem
1000 *
1001 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
1002 */
1003static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
1004{
1005	bool found_raid56 = false;
1006	bool found_raid1c34 = false;
1007
1008	if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
1009	    (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
1010	    (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
1011		struct list_head *head = &fs_info->space_info;
1012		struct btrfs_space_info *sinfo;
1013
1014		list_for_each_entry_rcu(sinfo, head, list) {
1015			down_read(&sinfo->groups_sem);
1016			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
1017				found_raid56 = true;
1018			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
1019				found_raid56 = true;
1020			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
1021				found_raid1c34 = true;
1022			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
1023				found_raid1c34 = true;
1024			up_read(&sinfo->groups_sem);
1025		}
1026		if (!found_raid56)
1027			btrfs_clear_fs_incompat(fs_info, RAID56);
1028		if (!found_raid1c34)
1029			btrfs_clear_fs_incompat(fs_info, RAID1C34);
1030	}
1031}
1032
1033static struct btrfs_root *btrfs_block_group_root(struct btrfs_fs_info *fs_info)
1034{
1035	if (btrfs_fs_compat_ro(fs_info, BLOCK_GROUP_TREE))
1036		return fs_info->block_group_root;
1037	return btrfs_extent_root(fs_info, 0);
1038}
1039
1040static int remove_block_group_item(struct btrfs_trans_handle *trans,
1041				   struct btrfs_path *path,
1042				   struct btrfs_block_group *block_group)
1043{
1044	struct btrfs_fs_info *fs_info = trans->fs_info;
1045	struct btrfs_root *root;
1046	struct btrfs_key key;
1047	int ret;
1048
1049	root = btrfs_block_group_root(fs_info);
1050	key.objectid = block_group->start;
1051	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1052	key.offset = block_group->length;
1053
1054	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1055	if (ret > 0)
1056		ret = -ENOENT;
1057	if (ret < 0)
1058		return ret;
1059
1060	ret = btrfs_del_item(trans, root, path);
1061	return ret;
1062}
1063
1064int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1065			     struct btrfs_chunk_map *map)
1066{
1067	struct btrfs_fs_info *fs_info = trans->fs_info;
1068	BTRFS_PATH_AUTO_FREE(path);
1069	struct btrfs_block_group *block_group;
1070	struct btrfs_free_cluster *cluster;
1071	struct inode *inode;
1072	struct kobject *kobj = NULL;
1073	int ret;
1074	int index;
1075	int factor;
1076	struct btrfs_caching_control *caching_ctl = NULL;
1077	bool remove_map;
1078	bool remove_rsv = false;
1079
1080	block_group = btrfs_lookup_block_group(fs_info, map->start);
1081	if (!block_group)
1082		return -ENOENT;
1083
1084	BUG_ON(!block_group->ro);
1085
1086	trace_btrfs_remove_block_group(block_group);
1087	/*
1088	 * Free the reserved super bytes from this block group before
1089	 * remove it.
1090	 */
1091	btrfs_free_excluded_extents(block_group);
1092	btrfs_free_ref_tree_range(fs_info, block_group->start,
1093				  block_group->length);
1094
1095	index = btrfs_bg_flags_to_raid_index(block_group->flags);
1096	factor = btrfs_bg_type_to_factor(block_group->flags);
1097
1098	/* make sure this block group isn't part of an allocation cluster */
1099	cluster = &fs_info->data_alloc_cluster;
1100	spin_lock(&cluster->refill_lock);
1101	btrfs_return_cluster_to_free_space(block_group, cluster);
1102	spin_unlock(&cluster->refill_lock);
1103
1104	/*
1105	 * make sure this block group isn't part of a metadata
1106	 * allocation cluster
1107	 */
1108	cluster = &fs_info->meta_alloc_cluster;
1109	spin_lock(&cluster->refill_lock);
1110	btrfs_return_cluster_to_free_space(block_group, cluster);
1111	spin_unlock(&cluster->refill_lock);
1112
1113	btrfs_clear_treelog_bg(block_group);
1114	btrfs_clear_data_reloc_bg(block_group);
1115
1116	path = btrfs_alloc_path();
1117	if (!path) {
1118		ret = -ENOMEM;
1119		goto out;
1120	}
1121
1122	/*
1123	 * get the inode first so any iput calls done for the io_list
1124	 * aren't the final iput (no unlinks allowed now)
1125	 */
1126	inode = lookup_free_space_inode(block_group, path);
1127
1128	mutex_lock(&trans->transaction->cache_write_mutex);
1129	/*
1130	 * Make sure our free space cache IO is done before removing the
1131	 * free space inode
1132	 */
1133	spin_lock(&trans->transaction->dirty_bgs_lock);
1134	if (!list_empty(&block_group->io_list)) {
1135		list_del_init(&block_group->io_list);
1136
1137		WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1138
1139		spin_unlock(&trans->transaction->dirty_bgs_lock);
1140		btrfs_wait_cache_io(trans, block_group, path);
1141		btrfs_put_block_group(block_group);
1142		spin_lock(&trans->transaction->dirty_bgs_lock);
1143	}
1144
1145	if (!list_empty(&block_group->dirty_list)) {
1146		list_del_init(&block_group->dirty_list);
1147		remove_rsv = true;
1148		btrfs_put_block_group(block_group);
1149	}
1150	spin_unlock(&trans->transaction->dirty_bgs_lock);
1151	mutex_unlock(&trans->transaction->cache_write_mutex);
1152
1153	ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1154	if (ret)
1155		goto out;
1156
1157	write_lock(&fs_info->block_group_cache_lock);
1158	rb_erase_cached(&block_group->cache_node,
1159			&fs_info->block_group_cache_tree);
1160	RB_CLEAR_NODE(&block_group->cache_node);
1161
1162	/* Once for the block groups rbtree */
1163	btrfs_put_block_group(block_group);
1164
1165	write_unlock(&fs_info->block_group_cache_lock);
1166
1167	down_write(&block_group->space_info->groups_sem);
1168	/*
1169	 * we must use list_del_init so people can check to see if they
1170	 * are still on the list after taking the semaphore
1171	 */
1172	list_del_init(&block_group->list);
1173	if (list_empty(&block_group->space_info->block_groups[index])) {
1174		kobj = block_group->space_info->block_group_kobjs[index];
1175		block_group->space_info->block_group_kobjs[index] = NULL;
1176		clear_avail_alloc_bits(fs_info, block_group->flags);
1177	}
1178	up_write(&block_group->space_info->groups_sem);
1179	clear_incompat_bg_bits(fs_info, block_group->flags);
1180	if (kobj) {
1181		kobject_del(kobj);
1182		kobject_put(kobj);
1183	}
1184
1185	if (block_group->cached == BTRFS_CACHE_STARTED)
1186		btrfs_wait_block_group_cache_done(block_group);
1187
1188	write_lock(&fs_info->block_group_cache_lock);
1189	caching_ctl = btrfs_get_caching_control(block_group);
1190	if (!caching_ctl) {
1191		struct btrfs_caching_control *ctl;
1192
1193		list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1194			if (ctl->block_group == block_group) {
1195				caching_ctl = ctl;
1196				refcount_inc(&caching_ctl->count);
1197				break;
1198			}
1199		}
1200	}
1201	if (caching_ctl)
1202		list_del_init(&caching_ctl->list);
1203	write_unlock(&fs_info->block_group_cache_lock);
1204
1205	if (caching_ctl) {
1206		/* Once for the caching bgs list and once for us. */
1207		btrfs_put_caching_control(caching_ctl);
1208		btrfs_put_caching_control(caching_ctl);
1209	}
1210
1211	spin_lock(&trans->transaction->dirty_bgs_lock);
1212	WARN_ON(!list_empty(&block_group->dirty_list));
1213	WARN_ON(!list_empty(&block_group->io_list));
1214	spin_unlock(&trans->transaction->dirty_bgs_lock);
1215
1216	btrfs_remove_free_space_cache(block_group);
1217
1218	spin_lock(&block_group->space_info->lock);
1219	list_del_init(&block_group->ro_list);
1220
1221	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1222		WARN_ON(block_group->space_info->total_bytes
1223			< block_group->length);
1224		WARN_ON(block_group->space_info->bytes_readonly
1225			< block_group->length - block_group->zone_unusable);
1226		WARN_ON(block_group->space_info->bytes_zone_unusable
1227			< block_group->zone_unusable);
1228		WARN_ON(block_group->space_info->disk_total
1229			< block_group->length * factor);
1230	}
1231	block_group->space_info->total_bytes -= block_group->length;
1232	block_group->space_info->bytes_readonly -=
1233		(block_group->length - block_group->zone_unusable);
1234	btrfs_space_info_update_bytes_zone_unusable(block_group->space_info,
1235						    -block_group->zone_unusable);
1236	block_group->space_info->disk_total -= block_group->length * factor;
1237
1238	spin_unlock(&block_group->space_info->lock);
1239
1240	/*
1241	 * Remove the free space for the block group from the free space tree
1242	 * and the block group's item from the extent tree before marking the
1243	 * block group as removed. This is to prevent races with tasks that
1244	 * freeze and unfreeze a block group, this task and another task
1245	 * allocating a new block group - the unfreeze task ends up removing
1246	 * the block group's extent map before the task calling this function
1247	 * deletes the block group item from the extent tree, allowing for
1248	 * another task to attempt to create another block group with the same
1249	 * item key (and failing with -EEXIST and a transaction abort).
1250	 */
1251	ret = btrfs_remove_block_group_free_space(trans, block_group);
1252	if (ret)
1253		goto out;
1254
1255	ret = remove_block_group_item(trans, path, block_group);
1256	if (ret < 0)
1257		goto out;
1258
1259	spin_lock(&block_group->lock);
1260	/*
1261	 * Hitting this WARN means we removed a block group with an unwritten
1262	 * region. It will cause "unable to find chunk map for logical" errors.
1263	 */
1264	if (WARN_ON(has_unwritten_metadata(block_group)))
1265		btrfs_warn(fs_info,
1266			   "block group %llu is removed before metadata write out",
1267			   block_group->start);
1268
1269	set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1270
1271	/*
1272	 * At this point trimming or scrub can't start on this block group,
1273	 * because we removed the block group from the rbtree
1274	 * fs_info->block_group_cache_tree so no one can't find it anymore and
1275	 * even if someone already got this block group before we removed it
1276	 * from the rbtree, they have already incremented block_group->frozen -
1277	 * if they didn't, for the trimming case they won't find any free space
1278	 * entries because we already removed them all when we called
1279	 * btrfs_remove_free_space_cache().
1280	 *
1281	 * And we must not remove the chunk map from the fs_info->mapping_tree
1282	 * to prevent the same logical address range and physical device space
1283	 * ranges from being reused for a new block group. This is needed to
1284	 * avoid races with trimming and scrub.
1285	 *
1286	 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1287	 * completely transactionless, so while it is trimming a range the
1288	 * currently running transaction might finish and a new one start,
1289	 * allowing for new block groups to be created that can reuse the same
1290	 * physical device locations unless we take this special care.
1291	 *
1292	 * There may also be an implicit trim operation if the file system
1293	 * is mounted with -odiscard. The same protections must remain
1294	 * in place until the extents have been discarded completely when
1295	 * the transaction commit has completed.
1296	 */
1297	remove_map = (atomic_read(&block_group->frozen) == 0);
1298	spin_unlock(&block_group->lock);
1299
1300	if (remove_map)
1301		btrfs_remove_chunk_map(fs_info, map);
1302
1303out:
1304	/* Once for the lookup reference */
1305	btrfs_put_block_group(block_group);
1306	if (remove_rsv)
1307		btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
1308	return ret;
1309}
1310
1311struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1312		struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1313{
1314	struct btrfs_root *root = btrfs_block_group_root(fs_info);
1315	struct btrfs_chunk_map *map;
1316	unsigned int num_items;
1317
1318	map = btrfs_find_chunk_map(fs_info, chunk_offset, 1);
1319	ASSERT(map != NULL);
1320	ASSERT(map->start == chunk_offset);
1321
1322	/*
1323	 * We need to reserve 3 + N units from the metadata space info in order
1324	 * to remove a block group (done at btrfs_remove_chunk() and at
1325	 * btrfs_remove_block_group()), which are used for:
1326	 *
1327	 * 1 unit for adding the free space inode's orphan (located in the tree
1328	 * of tree roots).
1329	 * 1 unit for deleting the block group item (located in the extent
1330	 * tree).
1331	 * 1 unit for deleting the free space item (located in tree of tree
1332	 * roots).
1333	 * N units for deleting N device extent items corresponding to each
1334	 * stripe (located in the device tree).
1335	 *
1336	 * In order to remove a block group we also need to reserve units in the
1337	 * system space info in order to update the chunk tree (update one or
1338	 * more device items and remove one chunk item), but this is done at
1339	 * btrfs_remove_chunk() through a call to check_system_chunk().
1340	 */
1341	num_items = 3 + map->num_stripes;
1342	btrfs_free_chunk_map(map);
1343
1344	return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1345}
1346
1347/*
1348 * Mark block group @cache read-only, so later write won't happen to block
1349 * group @cache.
1350 *
1351 * If @force is not set, this function will only mark the block group readonly
1352 * if we have enough free space (1M) in other metadata/system block groups.
1353 * If @force is not set, this function will mark the block group readonly
1354 * without checking free space.
1355 *
1356 * NOTE: This function doesn't care if other block groups can contain all the
1357 * data in this block group. That check should be done by relocation routine,
1358 * not this function.
1359 */
1360static int inc_block_group_ro(struct btrfs_block_group *cache, bool force)
1361{
1362	struct btrfs_space_info *sinfo = cache->space_info;
1363	u64 num_bytes;
1364	int ret = -ENOSPC;
1365
1366	spin_lock(&sinfo->lock);
1367	spin_lock(&cache->lock);
1368
1369	if (cache->swap_extents) {
1370		ret = -ETXTBSY;
1371		goto out;
1372	}
1373
1374	if (cache->ro) {
1375		cache->ro++;
1376		ret = 0;
1377		goto out;
1378	}
1379
1380	num_bytes = cache->length - cache->reserved - cache->pinned -
1381		    cache->bytes_super - cache->zone_unusable - cache->used;
1382
1383	/*
1384	 * Data never overcommits, even in mixed mode, so do just the straight
1385	 * check of left over space in how much we have allocated.
1386	 */
1387	if (force) {
1388		ret = 0;
1389	} else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1390		u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1391
1392		/*
1393		 * Here we make sure if we mark this bg RO, we still have enough
1394		 * free space as buffer.
1395		 */
1396		if (sinfo_used + num_bytes <= sinfo->total_bytes)
1397			ret = 0;
1398	} else {
1399		/*
1400		 * We overcommit metadata, so we need to do the
1401		 * btrfs_can_overcommit check here, and we need to pass in
1402		 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1403		 * leeway to allow us to mark this block group as read only.
1404		 */
1405		if (btrfs_can_overcommit(sinfo, num_bytes, BTRFS_RESERVE_NO_FLUSH))
1406			ret = 0;
1407	}
1408
1409	if (!ret) {
1410		sinfo->bytes_readonly += num_bytes;
1411		if (btrfs_is_zoned(cache->fs_info)) {
1412			/* Migrate zone_unusable bytes to readonly */
1413			sinfo->bytes_readonly += cache->zone_unusable;
1414			btrfs_space_info_update_bytes_zone_unusable(sinfo, -cache->zone_unusable);
1415			cache->zone_unusable = 0;
1416		}
1417		cache->ro++;
1418		list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1419	}
1420out:
1421	spin_unlock(&cache->lock);
1422	spin_unlock(&sinfo->lock);
1423	if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1424		btrfs_info(cache->fs_info,
1425			"unable to make block group %llu ro", cache->start);
1426		btrfs_dump_space_info(cache->space_info, 0, false);
1427	}
1428	return ret;
1429}
1430
1431static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1432				 const struct btrfs_block_group *bg)
1433{
1434	struct btrfs_fs_info *fs_info = trans->fs_info;
1435	struct btrfs_transaction *prev_trans = NULL;
1436	const u64 start = bg->start;
1437	const u64 end = start + bg->length - 1;
1438	int ret;
1439
1440	spin_lock(&fs_info->trans_lock);
1441	if (!list_is_first(&trans->transaction->list, &fs_info->trans_list)) {
1442		prev_trans = list_prev_entry(trans->transaction, list);
1443		refcount_inc(&prev_trans->use_count);
1444	}
1445	spin_unlock(&fs_info->trans_lock);
1446
1447	/*
1448	 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1449	 * btrfs_finish_extent_commit(). If we are at transaction N, another
1450	 * task might be running finish_extent_commit() for the previous
1451	 * transaction N - 1, and have seen a range belonging to the block
1452	 * group in pinned_extents before we were able to clear the whole block
1453	 * group range from pinned_extents. This means that task can lookup for
1454	 * the block group after we unpinned it from pinned_extents and removed
1455	 * it, leading to an error at unpin_extent_range().
1456	 */
1457	mutex_lock(&fs_info->unused_bg_unpin_mutex);
1458	if (prev_trans) {
1459		ret = btrfs_clear_extent_bit(&prev_trans->pinned_extents, start, end,
1460					     EXTENT_DIRTY, NULL);
1461		if (ret)
1462			goto out;
1463	}
1464
1465	ret = btrfs_clear_extent_bit(&trans->transaction->pinned_extents, start, end,
1466				     EXTENT_DIRTY, NULL);
1467out:
1468	mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1469	if (prev_trans)
1470		btrfs_put_transaction(prev_trans);
1471
1472	return ret == 0;
1473}
1474
1475/*
1476 * Link the block_group to a list via bg_list.
1477 *
1478 * @bg:       The block_group to link to the list.
1479 * @list:     The list to link it to.
1480 *
1481 * Use this rather than list_add_tail() directly to ensure proper respect
1482 * to locking and refcounting.
1483 *
1484 * Returns: true if the bg was linked with a refcount bump and false otherwise.
1485 */
1486static bool btrfs_link_bg_list(struct btrfs_block_group *bg, struct list_head *list)
1487{
1488	struct btrfs_fs_info *fs_info = bg->fs_info;
1489	bool added = false;
1490
1491	spin_lock(&fs_info->unused_bgs_lock);
1492	if (list_empty(&bg->bg_list)) {
1493		btrfs_get_block_group(bg);
1494		list_add_tail(&bg->bg_list, list);
1495		added = true;
1496	}
1497	spin_unlock(&fs_info->unused_bgs_lock);
1498	return added;
1499}
1500
1501/*
1502 * Process the unused_bgs list and remove any that don't have any allocated
1503 * space inside of them.
1504 */
1505void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1506{
1507	LIST_HEAD(retry_list);
1508	struct btrfs_block_group *block_group;
1509	struct btrfs_space_info *space_info;
1510	struct btrfs_trans_handle *trans;
1511	const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1512	int ret = 0;
1513
1514	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1515		return;
1516
1517	if (btrfs_fs_closing(fs_info))
1518		return;
1519
1520	/*
1521	 * Long running balances can keep us blocked here for eternity, so
1522	 * simply skip deletion if we're unable to get the mutex.
1523	 */
1524	if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1525		return;
1526
1527	spin_lock(&fs_info->unused_bgs_lock);
1528	while (!list_empty(&fs_info->unused_bgs)) {
1529		u64 used;
1530		int trimming;
1531
1532		block_group = list_first_entry(&fs_info->unused_bgs,
1533					       struct btrfs_block_group,
1534					       bg_list);
1535		list_del_init(&block_group->bg_list);
1536
1537		space_info = block_group->space_info;
1538
1539		if (ret || btrfs_mixed_space_info(space_info)) {
1540			btrfs_put_block_group(block_group);
1541			continue;
1542		}
1543		spin_unlock(&fs_info->unused_bgs_lock);
1544
1545		btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1546
1547		/* Don't want to race with allocators so take the groups_sem */
1548		down_write(&space_info->groups_sem);
1549
1550		/*
1551		 * Async discard moves the final block group discard to be prior
1552		 * to the unused_bgs code path.  Therefore, if it's not fully
1553		 * trimmed, punt it back to the async discard lists.
1554		 */
1555		if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1556		    !btrfs_is_free_space_trimmed(block_group)) {
1557			trace_btrfs_skip_unused_block_group(block_group);
1558			up_write(&space_info->groups_sem);
1559			/* Requeue if we failed because of async discard */
1560			btrfs_discard_queue_work(&fs_info->discard_ctl,
1561						 block_group);
1562			goto next;
1563		}
1564
1565		spin_lock(&space_info->lock);
1566		spin_lock(&block_group->lock);
1567		if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1568		    list_is_singular(&block_group->list)) {
1569			/*
1570			 * We want to bail if we made new allocations or have
1571			 * outstanding allocations in this block group.  We do
1572			 * the ro check in case balance is currently acting on
1573			 * this block group.
1574			 *
1575			 * Also bail out if this is the only block group for its
1576			 * type, because otherwise we would lose profile
1577			 * information from fs_info->avail_*_alloc_bits and the
1578			 * next block group of this type would be created with a
1579			 * "single" profile (even if we're in a raid fs) because
1580			 * fs_info->avail_*_alloc_bits would be 0.
1581			 */
1582			trace_btrfs_skip_unused_block_group(block_group);
1583			spin_unlock(&block_group->lock);
1584			spin_unlock(&space_info->lock);
1585			up_write(&space_info->groups_sem);
1586			goto next;
1587		}
1588
1589		/*
1590		 * The block group may be unused but there may be space reserved
1591		 * accounting with the existence of that block group, that is,
1592		 * space_info->bytes_may_use was incremented by a task but no
1593		 * space was yet allocated from the block group by the task.
1594		 * That space may or may not be allocated, as we are generally
1595		 * pessimistic about space reservation for metadata as well as
1596		 * for data when using compression (as we reserve space based on
1597		 * the worst case, when data can't be compressed, and before
1598		 * actually attempting compression, before starting writeback).
1599		 *
1600		 * So check if the total space of the space_info minus the size
1601		 * of this block group is less than the used space of the
1602		 * space_info - if that's the case, then it means we have tasks
1603		 * that might be relying on the block group in order to allocate
1604		 * extents, and add back the block group to the unused list when
1605		 * we finish, so that we retry later in case no tasks ended up
1606		 * needing to allocate extents from the block group.
1607		 */
1608		used = btrfs_space_info_used(space_info, true);
1609		if ((space_info->total_bytes - block_group->length < used &&
1610		     block_group->zone_unusable < block_group->length) ||
1611		    has_unwritten_metadata(block_group)) {
1612			/*
1613			 * Add a reference for the list, compensate for the ref
1614			 * drop under the "next" label for the
1615			 * fs_info->unused_bgs list.
1616			 */
1617			btrfs_link_bg_list(block_group, &retry_list);
1618
1619			trace_btrfs_skip_unused_block_group(block_group);
1620			spin_unlock(&block_group->lock);
1621			spin_unlock(&space_info->lock);
1622			up_write(&space_info->groups_sem);
1623			goto next;
1624		}
1625
1626		spin_unlock(&block_group->lock);
1627		spin_unlock(&space_info->lock);
1628
1629		/* We don't want to force the issue, only flip if it's ok. */
1630		ret = inc_block_group_ro(block_group, 0);
1631		up_write(&space_info->groups_sem);
1632		if (ret < 0) {
1633			ret = 0;
1634			goto next;
1635		}
1636
1637		ret = btrfs_zone_finish(block_group);
1638		if (ret < 0) {
1639			btrfs_dec_block_group_ro(block_group);
1640			if (ret == -EAGAIN) {
1641				btrfs_link_bg_list(block_group, &retry_list);
1642				ret = 0;
1643			}
1644			goto next;
1645		}
1646
1647		/*
1648		 * Want to do this before we do anything else so we can recover
1649		 * properly if we fail to join the transaction.
1650		 */
1651		trans = btrfs_start_trans_remove_block_group(fs_info,
1652						     block_group->start);
1653		if (IS_ERR(trans)) {
1654			btrfs_dec_block_group_ro(block_group);
1655			ret = PTR_ERR(trans);
1656			goto next;
1657		}
1658
1659		/*
1660		 * We could have pending pinned extents for this block group,
1661		 * just delete them, we don't care about them anymore.
1662		 */
1663		if (!clean_pinned_extents(trans, block_group)) {
1664			btrfs_dec_block_group_ro(block_group);
1665			goto end_trans;
1666		}
1667
1668		/*
1669		 * At this point, the block_group is read only and should fail
1670		 * new allocations.  However, btrfs_finish_extent_commit() can
1671		 * cause this block_group to be placed back on the discard
1672		 * lists because now the block_group isn't fully discarded.
1673		 * Bail here and try again later after discarding everything.
1674		 */
1675		spin_lock(&fs_info->discard_ctl.lock);
1676		if (!list_empty(&block_group->discard_list)) {
1677			spin_unlock(&fs_info->discard_ctl.lock);
1678			btrfs_dec_block_group_ro(block_group);
1679			btrfs_discard_queue_work(&fs_info->discard_ctl,
1680						 block_group);
1681			goto end_trans;
1682		}
1683		spin_unlock(&fs_info->discard_ctl.lock);
1684
1685		/* Reset pinned so btrfs_put_block_group doesn't complain */
1686		spin_lock(&space_info->lock);
1687		spin_lock(&block_group->lock);
1688
1689		btrfs_space_info_update_bytes_pinned(space_info, -block_group->pinned);
1690		space_info->bytes_readonly += block_group->pinned;
1691		block_group->pinned = 0;
1692
1693		spin_unlock(&block_group->lock);
1694		spin_unlock(&space_info->lock);
1695
1696		/*
1697		 * The normal path here is an unused block group is passed here,
1698		 * then trimming is handled in the transaction commit path.
1699		 * Async discard interposes before this to do the trimming
1700		 * before coming down the unused block group path as trimming
1701		 * will no longer be done later in the transaction commit path.
1702		 */
1703		if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1704			goto flip_async;
1705
1706		/*
1707		 * DISCARD can flip during remount. On zoned filesystems, we
1708		 * need to reset sequential-required zones.
1709		 */
1710		trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1711				btrfs_is_zoned(fs_info);
1712
1713		/* Implicit trim during transaction commit. */
1714		if (trimming)
1715			btrfs_freeze_block_group(block_group);
1716
1717		/*
1718		 * Btrfs_remove_chunk will abort the transaction if things go
1719		 * horribly wrong.
1720		 */
1721		ret = btrfs_remove_chunk(trans, block_group->start);
1722
1723		if (ret) {
1724			if (trimming)
1725				btrfs_unfreeze_block_group(block_group);
1726			goto end_trans;
1727		}
1728
1729		/*
1730		 * If we're not mounted with -odiscard, we can just forget
1731		 * about this block group. Otherwise we'll need to wait
1732		 * until transaction commit to do the actual discard.
1733		 */
1734		if (trimming) {
1735			spin_lock(&fs_info->unused_bgs_lock);
1736			/*
1737			 * A concurrent scrub might have added us to the list
1738			 * fs_info->unused_bgs, so use a list_move operation
1739			 * to add the block group to the deleted_bgs list.
1740			 */
1741			list_move(&block_group->bg_list,
1742				  &trans->transaction->deleted_bgs);
1743			spin_unlock(&fs_info->unused_bgs_lock);
1744			btrfs_get_block_group(block_group);
1745		}
1746end_trans:
1747		btrfs_end_transaction(trans);
1748next:
1749		btrfs_put_block_group(block_group);
1750		spin_lock(&fs_info->unused_bgs_lock);
1751	}
1752	list_splice_tail(&retry_list, &fs_info->unused_bgs);
1753	spin_unlock(&fs_info->unused_bgs_lock);
1754	mutex_unlock(&fs_info->reclaim_bgs_lock);
1755	return;
1756
1757flip_async:
1758	btrfs_end_transaction(trans);
1759	spin_lock(&fs_info->unused_bgs_lock);
1760	list_splice_tail(&retry_list, &fs_info->unused_bgs);
1761	spin_unlock(&fs_info->unused_bgs_lock);
1762	mutex_unlock(&fs_info->reclaim_bgs_lock);
1763	btrfs_put_block_group(block_group);
1764	btrfs_discard_punt_unused_bgs_list(fs_info);
1765}
1766
1767void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1768{
1769	struct btrfs_fs_info *fs_info = bg->fs_info;
1770
1771	spin_lock(&fs_info->unused_bgs_lock);
1772	if (list_empty(&bg->bg_list)) {
1773		btrfs_get_block_group(bg);
1774		trace_btrfs_add_unused_block_group(bg);
1775		list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1776	} else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1777		/* Pull out the block group from the reclaim_bgs list. */
1778		trace_btrfs_add_unused_block_group(bg);
1779		list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1780	}
1781	spin_unlock(&fs_info->unused_bgs_lock);
1782}
1783
1784/*
1785 * We want block groups with a low number of used bytes to be in the beginning
1786 * of the list, so they will get reclaimed first.
1787 */
1788static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1789			   const struct list_head *b)
1790{
1791	const struct btrfs_block_group *bg1, *bg2;
1792
1793	bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1794	bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1795
1796	/*
1797	 * Some other task may be updating the ->used field concurrently, but it
1798	 * is not serious if we get a stale value or load/store tearing issues,
1799	 * as sorting the list of block groups to reclaim is not critical and an
1800	 * occasional imperfect order is ok. So silence KCSAN and avoid the
1801	 * overhead of locking or any other synchronization.
1802	 */
1803	return data_race(bg1->used > bg2->used);
1804}
1805
1806static inline bool btrfs_should_reclaim(const struct btrfs_fs_info *fs_info)
1807{
1808	if (btrfs_is_zoned(fs_info))
1809		return btrfs_zoned_should_reclaim(fs_info);
1810	return true;
1811}
1812
1813static bool should_reclaim_block_group(const struct btrfs_block_group *bg, u64 bytes_freed)
1814{
1815	const int thresh_pct = btrfs_calc_reclaim_threshold(bg->space_info);
1816	u64 thresh_bytes = mult_perc(bg->length, thresh_pct);
1817	const u64 new_val = bg->used;
1818	const u64 old_val = new_val + bytes_freed;
1819
1820	if (thresh_bytes == 0)
1821		return false;
1822
1823	/*
1824	 * If we were below the threshold before don't reclaim, we are likely a
1825	 * brand new block group and we don't want to relocate new block groups.
1826	 */
1827	if (old_val < thresh_bytes)
1828		return false;
1829	if (new_val >= thresh_bytes)
1830		return false;
1831	return true;
1832}
1833
1834void btrfs_reclaim_bgs_work(struct work_struct *work)
1835{
1836	struct btrfs_fs_info *fs_info =
1837		container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1838	struct btrfs_block_group *bg;
1839	struct btrfs_space_info *space_info;
1840	LIST_HEAD(retry_list);
1841
1842	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1843		return;
1844
1845	if (btrfs_fs_closing(fs_info))
1846		return;
1847
1848	if (!btrfs_should_reclaim(fs_info))
1849		return;
1850
1851	guard(super_write)(fs_info->sb);
1852
1853	if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE))
1854		return;
1855
1856	/*
1857	 * Long running balances can keep us blocked here for eternity, so
1858	 * simply skip reclaim if we're unable to get the mutex.
1859	 */
1860	if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1861		btrfs_exclop_finish(fs_info);
1862		return;
1863	}
1864
1865	spin_lock(&fs_info->unused_bgs_lock);
1866	/*
1867	 * Sort happens under lock because we can't simply splice it and sort.
1868	 * The block groups might still be in use and reachable via bg_list,
1869	 * and their presence in the reclaim_bgs list must be preserved.
1870	 */
1871	list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1872	while (!list_empty(&fs_info->reclaim_bgs)) {
1873		u64 used;
1874		u64 reserved;
1875		u64 old_total;
1876		int ret = 0;
1877
1878		bg = list_first_entry(&fs_info->reclaim_bgs,
1879				      struct btrfs_block_group,
1880				      bg_list);
1881		list_del_init(&bg->bg_list);
1882
1883		space_info = bg->space_info;
1884		spin_unlock(&fs_info->unused_bgs_lock);
1885
1886		/* Don't race with allocators so take the groups_sem */
1887		down_write(&space_info->groups_sem);
1888
1889		spin_lock(&space_info->lock);
1890		spin_lock(&bg->lock);
1891		if (bg->reserved || bg->pinned || bg->ro) {
1892			/*
1893			 * We want to bail if we made new allocations or have
1894			 * outstanding allocations in this block group.  We do
1895			 * the ro check in case balance is currently acting on
1896			 * this block group.
1897			 */
1898			spin_unlock(&bg->lock);
1899			spin_unlock(&space_info->lock);
1900			up_write(&space_info->groups_sem);
1901			goto next;
1902		}
1903		if (bg->used == 0) {
1904			/*
1905			 * It is possible that we trigger relocation on a block
1906			 * group as its extents are deleted and it first goes
1907			 * below the threshold, then shortly after goes empty.
1908			 *
1909			 * In this case, relocating it does delete it, but has
1910			 * some overhead in relocation specific metadata, looking
1911			 * for the non-existent extents and running some extra
1912			 * transactions, which we can avoid by using one of the
1913			 * other mechanisms for dealing with empty block groups.
1914			 */
1915			if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1916				btrfs_mark_bg_unused(bg);
1917			spin_unlock(&bg->lock);
1918			spin_unlock(&space_info->lock);
1919			up_write(&space_info->groups_sem);
1920			goto next;
1921
1922		}
1923		/*
1924		 * The block group might no longer meet the reclaim condition by
1925		 * the time we get around to reclaiming it, so to avoid
1926		 * reclaiming overly full block_groups, skip reclaiming them.
1927		 *
1928		 * Since the decision making process also depends on the amount
1929		 * being freed, pass in a fake giant value to skip that extra
1930		 * check, which is more meaningful when adding to the list in
1931		 * the first place.
1932		 */
1933		if (!should_reclaim_block_group(bg, bg->length)) {
1934			spin_unlock(&bg->lock);
1935			spin_unlock(&space_info->lock);
1936			up_write(&space_info->groups_sem);
1937			goto next;
1938		}
1939
1940		spin_unlock(&bg->lock);
1941		old_total = space_info->total_bytes;
1942		spin_unlock(&space_info->lock);
1943
1944		/*
1945		 * Get out fast, in case we're read-only or unmounting the
1946		 * filesystem. It is OK to drop block groups from the list even
1947		 * for the read-only case. As we did take the super write lock,
1948		 * "mount -o remount,ro" won't happen and read-only filesystem
1949		 * means it is forced read-only due to a fatal error. So, it
1950		 * never gets back to read-write to let us reclaim again.
1951		 */
1952		if (btrfs_need_cleaner_sleep(fs_info)) {
1953			up_write(&space_info->groups_sem);
1954			goto next;
1955		}
1956
1957		ret = inc_block_group_ro(bg, 0);
1958		up_write(&space_info->groups_sem);
1959		if (ret < 0)
1960			goto next;
1961
1962		/*
1963		 * The amount of bytes reclaimed corresponds to the sum of the
1964		 * "used" and "reserved" counters. We have set the block group
1965		 * to RO above, which prevents reservations from happening but
1966		 * we may have existing reservations for which allocation has
1967		 * not yet been done - btrfs_update_block_group() was not yet
1968		 * called, which is where we will transfer a reserved extent's
1969		 * size from the "reserved" counter to the "used" counter - this
1970		 * happens when running delayed references. When we relocate the
1971		 * chunk below, relocation first flushes delalloc, waits for
1972		 * ordered extent completion (which is where we create delayed
1973		 * references for data extents) and commits the current
1974		 * transaction (which runs delayed references), and only after
1975		 * it does the actual work to move extents out of the block
1976		 * group. So the reported amount of reclaimed bytes is
1977		 * effectively the sum of the 'used' and 'reserved' counters.
1978		 */
1979		spin_lock(&bg->lock);
1980		used = bg->used;
1981		reserved = bg->reserved;
1982		spin_unlock(&bg->lock);
1983
1984		trace_btrfs_reclaim_block_group(bg);
1985		ret = btrfs_relocate_chunk(fs_info, bg->start, false);
1986		if (ret) {
1987			btrfs_dec_block_group_ro(bg);
1988			btrfs_err(fs_info, "error relocating chunk %llu",
1989				  bg->start);
1990			used = 0;
1991			reserved = 0;
1992			spin_lock(&space_info->lock);
1993			space_info->reclaim_errors++;
1994			spin_unlock(&space_info->lock);
1995		}
1996		spin_lock(&space_info->lock);
1997		space_info->reclaim_count++;
1998		space_info->reclaim_bytes += used;
1999		space_info->reclaim_bytes += reserved;
2000		if (space_info->total_bytes < old_total)
2001			btrfs_set_periodic_reclaim_ready(space_info, true);
2002		spin_unlock(&space_info->lock);
2003
2004next:
2005		if (ret && !READ_ONCE(space_info->periodic_reclaim))
2006			btrfs_link_bg_list(bg, &retry_list);
2007		btrfs_put_block_group(bg);
2008
2009		mutex_unlock(&fs_info->reclaim_bgs_lock);
2010		/*
2011		 * Reclaiming all the block groups in the list can take really
2012		 * long.  Prioritize cleaning up unused block groups.
2013		 */
2014		btrfs_delete_unused_bgs(fs_info);
2015		/*
2016		 * If we are interrupted by a balance, we can just bail out. The
2017		 * cleaner thread restart again if necessary.
2018		 */
2019		if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
2020			goto end;
2021		spin_lock(&fs_info->unused_bgs_lock);
2022	}
2023	spin_unlock(&fs_info->unused_bgs_lock);
2024	mutex_unlock(&fs_info->reclaim_bgs_lock);
2025end:
2026	spin_lock(&fs_info->unused_bgs_lock);
2027	list_splice_tail(&retry_list, &fs_info->reclaim_bgs);
2028	spin_unlock(&fs_info->unused_bgs_lock);
2029	btrfs_exclop_finish(fs_info);
2030}
2031
2032void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
2033{
2034	btrfs_reclaim_sweep(fs_info);
2035	spin_lock(&fs_info->unused_bgs_lock);
2036	if (!list_empty(&fs_info->reclaim_bgs))
2037		queue_work(system_dfl_wq, &fs_info->reclaim_bgs_work);
2038	spin_unlock(&fs_info->unused_bgs_lock);
2039}
2040
2041void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
2042{
2043	struct btrfs_fs_info *fs_info = bg->fs_info;
2044
2045	if (btrfs_link_bg_list(bg, &fs_info->reclaim_bgs))
2046		trace_btrfs_add_reclaim_block_group(bg);
2047}
2048
2049static int read_bg_from_eb(struct btrfs_fs_info *fs_info, const struct btrfs_key *key,
2050			   const struct btrfs_path *path)
2051{
2052	struct btrfs_chunk_map *map;
2053	struct btrfs_block_group_item bg;
2054	struct extent_buffer *leaf;
2055	int slot;
2056	u64 flags;
2057	int ret = 0;
2058
2059	slot = path->slots[0];
2060	leaf = path->nodes[0];
2061
2062	map = btrfs_find_chunk_map(fs_info, key->objectid, key->offset);
2063	if (!map) {
2064		btrfs_err(fs_info,
2065			  "logical %llu len %llu found bg but no related chunk",
2066			  key->objectid, key->offset);
2067		return -ENOENT;
2068	}
2069
2070	if (unlikely(map->start != key->objectid || map->chunk_len != key->offset)) {
2071		btrfs_err(fs_info,
2072			"block group %llu len %llu mismatch with chunk %llu len %llu",
2073			  key->objectid, key->offset, map->start, map->chunk_len);
2074		ret = -EUCLEAN;
2075		goto out_free_map;
2076	}
2077
2078	read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
2079			   sizeof(bg));
2080	flags = btrfs_stack_block_group_flags(&bg) &
2081		BTRFS_BLOCK_GROUP_TYPE_MASK;
2082
2083	if (unlikely(flags != (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK))) {
2084		btrfs_err(fs_info,
2085"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
2086			  key->objectid, key->offset, flags,
2087			  (BTRFS_BLOCK_GROUP_TYPE_MASK & map->type));
2088		ret = -EUCLEAN;
2089	}
2090
2091out_free_map:
2092	btrfs_free_chunk_map(map);
2093	return ret;
2094}
2095
2096static int find_first_block_group(struct btrfs_fs_info *fs_info,
2097				  struct btrfs_path *path,
2098				  const struct btrfs_key *key)
2099{
2100	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2101	int ret;
2102	struct btrfs_key found_key;
2103
2104	btrfs_for_each_slot(root, key, &found_key, path, ret) {
2105		if (found_key.objectid >= key->objectid &&
2106		    found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2107			return read_bg_from_eb(fs_info, &found_key, path);
2108		}
2109	}
2110	return ret;
2111}
2112
2113static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2114{
2115	u64 extra_flags = chunk_to_extended(flags) &
2116				BTRFS_EXTENDED_PROFILE_MASK;
2117
2118	write_seqlock(&fs_info->profiles_lock);
2119	if (flags & BTRFS_BLOCK_GROUP_DATA)
2120		fs_info->avail_data_alloc_bits |= extra_flags;
2121	if (flags & BTRFS_BLOCK_GROUP_METADATA)
2122		fs_info->avail_metadata_alloc_bits |= extra_flags;
2123	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2124		fs_info->avail_system_alloc_bits |= extra_flags;
2125	write_sequnlock(&fs_info->profiles_lock);
2126}
2127
2128/*
2129 * Map a physical disk address to a list of logical addresses.
2130 *
2131 * @fs_info:       the filesystem
2132 * @chunk_start:   logical address of block group
2133 * @physical:	   physical address to map to logical addresses
2134 * @logical:	   return array of logical addresses which map to @physical
2135 * @naddrs:	   length of @logical
2136 * @stripe_len:    size of IO stripe for the given block group
2137 *
2138 * Maps a particular @physical disk address to a list of @logical addresses.
2139 * Used primarily to exclude those portions of a block group that contain super
2140 * block copies.
2141 */
2142int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2143		     u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2144{
2145	struct btrfs_chunk_map *map;
2146	u64 *buf;
2147	u64 bytenr;
2148	u64 data_stripe_length;
2149	u64 io_stripe_size;
2150	int i, nr = 0;
2151	int ret = 0;
2152
2153	map = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2154	if (IS_ERR(map))
2155		return -EIO;
2156
2157	data_stripe_length = map->stripe_size;
2158	io_stripe_size = BTRFS_STRIPE_LEN;
2159	chunk_start = map->start;
2160
2161	/* For RAID5/6 adjust to a full IO stripe length */
2162	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2163		io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2164
2165	buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2166	if (!buf) {
2167		ret = -ENOMEM;
2168		goto out;
2169	}
2170
2171	for (i = 0; i < map->num_stripes; i++) {
2172		bool already_inserted = false;
2173		u32 stripe_nr;
2174		u32 offset;
2175		int j;
2176
2177		if (!in_range(physical, map->stripes[i].physical,
2178			      data_stripe_length))
2179			continue;
2180
2181		stripe_nr = (physical - map->stripes[i].physical) >>
2182			    BTRFS_STRIPE_LEN_SHIFT;
2183		offset = (physical - map->stripes[i].physical) &
2184			 BTRFS_STRIPE_LEN_MASK;
2185
2186		if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2187				 BTRFS_BLOCK_GROUP_RAID10))
2188			stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2189					    map->sub_stripes);
2190		/*
2191		 * The remaining case would be for RAID56, multiply by
2192		 * nr_data_stripes().  Alternatively, just use rmap_len below
2193		 * instead of map->stripe_len
2194		 */
2195		bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2196
2197		/* Ensure we don't add duplicate addresses */
2198		for (j = 0; j < nr; j++) {
2199			if (buf[j] == bytenr) {
2200				already_inserted = true;
2201				break;
2202			}
2203		}
2204
2205		if (!already_inserted)
2206			buf[nr++] = bytenr;
2207	}
2208
2209	*logical = buf;
2210	*naddrs = nr;
2211	*stripe_len = io_stripe_size;
2212out:
2213	btrfs_free_chunk_map(map);
2214	return ret;
2215}
2216
2217static int exclude_super_stripes(struct btrfs_block_group *cache)
2218{
2219	struct btrfs_fs_info *fs_info = cache->fs_info;
2220	const bool zoned = btrfs_is_zoned(fs_info);
2221	u64 bytenr;
2222	u64 *logical;
2223	int stripe_len;
2224	int i, nr, ret;
2225
2226	if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2227		stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2228		cache->bytes_super += stripe_len;
2229		ret = btrfs_set_extent_bit(&fs_info->excluded_extents, cache->start,
2230					   cache->start + stripe_len - 1,
2231					   EXTENT_DIRTY, NULL);
2232		if (ret)
2233			return ret;
2234	}
2235
2236	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2237		bytenr = btrfs_sb_offset(i);
2238		ret = btrfs_rmap_block(fs_info, cache->start,
2239				       bytenr, &logical, &nr, &stripe_len);
2240		if (ret)
2241			return ret;
2242
2243		/* Shouldn't have super stripes in sequential zones */
2244		if (unlikely(zoned && nr)) {
2245			kfree(logical);
2246			btrfs_err(fs_info,
2247			"zoned: block group %llu must not contain super block",
2248				  cache->start);
2249			return -EUCLEAN;
2250		}
2251
2252		while (nr--) {
2253			u64 len = min_t(u64, stripe_len,
2254				cache->start + cache->length - logical[nr]);
2255
2256			cache->bytes_super += len;
2257			ret = btrfs_set_extent_bit(&fs_info->excluded_extents,
2258						   logical[nr], logical[nr] + len - 1,
2259						   EXTENT_DIRTY, NULL);
2260			if (ret) {
2261				kfree(logical);
2262				return ret;
2263			}
2264		}
2265
2266		kfree(logical);
2267	}
2268	return 0;
2269}
2270
2271static struct btrfs_block_group *btrfs_create_block_group_cache(
2272		struct btrfs_fs_info *fs_info, u64 start)
2273{
2274	struct btrfs_block_group *cache;
2275
2276	cache = kzalloc(sizeof(*cache), GFP_NOFS);
2277	if (!cache)
2278		return NULL;
2279
2280	cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2281					GFP_NOFS);
2282	if (!cache->free_space_ctl) {
2283		kfree(cache);
2284		return NULL;
2285	}
2286
2287	cache->start = start;
2288
2289	cache->fs_info = fs_info;
2290	cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2291
2292	cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2293
2294	refcount_set(&cache->refs, 1);
2295	spin_lock_init(&cache->lock);
2296	init_rwsem(&cache->data_rwsem);
2297	INIT_LIST_HEAD(&cache->list);
2298	INIT_LIST_HEAD(&cache->cluster_list);
2299	INIT_LIST_HEAD(&cache->bg_list);
2300	INIT_LIST_HEAD(&cache->ro_list);
2301	INIT_LIST_HEAD(&cache->discard_list);
2302	INIT_LIST_HEAD(&cache->dirty_list);
2303	INIT_LIST_HEAD(&cache->io_list);
2304	INIT_LIST_HEAD(&cache->active_bg_list);
2305	btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2306	atomic_set(&cache->frozen, 0);
2307	mutex_init(&cache->free_space_lock);
2308
2309	return cache;
2310}
2311
2312/*
2313 * Iterate all chunks and verify that each of them has the corresponding block
2314 * group
2315 */
2316static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2317{
2318	u64 start = 0;
2319	int ret = 0;
2320
2321	while (1) {
2322		struct btrfs_chunk_map *map;
2323		struct btrfs_block_group *bg;
2324
2325		/*
2326		 * btrfs_find_chunk_map() will return the first chunk map
2327		 * intersecting the range, so setting @length to 1 is enough to
2328		 * get the first chunk.
2329		 */
2330		map = btrfs_find_chunk_map(fs_info, start, 1);
2331		if (!map)
2332			break;
2333
2334		bg = btrfs_lookup_block_group(fs_info, map->start);
2335		if (unlikely(!bg)) {
2336			btrfs_err(fs_info,
2337	"chunk start=%llu len=%llu doesn't have corresponding block group",
2338				     map->start, map->chunk_len);
2339			ret = -EUCLEAN;
2340			btrfs_free_chunk_map(map);
2341			break;
2342		}
2343		if (unlikely(bg->start != map->start || bg->length != map->chunk_len ||
2344			     (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2345			     (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK))) {
2346			btrfs_err(fs_info,
2347"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2348				map->start, map->chunk_len,
2349				map->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2350				bg->start, bg->length,
2351				bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2352			ret = -EUCLEAN;
2353			btrfs_free_chunk_map(map);
2354			btrfs_put_block_group(bg);
2355			break;
2356		}
2357		start = map->start + map->chunk_len;
2358		btrfs_free_chunk_map(map);
2359		btrfs_put_block_group(bg);
2360	}
2361	return ret;
2362}
2363
2364static int read_one_block_group(struct btrfs_fs_info *info,
2365				struct btrfs_block_group_item *bgi,
2366				const struct btrfs_key *key,
2367				int need_clear)
2368{
2369	struct btrfs_block_group *cache;
2370	const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2371	int ret;
2372
2373	ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2374
2375	cache = btrfs_create_block_group_cache(info, key->objectid);
2376	if (!cache)
2377		return -ENOMEM;
2378
2379	cache->length = key->offset;
2380	cache->used = btrfs_stack_block_group_used(bgi);
2381	cache->commit_used = cache->used;
2382	cache->flags = btrfs_stack_block_group_flags(bgi);
2383	cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2384	cache->space_info = btrfs_find_space_info(info, cache->flags);
2385
2386	btrfs_set_free_space_tree_thresholds(cache);
2387
2388	if (need_clear) {
2389		/*
2390		 * When we mount with old space cache, we need to
2391		 * set BTRFS_DC_CLEAR and set dirty flag.
2392		 *
2393		 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2394		 *    truncate the old free space cache inode and
2395		 *    setup a new one.
2396		 * b) Setting 'dirty flag' makes sure that we flush
2397		 *    the new space cache info onto disk.
2398		 */
2399		if (btrfs_test_opt(info, SPACE_CACHE))
2400			cache->disk_cache_state = BTRFS_DC_CLEAR;
2401	}
2402	if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2403	    (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2404			btrfs_err(info,
2405"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2406				  cache->start);
2407			ret = -EINVAL;
2408			goto error;
2409	}
2410
2411	ret = btrfs_load_block_group_zone_info(cache, false);
2412	if (ret) {
2413		btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2414			  cache->start);
2415		goto error;
2416	}
2417
2418	/*
2419	 * We need to exclude the super stripes now so that the space info has
2420	 * super bytes accounted for, otherwise we'll think we have more space
2421	 * than we actually do.
2422	 */
2423	ret = exclude_super_stripes(cache);
2424	if (ret) {
2425		/* We may have excluded something, so call this just in case. */
2426		btrfs_free_excluded_extents(cache);
2427		goto error;
2428	}
2429
2430	/*
2431	 * For zoned filesystem, space after the allocation offset is the only
2432	 * free space for a block group. So, we don't need any caching work.
2433	 * btrfs_calc_zone_unusable() will set the amount of free space and
2434	 * zone_unusable space.
2435	 *
2436	 * For regular filesystem, check for two cases, either we are full, and
2437	 * therefore don't need to bother with the caching work since we won't
2438	 * find any space, or we are empty, and we can just add all the space
2439	 * in and be done with it.  This saves us _a_lot_ of time, particularly
2440	 * in the full case.
2441	 */
2442	if (btrfs_is_zoned(info)) {
2443		btrfs_calc_zone_unusable(cache);
2444		/* Should not have any excluded extents. Just in case, though. */
2445		btrfs_free_excluded_extents(cache);
2446	} else if (cache->length == cache->used) {
2447		cache->cached = BTRFS_CACHE_FINISHED;
2448		btrfs_free_excluded_extents(cache);
2449	} else if (cache->used == 0) {
2450		cache->cached = BTRFS_CACHE_FINISHED;
2451		ret = btrfs_add_new_free_space(cache, cache->start,
2452					       cache->start + cache->length, NULL);
2453		btrfs_free_excluded_extents(cache);
2454		if (ret)
2455			goto error;
2456	}
2457
2458	ret = btrfs_add_block_group_cache(cache);
2459	if (ret) {
2460		btrfs_remove_free_space_cache(cache);
2461		goto error;
2462	}
2463
2464	trace_btrfs_add_block_group(info, cache, 0);
2465	btrfs_add_bg_to_space_info(info, cache);
2466
2467	set_avail_alloc_bits(info, cache->flags);
2468	if (btrfs_chunk_writeable(info, cache->start)) {
2469		if (cache->used == 0) {
2470			ASSERT(list_empty(&cache->bg_list));
2471			if (btrfs_test_opt(info, DISCARD_ASYNC))
2472				btrfs_discard_queue_work(&info->discard_ctl, cache);
2473			else
2474				btrfs_mark_bg_unused(cache);
2475		}
2476	} else {
2477		inc_block_group_ro(cache, 1);
2478	}
2479
2480	return 0;
2481error:
2482	btrfs_put_block_group(cache);
2483	return ret;
2484}
2485
2486static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2487{
2488	struct rb_node *node;
2489	int ret = 0;
2490
2491	for (node = rb_first_cached(&fs_info->mapping_tree); node; node = rb_next(node)) {
2492		struct btrfs_chunk_map *map;
2493		struct btrfs_block_group *bg;
2494
2495		map = rb_entry(node, struct btrfs_chunk_map, rb_node);
2496		bg = btrfs_create_block_group_cache(fs_info, map->start);
2497		if (!bg) {
2498			ret = -ENOMEM;
2499			break;
2500		}
2501
2502		/* Fill dummy cache as FULL */
2503		bg->length = map->chunk_len;
2504		bg->flags = map->type;
2505		bg->cached = BTRFS_CACHE_FINISHED;
2506		bg->used = map->chunk_len;
2507		bg->flags = map->type;
2508		bg->space_info = btrfs_find_space_info(fs_info, bg->flags);
2509		ret = btrfs_add_block_group_cache(bg);
2510		/*
2511		 * We may have some valid block group cache added already, in
2512		 * that case we skip to the next one.
2513		 */
2514		if (ret == -EEXIST) {
2515			ret = 0;
2516			btrfs_put_block_group(bg);
2517			continue;
2518		}
2519
2520		if (ret) {
2521			btrfs_remove_free_space_cache(bg);
2522			btrfs_put_block_group(bg);
2523			break;
2524		}
2525
2526		btrfs_add_bg_to_space_info(fs_info, bg);
2527
2528		set_avail_alloc_bits(fs_info, bg->flags);
2529	}
2530	if (!ret)
2531		btrfs_init_global_block_rsv(fs_info);
2532	return ret;
2533}
2534
2535int btrfs_read_block_groups(struct btrfs_fs_info *info)
2536{
2537	struct btrfs_root *root = btrfs_block_group_root(info);
2538	struct btrfs_path *path;
2539	int ret;
2540	struct btrfs_block_group *cache;
2541	struct btrfs_space_info *space_info;
2542	struct btrfs_key key;
2543	int need_clear = 0;
2544	u64 cache_gen;
2545
2546	/*
2547	 * Either no extent root (with ibadroots rescue option) or we have
2548	 * unsupported RO options. The fs can never be mounted read-write, so no
2549	 * need to waste time searching block group items.
2550	 *
2551	 * This also allows new extent tree related changes to be RO compat,
2552	 * no need for a full incompat flag.
2553	 */
2554	if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2555		      ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2556		return fill_dummy_bgs(info);
2557
2558	key.objectid = 0;
2559	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2560	key.offset = 0;
2561	path = btrfs_alloc_path();
2562	if (!path)
2563		return -ENOMEM;
2564
2565	cache_gen = btrfs_super_cache_generation(info->super_copy);
2566	if (btrfs_test_opt(info, SPACE_CACHE) &&
2567	    btrfs_super_generation(info->super_copy) != cache_gen)
2568		need_clear = 1;
2569	if (btrfs_test_opt(info, CLEAR_CACHE))
2570		need_clear = 1;
2571
2572	while (1) {
2573		struct btrfs_block_group_item bgi;
2574		struct extent_buffer *leaf;
2575		int slot;
2576
2577		ret = find_first_block_group(info, path, &key);
2578		if (ret > 0)
2579			break;
2580		if (ret != 0)
2581			goto error;
2582
2583		leaf = path->nodes[0];
2584		slot = path->slots[0];
2585
2586		read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2587				   sizeof(bgi));
2588
2589		btrfs_item_key_to_cpu(leaf, &key, slot);
2590		btrfs_release_path(path);
2591		ret = read_one_block_group(info, &bgi, &key, need_clear);
2592		if (ret < 0)
2593			goto error;
2594		key.objectid += key.offset;
2595		key.offset = 0;
2596	}
2597	btrfs_release_path(path);
2598
2599	list_for_each_entry(space_info, &info->space_info, list) {
2600		int i;
2601
2602		for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2603			if (list_empty(&space_info->block_groups[i]))
2604				continue;
2605			cache = list_first_entry(&space_info->block_groups[i],
2606						 struct btrfs_block_group,
2607						 list);
2608			btrfs_sysfs_add_block_group_type(cache);
2609		}
2610
2611		if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2612		      (BTRFS_BLOCK_GROUP_RAID10 |
2613		       BTRFS_BLOCK_GROUP_RAID1_MASK |
2614		       BTRFS_BLOCK_GROUP_RAID56_MASK |
2615		       BTRFS_BLOCK_GROUP_DUP)))
2616			continue;
2617		/*
2618		 * Avoid allocating from un-mirrored block group if there are
2619		 * mirrored block groups.
2620		 */
2621		list_for_each_entry(cache,
2622				&space_info->block_groups[BTRFS_RAID_RAID0],
2623				list)
2624			inc_block_group_ro(cache, 1);
2625		list_for_each_entry(cache,
2626				&space_info->block_groups[BTRFS_RAID_SINGLE],
2627				list)
2628			inc_block_group_ro(cache, 1);
2629	}
2630
2631	btrfs_init_global_block_rsv(info);
2632	ret = check_chunk_block_group_mappings(info);
2633error:
2634	btrfs_free_path(path);
2635	/*
2636	 * We've hit some error while reading the extent tree, and have
2637	 * rescue=ibadroots mount option.
2638	 * Try to fill the tree using dummy block groups so that the user can
2639	 * continue to mount and grab their data.
2640	 */
2641	if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2642		ret = fill_dummy_bgs(info);
2643	return ret;
2644}
2645
2646/*
2647 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2648 * allocation.
2649 *
2650 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2651 * phases.
2652 */
2653static int insert_block_group_item(struct btrfs_trans_handle *trans,
2654				   struct btrfs_block_group *block_group)
2655{
2656	struct btrfs_fs_info *fs_info = trans->fs_info;
2657	struct btrfs_block_group_item bgi;
2658	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2659	struct btrfs_key key;
2660	u64 old_commit_used;
2661	int ret;
2662
2663	spin_lock(&block_group->lock);
2664	btrfs_set_stack_block_group_used(&bgi, block_group->used);
2665	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2666						   block_group->global_root_id);
2667	btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2668	old_commit_used = block_group->commit_used;
2669	block_group->commit_used = block_group->used;
2670	key.objectid = block_group->start;
2671	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2672	key.offset = block_group->length;
2673	spin_unlock(&block_group->lock);
2674
2675	ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2676	if (ret < 0) {
2677		spin_lock(&block_group->lock);
2678		block_group->commit_used = old_commit_used;
2679		spin_unlock(&block_group->lock);
2680	}
2681
2682	return ret;
2683}
2684
2685static int insert_dev_extent(struct btrfs_trans_handle *trans,
2686			     const struct btrfs_device *device, u64 chunk_offset,
2687			     u64 start, u64 num_bytes)
2688{
2689	struct btrfs_fs_info *fs_info = device->fs_info;
2690	struct btrfs_root *root = fs_info->dev_root;
2691	BTRFS_PATH_AUTO_FREE(path);
2692	struct btrfs_dev_extent *extent;
2693	struct extent_buffer *leaf;
2694	struct btrfs_key key;
2695	int ret;
2696
2697	WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2698	WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2699	path = btrfs_alloc_path();
2700	if (!path)
2701		return -ENOMEM;
2702
2703	key.objectid = device->devid;
2704	key.type = BTRFS_DEV_EXTENT_KEY;
2705	key.offset = start;
2706	ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2707	if (ret)
2708		return ret;
2709
2710	leaf = path->nodes[0];
2711	extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2712	btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2713	btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2714					    BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2715	btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2716	btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2717
2718	return ret;
2719}
2720
2721/*
2722 * This function belongs to phase 2.
2723 *
2724 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2725 * phases.
2726 */
2727static int insert_dev_extents(struct btrfs_trans_handle *trans,
2728				   u64 chunk_offset, u64 chunk_size)
2729{
2730	struct btrfs_fs_info *fs_info = trans->fs_info;
2731	struct btrfs_device *device;
2732	struct btrfs_chunk_map *map;
2733	u64 dev_offset;
2734	int i;
2735	int ret = 0;
2736
2737	map = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2738	if (IS_ERR(map))
2739		return PTR_ERR(map);
2740
2741	/*
2742	 * Take the device list mutex to prevent races with the final phase of
2743	 * a device replace operation that replaces the device object associated
2744	 * with the map's stripes, because the device object's id can change
2745	 * at any time during that final phase of the device replace operation
2746	 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2747	 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2748	 * resulting in persisting a device extent item with such ID.
2749	 */
2750	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2751	for (i = 0; i < map->num_stripes; i++) {
2752		device = map->stripes[i].dev;
2753		dev_offset = map->stripes[i].physical;
2754
2755		ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2756					map->stripe_size);
2757		if (ret)
2758			break;
2759	}
2760	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2761
2762	btrfs_free_chunk_map(map);
2763	return ret;
2764}
2765
2766/*
2767 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2768 * chunk allocation.
2769 *
2770 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2771 * phases.
2772 */
2773void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2774{
2775	struct btrfs_fs_info *fs_info = trans->fs_info;
2776	struct btrfs_block_group *block_group;
2777	int ret = 0;
2778
2779	while (!list_empty(&trans->new_bgs)) {
2780		int index;
2781
2782		block_group = list_first_entry(&trans->new_bgs,
2783					       struct btrfs_block_group,
2784					       bg_list);
2785		if (ret)
2786			goto next;
2787
2788		index = btrfs_bg_flags_to_raid_index(block_group->flags);
2789
2790		ret = insert_block_group_item(trans, block_group);
2791		if (ret)
2792			btrfs_abort_transaction(trans, ret);
2793		if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2794			      &block_group->runtime_flags)) {
2795			mutex_lock(&fs_info->chunk_mutex);
2796			ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2797			mutex_unlock(&fs_info->chunk_mutex);
2798			if (ret)
2799				btrfs_abort_transaction(trans, ret);
2800		}
2801		ret = insert_dev_extents(trans, block_group->start,
2802					 block_group->length);
2803		if (ret)
2804			btrfs_abort_transaction(trans, ret);
2805		btrfs_add_block_group_free_space(trans, block_group);
2806
2807		/*
2808		 * If we restriped during balance, we may have added a new raid
2809		 * type, so now add the sysfs entries when it is safe to do so.
2810		 * We don't have to worry about locking here as it's handled in
2811		 * btrfs_sysfs_add_block_group_type.
2812		 */
2813		if (block_group->space_info->block_group_kobjs[index] == NULL)
2814			btrfs_sysfs_add_block_group_type(block_group);
2815
2816		/* Already aborted the transaction if it failed. */
2817next:
2818		btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2819
2820		spin_lock(&fs_info->unused_bgs_lock);
2821		list_del_init(&block_group->bg_list);
2822		clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2823		btrfs_put_block_group(block_group);
2824		spin_unlock(&fs_info->unused_bgs_lock);
2825
2826		/*
2827		 * If the block group is still unused, add it to the list of
2828		 * unused block groups. The block group may have been created in
2829		 * order to satisfy a space reservation, in which case the
2830		 * extent allocation only happens later. But often we don't
2831		 * actually need to allocate space that we previously reserved,
2832		 * so the block group may become unused for a long time. For
2833		 * example for metadata we generally reserve space for a worst
2834		 * possible scenario, but then don't end up allocating all that
2835		 * space or none at all (due to no need to COW, extent buffers
2836		 * were already COWed in the current transaction and still
2837		 * unwritten, tree heights lower than the maximum possible
2838		 * height, etc). For data we generally reserve the exact amount
2839		 * of space we are going to allocate later, the exception is
2840		 * when using compression, as we must reserve space based on the
2841		 * uncompressed data size, because the compression is only done
2842		 * when writeback triggered and we don't know how much space we
2843		 * are actually going to need, so we reserve the uncompressed
2844		 * size because the data may be incompressible in the worst case.
2845		 */
2846		if (ret == 0) {
2847			bool used;
2848
2849			spin_lock(&block_group->lock);
2850			used = btrfs_is_block_group_used(block_group);
2851			spin_unlock(&block_group->lock);
2852
2853			if (!used)
2854				btrfs_mark_bg_unused(block_group);
2855		}
2856	}
2857	btrfs_trans_release_chunk_metadata(trans);
2858}
2859
2860/*
2861 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2862 * global root id.  For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2863 */
2864static u64 calculate_global_root_id(const struct btrfs_fs_info *fs_info, u64 offset)
2865{
2866	u64 div = SZ_1G;
2867	u64 index;
2868
2869	if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2870		return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2871
2872	/* If we have a smaller fs index based on 128MiB. */
2873	if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2874		div = SZ_128M;
2875
2876	offset = div64_u64(offset, div);
2877	div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2878	return index;
2879}
2880
2881struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2882						 struct btrfs_space_info *space_info,
2883						 u64 type, u64 chunk_offset, u64 size)
2884{
2885	struct btrfs_fs_info *fs_info = trans->fs_info;
2886	struct btrfs_block_group *cache;
2887	int ret;
2888
2889	btrfs_set_log_full_commit(trans);
2890
2891	cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2892	if (!cache)
2893		return ERR_PTR(-ENOMEM);
2894
2895	/*
2896	 * Mark it as new before adding it to the rbtree of block groups or any
2897	 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2898	 * before the new flag is set.
2899	 */
2900	set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2901
2902	cache->length = size;
2903	btrfs_set_free_space_tree_thresholds(cache);
2904	cache->flags = type;
2905	cache->cached = BTRFS_CACHE_FINISHED;
2906	cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2907
2908	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2909		set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2910
2911	ret = btrfs_load_block_group_zone_info(cache, true);
2912	if (ret) {
2913		btrfs_put_block_group(cache);
2914		return ERR_PTR(ret);
2915	}
2916
2917	ret = exclude_super_stripes(cache);
2918	if (ret) {
2919		/* We may have excluded something, so call this just in case */
2920		btrfs_free_excluded_extents(cache);
2921		btrfs_put_block_group(cache);
2922		return ERR_PTR(ret);
2923	}
2924
2925	ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2926	btrfs_free_excluded_extents(cache);
2927	if (ret) {
2928		btrfs_put_block_group(cache);
2929		return ERR_PTR(ret);
2930	}
2931
2932	/*
2933	 * Ensure the corresponding space_info object is created and
2934	 * assigned to our block group. We want our bg to be added to the rbtree
2935	 * with its ->space_info set.
2936	 */
2937	cache->space_info = space_info;
2938	ASSERT(cache->space_info);
2939
2940	ret = btrfs_add_block_group_cache(cache);
2941	if (ret) {
2942		btrfs_remove_free_space_cache(cache);
2943		btrfs_put_block_group(cache);
2944		return ERR_PTR(ret);
2945	}
2946
2947	/*
2948	 * Now that our block group has its ->space_info set and is inserted in
2949	 * the rbtree, update the space info's counters.
2950	 */
2951	trace_btrfs_add_block_group(fs_info, cache, 1);
2952	btrfs_add_bg_to_space_info(fs_info, cache);
2953	btrfs_update_global_block_rsv(fs_info);
2954
2955#ifdef CONFIG_BTRFS_DEBUG
2956	if (btrfs_should_fragment_free_space(cache)) {
2957		cache->space_info->bytes_used += size >> 1;
2958		fragment_free_space(cache);
2959	}
2960#endif
2961
2962	btrfs_link_bg_list(cache, &trans->new_bgs);
2963	btrfs_inc_delayed_refs_rsv_bg_inserts(fs_info);
2964
2965	set_avail_alloc_bits(fs_info, type);
2966	return cache;
2967}
2968
2969/*
2970 * Mark one block group RO, can be called several times for the same block
2971 * group.
2972 *
2973 * @cache:		the destination block group
2974 * @do_chunk_alloc:	whether need to do chunk pre-allocation, this is to
2975 * 			ensure we still have some free space after marking this
2976 * 			block group RO.
2977 */
2978int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2979			     bool do_chunk_alloc)
2980{
2981	struct btrfs_fs_info *fs_info = cache->fs_info;
2982	struct btrfs_space_info *space_info = cache->space_info;
2983	struct btrfs_trans_handle *trans;
2984	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2985	u64 alloc_flags;
2986	int ret;
2987	bool dirty_bg_running;
2988
2989	/*
2990	 * This can only happen when we are doing read-only scrub on read-only
2991	 * mount.
2992	 * In that case we should not start a new transaction on read-only fs.
2993	 * Thus here we skip all chunk allocations.
2994	 */
2995	if (sb_rdonly(fs_info->sb)) {
2996		mutex_lock(&fs_info->ro_block_group_mutex);
2997		ret = inc_block_group_ro(cache, 0);
2998		mutex_unlock(&fs_info->ro_block_group_mutex);
2999		return ret;
3000	}
3001
3002	do {
3003		trans = btrfs_join_transaction(root);
3004		if (IS_ERR(trans))
3005			return PTR_ERR(trans);
3006
3007		dirty_bg_running = false;
3008
3009		/*
3010		 * We're not allowed to set block groups readonly after the dirty
3011		 * block group cache has started writing.  If it already started,
3012		 * back off and let this transaction commit.
3013		 */
3014		mutex_lock(&fs_info->ro_block_group_mutex);
3015		if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
3016			u64 transid = trans->transid;
3017
3018			mutex_unlock(&fs_info->ro_block_group_mutex);
3019			btrfs_end_transaction(trans);
3020
3021			ret = btrfs_wait_for_commit(fs_info, transid);
3022			if (ret)
3023				return ret;
3024			dirty_bg_running = true;
3025		}
3026	} while (dirty_bg_running);
3027
3028	if (do_chunk_alloc) {
3029		/*
3030		 * If we are changing raid levels, try to allocate a
3031		 * corresponding block group with the new raid level.
3032		 */
3033		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
3034		if (alloc_flags != cache->flags) {
3035			ret = btrfs_chunk_alloc(trans, space_info, alloc_flags,
3036						CHUNK_ALLOC_FORCE);
3037			/*
3038			 * ENOSPC is allowed here, we may have enough space
3039			 * already allocated at the new raid level to carry on
3040			 */
3041			if (ret == -ENOSPC)
3042				ret = 0;
3043			if (ret < 0)
3044				goto out;
3045		}
3046	}
3047
3048	ret = inc_block_group_ro(cache, 0);
3049	if (!ret)
3050		goto out;
3051	if (ret == -ETXTBSY)
3052		goto unlock_out;
3053
3054	/*
3055	 * Skip chunk allocation if the bg is SYSTEM, this is to avoid system
3056	 * chunk allocation storm to exhaust the system chunk array.  Otherwise
3057	 * we still want to try our best to mark the block group read-only.
3058	 */
3059	if (!do_chunk_alloc && ret == -ENOSPC &&
3060	    (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
3061		goto unlock_out;
3062
3063	alloc_flags = btrfs_get_alloc_profile(fs_info, space_info->flags);
3064	ret = btrfs_chunk_alloc(trans, space_info, alloc_flags, CHUNK_ALLOC_FORCE);
3065	if (ret < 0)
3066		goto out;
3067	/*
3068	 * We have allocated a new chunk. We also need to activate that chunk to
3069	 * grant metadata tickets for zoned filesystem.
3070	 */
3071	ret = btrfs_zoned_activate_one_bg(space_info, true);
3072	if (ret < 0)
3073		goto out;
3074
3075	ret = inc_block_group_ro(cache, 0);
3076	if (ret == -ETXTBSY)
3077		goto unlock_out;
3078out:
3079	if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
3080		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
3081		mutex_lock(&fs_info->chunk_mutex);
3082		check_system_chunk(trans, alloc_flags);
3083		mutex_unlock(&fs_info->chunk_mutex);
3084	}
3085unlock_out:
3086	mutex_unlock(&fs_info->ro_block_group_mutex);
3087
3088	btrfs_end_transaction(trans);
3089	return ret;
3090}
3091
3092void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
3093{
3094	struct btrfs_space_info *sinfo = cache->space_info;
3095	u64 num_bytes;
3096
3097	BUG_ON(!cache->ro);
3098
3099	spin_lock(&sinfo->lock);
3100	spin_lock(&cache->lock);
3101	if (!--cache->ro) {
3102		if (btrfs_is_zoned(cache->fs_info)) {
3103			/* Migrate zone_unusable bytes back */
3104			cache->zone_unusable =
3105				(cache->alloc_offset - cache->used - cache->pinned -
3106				 cache->reserved) +
3107				(cache->length - cache->zone_capacity);
3108			btrfs_space_info_update_bytes_zone_unusable(sinfo, cache->zone_unusable);
3109			sinfo->bytes_readonly -= cache->zone_unusable;
3110		}
3111		num_bytes = cache->length - cache->reserved -
3112			    cache->pinned - cache->bytes_super -
3113			    cache->zone_unusable - cache->used;
3114		sinfo->bytes_readonly -= num_bytes;
3115		list_del_init(&cache->ro_list);
3116	}
3117	spin_unlock(&cache->lock);
3118	spin_unlock(&sinfo->lock);
3119}
3120
3121static int update_block_group_item(struct btrfs_trans_handle *trans,
3122				   struct btrfs_path *path,
3123				   struct btrfs_block_group *cache)
3124{
3125	struct btrfs_fs_info *fs_info = trans->fs_info;
3126	int ret;
3127	struct btrfs_root *root = btrfs_block_group_root(fs_info);
3128	unsigned long bi;
3129	struct extent_buffer *leaf;
3130	struct btrfs_block_group_item bgi;
3131	struct btrfs_key key;
3132	u64 old_commit_used;
3133	u64 used;
3134
3135	/*
3136	 * Block group items update can be triggered out of commit transaction
3137	 * critical section, thus we need a consistent view of used bytes.
3138	 * We cannot use cache->used directly outside of the spin lock, as it
3139	 * may be changed.
3140	 */
3141	spin_lock(&cache->lock);
3142	old_commit_used = cache->commit_used;
3143	used = cache->used;
3144	/* No change in used bytes, can safely skip it. */
3145	if (cache->commit_used == used) {
3146		spin_unlock(&cache->lock);
3147		return 0;
3148	}
3149	cache->commit_used = used;
3150	spin_unlock(&cache->lock);
3151
3152	key.objectid = cache->start;
3153	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3154	key.offset = cache->length;
3155
3156	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3157	if (ret) {
3158		if (ret > 0)
3159			ret = -ENOENT;
3160		goto fail;
3161	}
3162
3163	leaf = path->nodes[0];
3164	bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3165	btrfs_set_stack_block_group_used(&bgi, used);
3166	btrfs_set_stack_block_group_chunk_objectid(&bgi,
3167						   cache->global_root_id);
3168	btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3169	write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3170fail:
3171	btrfs_release_path(path);
3172	/*
3173	 * We didn't update the block group item, need to revert commit_used
3174	 * unless the block group item didn't exist yet - this is to prevent a
3175	 * race with a concurrent insertion of the block group item, with
3176	 * insert_block_group_item(), that happened just after we attempted to
3177	 * update. In that case we would reset commit_used to 0 just after the
3178	 * insertion set it to a value greater than 0 - if the block group later
3179	 * becomes with 0 used bytes, we would incorrectly skip its update.
3180	 */
3181	if (ret < 0 && ret != -ENOENT) {
3182		spin_lock(&cache->lock);
3183		cache->commit_used = old_commit_used;
3184		spin_unlock(&cache->lock);
3185	}
3186	return ret;
3187
3188}
3189
3190static int cache_save_setup(struct btrfs_block_group *block_group,
3191			    struct btrfs_trans_handle *trans,
3192			    struct btrfs_path *path)
3193{
3194	struct btrfs_fs_info *fs_info = block_group->fs_info;
3195	struct inode *inode = NULL;
3196	struct extent_changeset *data_reserved = NULL;
3197	u64 alloc_hint = 0;
3198	int dcs = BTRFS_DC_ERROR;
3199	u64 cache_size = 0;
3200	int retries = 0;
3201	int ret = 0;
3202
3203	if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3204		return 0;
3205
3206	/*
3207	 * If this block group is smaller than 100 megs don't bother caching the
3208	 * block group.
3209	 */
3210	if (block_group->length < (100 * SZ_1M)) {
3211		spin_lock(&block_group->lock);
3212		block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3213		spin_unlock(&block_group->lock);
3214		return 0;
3215	}
3216
3217	if (TRANS_ABORTED(trans))
3218		return 0;
3219again:
3220	inode = lookup_free_space_inode(block_group, path);
3221	if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3222		ret = PTR_ERR(inode);
3223		btrfs_release_path(path);
3224		goto out;
3225	}
3226
3227	if (IS_ERR(inode)) {
3228		BUG_ON(retries);
3229		retries++;
3230
3231		if (block_group->ro)
3232			goto out_free;
3233
3234		ret = create_free_space_inode(trans, block_group, path);
3235		if (ret)
3236			goto out_free;
3237		goto again;
3238	}
3239
3240	/*
3241	 * We want to set the generation to 0, that way if anything goes wrong
3242	 * from here on out we know not to trust this cache when we load up next
3243	 * time.
3244	 */
3245	BTRFS_I(inode)->generation = 0;
3246	ret = btrfs_update_inode(trans, BTRFS_I(inode));
3247	if (unlikely(ret)) {
3248		/*
3249		 * So theoretically we could recover from this, simply set the
3250		 * super cache generation to 0 so we know to invalidate the
3251		 * cache, but then we'd have to keep track of the block groups
3252		 * that fail this way so we know we _have_ to reset this cache
3253		 * before the next commit or risk reading stale cache.  So to
3254		 * limit our exposure to horrible edge cases lets just abort the
3255		 * transaction, this only happens in really bad situations
3256		 * anyway.
3257		 */
3258		btrfs_abort_transaction(trans, ret);
3259		goto out_put;
3260	}
3261	WARN_ON(ret);
3262
3263	/* We've already setup this transaction, go ahead and exit */
3264	if (block_group->cache_generation == trans->transid &&
3265	    i_size_read(inode)) {
3266		dcs = BTRFS_DC_SETUP;
3267		goto out_put;
3268	}
3269
3270	if (i_size_read(inode) > 0) {
3271		ret = btrfs_check_trunc_cache_free_space(fs_info,
3272					&fs_info->global_block_rsv);
3273		if (ret)
3274			goto out_put;
3275
3276		ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3277		if (ret)
3278			goto out_put;
3279	}
3280
3281	spin_lock(&block_group->lock);
3282	if (block_group->cached != BTRFS_CACHE_FINISHED ||
3283	    !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3284		/*
3285		 * don't bother trying to write stuff out _if_
3286		 * a) we're not cached,
3287		 * b) we're with nospace_cache mount option,
3288		 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3289		 */
3290		dcs = BTRFS_DC_WRITTEN;
3291		spin_unlock(&block_group->lock);
3292		goto out_put;
3293	}
3294	spin_unlock(&block_group->lock);
3295
3296	/*
3297	 * We hit an ENOSPC when setting up the cache in this transaction, just
3298	 * skip doing the setup, we've already cleared the cache so we're safe.
3299	 */
3300	if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3301		ret = -ENOSPC;
3302		goto out_put;
3303	}
3304
3305	/*
3306	 * Try to preallocate enough space based on how big the block group is.
3307	 * Keep in mind this has to include any pinned space which could end up
3308	 * taking up quite a bit since it's not folded into the other space
3309	 * cache.
3310	 */
3311	cache_size = div_u64(block_group->length, SZ_256M);
3312	if (!cache_size)
3313		cache_size = 1;
3314
3315	cache_size *= 16;
3316	cache_size *= fs_info->sectorsize;
3317
3318	ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3319					  cache_size, false);
3320	if (ret)
3321		goto out_put;
3322
3323	ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3324					      cache_size, cache_size,
3325					      &alloc_hint);
3326	/*
3327	 * Our cache requires contiguous chunks so that we don't modify a bunch
3328	 * of metadata or split extents when writing the cache out, which means
3329	 * we can enospc if we are heavily fragmented in addition to just normal
3330	 * out of space conditions.  So if we hit this just skip setting up any
3331	 * other block groups for this transaction, maybe we'll unpin enough
3332	 * space the next time around.
3333	 */
3334	if (!ret)
3335		dcs = BTRFS_DC_SETUP;
3336	else if (ret == -ENOSPC)
3337		set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3338
3339out_put:
3340	iput(inode);
3341out_free:
3342	btrfs_release_path(path);
3343out:
3344	spin_lock(&block_group->lock);
3345	if (!ret && dcs == BTRFS_DC_SETUP)
3346		block_group->cache_generation = trans->transid;
3347	block_group->disk_cache_state = dcs;
3348	spin_unlock(&block_group->lock);
3349
3350	extent_changeset_free(data_reserved);
3351	return ret;
3352}
3353
3354int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3355{
3356	struct btrfs_fs_info *fs_info = trans->fs_info;
3357	struct btrfs_block_group *cache, *tmp;
3358	struct btrfs_transaction *cur_trans = trans->transaction;
3359	BTRFS_PATH_AUTO_FREE(path);
3360
3361	if (list_empty(&cur_trans->dirty_bgs) ||
3362	    !btrfs_test_opt(fs_info, SPACE_CACHE))
3363		return 0;
3364
3365	path = btrfs_alloc_path();
3366	if (!path)
3367		return -ENOMEM;
3368
3369	/* Could add new block groups, use _safe just in case */
3370	list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3371				 dirty_list) {
3372		if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3373			cache_save_setup(cache, trans, path);
3374	}
3375
3376	return 0;
3377}
3378
3379/*
3380 * Transaction commit does final block group cache writeback during a critical
3381 * section where nothing is allowed to change the FS.  This is required in
3382 * order for the cache to actually match the block group, but can introduce a
3383 * lot of latency into the commit.
3384 *
3385 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3386 * There's a chance we'll have to redo some of it if the block group changes
3387 * again during the commit, but it greatly reduces the commit latency by
3388 * getting rid of the easy block groups while we're still allowing others to
3389 * join the commit.
3390 */
3391int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3392{
3393	struct btrfs_fs_info *fs_info = trans->fs_info;
3394	struct btrfs_block_group *cache;
3395	struct btrfs_transaction *cur_trans = trans->transaction;
3396	int ret = 0;
3397	int should_put;
3398	BTRFS_PATH_AUTO_FREE(path);
3399	LIST_HEAD(dirty);
3400	struct list_head *io = &cur_trans->io_bgs;
3401	int loops = 0;
3402
3403	spin_lock(&cur_trans->dirty_bgs_lock);
3404	if (list_empty(&cur_trans->dirty_bgs)) {
3405		spin_unlock(&cur_trans->dirty_bgs_lock);
3406		return 0;
3407	}
3408	list_splice_init(&cur_trans->dirty_bgs, &dirty);
3409	spin_unlock(&cur_trans->dirty_bgs_lock);
3410
3411again:
3412	/* Make sure all the block groups on our dirty list actually exist */
3413	btrfs_create_pending_block_groups(trans);
3414
3415	if (!path) {
3416		path = btrfs_alloc_path();
3417		if (!path) {
3418			ret = -ENOMEM;
3419			goto out;
3420		}
3421	}
3422
3423	/*
3424	 * cache_write_mutex is here only to save us from balance or automatic
3425	 * removal of empty block groups deleting this block group while we are
3426	 * writing out the cache
3427	 */
3428	mutex_lock(&trans->transaction->cache_write_mutex);
3429	while (!list_empty(&dirty)) {
3430		bool drop_reserve = true;
3431
3432		cache = list_first_entry(&dirty, struct btrfs_block_group,
3433					 dirty_list);
3434		/*
3435		 * This can happen if something re-dirties a block group that
3436		 * is already under IO.  Just wait for it to finish and then do
3437		 * it all again
3438		 */
3439		if (!list_empty(&cache->io_list)) {
3440			list_del_init(&cache->io_list);
3441			btrfs_wait_cache_io(trans, cache, path);
3442			btrfs_put_block_group(cache);
3443		}
3444
3445
3446		/*
3447		 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3448		 * it should update the cache_state.  Don't delete until after
3449		 * we wait.
3450		 *
3451		 * Since we're not running in the commit critical section
3452		 * we need the dirty_bgs_lock to protect from update_block_group
3453		 */
3454		spin_lock(&cur_trans->dirty_bgs_lock);
3455		list_del_init(&cache->dirty_list);
3456		spin_unlock(&cur_trans->dirty_bgs_lock);
3457
3458		should_put = 1;
3459
3460		cache_save_setup(cache, trans, path);
3461
3462		if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3463			cache->io_ctl.inode = NULL;
3464			ret = btrfs_write_out_cache(trans, cache, path);
3465			if (ret == 0 && cache->io_ctl.inode) {
3466				should_put = 0;
3467
3468				/*
3469				 * The cache_write_mutex is protecting the
3470				 * io_list, also refer to the definition of
3471				 * btrfs_transaction::io_bgs for more details
3472				 */
3473				list_add_tail(&cache->io_list, io);
3474			} else {
3475				/*
3476				 * If we failed to write the cache, the
3477				 * generation will be bad and life goes on
3478				 */
3479				ret = 0;
3480			}
3481		}
3482		if (!ret) {
3483			ret = update_block_group_item(trans, path, cache);
3484			/*
3485			 * Our block group might still be attached to the list
3486			 * of new block groups in the transaction handle of some
3487			 * other task (struct btrfs_trans_handle->new_bgs). This
3488			 * means its block group item isn't yet in the extent
3489			 * tree. If this happens ignore the error, as we will
3490			 * try again later in the critical section of the
3491			 * transaction commit.
3492			 */
3493			if (ret == -ENOENT) {
3494				ret = 0;
3495				spin_lock(&cur_trans->dirty_bgs_lock);
3496				if (list_empty(&cache->dirty_list)) {
3497					list_add_tail(&cache->dirty_list,
3498						      &cur_trans->dirty_bgs);
3499					btrfs_get_block_group(cache);
3500					drop_reserve = false;
3501				}
3502				spin_unlock(&cur_trans->dirty_bgs_lock);
3503			} else if (ret) {
3504				btrfs_abort_transaction(trans, ret);
3505			}
3506		}
3507
3508		/* If it's not on the io list, we need to put the block group */
3509		if (should_put)
3510			btrfs_put_block_group(cache);
3511		if (drop_reserve)
3512			btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3513		/*
3514		 * Avoid blocking other tasks for too long. It might even save
3515		 * us from writing caches for block groups that are going to be
3516		 * removed.
3517		 */
3518		mutex_unlock(&trans->transaction->cache_write_mutex);
3519		if (ret)
3520			goto out;
3521		mutex_lock(&trans->transaction->cache_write_mutex);
3522	}
3523	mutex_unlock(&trans->transaction->cache_write_mutex);
3524
3525	/*
3526	 * Go through delayed refs for all the stuff we've just kicked off
3527	 * and then loop back (just once)
3528	 */
3529	if (!ret)
3530		ret = btrfs_run_delayed_refs(trans, 0);
3531	if (!ret && loops == 0) {
3532		loops++;
3533		spin_lock(&cur_trans->dirty_bgs_lock);
3534		list_splice_init(&cur_trans->dirty_bgs, &dirty);
3535		/*
3536		 * dirty_bgs_lock protects us from concurrent block group
3537		 * deletes too (not just cache_write_mutex).
3538		 */
3539		if (!list_empty(&dirty)) {
3540			spin_unlock(&cur_trans->dirty_bgs_lock);
3541			goto again;
3542		}
3543		spin_unlock(&cur_trans->dirty_bgs_lock);
3544	}
3545out:
3546	if (ret < 0) {
3547		spin_lock(&cur_trans->dirty_bgs_lock);
3548		list_splice_init(&dirty, &cur_trans->dirty_bgs);
3549		spin_unlock(&cur_trans->dirty_bgs_lock);
3550		btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3551	}
3552
3553	return ret;
3554}
3555
3556int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3557{
3558	struct btrfs_fs_info *fs_info = trans->fs_info;
3559	struct btrfs_block_group *cache;
3560	struct btrfs_transaction *cur_trans = trans->transaction;
3561	int ret = 0;
3562	int should_put;
3563	BTRFS_PATH_AUTO_FREE(path);
3564	struct list_head *io = &cur_trans->io_bgs;
3565
3566	path = btrfs_alloc_path();
3567	if (!path)
3568		return -ENOMEM;
3569
3570	/*
3571	 * Even though we are in the critical section of the transaction commit,
3572	 * we can still have concurrent tasks adding elements to this
3573	 * transaction's list of dirty block groups. These tasks correspond to
3574	 * endio free space workers started when writeback finishes for a
3575	 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3576	 * allocate new block groups as a result of COWing nodes of the root
3577	 * tree when updating the free space inode. The writeback for the space
3578	 * caches is triggered by an earlier call to
3579	 * btrfs_start_dirty_block_groups() and iterations of the following
3580	 * loop.
3581	 * Also we want to do the cache_save_setup first and then run the
3582	 * delayed refs to make sure we have the best chance at doing this all
3583	 * in one shot.
3584	 */
3585	spin_lock(&cur_trans->dirty_bgs_lock);
3586	while (!list_empty(&cur_trans->dirty_bgs)) {
3587		cache = list_first_entry(&cur_trans->dirty_bgs,
3588					 struct btrfs_block_group,
3589					 dirty_list);
3590
3591		/*
3592		 * This can happen if cache_save_setup re-dirties a block group
3593		 * that is already under IO.  Just wait for it to finish and
3594		 * then do it all again
3595		 */
3596		if (!list_empty(&cache->io_list)) {
3597			spin_unlock(&cur_trans->dirty_bgs_lock);
3598			list_del_init(&cache->io_list);
3599			btrfs_wait_cache_io(trans, cache, path);
3600			btrfs_put_block_group(cache);
3601			spin_lock(&cur_trans->dirty_bgs_lock);
3602		}
3603
3604		/*
3605		 * Don't remove from the dirty list until after we've waited on
3606		 * any pending IO
3607		 */
3608		list_del_init(&cache->dirty_list);
3609		spin_unlock(&cur_trans->dirty_bgs_lock);
3610		should_put = 1;
3611
3612		cache_save_setup(cache, trans, path);
3613
3614		if (!ret)
3615			ret = btrfs_run_delayed_refs(trans, U64_MAX);
3616
3617		if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3618			cache->io_ctl.inode = NULL;
3619			ret = btrfs_write_out_cache(trans, cache, path);
3620			if (ret == 0 && cache->io_ctl.inode) {
3621				should_put = 0;
3622				list_add_tail(&cache->io_list, io);
3623			} else {
3624				/*
3625				 * If we failed to write the cache, the
3626				 * generation will be bad and life goes on
3627				 */
3628				ret = 0;
3629			}
3630		}
3631		if (!ret) {
3632			ret = update_block_group_item(trans, path, cache);
3633			/*
3634			 * One of the free space endio workers might have
3635			 * created a new block group while updating a free space
3636			 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3637			 * and hasn't released its transaction handle yet, in
3638			 * which case the new block group is still attached to
3639			 * its transaction handle and its creation has not
3640			 * finished yet (no block group item in the extent tree
3641			 * yet, etc). If this is the case, wait for all free
3642			 * space endio workers to finish and retry. This is a
3643			 * very rare case so no need for a more efficient and
3644			 * complex approach.
3645			 */
3646			if (ret == -ENOENT) {
3647				wait_event(cur_trans->writer_wait,
3648				   atomic_read(&cur_trans->num_writers) == 1);
3649				ret = update_block_group_item(trans, path, cache);
3650				if (ret)
3651					btrfs_abort_transaction(trans, ret);
3652			} else if (ret) {
3653				btrfs_abort_transaction(trans, ret);
3654			}
3655		}
3656
3657		/* If its not on the io list, we need to put the block group */
3658		if (should_put)
3659			btrfs_put_block_group(cache);
3660		btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3661		spin_lock(&cur_trans->dirty_bgs_lock);
3662	}
3663	spin_unlock(&cur_trans->dirty_bgs_lock);
3664
3665	/*
3666	 * Refer to the definition of io_bgs member for details why it's safe
3667	 * to use it without any locking
3668	 */
3669	while (!list_empty(io)) {
3670		cache = list_first_entry(io, struct btrfs_block_group,
3671					 io_list);
3672		list_del_init(&cache->io_list);
3673		btrfs_wait_cache_io(trans, cache, path);
3674		btrfs_put_block_group(cache);
3675	}
3676
3677	return ret;
3678}
3679
3680static void btrfs_maybe_reset_size_class(struct btrfs_block_group *bg)
3681{
3682	lockdep_assert_held(&bg->lock);
3683	if (btrfs_block_group_should_use_size_class(bg) &&
3684	    bg->used == 0 && bg->reserved == 0)
3685		bg->size_class = BTRFS_BG_SZ_NONE;
3686}
3687
3688int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3689			     u64 bytenr, u64 num_bytes, bool alloc)
3690{
3691	struct btrfs_fs_info *info = trans->fs_info;
3692	struct btrfs_space_info *space_info;
3693	struct btrfs_block_group *cache;
3694	u64 old_val;
3695	bool reclaim = false;
3696	bool bg_already_dirty = true;
3697	int factor;
3698
3699	/* Block accounting for super block */
3700	spin_lock(&info->delalloc_root_lock);
3701	old_val = btrfs_super_bytes_used(info->super_copy);
3702	if (alloc)
3703		old_val += num_bytes;
3704	else
3705		old_val -= num_bytes;
3706	btrfs_set_super_bytes_used(info->super_copy, old_val);
3707	spin_unlock(&info->delalloc_root_lock);
3708
3709	cache = btrfs_lookup_block_group(info, bytenr);
3710	if (!cache)
3711		return -ENOENT;
3712
3713	/* An extent can not span multiple block groups. */
3714	ASSERT(bytenr + num_bytes <= cache->start + cache->length);
3715
3716	space_info = cache->space_info;
3717	factor = btrfs_bg_type_to_factor(cache->flags);
3718
3719	/*
3720	 * If this block group has free space cache written out, we need to make
3721	 * sure to load it if we are removing space.  This is because we need
3722	 * the unpinning stage to actually add the space back to the block group,
3723	 * otherwise we will leak space.
3724	 */
3725	if (!alloc && !btrfs_block_group_done(cache))
3726		btrfs_cache_block_group(cache, true);
3727
3728	spin_lock(&space_info->lock);
3729	spin_lock(&cache->lock);
3730
3731	if (btrfs_test_opt(info, SPACE_CACHE) &&
3732	    cache->disk_cache_state < BTRFS_DC_CLEAR)
3733		cache->disk_cache_state = BTRFS_DC_CLEAR;
3734
3735	old_val = cache->used;
3736	if (alloc) {
3737		old_val += num_bytes;
3738		cache->used = old_val;
3739		cache->reserved -= num_bytes;
3740		cache->reclaim_mark = 0;
3741		space_info->bytes_reserved -= num_bytes;
3742		space_info->bytes_used += num_bytes;
3743		space_info->disk_used += num_bytes * factor;
3744		if (READ_ONCE(space_info->periodic_reclaim))
3745			btrfs_space_info_update_reclaimable(space_info, -num_bytes);
3746		spin_unlock(&cache->lock);
3747		spin_unlock(&space_info->lock);
3748	} else {
3749		old_val -= num_bytes;
3750		cache->used = old_val;
3751		cache->pinned += num_bytes;
3752		btrfs_maybe_reset_size_class(cache);
3753		btrfs_space_info_update_bytes_pinned(space_info, num_bytes);
3754		space_info->bytes_used -= num_bytes;
3755		space_info->disk_used -= num_bytes * factor;
3756		if (READ_ONCE(space_info->periodic_reclaim))
3757			btrfs_space_info_update_reclaimable(space_info, num_bytes);
3758		else
3759			reclaim = should_reclaim_block_group(cache, num_bytes);
3760
3761		spin_unlock(&cache->lock);
3762		spin_unlock(&space_info->lock);
3763
3764		btrfs_set_extent_bit(&trans->transaction->pinned_extents, bytenr,
3765				     bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
3766	}
3767
3768	spin_lock(&trans->transaction->dirty_bgs_lock);
3769	if (list_empty(&cache->dirty_list)) {
3770		list_add_tail(&cache->dirty_list, &trans->transaction->dirty_bgs);
3771		bg_already_dirty = false;
3772		btrfs_get_block_group(cache);
3773	}
3774	spin_unlock(&trans->transaction->dirty_bgs_lock);
3775
3776	/*
3777	 * No longer have used bytes in this block group, queue it for deletion.
3778	 * We do this after adding the block group to the dirty list to avoid
3779	 * races between cleaner kthread and space cache writeout.
3780	 */
3781	if (!alloc && old_val == 0) {
3782		if (!btrfs_test_opt(info, DISCARD_ASYNC))
3783			btrfs_mark_bg_unused(cache);
3784	} else if (!alloc && reclaim) {
3785		btrfs_mark_bg_to_reclaim(cache);
3786	}
3787
3788	btrfs_put_block_group(cache);
3789
3790	/* Modified block groups are accounted for in the delayed_refs_rsv. */
3791	if (!bg_already_dirty)
3792		btrfs_inc_delayed_refs_rsv_bg_updates(info);
3793
3794	return 0;
3795}
3796
3797/*
3798 * Update the block_group and space info counters.
3799 *
3800 * @cache:	The cache we are manipulating
3801 * @ram_bytes:  The number of bytes of file content, and will be same to
3802 *              @num_bytes except for the compress path.
3803 * @num_bytes:	The number of bytes in question
3804 * @delalloc:   The blocks are allocated for the delalloc write
3805 *
3806 * This is called by the allocator when it reserves space. If this is a
3807 * reservation and the block group has become read only we cannot make the
3808 * reservation and return -EAGAIN, otherwise this function always succeeds.
3809 */
3810int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3811			     u64 ram_bytes, u64 num_bytes, bool delalloc,
3812			     bool force_wrong_size_class)
3813{
3814	struct btrfs_space_info *space_info = cache->space_info;
3815	enum btrfs_block_group_size_class size_class;
3816	int ret = 0;
3817
3818	spin_lock(&space_info->lock);
3819	spin_lock(&cache->lock);
3820	if (cache->ro) {
3821		ret = -EAGAIN;
3822		goto out_error;
3823	}
3824
3825	if (btrfs_block_group_should_use_size_class(cache)) {
3826		size_class = btrfs_calc_block_group_size_class(num_bytes);
3827		ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3828		if (ret)
3829			goto out_error;
3830	}
3831
3832	cache->reserved += num_bytes;
3833	if (delalloc)
3834		cache->delalloc_bytes += num_bytes;
3835
3836	trace_btrfs_space_reservation(cache->fs_info, "space_info",
3837				      space_info->flags, num_bytes, 1);
3838	spin_unlock(&cache->lock);
3839
3840	space_info->bytes_reserved += num_bytes;
3841	btrfs_space_info_update_bytes_may_use(space_info, -ram_bytes);
3842
3843	/*
3844	 * Compression can use less space than we reserved, so wake tickets if
3845	 * that happens.
3846	 */
3847	if (num_bytes < ram_bytes)
3848		btrfs_try_granting_tickets(space_info);
3849	spin_unlock(&space_info->lock);
3850
3851	return 0;
3852
3853out_error:
3854	spin_unlock(&cache->lock);
3855	spin_unlock(&space_info->lock);
3856	return ret;
3857}
3858
3859/*
3860 * Update the block_group and space info counters.
3861 *
3862 * @cache:       The cache we are manipulating.
3863 * @num_bytes:   The number of bytes in question.
3864 * @is_delalloc: Whether the blocks are allocated for a delalloc write.
3865 *
3866 * This is called by somebody who is freeing space that was never actually used
3867 * on disk.  For example if you reserve some space for a new leaf in transaction
3868 * A and before transaction A commits you free that leaf, you call this with
3869 * reserve set to 0 in order to clear the reservation.
3870 */
3871void btrfs_free_reserved_bytes(struct btrfs_block_group *cache, u64 num_bytes,
3872			       bool is_delalloc)
3873{
3874	struct btrfs_space_info *space_info = cache->space_info;
3875	bool bg_ro;
3876
3877	spin_lock(&space_info->lock);
3878	spin_lock(&cache->lock);
3879	bg_ro = cache->ro;
3880	cache->reserved -= num_bytes;
3881	btrfs_maybe_reset_size_class(cache);
3882	if (is_delalloc)
3883		cache->delalloc_bytes -= num_bytes;
3884	spin_unlock(&cache->lock);
3885
3886	if (bg_ro)
3887		space_info->bytes_readonly += num_bytes;
3888	else if (btrfs_is_zoned(cache->fs_info))
3889		space_info->bytes_zone_unusable += num_bytes;
3890
3891	space_info->bytes_reserved -= num_bytes;
3892	space_info->max_extent_size = 0;
3893
3894	btrfs_try_granting_tickets(space_info);
3895	spin_unlock(&space_info->lock);
3896}
3897
3898static void force_metadata_allocation(struct btrfs_fs_info *info)
3899{
3900	struct list_head *head = &info->space_info;
3901	struct btrfs_space_info *found;
3902
3903	list_for_each_entry(found, head, list) {
3904		if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3905			found->force_alloc = CHUNK_ALLOC_FORCE;
3906	}
3907}
3908
3909static bool should_alloc_chunk(const struct btrfs_fs_info *fs_info,
3910			       const struct btrfs_space_info *sinfo, int force)
3911{
3912	u64 bytes_used = btrfs_space_info_used(sinfo, false);
3913	u64 thresh;
3914
3915	if (force == CHUNK_ALLOC_FORCE)
3916		return true;
3917
3918	/*
3919	 * in limited mode, we want to have some free space up to
3920	 * about 1% of the FS size.
3921	 */
3922	if (force == CHUNK_ALLOC_LIMITED) {
3923		thresh = btrfs_super_total_bytes(fs_info->super_copy);
3924		thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3925
3926		if (sinfo->total_bytes - bytes_used < thresh)
3927			return true;
3928	}
3929
3930	if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3931		return false;
3932	return true;
3933}
3934
3935int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3936{
3937	u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3938	struct btrfs_space_info *space_info;
3939
3940	space_info = btrfs_find_space_info(trans->fs_info, type);
3941	if (!space_info) {
3942		DEBUG_WARN();
3943		return -EINVAL;
3944	}
3945
3946	return btrfs_chunk_alloc(trans, space_info, alloc_flags, CHUNK_ALLOC_FORCE);
3947}
3948
3949static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans,
3950						struct btrfs_space_info *space_info,
3951						u64 flags)
3952{
3953	struct btrfs_block_group *bg;
3954	int ret;
3955
3956	/*
3957	 * Check if we have enough space in the system space info because we
3958	 * will need to update device items in the chunk btree and insert a new
3959	 * chunk item in the chunk btree as well. This will allocate a new
3960	 * system block group if needed.
3961	 */
3962	check_system_chunk(trans, flags);
3963
3964	bg = btrfs_create_chunk(trans, space_info, flags);
3965	if (IS_ERR(bg)) {
3966		ret = PTR_ERR(bg);
3967		goto out;
3968	}
3969
3970	ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3971	/*
3972	 * Normally we are not expected to fail with -ENOSPC here, since we have
3973	 * previously reserved space in the system space_info and allocated one
3974	 * new system chunk if necessary. However there are three exceptions:
3975	 *
3976	 * 1) We may have enough free space in the system space_info but all the
3977	 *    existing system block groups have a profile which can not be used
3978	 *    for extent allocation.
3979	 *
3980	 *    This happens when mounting in degraded mode. For example we have a
3981	 *    RAID1 filesystem with 2 devices, lose one device and mount the fs
3982	 *    using the other device in degraded mode. If we then allocate a chunk,
3983	 *    we may have enough free space in the existing system space_info, but
3984	 *    none of the block groups can be used for extent allocation since they
3985	 *    have a RAID1 profile, and because we are in degraded mode with a
3986	 *    single device, we are forced to allocate a new system chunk with a
3987	 *    SINGLE profile. Making check_system_chunk() iterate over all system
3988	 *    block groups and check if they have a usable profile and enough space
3989	 *    can be slow on very large filesystems, so we tolerate the -ENOSPC and
3990	 *    try again after forcing allocation of a new system chunk. Like this
3991	 *    we avoid paying the cost of that search in normal circumstances, when
3992	 *    we were not mounted in degraded mode;
3993	 *
3994	 * 2) We had enough free space info the system space_info, and one suitable
3995	 *    block group to allocate from when we called check_system_chunk()
3996	 *    above. However right after we called it, the only system block group
3997	 *    with enough free space got turned into RO mode by a running scrub,
3998	 *    and in this case we have to allocate a new one and retry. We only
3999	 *    need do this allocate and retry once, since we have a transaction
4000	 *    handle and scrub uses the commit root to search for block groups;
4001	 *
4002	 * 3) We had one system block group with enough free space when we called
4003	 *    check_system_chunk(), but after that, right before we tried to
4004	 *    allocate the last extent buffer we needed, a discard operation came
4005	 *    in and it temporarily removed the last free space entry from the
4006	 *    block group (discard removes a free space entry, discards it, and
4007	 *    then adds back the entry to the block group cache).
4008	 */
4009	if (ret == -ENOSPC) {
4010		const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
4011		struct btrfs_block_group *sys_bg;
4012		struct btrfs_space_info *sys_space_info;
4013
4014		sys_space_info = btrfs_find_space_info(trans->fs_info, sys_flags);
4015		if (unlikely(!sys_space_info)) {
4016			ret = -EINVAL;
4017			btrfs_abort_transaction(trans, ret);
4018			goto out;
4019		}
4020
4021		sys_bg = btrfs_create_chunk(trans, sys_space_info, sys_flags);
4022		if (IS_ERR(sys_bg)) {
4023			ret = PTR_ERR(sys_bg);
4024			btrfs_abort_transaction(trans, ret);
4025			goto out;
4026		}
4027
4028		ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
4029		if (unlikely(ret)) {
4030			btrfs_abort_transaction(trans, ret);
4031			goto out;
4032		}
4033
4034		ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
4035		if (unlikely(ret)) {
4036			btrfs_abort_transaction(trans, ret);
4037			goto out;
4038		}
4039	} else if (unlikely(ret)) {
4040		btrfs_abort_transaction(trans, ret);
4041		goto out;
4042	}
4043out:
4044	btrfs_trans_release_chunk_metadata(trans);
4045
4046	if (ret)
4047		return ERR_PTR(ret);
4048
4049	btrfs_get_block_group(bg);
4050	return bg;
4051}
4052
4053/*
4054 * Chunk allocation is done in 2 phases:
4055 *
4056 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
4057 *    the chunk, the chunk mapping, create its block group and add the items
4058 *    that belong in the chunk btree to it - more specifically, we need to
4059 *    update device items in the chunk btree and add a new chunk item to it.
4060 *
4061 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
4062 *    group item to the extent btree and the device extent items to the devices
4063 *    btree.
4064 *
4065 * This is done to prevent deadlocks. For example when COWing a node from the
4066 * extent btree we are holding a write lock on the node's parent and if we
4067 * trigger chunk allocation and attempted to insert the new block group item
4068 * in the extent btree right way, we could deadlock because the path for the
4069 * insertion can include that parent node. At first glance it seems impossible
4070 * to trigger chunk allocation after starting a transaction since tasks should
4071 * reserve enough transaction units (metadata space), however while that is true
4072 * most of the time, chunk allocation may still be triggered for several reasons:
4073 *
4074 * 1) When reserving metadata, we check if there is enough free space in the
4075 *    metadata space_info and therefore don't trigger allocation of a new chunk.
4076 *    However later when the task actually tries to COW an extent buffer from
4077 *    the extent btree or from the device btree for example, it is forced to
4078 *    allocate a new block group (chunk) because the only one that had enough
4079 *    free space was just turned to RO mode by a running scrub for example (or
4080 *    device replace, block group reclaim thread, etc), so we can not use it
4081 *    for allocating an extent and end up being forced to allocate a new one;
4082 *
4083 * 2) Because we only check that the metadata space_info has enough free bytes,
4084 *    we end up not allocating a new metadata chunk in that case. However if
4085 *    the filesystem was mounted in degraded mode, none of the existing block
4086 *    groups might be suitable for extent allocation due to their incompatible
4087 *    profile (for e.g. mounting a 2 devices filesystem, where all block groups
4088 *    use a RAID1 profile, in degraded mode using a single device). In this case
4089 *    when the task attempts to COW some extent buffer of the extent btree for
4090 *    example, it will trigger allocation of a new metadata block group with a
4091 *    suitable profile (SINGLE profile in the example of the degraded mount of
4092 *    the RAID1 filesystem);
4093 *
4094 * 3) The task has reserved enough transaction units / metadata space, but when
4095 *    it attempts to COW an extent buffer from the extent or device btree for
4096 *    example, it does not find any free extent in any metadata block group,
4097 *    therefore forced to try to allocate a new metadata block group.
4098 *    This is because some other task allocated all available extents in the
4099 *    meanwhile - this typically happens with tasks that don't reserve space
4100 *    properly, either intentionally or as a bug. One example where this is
4101 *    done intentionally is fsync, as it does not reserve any transaction units
4102 *    and ends up allocating a variable number of metadata extents for log
4103 *    tree extent buffers;
4104 *
4105 * 4) The task has reserved enough transaction units / metadata space, but right
4106 *    before it tries to allocate the last extent buffer it needs, a discard
4107 *    operation comes in and, temporarily, removes the last free space entry from
4108 *    the only metadata block group that had free space (discard starts by
4109 *    removing a free space entry from a block group, then does the discard
4110 *    operation and, once it's done, it adds back the free space entry to the
4111 *    block group).
4112 *
4113 * We also need this 2 phases setup when adding a device to a filesystem with
4114 * a seed device - we must create new metadata and system chunks without adding
4115 * any of the block group items to the chunk, extent and device btrees. If we
4116 * did not do it this way, we would get ENOSPC when attempting to update those
4117 * btrees, since all the chunks from the seed device are read-only.
4118 *
4119 * Phase 1 does the updates and insertions to the chunk btree because if we had
4120 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
4121 * parallel, we risk having too many system chunks allocated by many tasks if
4122 * many tasks reach phase 1 without the previous ones completing phase 2. In the
4123 * extreme case this leads to exhaustion of the system chunk array in the
4124 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
4125 * and with RAID filesystems (so we have more device items in the chunk btree).
4126 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
4127 * the system chunk array due to concurrent allocations") provides more details.
4128 *
4129 * Allocation of system chunks does not happen through this function. A task that
4130 * needs to update the chunk btree (the only btree that uses system chunks), must
4131 * preallocate chunk space by calling either check_system_chunk() or
4132 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
4133 * metadata chunk or when removing a chunk, while the later is used before doing
4134 * a modification to the chunk btree - use cases for the later are adding,
4135 * removing and resizing a device as well as relocation of a system chunk.
4136 * See the comment below for more details.
4137 *
4138 * The reservation of system space, done through check_system_chunk(), as well
4139 * as all the updates and insertions into the chunk btree must be done while
4140 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4141 * an extent buffer from the chunks btree we never trigger allocation of a new
4142 * system chunk, which would result in a deadlock (trying to lock twice an
4143 * extent buffer of the chunk btree, first time before triggering the chunk
4144 * allocation and the second time during chunk allocation while attempting to
4145 * update the chunks btree). The system chunk array is also updated while holding
4146 * that mutex. The same logic applies to removing chunks - we must reserve system
4147 * space, update the chunk btree and the system chunk array in the superblock
4148 * while holding fs_info->chunk_mutex.
4149 *
4150 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4151 *
4152 * @space_info: specify which space_info the new chunk should belong to.
4153 *
4154 * If @force is CHUNK_ALLOC_FORCE:
4155 *    - return 1 if it successfully allocates a chunk,
4156 *    - return errors including -ENOSPC otherwise.
4157 * If @force is NOT CHUNK_ALLOC_FORCE:
4158 *    - return 0 if it doesn't need to allocate a new chunk,
4159 *    - return 1 if it successfully allocates a chunk,
4160 *    - return errors including -ENOSPC otherwise.
4161 */
4162int btrfs_chunk_alloc(struct btrfs_trans_handle *trans,
4163		      struct btrfs_space_info *space_info, u64 flags,
4164		      enum btrfs_chunk_alloc_enum force)
4165{
4166	struct btrfs_fs_info *fs_info = trans->fs_info;
4167	struct btrfs_block_group *ret_bg;
4168	bool wait_for_alloc = false;
4169	bool should_alloc = false;
4170	bool from_extent_allocation = false;
4171	int ret = 0;
4172
4173	if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4174		from_extent_allocation = true;
4175		force = CHUNK_ALLOC_FORCE;
4176	}
4177
4178	/* Don't re-enter if we're already allocating a chunk */
4179	if (trans->allocating_chunk)
4180		return -ENOSPC;
4181	/*
4182	 * Allocation of system chunks can not happen through this path, as we
4183	 * could end up in a deadlock if we are allocating a data or metadata
4184	 * chunk and there is another task modifying the chunk btree.
4185	 *
4186	 * This is because while we are holding the chunk mutex, we will attempt
4187	 * to add the new chunk item to the chunk btree or update an existing
4188	 * device item in the chunk btree, while the other task that is modifying
4189	 * the chunk btree is attempting to COW an extent buffer while holding a
4190	 * lock on it and on its parent - if the COW operation triggers a system
4191	 * chunk allocation, then we can deadlock because we are holding the
4192	 * chunk mutex and we may need to access that extent buffer or its parent
4193	 * in order to add the chunk item or update a device item.
4194	 *
4195	 * Tasks that want to modify the chunk tree should reserve system space
4196	 * before updating the chunk btree, by calling either
4197	 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4198	 * It's possible that after a task reserves the space, it still ends up
4199	 * here - this happens in the cases described above at do_chunk_alloc().
4200	 * The task will have to either retry or fail.
4201	 */
4202	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4203		return -ENOSPC;
4204
4205	do {
4206		spin_lock(&space_info->lock);
4207		if (force < space_info->force_alloc)
4208			force = space_info->force_alloc;
4209		should_alloc = should_alloc_chunk(fs_info, space_info, force);
4210		if (space_info->full) {
4211			/* No more free physical space */
4212			spin_unlock(&space_info->lock);
4213			if (should_alloc)
4214				ret = -ENOSPC;
4215			else
4216				ret = 0;
4217			return ret;
4218		} else if (!should_alloc) {
4219			spin_unlock(&space_info->lock);
4220			return 0;
4221		} else if (space_info->chunk_alloc) {
4222			/*
4223			 * Someone is already allocating, so we need to block
4224			 * until this someone is finished and then loop to
4225			 * recheck if we should continue with our allocation
4226			 * attempt.
4227			 */
4228			spin_unlock(&space_info->lock);
4229			wait_for_alloc = true;
4230			force = CHUNK_ALLOC_NO_FORCE;
4231			mutex_lock(&fs_info->chunk_mutex);
4232			mutex_unlock(&fs_info->chunk_mutex);
4233		} else {
4234			/* Proceed with allocation */
4235			space_info->chunk_alloc = true;
4236			spin_unlock(&space_info->lock);
4237			wait_for_alloc = false;
4238		}
4239
4240		cond_resched();
4241	} while (wait_for_alloc);
4242
4243	mutex_lock(&fs_info->chunk_mutex);
4244	trans->allocating_chunk = true;
4245
4246	/*
4247	 * If we have mixed data/metadata chunks we want to make sure we keep
4248	 * allocating mixed chunks instead of individual chunks.
4249	 */
4250	if (btrfs_mixed_space_info(space_info))
4251		flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4252
4253	/*
4254	 * if we're doing a data chunk, go ahead and make sure that
4255	 * we keep a reasonable number of metadata chunks allocated in the
4256	 * FS as well.
4257	 */
4258	if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4259		fs_info->data_chunk_allocations++;
4260		if (!(fs_info->data_chunk_allocations %
4261		      fs_info->metadata_ratio))
4262			force_metadata_allocation(fs_info);
4263	}
4264
4265	ret_bg = do_chunk_alloc(trans, space_info, flags);
4266	trans->allocating_chunk = false;
4267
4268	if (IS_ERR(ret_bg)) {
4269		ret = PTR_ERR(ret_bg);
4270	} else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4271		/*
4272		 * New block group is likely to be used soon. Try to activate
4273		 * it now. Failure is OK for now.
4274		 */
4275		btrfs_zone_activate(ret_bg);
4276	}
4277
4278	if (!ret)
4279		btrfs_put_block_group(ret_bg);
4280
4281	spin_lock(&space_info->lock);
4282	if (ret < 0) {
4283		if (ret == -ENOSPC)
4284			space_info->full = true;
4285		else
4286			goto out;
4287	} else {
4288		ret = 1;
4289		space_info->max_extent_size = 0;
4290	}
4291
4292	space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4293out:
4294	space_info->chunk_alloc = false;
4295	spin_unlock(&space_info->lock);
4296	mutex_unlock(&fs_info->chunk_mutex);
4297
4298	return ret;
4299}
4300
4301static u64 get_profile_num_devs(const struct btrfs_fs_info *fs_info, u64 type)
4302{
4303	u64 num_dev;
4304
4305	num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4306	if (!num_dev)
4307		num_dev = fs_info->fs_devices->rw_devices;
4308
4309	return num_dev;
4310}
4311
4312static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4313				u64 bytes,
4314				u64 type)
4315{
4316	struct btrfs_fs_info *fs_info = trans->fs_info;
4317	struct btrfs_space_info *info;
4318	u64 left;
4319	int ret = 0;
4320
4321	/*
4322	 * Needed because we can end up allocating a system chunk and for an
4323	 * atomic and race free space reservation in the chunk block reserve.
4324	 */
4325	lockdep_assert_held(&fs_info->chunk_mutex);
4326
4327	info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4328	spin_lock(&info->lock);
4329	left = info->total_bytes - btrfs_space_info_used(info, true);
4330	spin_unlock(&info->lock);
4331
4332	if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4333		btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4334			   left, bytes, type);
4335		btrfs_dump_space_info(info, 0, false);
4336	}
4337
4338	if (left < bytes) {
4339		u64 flags = btrfs_system_alloc_profile(fs_info);
4340		struct btrfs_block_group *bg;
4341		struct btrfs_space_info *space_info;
4342
4343		space_info = btrfs_find_space_info(fs_info, flags);
4344		ASSERT(space_info);
4345
4346		/*
4347		 * Ignore failure to create system chunk. We might end up not
4348		 * needing it, as we might not need to COW all nodes/leafs from
4349		 * the paths we visit in the chunk tree (they were already COWed
4350		 * or created in the current transaction for example).
4351		 */
4352		bg = btrfs_create_chunk(trans, space_info, flags);
4353		if (IS_ERR(bg)) {
4354			ret = PTR_ERR(bg);
4355		} else {
4356			/*
4357			 * We have a new chunk. We also need to activate it for
4358			 * zoned filesystem.
4359			 */
4360			ret = btrfs_zoned_activate_one_bg(info, true);
4361			if (ret < 0)
4362				return;
4363
4364			/*
4365			 * If we fail to add the chunk item here, we end up
4366			 * trying again at phase 2 of chunk allocation, at
4367			 * btrfs_create_pending_block_groups(). So ignore
4368			 * any error here. An ENOSPC here could happen, due to
4369			 * the cases described at do_chunk_alloc() - the system
4370			 * block group we just created was just turned into RO
4371			 * mode by a scrub for example, or a running discard
4372			 * temporarily removed its free space entries, etc.
4373			 */
4374			btrfs_chunk_alloc_add_chunk_item(trans, bg);
4375		}
4376	}
4377
4378	if (!ret) {
4379		ret = btrfs_block_rsv_add(fs_info,
4380					  &fs_info->chunk_block_rsv,
4381					  bytes, BTRFS_RESERVE_NO_FLUSH);
4382		if (!ret)
4383			trans->chunk_bytes_reserved += bytes;
4384	}
4385}
4386
4387/*
4388 * Reserve space in the system space for allocating or removing a chunk.
4389 * The caller must be holding fs_info->chunk_mutex.
4390 */
4391void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4392{
4393	struct btrfs_fs_info *fs_info = trans->fs_info;
4394	const u64 num_devs = get_profile_num_devs(fs_info, type);
4395	u64 bytes;
4396
4397	/* num_devs device items to update and 1 chunk item to add or remove. */
4398	bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4399		btrfs_calc_insert_metadata_size(fs_info, 1);
4400
4401	reserve_chunk_space(trans, bytes, type);
4402}
4403
4404/*
4405 * Reserve space in the system space, if needed, for doing a modification to the
4406 * chunk btree.
4407 *
4408 * @trans:		A transaction handle.
4409 * @is_item_insertion:	Indicate if the modification is for inserting a new item
4410 *			in the chunk btree or if it's for the deletion or update
4411 *			of an existing item.
4412 *
4413 * This is used in a context where we need to update the chunk btree outside
4414 * block group allocation and removal, to avoid a deadlock with a concurrent
4415 * task that is allocating a metadata or data block group and therefore needs to
4416 * update the chunk btree while holding the chunk mutex. After the update to the
4417 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4418 *
4419 */
4420void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4421				  bool is_item_insertion)
4422{
4423	struct btrfs_fs_info *fs_info = trans->fs_info;
4424	u64 bytes;
4425
4426	if (is_item_insertion)
4427		bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4428	else
4429		bytes = btrfs_calc_metadata_size(fs_info, 1);
4430
4431	mutex_lock(&fs_info->chunk_mutex);
4432	reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4433	mutex_unlock(&fs_info->chunk_mutex);
4434}
4435
4436void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4437{
4438	struct btrfs_block_group *block_group;
4439
4440	block_group = btrfs_lookup_first_block_group(info, 0);
4441	while (block_group) {
4442		btrfs_wait_block_group_cache_done(block_group);
4443		spin_lock(&block_group->lock);
4444		if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4445				       &block_group->runtime_flags)) {
4446			struct btrfs_inode *inode = block_group->inode;
4447
4448			block_group->inode = NULL;
4449			spin_unlock(&block_group->lock);
4450
4451			ASSERT(block_group->io_ctl.inode == NULL);
4452			iput(&inode->vfs_inode);
4453		} else {
4454			spin_unlock(&block_group->lock);
4455		}
4456		block_group = btrfs_next_block_group(block_group);
4457	}
4458}
4459
4460static void check_removing_space_info(struct btrfs_space_info *space_info)
4461{
4462	struct btrfs_fs_info *info = space_info->fs_info;
4463
4464	if (space_info->subgroup_id == BTRFS_SUB_GROUP_PRIMARY) {
4465		/* This is a top space_info, proceed with its children first. */
4466		for (int i = 0; i < BTRFS_SPACE_INFO_SUB_GROUP_MAX; i++) {
4467			if (space_info->sub_group[i]) {
4468				check_removing_space_info(space_info->sub_group[i]);
4469				btrfs_sysfs_remove_space_info(space_info->sub_group[i]);
4470				space_info->sub_group[i] = NULL;
4471			}
4472		}
4473	}
4474
4475	/*
4476	 * Do not hide this behind enospc_debug, this is actually important and
4477	 * indicates a real bug if this happens.
4478	 */
4479	if (WARN_ON(space_info->bytes_pinned > 0 || space_info->bytes_may_use > 0))
4480		btrfs_dump_space_info(space_info, 0, false);
4481
4482	/*
4483	 * If there was a failure to cleanup a log tree, very likely due to an
4484	 * IO failure on a writeback attempt of one or more of its extent
4485	 * buffers, we could not do proper (and cheap) unaccounting of their
4486	 * reserved space, so don't warn on bytes_reserved > 0 in that case.
4487	 */
4488	if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4489	    !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4490		if (WARN_ON(space_info->bytes_reserved > 0))
4491			btrfs_dump_space_info(space_info, 0, false);
4492	}
4493
4494	WARN_ON(space_info->reclaim_size > 0);
4495}
4496
4497/*
4498 * Must be called only after stopping all workers, since we could have block
4499 * group caching kthreads running, and therefore they could race with us if we
4500 * freed the block groups before stopping them.
4501 */
4502int btrfs_free_block_groups(struct btrfs_fs_info *info)
4503{
4504	struct btrfs_block_group *block_group;
4505	struct btrfs_space_info *space_info;
4506	struct btrfs_caching_control *caching_ctl;
4507	struct rb_node *n;
4508
4509	if (btrfs_is_zoned(info)) {
4510		if (info->active_meta_bg) {
4511			btrfs_put_block_group(info->active_meta_bg);
4512			info->active_meta_bg = NULL;
4513		}
4514		if (info->active_system_bg) {
4515			btrfs_put_block_group(info->active_system_bg);
4516			info->active_system_bg = NULL;
4517		}
4518	}
4519
4520	write_lock(&info->block_group_cache_lock);
4521	while (!list_empty(&info->caching_block_groups)) {
4522		caching_ctl = list_first_entry(&info->caching_block_groups,
4523					       struct btrfs_caching_control, list);
4524		list_del(&caching_ctl->list);
4525		btrfs_put_caching_control(caching_ctl);
4526	}
4527	write_unlock(&info->block_group_cache_lock);
4528
4529	spin_lock(&info->unused_bgs_lock);
4530	while (!list_empty(&info->unused_bgs)) {
4531		block_group = list_first_entry(&info->unused_bgs,
4532					       struct btrfs_block_group,
4533					       bg_list);
4534		list_del_init(&block_group->bg_list);
4535		btrfs_put_block_group(block_group);
4536	}
4537
4538	while (!list_empty(&info->reclaim_bgs)) {
4539		block_group = list_first_entry(&info->reclaim_bgs,
4540					       struct btrfs_block_group,
4541					       bg_list);
4542		list_del_init(&block_group->bg_list);
4543		btrfs_put_block_group(block_group);
4544	}
4545	spin_unlock(&info->unused_bgs_lock);
4546
4547	spin_lock(&info->zone_active_bgs_lock);
4548	while (!list_empty(&info->zone_active_bgs)) {
4549		block_group = list_first_entry(&info->zone_active_bgs,
4550					       struct btrfs_block_group,
4551					       active_bg_list);
4552		list_del_init(&block_group->active_bg_list);
4553		btrfs_put_block_group(block_group);
4554	}
4555	spin_unlock(&info->zone_active_bgs_lock);
4556
4557	write_lock(&info->block_group_cache_lock);
4558	while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4559		block_group = rb_entry(n, struct btrfs_block_group,
4560				       cache_node);
4561		rb_erase_cached(&block_group->cache_node,
4562				&info->block_group_cache_tree);
4563		RB_CLEAR_NODE(&block_group->cache_node);
4564		write_unlock(&info->block_group_cache_lock);
4565
4566		down_write(&block_group->space_info->groups_sem);
4567		list_del(&block_group->list);
4568		up_write(&block_group->space_info->groups_sem);
4569
4570		/*
4571		 * We haven't cached this block group, which means we could
4572		 * possibly have excluded extents on this block group.
4573		 */
4574		if (block_group->cached == BTRFS_CACHE_NO ||
4575		    block_group->cached == BTRFS_CACHE_ERROR)
4576			btrfs_free_excluded_extents(block_group);
4577
4578		btrfs_remove_free_space_cache(block_group);
4579		ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4580		ASSERT(list_empty(&block_group->dirty_list));
4581		ASSERT(list_empty(&block_group->io_list));
4582		ASSERT(list_empty(&block_group->bg_list));
4583		ASSERT(refcount_read(&block_group->refs) == 1);
4584		ASSERT(block_group->swap_extents == 0);
4585		btrfs_put_block_group(block_group);
4586
4587		write_lock(&info->block_group_cache_lock);
4588	}
4589	write_unlock(&info->block_group_cache_lock);
4590
4591	btrfs_release_global_block_rsv(info);
4592
4593	while (!list_empty(&info->space_info)) {
4594		space_info = list_first_entry(&info->space_info,
4595					      struct btrfs_space_info, list);
4596
4597		check_removing_space_info(space_info);
4598		list_del(&space_info->list);
4599		btrfs_sysfs_remove_space_info(space_info);
4600	}
4601	return 0;
4602}
4603
4604void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4605{
4606	atomic_inc(&cache->frozen);
4607}
4608
4609void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4610{
4611	struct btrfs_fs_info *fs_info = block_group->fs_info;
4612	bool cleanup;
4613
4614	spin_lock(&block_group->lock);
4615	cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4616		   test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4617	spin_unlock(&block_group->lock);
4618
4619	if (cleanup) {
4620		struct btrfs_chunk_map *map;
4621
4622		map = btrfs_find_chunk_map(fs_info, block_group->start, 1);
4623		/* Logic error, can't happen. */
4624		ASSERT(map);
4625
4626		btrfs_remove_chunk_map(fs_info, map);
4627
4628		/* Once for our lookup reference. */
4629		btrfs_free_chunk_map(map);
4630
4631		/*
4632		 * We may have left one free space entry and other possible
4633		 * tasks trimming this block group have left 1 entry each one.
4634		 * Free them if any.
4635		 */
4636		btrfs_remove_free_space_cache(block_group);
4637	}
4638}
4639
4640bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4641{
4642	bool ret = true;
4643
4644	spin_lock(&bg->lock);
4645	if (bg->ro)
4646		ret = false;
4647	else
4648		bg->swap_extents++;
4649	spin_unlock(&bg->lock);
4650
4651	return ret;
4652}
4653
4654void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4655{
4656	spin_lock(&bg->lock);
4657	ASSERT(!bg->ro);
4658	ASSERT(bg->swap_extents >= amount);
4659	bg->swap_extents -= amount;
4660	spin_unlock(&bg->lock);
4661}
4662
4663enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4664{
4665	if (size <= SZ_128K)
4666		return BTRFS_BG_SZ_SMALL;
4667	if (size <= SZ_8M)
4668		return BTRFS_BG_SZ_MEDIUM;
4669	return BTRFS_BG_SZ_LARGE;
4670}
4671
4672/*
4673 * Handle a block group allocating an extent in a size class
4674 *
4675 * @bg:				The block group we allocated in.
4676 * @size_class:			The size class of the allocation.
4677 * @force_wrong_size_class:	Whether we are desperate enough to allow
4678 *				mismatched size classes.
4679 *
4680 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4681 * case of a race that leads to the wrong size class without
4682 * force_wrong_size_class set.
4683 *
4684 * find_free_extent will skip block groups with a mismatched size class until
4685 * it really needs to avoid ENOSPC. In that case it will set
4686 * force_wrong_size_class. However, if a block group is newly allocated and
4687 * doesn't yet have a size class, then it is possible for two allocations of
4688 * different sizes to race and both try to use it. The loser is caught here and
4689 * has to retry.
4690 */
4691int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4692				     enum btrfs_block_group_size_class size_class,
4693				     bool force_wrong_size_class)
4694{
4695	ASSERT(size_class != BTRFS_BG_SZ_NONE);
4696
4697	/* The new allocation is in the right size class, do nothing */
4698	if (bg->size_class == size_class)
4699		return 0;
4700	/*
4701	 * The new allocation is in a mismatched size class.
4702	 * This means one of two things:
4703	 *
4704	 * 1. Two tasks in find_free_extent for different size_classes raced
4705	 *    and hit the same empty block_group. Make the loser try again.
4706	 * 2. A call to find_free_extent got desperate enough to set
4707	 *    'force_wrong_slab'. Don't change the size_class, but allow the
4708	 *    allocation.
4709	 */
4710	if (bg->size_class != BTRFS_BG_SZ_NONE) {
4711		if (force_wrong_size_class)
4712			return 0;
4713		return -EAGAIN;
4714	}
4715	/*
4716	 * The happy new block group case: the new allocation is the first
4717	 * one in the block_group so we set size_class.
4718	 */
4719	bg->size_class = size_class;
4720
4721	return 0;
4722}
4723
4724bool btrfs_block_group_should_use_size_class(const struct btrfs_block_group *bg)
4725{
4726	if (btrfs_is_zoned(bg->fs_info))
4727		return false;
4728	if (!btrfs_is_block_group_data_only(bg))
4729		return false;
4730	return true;
4731}