Linux Audio

Check our new training course

Loading...
v6.8
   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(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
  37/*
  38 * Return target flags in extended format or 0 if restripe for this chunk_type
  39 * is not in progress
  40 *
  41 * Should be called with balance_lock held
  42 */
  43static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
  44{
  45	struct btrfs_balance_control *bctl = fs_info->balance_ctl;
  46	u64 target = 0;
  47
  48	if (!bctl)
  49		return 0;
  50
  51	if (flags & BTRFS_BLOCK_GROUP_DATA &&
  52	    bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  53		target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
  54	} else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
  55		   bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  56		target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
  57	} else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
  58		   bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  59		target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
  60	}
  61
  62	return target;
  63}
  64
  65/*
  66 * @flags: available profiles in extended format (see ctree.h)
  67 *
  68 * Return reduced profile in chunk format.  If profile changing is in progress
  69 * (either running or paused) picks the target profile (if it's already
  70 * available), otherwise falls back to plain reducing.
  71 */
  72static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
  73{
  74	u64 num_devices = fs_info->fs_devices->rw_devices;
  75	u64 target;
  76	u64 raid_type;
  77	u64 allowed = 0;
  78
  79	/*
  80	 * See if restripe for this chunk_type is in progress, if so try to
  81	 * reduce to the target profile
  82	 */
  83	spin_lock(&fs_info->balance_lock);
  84	target = get_restripe_target(fs_info, flags);
  85	if (target) {
  86		spin_unlock(&fs_info->balance_lock);
  87		return extended_to_chunk(target);
  88	}
  89	spin_unlock(&fs_info->balance_lock);
  90
  91	/* First, mask out the RAID levels which aren't possible */
  92	for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
  93		if (num_devices >= btrfs_raid_array[raid_type].devs_min)
  94			allowed |= btrfs_raid_array[raid_type].bg_flag;
  95	}
  96	allowed &= flags;
  97
  98	/* Select the highest-redundancy RAID level. */
  99	if (allowed & BTRFS_BLOCK_GROUP_RAID1C4)
 100		allowed = BTRFS_BLOCK_GROUP_RAID1C4;
 101	else if (allowed & BTRFS_BLOCK_GROUP_RAID6)
 102		allowed = BTRFS_BLOCK_GROUP_RAID6;
 103	else if (allowed & BTRFS_BLOCK_GROUP_RAID1C3)
 104		allowed = BTRFS_BLOCK_GROUP_RAID1C3;
 105	else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
 106		allowed = BTRFS_BLOCK_GROUP_RAID5;
 107	else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
 108		allowed = BTRFS_BLOCK_GROUP_RAID10;
 109	else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
 110		allowed = BTRFS_BLOCK_GROUP_RAID1;
 111	else if (allowed & BTRFS_BLOCK_GROUP_DUP)
 112		allowed = BTRFS_BLOCK_GROUP_DUP;
 113	else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
 114		allowed = BTRFS_BLOCK_GROUP_RAID0;
 115
 116	flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
 117
 118	return extended_to_chunk(flags | allowed);
 119}
 120
 121u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
 122{
 123	unsigned seq;
 124	u64 flags;
 125
 126	do {
 127		flags = orig_flags;
 128		seq = read_seqbegin(&fs_info->profiles_lock);
 129
 130		if (flags & BTRFS_BLOCK_GROUP_DATA)
 131			flags |= fs_info->avail_data_alloc_bits;
 132		else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 133			flags |= fs_info->avail_system_alloc_bits;
 134		else if (flags & BTRFS_BLOCK_GROUP_METADATA)
 135			flags |= fs_info->avail_metadata_alloc_bits;
 136	} while (read_seqretry(&fs_info->profiles_lock, seq));
 137
 138	return btrfs_reduce_alloc_profile(fs_info, flags);
 139}
 140
 141void btrfs_get_block_group(struct btrfs_block_group *cache)
 142{
 143	refcount_inc(&cache->refs);
 144}
 145
 146void btrfs_put_block_group(struct btrfs_block_group *cache)
 147{
 148	if (refcount_dec_and_test(&cache->refs)) {
 149		WARN_ON(cache->pinned > 0);
 150		/*
 151		 * If there was a failure to cleanup a log tree, very likely due
 152		 * to an IO failure on a writeback attempt of one or more of its
 153		 * extent buffers, we could not do proper (and cheap) unaccounting
 154		 * of their reserved space, so don't warn on reserved > 0 in that
 155		 * case.
 156		 */
 157		if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
 158		    !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
 159			WARN_ON(cache->reserved > 0);
 160
 161		/*
 162		 * A block_group shouldn't be on the discard_list anymore.
 163		 * Remove the block_group from the discard_list to prevent us
 164		 * from causing a panic due to NULL pointer dereference.
 165		 */
 166		if (WARN_ON(!list_empty(&cache->discard_list)))
 167			btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
 168						  cache);
 169
 170		kfree(cache->free_space_ctl);
 171		btrfs_free_chunk_map(cache->physical_map);
 172		kfree(cache);
 173	}
 174}
 175
 176/*
 177 * This adds the block group to the fs_info rb tree for the block group cache
 178 */
 179static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 180				       struct btrfs_block_group *block_group)
 181{
 182	struct rb_node **p;
 183	struct rb_node *parent = NULL;
 184	struct btrfs_block_group *cache;
 185	bool leftmost = true;
 186
 187	ASSERT(block_group->length != 0);
 188
 189	write_lock(&info->block_group_cache_lock);
 190	p = &info->block_group_cache_tree.rb_root.rb_node;
 191
 192	while (*p) {
 193		parent = *p;
 194		cache = rb_entry(parent, struct btrfs_block_group, cache_node);
 195		if (block_group->start < cache->start) {
 196			p = &(*p)->rb_left;
 197		} else if (block_group->start > cache->start) {
 198			p = &(*p)->rb_right;
 199			leftmost = false;
 200		} else {
 201			write_unlock(&info->block_group_cache_lock);
 202			return -EEXIST;
 203		}
 204	}
 205
 206	rb_link_node(&block_group->cache_node, parent, p);
 207	rb_insert_color_cached(&block_group->cache_node,
 208			       &info->block_group_cache_tree, leftmost);
 209
 210	write_unlock(&info->block_group_cache_lock);
 211
 212	return 0;
 213}
 214
 215/*
 216 * This will return the block group at or after bytenr if contains is 0, else
 217 * it will return the block group that contains the bytenr
 218 */
 219static struct btrfs_block_group *block_group_cache_tree_search(
 220		struct btrfs_fs_info *info, u64 bytenr, int contains)
 221{
 222	struct btrfs_block_group *cache, *ret = NULL;
 223	struct rb_node *n;
 224	u64 end, start;
 225
 226	read_lock(&info->block_group_cache_lock);
 227	n = info->block_group_cache_tree.rb_root.rb_node;
 228
 229	while (n) {
 230		cache = rb_entry(n, struct btrfs_block_group, cache_node);
 231		end = cache->start + cache->length - 1;
 232		start = cache->start;
 233
 234		if (bytenr < start) {
 235			if (!contains && (!ret || start < ret->start))
 236				ret = cache;
 237			n = n->rb_left;
 238		} else if (bytenr > start) {
 239			if (contains && bytenr <= end) {
 240				ret = cache;
 241				break;
 242			}
 243			n = n->rb_right;
 244		} else {
 245			ret = cache;
 246			break;
 247		}
 248	}
 249	if (ret)
 250		btrfs_get_block_group(ret);
 251	read_unlock(&info->block_group_cache_lock);
 252
 253	return ret;
 254}
 255
 256/*
 257 * Return the block group that starts at or after bytenr
 258 */
 259struct btrfs_block_group *btrfs_lookup_first_block_group(
 260		struct btrfs_fs_info *info, u64 bytenr)
 261{
 262	return block_group_cache_tree_search(info, bytenr, 0);
 263}
 264
 265/*
 266 * Return the block group that contains the given bytenr
 267 */
 268struct btrfs_block_group *btrfs_lookup_block_group(
 269		struct btrfs_fs_info *info, u64 bytenr)
 270{
 271	return block_group_cache_tree_search(info, bytenr, 1);
 272}
 273
 274struct btrfs_block_group *btrfs_next_block_group(
 275		struct btrfs_block_group *cache)
 276{
 277	struct btrfs_fs_info *fs_info = cache->fs_info;
 278	struct rb_node *node;
 279
 280	read_lock(&fs_info->block_group_cache_lock);
 281
 282	/* If our block group was removed, we need a full search. */
 283	if (RB_EMPTY_NODE(&cache->cache_node)) {
 284		const u64 next_bytenr = cache->start + cache->length;
 285
 286		read_unlock(&fs_info->block_group_cache_lock);
 287		btrfs_put_block_group(cache);
 288		return btrfs_lookup_first_block_group(fs_info, next_bytenr);
 289	}
 290	node = rb_next(&cache->cache_node);
 291	btrfs_put_block_group(cache);
 292	if (node) {
 293		cache = rb_entry(node, struct btrfs_block_group, cache_node);
 294		btrfs_get_block_group(cache);
 295	} else
 296		cache = NULL;
 297	read_unlock(&fs_info->block_group_cache_lock);
 298	return cache;
 299}
 300
 301/*
 302 * Check if we can do a NOCOW write for a given extent.
 303 *
 304 * @fs_info:       The filesystem information object.
 305 * @bytenr:        Logical start address of the extent.
 306 *
 307 * Check if we can do a NOCOW write for the given extent, and increments the
 308 * number of NOCOW writers in the block group that contains the extent, as long
 309 * as the block group exists and it's currently not in read-only mode.
 310 *
 311 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
 312 *          is responsible for calling btrfs_dec_nocow_writers() later.
 313 *
 314 *          Or NULL if we can not do a NOCOW write
 315 */
 316struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
 317						  u64 bytenr)
 318{
 319	struct btrfs_block_group *bg;
 320	bool can_nocow = true;
 321
 322	bg = btrfs_lookup_block_group(fs_info, bytenr);
 323	if (!bg)
 324		return NULL;
 325
 326	spin_lock(&bg->lock);
 327	if (bg->ro)
 328		can_nocow = false;
 329	else
 330		atomic_inc(&bg->nocow_writers);
 331	spin_unlock(&bg->lock);
 332
 333	if (!can_nocow) {
 334		btrfs_put_block_group(bg);
 335		return NULL;
 336	}
 337
 338	/* No put on block group, done by btrfs_dec_nocow_writers(). */
 339	return bg;
 340}
 341
 342/*
 343 * Decrement the number of NOCOW writers in a block group.
 344 *
 345 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
 346 * and on the block group returned by that call. Typically this is called after
 347 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
 348 * relocation.
 349 *
 350 * After this call, the caller should not use the block group anymore. It it wants
 351 * to use it, then it should get a reference on it before calling this function.
 352 */
 353void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
 354{
 355	if (atomic_dec_and_test(&bg->nocow_writers))
 356		wake_up_var(&bg->nocow_writers);
 357
 358	/* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
 359	btrfs_put_block_group(bg);
 360}
 361
 362void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
 363{
 364	wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
 365}
 366
 367void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
 368					const u64 start)
 369{
 370	struct btrfs_block_group *bg;
 371
 372	bg = btrfs_lookup_block_group(fs_info, start);
 373	ASSERT(bg);
 374	if (atomic_dec_and_test(&bg->reservations))
 375		wake_up_var(&bg->reservations);
 376	btrfs_put_block_group(bg);
 377}
 378
 379void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
 380{
 381	struct btrfs_space_info *space_info = bg->space_info;
 382
 383	ASSERT(bg->ro);
 384
 385	if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
 386		return;
 387
 388	/*
 389	 * Our block group is read only but before we set it to read only,
 390	 * some task might have had allocated an extent from it already, but it
 391	 * has not yet created a respective ordered extent (and added it to a
 392	 * root's list of ordered extents).
 393	 * Therefore wait for any task currently allocating extents, since the
 394	 * block group's reservations counter is incremented while a read lock
 395	 * on the groups' semaphore is held and decremented after releasing
 396	 * the read access on that semaphore and creating the ordered extent.
 397	 */
 398	down_write(&space_info->groups_sem);
 399	up_write(&space_info->groups_sem);
 400
 401	wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
 402}
 403
 404struct btrfs_caching_control *btrfs_get_caching_control(
 405		struct btrfs_block_group *cache)
 406{
 407	struct btrfs_caching_control *ctl;
 408
 409	spin_lock(&cache->lock);
 410	if (!cache->caching_ctl) {
 411		spin_unlock(&cache->lock);
 412		return NULL;
 413	}
 414
 415	ctl = cache->caching_ctl;
 416	refcount_inc(&ctl->count);
 417	spin_unlock(&cache->lock);
 418	return ctl;
 419}
 420
 421void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
 422{
 423	if (refcount_dec_and_test(&ctl->count))
 424		kfree(ctl);
 425}
 426
 427/*
 428 * When we wait for progress in the block group caching, its because our
 429 * allocation attempt failed at least once.  So, we must sleep and let some
 430 * progress happen before we try again.
 431 *
 432 * This function will sleep at least once waiting for new free space to show
 433 * up, and then it will check the block group free space numbers for our min
 434 * num_bytes.  Another option is to have it go ahead and look in the rbtree for
 435 * a free extent of a given size, but this is a good start.
 436 *
 437 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
 438 * any of the information in this block group.
 439 */
 440void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
 441					   u64 num_bytes)
 442{
 443	struct btrfs_caching_control *caching_ctl;
 444	int progress;
 445
 446	caching_ctl = btrfs_get_caching_control(cache);
 447	if (!caching_ctl)
 448		return;
 449
 450	/*
 451	 * We've already failed to allocate from this block group, so even if
 452	 * there's enough space in the block group it isn't contiguous enough to
 453	 * allow for an allocation, so wait for at least the next wakeup tick,
 454	 * or for the thing to be done.
 455	 */
 456	progress = atomic_read(&caching_ctl->progress);
 457
 458	wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
 459		   (progress != atomic_read(&caching_ctl->progress) &&
 460		    (cache->free_space_ctl->free_space >= num_bytes)));
 461
 462	btrfs_put_caching_control(caching_ctl);
 463}
 464
 465static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
 466				       struct btrfs_caching_control *caching_ctl)
 467{
 468	wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
 469	return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
 470}
 471
 472static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
 473{
 474	struct btrfs_caching_control *caching_ctl;
 475	int ret;
 476
 477	caching_ctl = btrfs_get_caching_control(cache);
 478	if (!caching_ctl)
 479		return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
 480	ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
 481	btrfs_put_caching_control(caching_ctl);
 482	return ret;
 483}
 484
 485#ifdef CONFIG_BTRFS_DEBUG
 486static void fragment_free_space(struct btrfs_block_group *block_group)
 487{
 488	struct btrfs_fs_info *fs_info = block_group->fs_info;
 489	u64 start = block_group->start;
 490	u64 len = block_group->length;
 491	u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 492		fs_info->nodesize : fs_info->sectorsize;
 493	u64 step = chunk << 1;
 494
 495	while (len > chunk) {
 496		btrfs_remove_free_space(block_group, start, chunk);
 497		start += step;
 498		if (len < step)
 499			len = 0;
 500		else
 501			len -= step;
 502	}
 503}
 504#endif
 505
 506/*
 507 * Add a free space range to the in memory free space cache of a block group.
 508 * This checks if the range contains super block locations and any such
 509 * locations are not added to the free space cache.
 510 *
 511 * @block_group:      The target block group.
 512 * @start:            Start offset of the range.
 513 * @end:              End offset of the range (exclusive).
 514 * @total_added_ret:  Optional pointer to return the total amount of space
 515 *                    added to the block group's free space cache.
 516 *
 517 * Returns 0 on success or < 0 on error.
 518 */
 519int btrfs_add_new_free_space(struct btrfs_block_group *block_group, u64 start,
 520			     u64 end, u64 *total_added_ret)
 521{
 522	struct btrfs_fs_info *info = block_group->fs_info;
 523	u64 extent_start, extent_end, size;
 524	int ret;
 525
 526	if (total_added_ret)
 527		*total_added_ret = 0;
 528
 529	while (start < end) {
 530		if (!find_first_extent_bit(&info->excluded_extents, start,
 531					   &extent_start, &extent_end,
 532					   EXTENT_DIRTY | EXTENT_UPTODATE,
 533					   NULL))
 534			break;
 535
 536		if (extent_start <= start) {
 537			start = extent_end + 1;
 538		} else if (extent_start > start && extent_start < end) {
 539			size = extent_start - start;
 540			ret = btrfs_add_free_space_async_trimmed(block_group,
 541								 start, size);
 542			if (ret)
 543				return ret;
 544			if (total_added_ret)
 545				*total_added_ret += size;
 546			start = extent_end + 1;
 547		} else {
 548			break;
 549		}
 550	}
 551
 552	if (start < end) {
 553		size = end - start;
 554		ret = btrfs_add_free_space_async_trimmed(block_group, start,
 555							 size);
 556		if (ret)
 557			return ret;
 558		if (total_added_ret)
 559			*total_added_ret += size;
 560	}
 561
 562	return 0;
 563}
 564
 565/*
 566 * Get an arbitrary extent item index / max_index through the block group
 567 *
 568 * @block_group   the block group to sample from
 569 * @index:        the integral step through the block group to grab from
 570 * @max_index:    the granularity of the sampling
 571 * @key:          return value parameter for the item we find
 572 *
 573 * Pre-conditions on indices:
 574 * 0 <= index <= max_index
 575 * 0 < max_index
 576 *
 577 * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
 578 * error code on error.
 579 */
 580static int sample_block_group_extent_item(struct btrfs_caching_control *caching_ctl,
 581					  struct btrfs_block_group *block_group,
 582					  int index, int max_index,
 583					  struct btrfs_key *found_key)
 584{
 585	struct btrfs_fs_info *fs_info = block_group->fs_info;
 586	struct btrfs_root *extent_root;
 587	u64 search_offset;
 588	u64 search_end = block_group->start + block_group->length;
 589	struct btrfs_path *path;
 590	struct btrfs_key search_key;
 591	int ret = 0;
 592
 593	ASSERT(index >= 0);
 594	ASSERT(index <= max_index);
 595	ASSERT(max_index > 0);
 596	lockdep_assert_held(&caching_ctl->mutex);
 597	lockdep_assert_held_read(&fs_info->commit_root_sem);
 598
 599	path = btrfs_alloc_path();
 600	if (!path)
 601		return -ENOMEM;
 602
 603	extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
 604						       BTRFS_SUPER_INFO_OFFSET));
 605
 606	path->skip_locking = 1;
 607	path->search_commit_root = 1;
 608	path->reada = READA_FORWARD;
 609
 610	search_offset = index * div_u64(block_group->length, max_index);
 611	search_key.objectid = block_group->start + search_offset;
 612	search_key.type = BTRFS_EXTENT_ITEM_KEY;
 613	search_key.offset = 0;
 614
 615	btrfs_for_each_slot(extent_root, &search_key, found_key, path, ret) {
 616		/* Success; sampled an extent item in the block group */
 617		if (found_key->type == BTRFS_EXTENT_ITEM_KEY &&
 618		    found_key->objectid >= block_group->start &&
 619		    found_key->objectid + found_key->offset <= search_end)
 620			break;
 621
 622		/* We can't possibly find a valid extent item anymore */
 623		if (found_key->objectid >= search_end) {
 624			ret = 1;
 625			break;
 626		}
 627	}
 628
 629	lockdep_assert_held(&caching_ctl->mutex);
 630	lockdep_assert_held_read(&fs_info->commit_root_sem);
 631	btrfs_free_path(path);
 632	return ret;
 633}
 634
 635/*
 636 * Best effort attempt to compute a block group's size class while caching it.
 637 *
 638 * @block_group: the block group we are caching
 639 *
 640 * We cannot infer the size class while adding free space extents, because that
 641 * logic doesn't care about contiguous file extents (it doesn't differentiate
 642 * between a 100M extent and 100 contiguous 1M extents). So we need to read the
 643 * file extent items. Reading all of them is quite wasteful, because usually
 644 * only a handful are enough to give a good answer. Therefore, we just grab 5 of
 645 * them at even steps through the block group and pick the smallest size class
 646 * we see. Since size class is best effort, and not guaranteed in general,
 647 * inaccuracy is acceptable.
 648 *
 649 * To be more explicit about why this algorithm makes sense:
 650 *
 651 * If we are caching in a block group from disk, then there are three major cases
 652 * to consider:
 653 * 1. the block group is well behaved and all extents in it are the same size
 654 *    class.
 655 * 2. the block group is mostly one size class with rare exceptions for last
 656 *    ditch allocations
 657 * 3. the block group was populated before size classes and can have a totally
 658 *    arbitrary mix of size classes.
 659 *
 660 * In case 1, looking at any extent in the block group will yield the correct
 661 * result. For the mixed cases, taking the minimum size class seems like a good
 662 * approximation, since gaps from frees will be usable to the size class. For
 663 * 2., a small handful of file extents is likely to yield the right answer. For
 664 * 3, we can either read every file extent, or admit that this is best effort
 665 * anyway and try to stay fast.
 666 *
 667 * Returns: 0 on success, negative error code on error.
 668 */
 669static int load_block_group_size_class(struct btrfs_caching_control *caching_ctl,
 670				       struct btrfs_block_group *block_group)
 671{
 672	struct btrfs_fs_info *fs_info = block_group->fs_info;
 673	struct btrfs_key key;
 674	int i;
 675	u64 min_size = block_group->length;
 676	enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
 677	int ret;
 678
 679	if (!btrfs_block_group_should_use_size_class(block_group))
 680		return 0;
 681
 682	lockdep_assert_held(&caching_ctl->mutex);
 683	lockdep_assert_held_read(&fs_info->commit_root_sem);
 684	for (i = 0; i < 5; ++i) {
 685		ret = sample_block_group_extent_item(caching_ctl, block_group, i, 5, &key);
 686		if (ret < 0)
 687			goto out;
 688		if (ret > 0)
 689			continue;
 690		min_size = min_t(u64, min_size, key.offset);
 691		size_class = btrfs_calc_block_group_size_class(min_size);
 692	}
 693	if (size_class != BTRFS_BG_SZ_NONE) {
 694		spin_lock(&block_group->lock);
 695		block_group->size_class = size_class;
 696		spin_unlock(&block_group->lock);
 697	}
 698out:
 699	return ret;
 700}
 701
 702static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 703{
 704	struct btrfs_block_group *block_group = caching_ctl->block_group;
 705	struct btrfs_fs_info *fs_info = block_group->fs_info;
 706	struct btrfs_root *extent_root;
 707	struct btrfs_path *path;
 708	struct extent_buffer *leaf;
 709	struct btrfs_key key;
 710	u64 total_found = 0;
 711	u64 last = 0;
 712	u32 nritems;
 713	int ret;
 714	bool wakeup = true;
 715
 716	path = btrfs_alloc_path();
 717	if (!path)
 718		return -ENOMEM;
 719
 720	last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
 721	extent_root = btrfs_extent_root(fs_info, last);
 722
 723#ifdef CONFIG_BTRFS_DEBUG
 724	/*
 725	 * If we're fragmenting we don't want to make anybody think we can
 726	 * allocate from this block group until we've had a chance to fragment
 727	 * the free space.
 728	 */
 729	if (btrfs_should_fragment_free_space(block_group))
 730		wakeup = false;
 731#endif
 732	/*
 733	 * We don't want to deadlock with somebody trying to allocate a new
 734	 * extent for the extent root while also trying to search the extent
 735	 * root to add free space.  So we skip locking and search the commit
 736	 * root, since its read-only
 737	 */
 738	path->skip_locking = 1;
 739	path->search_commit_root = 1;
 740	path->reada = READA_FORWARD;
 741
 742	key.objectid = last;
 743	key.offset = 0;
 744	key.type = BTRFS_EXTENT_ITEM_KEY;
 745
 746next:
 747	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 748	if (ret < 0)
 749		goto out;
 750
 751	leaf = path->nodes[0];
 752	nritems = btrfs_header_nritems(leaf);
 753
 754	while (1) {
 755		if (btrfs_fs_closing(fs_info) > 1) {
 756			last = (u64)-1;
 757			break;
 758		}
 759
 760		if (path->slots[0] < nritems) {
 761			btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 762		} else {
 763			ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
 764			if (ret)
 765				break;
 766
 767			if (need_resched() ||
 768			    rwsem_is_contended(&fs_info->commit_root_sem)) {
 769				btrfs_release_path(path);
 770				up_read(&fs_info->commit_root_sem);
 771				mutex_unlock(&caching_ctl->mutex);
 772				cond_resched();
 773				mutex_lock(&caching_ctl->mutex);
 774				down_read(&fs_info->commit_root_sem);
 775				goto next;
 776			}
 777
 778			ret = btrfs_next_leaf(extent_root, path);
 779			if (ret < 0)
 780				goto out;
 781			if (ret)
 782				break;
 783			leaf = path->nodes[0];
 784			nritems = btrfs_header_nritems(leaf);
 785			continue;
 786		}
 787
 788		if (key.objectid < last) {
 789			key.objectid = last;
 790			key.offset = 0;
 791			key.type = BTRFS_EXTENT_ITEM_KEY;
 792			btrfs_release_path(path);
 793			goto next;
 794		}
 795
 796		if (key.objectid < block_group->start) {
 797			path->slots[0]++;
 798			continue;
 799		}
 800
 801		if (key.objectid >= block_group->start + block_group->length)
 802			break;
 803
 804		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 805		    key.type == BTRFS_METADATA_ITEM_KEY) {
 806			u64 space_added;
 807
 808			ret = btrfs_add_new_free_space(block_group, last,
 809						       key.objectid, &space_added);
 810			if (ret)
 811				goto out;
 812			total_found += space_added;
 813			if (key.type == BTRFS_METADATA_ITEM_KEY)
 814				last = key.objectid +
 815					fs_info->nodesize;
 816			else
 817				last = key.objectid + key.offset;
 818
 819			if (total_found > CACHING_CTL_WAKE_UP) {
 820				total_found = 0;
 821				if (wakeup) {
 822					atomic_inc(&caching_ctl->progress);
 823					wake_up(&caching_ctl->wait);
 824				}
 825			}
 826		}
 827		path->slots[0]++;
 828	}
 829
 830	ret = btrfs_add_new_free_space(block_group, last,
 831				       block_group->start + block_group->length,
 832				       NULL);
 833out:
 834	btrfs_free_path(path);
 835	return ret;
 836}
 837
 838static inline void btrfs_free_excluded_extents(const struct btrfs_block_group *bg)
 839{
 840	clear_extent_bits(&bg->fs_info->excluded_extents, bg->start,
 841			  bg->start + bg->length - 1, EXTENT_UPTODATE);
 842}
 843
 844static noinline void caching_thread(struct btrfs_work *work)
 845{
 846	struct btrfs_block_group *block_group;
 847	struct btrfs_fs_info *fs_info;
 848	struct btrfs_caching_control *caching_ctl;
 849	int ret;
 850
 851	caching_ctl = container_of(work, struct btrfs_caching_control, work);
 852	block_group = caching_ctl->block_group;
 853	fs_info = block_group->fs_info;
 854
 855	mutex_lock(&caching_ctl->mutex);
 856	down_read(&fs_info->commit_root_sem);
 857
 858	load_block_group_size_class(caching_ctl, block_group);
 859	if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 860		ret = load_free_space_cache(block_group);
 861		if (ret == 1) {
 862			ret = 0;
 863			goto done;
 864		}
 865
 866		/*
 867		 * We failed to load the space cache, set ourselves to
 868		 * CACHE_STARTED and carry on.
 869		 */
 870		spin_lock(&block_group->lock);
 871		block_group->cached = BTRFS_CACHE_STARTED;
 872		spin_unlock(&block_group->lock);
 873		wake_up(&caching_ctl->wait);
 874	}
 875
 876	/*
 877	 * If we are in the transaction that populated the free space tree we
 878	 * can't actually cache from the free space tree as our commit root and
 879	 * real root are the same, so we could change the contents of the blocks
 880	 * while caching.  Instead do the slow caching in this case, and after
 881	 * the transaction has committed we will be safe.
 882	 */
 883	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
 884	    !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
 885		ret = load_free_space_tree(caching_ctl);
 886	else
 887		ret = load_extent_tree_free(caching_ctl);
 888done:
 889	spin_lock(&block_group->lock);
 890	block_group->caching_ctl = NULL;
 891	block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 892	spin_unlock(&block_group->lock);
 893
 894#ifdef CONFIG_BTRFS_DEBUG
 895	if (btrfs_should_fragment_free_space(block_group)) {
 896		u64 bytes_used;
 897
 898		spin_lock(&block_group->space_info->lock);
 899		spin_lock(&block_group->lock);
 900		bytes_used = block_group->length - block_group->used;
 901		block_group->space_info->bytes_used += bytes_used >> 1;
 902		spin_unlock(&block_group->lock);
 903		spin_unlock(&block_group->space_info->lock);
 904		fragment_free_space(block_group);
 905	}
 906#endif
 907
 908	up_read(&fs_info->commit_root_sem);
 909	btrfs_free_excluded_extents(block_group);
 910	mutex_unlock(&caching_ctl->mutex);
 911
 912	wake_up(&caching_ctl->wait);
 913
 914	btrfs_put_caching_control(caching_ctl);
 915	btrfs_put_block_group(block_group);
 916}
 917
 918int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
 919{
 920	struct btrfs_fs_info *fs_info = cache->fs_info;
 921	struct btrfs_caching_control *caching_ctl = NULL;
 922	int ret = 0;
 923
 924	/* Allocator for zoned filesystems does not use the cache at all */
 925	if (btrfs_is_zoned(fs_info))
 926		return 0;
 927
 928	caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 929	if (!caching_ctl)
 930		return -ENOMEM;
 931
 932	INIT_LIST_HEAD(&caching_ctl->list);
 933	mutex_init(&caching_ctl->mutex);
 934	init_waitqueue_head(&caching_ctl->wait);
 935	caching_ctl->block_group = cache;
 936	refcount_set(&caching_ctl->count, 2);
 937	atomic_set(&caching_ctl->progress, 0);
 938	btrfs_init_work(&caching_ctl->work, caching_thread, NULL);
 939
 940	spin_lock(&cache->lock);
 941	if (cache->cached != BTRFS_CACHE_NO) {
 942		kfree(caching_ctl);
 943
 944		caching_ctl = cache->caching_ctl;
 945		if (caching_ctl)
 946			refcount_inc(&caching_ctl->count);
 947		spin_unlock(&cache->lock);
 948		goto out;
 949	}
 950	WARN_ON(cache->caching_ctl);
 951	cache->caching_ctl = caching_ctl;
 952	cache->cached = BTRFS_CACHE_STARTED;
 953	spin_unlock(&cache->lock);
 954
 955	write_lock(&fs_info->block_group_cache_lock);
 956	refcount_inc(&caching_ctl->count);
 957	list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 958	write_unlock(&fs_info->block_group_cache_lock);
 959
 960	btrfs_get_block_group(cache);
 961
 962	btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 963out:
 964	if (wait && caching_ctl)
 965		ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
 966	if (caching_ctl)
 967		btrfs_put_caching_control(caching_ctl);
 968
 969	return ret;
 970}
 971
 972static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
 973{
 974	u64 extra_flags = chunk_to_extended(flags) &
 975				BTRFS_EXTENDED_PROFILE_MASK;
 976
 977	write_seqlock(&fs_info->profiles_lock);
 978	if (flags & BTRFS_BLOCK_GROUP_DATA)
 979		fs_info->avail_data_alloc_bits &= ~extra_flags;
 980	if (flags & BTRFS_BLOCK_GROUP_METADATA)
 981		fs_info->avail_metadata_alloc_bits &= ~extra_flags;
 982	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 983		fs_info->avail_system_alloc_bits &= ~extra_flags;
 984	write_sequnlock(&fs_info->profiles_lock);
 985}
 986
 987/*
 988 * Clear incompat bits for the following feature(s):
 989 *
 990 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
 991 *            in the whole filesystem
 992 *
 993 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
 994 */
 995static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
 996{
 997	bool found_raid56 = false;
 998	bool found_raid1c34 = false;
 999
1000	if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
1001	    (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
1002	    (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
1003		struct list_head *head = &fs_info->space_info;
1004		struct btrfs_space_info *sinfo;
1005
1006		list_for_each_entry_rcu(sinfo, head, list) {
1007			down_read(&sinfo->groups_sem);
1008			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
1009				found_raid56 = true;
1010			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
1011				found_raid56 = true;
1012			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
1013				found_raid1c34 = true;
1014			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
1015				found_raid1c34 = true;
1016			up_read(&sinfo->groups_sem);
1017		}
1018		if (!found_raid56)
1019			btrfs_clear_fs_incompat(fs_info, RAID56);
1020		if (!found_raid1c34)
1021			btrfs_clear_fs_incompat(fs_info, RAID1C34);
1022	}
1023}
1024
1025static int remove_block_group_item(struct btrfs_trans_handle *trans,
1026				   struct btrfs_path *path,
1027				   struct btrfs_block_group *block_group)
1028{
1029	struct btrfs_fs_info *fs_info = trans->fs_info;
1030	struct btrfs_root *root;
1031	struct btrfs_key key;
1032	int ret;
1033
1034	root = btrfs_block_group_root(fs_info);
1035	key.objectid = block_group->start;
1036	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1037	key.offset = block_group->length;
1038
1039	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1040	if (ret > 0)
1041		ret = -ENOENT;
1042	if (ret < 0)
1043		return ret;
1044
1045	ret = btrfs_del_item(trans, root, path);
1046	return ret;
1047}
1048
1049int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1050			     struct btrfs_chunk_map *map)
1051{
1052	struct btrfs_fs_info *fs_info = trans->fs_info;
1053	struct btrfs_path *path;
1054	struct btrfs_block_group *block_group;
1055	struct btrfs_free_cluster *cluster;
1056	struct inode *inode;
1057	struct kobject *kobj = NULL;
1058	int ret;
1059	int index;
1060	int factor;
1061	struct btrfs_caching_control *caching_ctl = NULL;
1062	bool remove_map;
1063	bool remove_rsv = false;
1064
1065	block_group = btrfs_lookup_block_group(fs_info, map->start);
1066	BUG_ON(!block_group);
 
 
1067	BUG_ON(!block_group->ro);
1068
1069	trace_btrfs_remove_block_group(block_group);
1070	/*
1071	 * Free the reserved super bytes from this block group before
1072	 * remove it.
1073	 */
1074	btrfs_free_excluded_extents(block_group);
1075	btrfs_free_ref_tree_range(fs_info, block_group->start,
1076				  block_group->length);
1077
1078	index = btrfs_bg_flags_to_raid_index(block_group->flags);
1079	factor = btrfs_bg_type_to_factor(block_group->flags);
1080
1081	/* make sure this block group isn't part of an allocation cluster */
1082	cluster = &fs_info->data_alloc_cluster;
1083	spin_lock(&cluster->refill_lock);
1084	btrfs_return_cluster_to_free_space(block_group, cluster);
1085	spin_unlock(&cluster->refill_lock);
1086
1087	/*
1088	 * make sure this block group isn't part of a metadata
1089	 * allocation cluster
1090	 */
1091	cluster = &fs_info->meta_alloc_cluster;
1092	spin_lock(&cluster->refill_lock);
1093	btrfs_return_cluster_to_free_space(block_group, cluster);
1094	spin_unlock(&cluster->refill_lock);
1095
1096	btrfs_clear_treelog_bg(block_group);
1097	btrfs_clear_data_reloc_bg(block_group);
1098
1099	path = btrfs_alloc_path();
1100	if (!path) {
1101		ret = -ENOMEM;
1102		goto out;
1103	}
1104
1105	/*
1106	 * get the inode first so any iput calls done for the io_list
1107	 * aren't the final iput (no unlinks allowed now)
1108	 */
1109	inode = lookup_free_space_inode(block_group, path);
1110
1111	mutex_lock(&trans->transaction->cache_write_mutex);
1112	/*
1113	 * Make sure our free space cache IO is done before removing the
1114	 * free space inode
1115	 */
1116	spin_lock(&trans->transaction->dirty_bgs_lock);
1117	if (!list_empty(&block_group->io_list)) {
1118		list_del_init(&block_group->io_list);
1119
1120		WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1121
1122		spin_unlock(&trans->transaction->dirty_bgs_lock);
1123		btrfs_wait_cache_io(trans, block_group, path);
1124		btrfs_put_block_group(block_group);
1125		spin_lock(&trans->transaction->dirty_bgs_lock);
1126	}
1127
1128	if (!list_empty(&block_group->dirty_list)) {
1129		list_del_init(&block_group->dirty_list);
1130		remove_rsv = true;
1131		btrfs_put_block_group(block_group);
1132	}
1133	spin_unlock(&trans->transaction->dirty_bgs_lock);
1134	mutex_unlock(&trans->transaction->cache_write_mutex);
1135
1136	ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1137	if (ret)
1138		goto out;
1139
1140	write_lock(&fs_info->block_group_cache_lock);
1141	rb_erase_cached(&block_group->cache_node,
1142			&fs_info->block_group_cache_tree);
1143	RB_CLEAR_NODE(&block_group->cache_node);
1144
1145	/* Once for the block groups rbtree */
1146	btrfs_put_block_group(block_group);
1147
1148	write_unlock(&fs_info->block_group_cache_lock);
1149
1150	down_write(&block_group->space_info->groups_sem);
1151	/*
1152	 * we must use list_del_init so people can check to see if they
1153	 * are still on the list after taking the semaphore
1154	 */
1155	list_del_init(&block_group->list);
1156	if (list_empty(&block_group->space_info->block_groups[index])) {
1157		kobj = block_group->space_info->block_group_kobjs[index];
1158		block_group->space_info->block_group_kobjs[index] = NULL;
1159		clear_avail_alloc_bits(fs_info, block_group->flags);
1160	}
1161	up_write(&block_group->space_info->groups_sem);
1162	clear_incompat_bg_bits(fs_info, block_group->flags);
1163	if (kobj) {
1164		kobject_del(kobj);
1165		kobject_put(kobj);
1166	}
1167
1168	if (block_group->cached == BTRFS_CACHE_STARTED)
1169		btrfs_wait_block_group_cache_done(block_group);
1170
1171	write_lock(&fs_info->block_group_cache_lock);
1172	caching_ctl = btrfs_get_caching_control(block_group);
1173	if (!caching_ctl) {
1174		struct btrfs_caching_control *ctl;
1175
1176		list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1177			if (ctl->block_group == block_group) {
1178				caching_ctl = ctl;
1179				refcount_inc(&caching_ctl->count);
1180				break;
1181			}
1182		}
1183	}
1184	if (caching_ctl)
1185		list_del_init(&caching_ctl->list);
1186	write_unlock(&fs_info->block_group_cache_lock);
1187
1188	if (caching_ctl) {
1189		/* Once for the caching bgs list and once for us. */
1190		btrfs_put_caching_control(caching_ctl);
1191		btrfs_put_caching_control(caching_ctl);
1192	}
1193
1194	spin_lock(&trans->transaction->dirty_bgs_lock);
1195	WARN_ON(!list_empty(&block_group->dirty_list));
1196	WARN_ON(!list_empty(&block_group->io_list));
1197	spin_unlock(&trans->transaction->dirty_bgs_lock);
1198
1199	btrfs_remove_free_space_cache(block_group);
1200
1201	spin_lock(&block_group->space_info->lock);
1202	list_del_init(&block_group->ro_list);
1203
1204	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1205		WARN_ON(block_group->space_info->total_bytes
1206			< block_group->length);
1207		WARN_ON(block_group->space_info->bytes_readonly
1208			< block_group->length - block_group->zone_unusable);
1209		WARN_ON(block_group->space_info->bytes_zone_unusable
1210			< block_group->zone_unusable);
1211		WARN_ON(block_group->space_info->disk_total
1212			< block_group->length * factor);
1213	}
1214	block_group->space_info->total_bytes -= block_group->length;
1215	block_group->space_info->bytes_readonly -=
1216		(block_group->length - block_group->zone_unusable);
1217	block_group->space_info->bytes_zone_unusable -=
1218		block_group->zone_unusable;
1219	block_group->space_info->disk_total -= block_group->length * factor;
1220
1221	spin_unlock(&block_group->space_info->lock);
1222
1223	/*
1224	 * Remove the free space for the block group from the free space tree
1225	 * and the block group's item from the extent tree before marking the
1226	 * block group as removed. This is to prevent races with tasks that
1227	 * freeze and unfreeze a block group, this task and another task
1228	 * allocating a new block group - the unfreeze task ends up removing
1229	 * the block group's extent map before the task calling this function
1230	 * deletes the block group item from the extent tree, allowing for
1231	 * another task to attempt to create another block group with the same
1232	 * item key (and failing with -EEXIST and a transaction abort).
1233	 */
1234	ret = remove_block_group_free_space(trans, block_group);
1235	if (ret)
1236		goto out;
1237
1238	ret = remove_block_group_item(trans, path, block_group);
1239	if (ret < 0)
1240		goto out;
1241
1242	spin_lock(&block_group->lock);
1243	set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1244
1245	/*
1246	 * At this point trimming or scrub can't start on this block group,
1247	 * because we removed the block group from the rbtree
1248	 * fs_info->block_group_cache_tree so no one can't find it anymore and
1249	 * even if someone already got this block group before we removed it
1250	 * from the rbtree, they have already incremented block_group->frozen -
1251	 * if they didn't, for the trimming case they won't find any free space
1252	 * entries because we already removed them all when we called
1253	 * btrfs_remove_free_space_cache().
1254	 *
1255	 * And we must not remove the chunk map from the fs_info->mapping_tree
1256	 * to prevent the same logical address range and physical device space
1257	 * ranges from being reused for a new block group. This is needed to
1258	 * avoid races with trimming and scrub.
1259	 *
1260	 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1261	 * completely transactionless, so while it is trimming a range the
1262	 * currently running transaction might finish and a new one start,
1263	 * allowing for new block groups to be created that can reuse the same
1264	 * physical device locations unless we take this special care.
1265	 *
1266	 * There may also be an implicit trim operation if the file system
1267	 * is mounted with -odiscard. The same protections must remain
1268	 * in place until the extents have been discarded completely when
1269	 * the transaction commit has completed.
1270	 */
1271	remove_map = (atomic_read(&block_group->frozen) == 0);
1272	spin_unlock(&block_group->lock);
1273
1274	if (remove_map)
1275		btrfs_remove_chunk_map(fs_info, map);
1276
1277out:
1278	/* Once for the lookup reference */
1279	btrfs_put_block_group(block_group);
1280	if (remove_rsv)
1281		btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
1282	btrfs_free_path(path);
1283	return ret;
1284}
1285
1286struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1287		struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1288{
1289	struct btrfs_root *root = btrfs_block_group_root(fs_info);
1290	struct btrfs_chunk_map *map;
1291	unsigned int num_items;
1292
1293	map = btrfs_find_chunk_map(fs_info, chunk_offset, 1);
1294	ASSERT(map != NULL);
1295	ASSERT(map->start == chunk_offset);
1296
1297	/*
1298	 * We need to reserve 3 + N units from the metadata space info in order
1299	 * to remove a block group (done at btrfs_remove_chunk() and at
1300	 * btrfs_remove_block_group()), which are used for:
1301	 *
1302	 * 1 unit for adding the free space inode's orphan (located in the tree
1303	 * of tree roots).
1304	 * 1 unit for deleting the block group item (located in the extent
1305	 * tree).
1306	 * 1 unit for deleting the free space item (located in tree of tree
1307	 * roots).
1308	 * N units for deleting N device extent items corresponding to each
1309	 * stripe (located in the device tree).
1310	 *
1311	 * In order to remove a block group we also need to reserve units in the
1312	 * system space info in order to update the chunk tree (update one or
1313	 * more device items and remove one chunk item), but this is done at
1314	 * btrfs_remove_chunk() through a call to check_system_chunk().
1315	 */
1316	num_items = 3 + map->num_stripes;
1317	btrfs_free_chunk_map(map);
1318
1319	return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1320}
1321
1322/*
1323 * Mark block group @cache read-only, so later write won't happen to block
1324 * group @cache.
1325 *
1326 * If @force is not set, this function will only mark the block group readonly
1327 * if we have enough free space (1M) in other metadata/system block groups.
1328 * If @force is not set, this function will mark the block group readonly
1329 * without checking free space.
1330 *
1331 * NOTE: This function doesn't care if other block groups can contain all the
1332 * data in this block group. That check should be done by relocation routine,
1333 * not this function.
1334 */
1335static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1336{
1337	struct btrfs_space_info *sinfo = cache->space_info;
1338	u64 num_bytes;
1339	int ret = -ENOSPC;
1340
1341	spin_lock(&sinfo->lock);
1342	spin_lock(&cache->lock);
1343
1344	if (cache->swap_extents) {
1345		ret = -ETXTBSY;
1346		goto out;
1347	}
1348
1349	if (cache->ro) {
1350		cache->ro++;
1351		ret = 0;
1352		goto out;
1353	}
1354
1355	num_bytes = cache->length - cache->reserved - cache->pinned -
1356		    cache->bytes_super - cache->zone_unusable - cache->used;
1357
1358	/*
1359	 * Data never overcommits, even in mixed mode, so do just the straight
1360	 * check of left over space in how much we have allocated.
1361	 */
1362	if (force) {
1363		ret = 0;
1364	} else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1365		u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1366
1367		/*
1368		 * Here we make sure if we mark this bg RO, we still have enough
1369		 * free space as buffer.
1370		 */
1371		if (sinfo_used + num_bytes <= sinfo->total_bytes)
1372			ret = 0;
1373	} else {
1374		/*
1375		 * We overcommit metadata, so we need to do the
1376		 * btrfs_can_overcommit check here, and we need to pass in
1377		 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1378		 * leeway to allow us to mark this block group as read only.
1379		 */
1380		if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1381					 BTRFS_RESERVE_NO_FLUSH))
1382			ret = 0;
1383	}
1384
1385	if (!ret) {
1386		sinfo->bytes_readonly += num_bytes;
1387		if (btrfs_is_zoned(cache->fs_info)) {
1388			/* Migrate zone_unusable bytes to readonly */
1389			sinfo->bytes_readonly += cache->zone_unusable;
1390			sinfo->bytes_zone_unusable -= cache->zone_unusable;
1391			cache->zone_unusable = 0;
1392		}
1393		cache->ro++;
1394		list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1395	}
1396out:
1397	spin_unlock(&cache->lock);
1398	spin_unlock(&sinfo->lock);
1399	if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1400		btrfs_info(cache->fs_info,
1401			"unable to make block group %llu ro", cache->start);
1402		btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1403	}
1404	return ret;
1405}
1406
1407static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1408				 struct btrfs_block_group *bg)
1409{
1410	struct btrfs_fs_info *fs_info = bg->fs_info;
1411	struct btrfs_transaction *prev_trans = NULL;
1412	const u64 start = bg->start;
1413	const u64 end = start + bg->length - 1;
1414	int ret;
1415
1416	spin_lock(&fs_info->trans_lock);
1417	if (trans->transaction->list.prev != &fs_info->trans_list) {
1418		prev_trans = list_last_entry(&trans->transaction->list,
1419					     struct btrfs_transaction, list);
1420		refcount_inc(&prev_trans->use_count);
1421	}
1422	spin_unlock(&fs_info->trans_lock);
1423
1424	/*
1425	 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1426	 * btrfs_finish_extent_commit(). If we are at transaction N, another
1427	 * task might be running finish_extent_commit() for the previous
1428	 * transaction N - 1, and have seen a range belonging to the block
1429	 * group in pinned_extents before we were able to clear the whole block
1430	 * group range from pinned_extents. This means that task can lookup for
1431	 * the block group after we unpinned it from pinned_extents and removed
1432	 * it, leading to a BUG_ON() at unpin_extent_range().
1433	 */
1434	mutex_lock(&fs_info->unused_bg_unpin_mutex);
1435	if (prev_trans) {
1436		ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1437					EXTENT_DIRTY);
1438		if (ret)
1439			goto out;
1440	}
1441
1442	ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1443				EXTENT_DIRTY);
1444out:
1445	mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1446	if (prev_trans)
1447		btrfs_put_transaction(prev_trans);
1448
1449	return ret == 0;
1450}
1451
1452/*
1453 * Process the unused_bgs list and remove any that don't have any allocated
1454 * space inside of them.
1455 */
1456void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1457{
1458	LIST_HEAD(retry_list);
1459	struct btrfs_block_group *block_group;
1460	struct btrfs_space_info *space_info;
1461	struct btrfs_trans_handle *trans;
1462	const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1463	int ret = 0;
1464
1465	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1466		return;
1467
1468	if (btrfs_fs_closing(fs_info))
1469		return;
1470
1471	/*
1472	 * Long running balances can keep us blocked here for eternity, so
1473	 * simply skip deletion if we're unable to get the mutex.
1474	 */
1475	if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1476		return;
1477
1478	spin_lock(&fs_info->unused_bgs_lock);
1479	while (!list_empty(&fs_info->unused_bgs)) {
1480		u64 used;
1481		int trimming;
1482
1483		block_group = list_first_entry(&fs_info->unused_bgs,
1484					       struct btrfs_block_group,
1485					       bg_list);
1486		list_del_init(&block_group->bg_list);
1487
1488		space_info = block_group->space_info;
1489
1490		if (ret || btrfs_mixed_space_info(space_info)) {
1491			btrfs_put_block_group(block_group);
1492			continue;
1493		}
1494		spin_unlock(&fs_info->unused_bgs_lock);
1495
1496		btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1497
1498		/* Don't want to race with allocators so take the groups_sem */
1499		down_write(&space_info->groups_sem);
1500
1501		/*
1502		 * Async discard moves the final block group discard to be prior
1503		 * to the unused_bgs code path.  Therefore, if it's not fully
1504		 * trimmed, punt it back to the async discard lists.
1505		 */
1506		if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1507		    !btrfs_is_free_space_trimmed(block_group)) {
1508			trace_btrfs_skip_unused_block_group(block_group);
1509			up_write(&space_info->groups_sem);
1510			/* Requeue if we failed because of async discard */
1511			btrfs_discard_queue_work(&fs_info->discard_ctl,
1512						 block_group);
1513			goto next;
1514		}
1515
1516		spin_lock(&space_info->lock);
1517		spin_lock(&block_group->lock);
1518		if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1519		    list_is_singular(&block_group->list)) {
1520			/*
1521			 * We want to bail if we made new allocations or have
1522			 * outstanding allocations in this block group.  We do
1523			 * the ro check in case balance is currently acting on
1524			 * this block group.
 
 
 
 
 
 
 
1525			 */
1526			trace_btrfs_skip_unused_block_group(block_group);
1527			spin_unlock(&block_group->lock);
1528			spin_unlock(&space_info->lock);
1529			up_write(&space_info->groups_sem);
1530			goto next;
1531		}
1532
1533		/*
1534		 * The block group may be unused but there may be space reserved
1535		 * accounting with the existence of that block group, that is,
1536		 * space_info->bytes_may_use was incremented by a task but no
1537		 * space was yet allocated from the block group by the task.
1538		 * That space may or may not be allocated, as we are generally
1539		 * pessimistic about space reservation for metadata as well as
1540		 * for data when using compression (as we reserve space based on
1541		 * the worst case, when data can't be compressed, and before
1542		 * actually attempting compression, before starting writeback).
1543		 *
1544		 * So check if the total space of the space_info minus the size
1545		 * of this block group is less than the used space of the
1546		 * space_info - if that's the case, then it means we have tasks
1547		 * that might be relying on the block group in order to allocate
1548		 * extents, and add back the block group to the unused list when
1549		 * we finish, so that we retry later in case no tasks ended up
1550		 * needing to allocate extents from the block group.
1551		 */
1552		used = btrfs_space_info_used(space_info, true);
1553		if (space_info->total_bytes - block_group->length < used) {
 
1554			/*
1555			 * Add a reference for the list, compensate for the ref
1556			 * drop under the "next" label for the
1557			 * fs_info->unused_bgs list.
1558			 */
1559			btrfs_get_block_group(block_group);
1560			list_add_tail(&block_group->bg_list, &retry_list);
1561
1562			trace_btrfs_skip_unused_block_group(block_group);
1563			spin_unlock(&block_group->lock);
1564			spin_unlock(&space_info->lock);
1565			up_write(&space_info->groups_sem);
1566			goto next;
1567		}
1568
1569		spin_unlock(&block_group->lock);
1570		spin_unlock(&space_info->lock);
1571
1572		/* We don't want to force the issue, only flip if it's ok. */
1573		ret = inc_block_group_ro(block_group, 0);
1574		up_write(&space_info->groups_sem);
1575		if (ret < 0) {
1576			ret = 0;
1577			goto next;
1578		}
1579
1580		ret = btrfs_zone_finish(block_group);
1581		if (ret < 0) {
1582			btrfs_dec_block_group_ro(block_group);
1583			if (ret == -EAGAIN)
1584				ret = 0;
1585			goto next;
1586		}
1587
1588		/*
1589		 * Want to do this before we do anything else so we can recover
1590		 * properly if we fail to join the transaction.
1591		 */
1592		trans = btrfs_start_trans_remove_block_group(fs_info,
1593						     block_group->start);
1594		if (IS_ERR(trans)) {
1595			btrfs_dec_block_group_ro(block_group);
1596			ret = PTR_ERR(trans);
1597			goto next;
1598		}
1599
1600		/*
1601		 * We could have pending pinned extents for this block group,
1602		 * just delete them, we don't care about them anymore.
1603		 */
1604		if (!clean_pinned_extents(trans, block_group)) {
1605			btrfs_dec_block_group_ro(block_group);
1606			goto end_trans;
1607		}
1608
1609		/*
1610		 * At this point, the block_group is read only and should fail
1611		 * new allocations.  However, btrfs_finish_extent_commit() can
1612		 * cause this block_group to be placed back on the discard
1613		 * lists because now the block_group isn't fully discarded.
1614		 * Bail here and try again later after discarding everything.
1615		 */
1616		spin_lock(&fs_info->discard_ctl.lock);
1617		if (!list_empty(&block_group->discard_list)) {
1618			spin_unlock(&fs_info->discard_ctl.lock);
1619			btrfs_dec_block_group_ro(block_group);
1620			btrfs_discard_queue_work(&fs_info->discard_ctl,
1621						 block_group);
1622			goto end_trans;
1623		}
1624		spin_unlock(&fs_info->discard_ctl.lock);
1625
1626		/* Reset pinned so btrfs_put_block_group doesn't complain */
1627		spin_lock(&space_info->lock);
1628		spin_lock(&block_group->lock);
1629
1630		btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1631						     -block_group->pinned);
1632		space_info->bytes_readonly += block_group->pinned;
1633		block_group->pinned = 0;
1634
1635		spin_unlock(&block_group->lock);
1636		spin_unlock(&space_info->lock);
1637
1638		/*
1639		 * The normal path here is an unused block group is passed here,
1640		 * then trimming is handled in the transaction commit path.
1641		 * Async discard interposes before this to do the trimming
1642		 * before coming down the unused block group path as trimming
1643		 * will no longer be done later in the transaction commit path.
1644		 */
1645		if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1646			goto flip_async;
1647
1648		/*
1649		 * DISCARD can flip during remount. On zoned filesystems, we
1650		 * need to reset sequential-required zones.
1651		 */
1652		trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1653				btrfs_is_zoned(fs_info);
1654
1655		/* Implicit trim during transaction commit. */
1656		if (trimming)
1657			btrfs_freeze_block_group(block_group);
1658
1659		/*
1660		 * Btrfs_remove_chunk will abort the transaction if things go
1661		 * horribly wrong.
1662		 */
1663		ret = btrfs_remove_chunk(trans, block_group->start);
1664
1665		if (ret) {
1666			if (trimming)
1667				btrfs_unfreeze_block_group(block_group);
1668			goto end_trans;
1669		}
1670
1671		/*
1672		 * If we're not mounted with -odiscard, we can just forget
1673		 * about this block group. Otherwise we'll need to wait
1674		 * until transaction commit to do the actual discard.
1675		 */
1676		if (trimming) {
1677			spin_lock(&fs_info->unused_bgs_lock);
1678			/*
1679			 * A concurrent scrub might have added us to the list
1680			 * fs_info->unused_bgs, so use a list_move operation
1681			 * to add the block group to the deleted_bgs list.
1682			 */
1683			list_move(&block_group->bg_list,
1684				  &trans->transaction->deleted_bgs);
1685			spin_unlock(&fs_info->unused_bgs_lock);
1686			btrfs_get_block_group(block_group);
1687		}
1688end_trans:
1689		btrfs_end_transaction(trans);
1690next:
1691		btrfs_put_block_group(block_group);
1692		spin_lock(&fs_info->unused_bgs_lock);
1693	}
1694	list_splice_tail(&retry_list, &fs_info->unused_bgs);
1695	spin_unlock(&fs_info->unused_bgs_lock);
1696	mutex_unlock(&fs_info->reclaim_bgs_lock);
1697	return;
1698
1699flip_async:
1700	btrfs_end_transaction(trans);
1701	spin_lock(&fs_info->unused_bgs_lock);
1702	list_splice_tail(&retry_list, &fs_info->unused_bgs);
1703	spin_unlock(&fs_info->unused_bgs_lock);
1704	mutex_unlock(&fs_info->reclaim_bgs_lock);
1705	btrfs_put_block_group(block_group);
1706	btrfs_discard_punt_unused_bgs_list(fs_info);
1707}
1708
1709void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1710{
1711	struct btrfs_fs_info *fs_info = bg->fs_info;
1712
1713	spin_lock(&fs_info->unused_bgs_lock);
1714	if (list_empty(&bg->bg_list)) {
1715		btrfs_get_block_group(bg);
1716		trace_btrfs_add_unused_block_group(bg);
1717		list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1718	} else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1719		/* Pull out the block group from the reclaim_bgs list. */
1720		trace_btrfs_add_unused_block_group(bg);
1721		list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1722	}
1723	spin_unlock(&fs_info->unused_bgs_lock);
1724}
1725
1726/*
1727 * We want block groups with a low number of used bytes to be in the beginning
1728 * of the list, so they will get reclaimed first.
1729 */
1730static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1731			   const struct list_head *b)
1732{
1733	const struct btrfs_block_group *bg1, *bg2;
1734
1735	bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1736	bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1737
1738	return bg1->used > bg2->used;
1739}
1740
1741static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1742{
1743	if (btrfs_is_zoned(fs_info))
1744		return btrfs_zoned_should_reclaim(fs_info);
1745	return true;
1746}
1747
1748static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1749{
1750	const struct btrfs_space_info *space_info = bg->space_info;
1751	const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1752	const u64 new_val = bg->used;
1753	const u64 old_val = new_val + bytes_freed;
1754	u64 thresh;
1755
1756	if (reclaim_thresh == 0)
1757		return false;
1758
1759	thresh = mult_perc(bg->length, reclaim_thresh);
1760
1761	/*
1762	 * If we were below the threshold before don't reclaim, we are likely a
1763	 * brand new block group and we don't want to relocate new block groups.
1764	 */
1765	if (old_val < thresh)
1766		return false;
1767	if (new_val >= thresh)
1768		return false;
1769	return true;
1770}
1771
1772void btrfs_reclaim_bgs_work(struct work_struct *work)
1773{
1774	struct btrfs_fs_info *fs_info =
1775		container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1776	struct btrfs_block_group *bg;
1777	struct btrfs_space_info *space_info;
1778
1779	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1780		return;
1781
1782	if (btrfs_fs_closing(fs_info))
1783		return;
1784
1785	if (!btrfs_should_reclaim(fs_info))
1786		return;
1787
1788	sb_start_write(fs_info->sb);
1789
1790	if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1791		sb_end_write(fs_info->sb);
1792		return;
1793	}
1794
1795	/*
1796	 * Long running balances can keep us blocked here for eternity, so
1797	 * simply skip reclaim if we're unable to get the mutex.
1798	 */
1799	if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1800		btrfs_exclop_finish(fs_info);
1801		sb_end_write(fs_info->sb);
1802		return;
1803	}
1804
1805	spin_lock(&fs_info->unused_bgs_lock);
1806	/*
1807	 * Sort happens under lock because we can't simply splice it and sort.
1808	 * The block groups might still be in use and reachable via bg_list,
1809	 * and their presence in the reclaim_bgs list must be preserved.
1810	 */
1811	list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1812	while (!list_empty(&fs_info->reclaim_bgs)) {
1813		u64 zone_unusable;
1814		int ret = 0;
1815
1816		bg = list_first_entry(&fs_info->reclaim_bgs,
1817				      struct btrfs_block_group,
1818				      bg_list);
1819		list_del_init(&bg->bg_list);
1820
1821		space_info = bg->space_info;
1822		spin_unlock(&fs_info->unused_bgs_lock);
1823
1824		/* Don't race with allocators so take the groups_sem */
1825		down_write(&space_info->groups_sem);
1826
1827		spin_lock(&bg->lock);
1828		if (bg->reserved || bg->pinned || bg->ro) {
1829			/*
1830			 * We want to bail if we made new allocations or have
1831			 * outstanding allocations in this block group.  We do
1832			 * the ro check in case balance is currently acting on
1833			 * this block group.
1834			 */
1835			spin_unlock(&bg->lock);
1836			up_write(&space_info->groups_sem);
1837			goto next;
1838		}
1839		if (bg->used == 0) {
1840			/*
1841			 * It is possible that we trigger relocation on a block
1842			 * group as its extents are deleted and it first goes
1843			 * below the threshold, then shortly after goes empty.
1844			 *
1845			 * In this case, relocating it does delete it, but has
1846			 * some overhead in relocation specific metadata, looking
1847			 * for the non-existent extents and running some extra
1848			 * transactions, which we can avoid by using one of the
1849			 * other mechanisms for dealing with empty block groups.
1850			 */
1851			if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1852				btrfs_mark_bg_unused(bg);
1853			spin_unlock(&bg->lock);
1854			up_write(&space_info->groups_sem);
1855			goto next;
1856
1857		}
1858		/*
1859		 * The block group might no longer meet the reclaim condition by
1860		 * the time we get around to reclaiming it, so to avoid
1861		 * reclaiming overly full block_groups, skip reclaiming them.
1862		 *
1863		 * Since the decision making process also depends on the amount
1864		 * being freed, pass in a fake giant value to skip that extra
1865		 * check, which is more meaningful when adding to the list in
1866		 * the first place.
1867		 */
1868		if (!should_reclaim_block_group(bg, bg->length)) {
1869			spin_unlock(&bg->lock);
1870			up_write(&space_info->groups_sem);
1871			goto next;
1872		}
1873		spin_unlock(&bg->lock);
1874
1875		/*
1876		 * Get out fast, in case we're read-only or unmounting the
1877		 * filesystem. It is OK to drop block groups from the list even
1878		 * for the read-only case. As we did sb_start_write(),
1879		 * "mount -o remount,ro" won't happen and read-only filesystem
1880		 * means it is forced read-only due to a fatal error. So, it
1881		 * never gets back to read-write to let us reclaim again.
1882		 */
1883		if (btrfs_need_cleaner_sleep(fs_info)) {
1884			up_write(&space_info->groups_sem);
1885			goto next;
1886		}
1887
1888		/*
1889		 * Cache the zone_unusable value before turning the block group
1890		 * to read only. As soon as the blog group is read only it's
1891		 * zone_unusable value gets moved to the block group's read-only
1892		 * bytes and isn't available for calculations anymore.
1893		 */
1894		zone_unusable = bg->zone_unusable;
1895		ret = inc_block_group_ro(bg, 0);
1896		up_write(&space_info->groups_sem);
1897		if (ret < 0)
1898			goto next;
1899
1900		btrfs_info(fs_info,
1901			"reclaiming chunk %llu with %llu%% used %llu%% unusable",
1902				bg->start,
1903				div64_u64(bg->used * 100, bg->length),
1904				div64_u64(zone_unusable * 100, bg->length));
1905		trace_btrfs_reclaim_block_group(bg);
1906		ret = btrfs_relocate_chunk(fs_info, bg->start);
1907		if (ret) {
1908			btrfs_dec_block_group_ro(bg);
1909			btrfs_err(fs_info, "error relocating chunk %llu",
1910				  bg->start);
1911		}
1912
1913next:
1914		if (ret)
1915			btrfs_mark_bg_to_reclaim(bg);
1916		btrfs_put_block_group(bg);
1917
1918		mutex_unlock(&fs_info->reclaim_bgs_lock);
1919		/*
1920		 * Reclaiming all the block groups in the list can take really
1921		 * long.  Prioritize cleaning up unused block groups.
1922		 */
1923		btrfs_delete_unused_bgs(fs_info);
1924		/*
1925		 * If we are interrupted by a balance, we can just bail out. The
1926		 * cleaner thread restart again if necessary.
1927		 */
1928		if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1929			goto end;
1930		spin_lock(&fs_info->unused_bgs_lock);
1931	}
1932	spin_unlock(&fs_info->unused_bgs_lock);
1933	mutex_unlock(&fs_info->reclaim_bgs_lock);
1934end:
1935	btrfs_exclop_finish(fs_info);
1936	sb_end_write(fs_info->sb);
1937}
1938
1939void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1940{
1941	spin_lock(&fs_info->unused_bgs_lock);
1942	if (!list_empty(&fs_info->reclaim_bgs))
1943		queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1944	spin_unlock(&fs_info->unused_bgs_lock);
1945}
1946
1947void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1948{
1949	struct btrfs_fs_info *fs_info = bg->fs_info;
1950
1951	spin_lock(&fs_info->unused_bgs_lock);
1952	if (list_empty(&bg->bg_list)) {
1953		btrfs_get_block_group(bg);
1954		trace_btrfs_add_reclaim_block_group(bg);
1955		list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1956	}
1957	spin_unlock(&fs_info->unused_bgs_lock);
1958}
1959
1960static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1961			   struct btrfs_path *path)
1962{
1963	struct btrfs_chunk_map *map;
1964	struct btrfs_block_group_item bg;
1965	struct extent_buffer *leaf;
1966	int slot;
1967	u64 flags;
1968	int ret = 0;
1969
1970	slot = path->slots[0];
1971	leaf = path->nodes[0];
1972
1973	map = btrfs_find_chunk_map(fs_info, key->objectid, key->offset);
1974	if (!map) {
1975		btrfs_err(fs_info,
1976			  "logical %llu len %llu found bg but no related chunk",
1977			  key->objectid, key->offset);
1978		return -ENOENT;
1979	}
1980
1981	if (map->start != key->objectid || map->chunk_len != key->offset) {
1982		btrfs_err(fs_info,
1983			"block group %llu len %llu mismatch with chunk %llu len %llu",
1984			  key->objectid, key->offset, map->start, map->chunk_len);
1985		ret = -EUCLEAN;
1986		goto out_free_map;
1987	}
1988
1989	read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1990			   sizeof(bg));
1991	flags = btrfs_stack_block_group_flags(&bg) &
1992		BTRFS_BLOCK_GROUP_TYPE_MASK;
1993
1994	if (flags != (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1995		btrfs_err(fs_info,
1996"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1997			  key->objectid, key->offset, flags,
1998			  (BTRFS_BLOCK_GROUP_TYPE_MASK & map->type));
1999		ret = -EUCLEAN;
2000	}
2001
2002out_free_map:
2003	btrfs_free_chunk_map(map);
2004	return ret;
2005}
2006
2007static int find_first_block_group(struct btrfs_fs_info *fs_info,
2008				  struct btrfs_path *path,
2009				  struct btrfs_key *key)
2010{
2011	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2012	int ret;
2013	struct btrfs_key found_key;
2014
2015	btrfs_for_each_slot(root, key, &found_key, path, ret) {
2016		if (found_key.objectid >= key->objectid &&
2017		    found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2018			return read_bg_from_eb(fs_info, &found_key, path);
2019		}
2020	}
2021	return ret;
2022}
2023
2024static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2025{
2026	u64 extra_flags = chunk_to_extended(flags) &
2027				BTRFS_EXTENDED_PROFILE_MASK;
2028
2029	write_seqlock(&fs_info->profiles_lock);
2030	if (flags & BTRFS_BLOCK_GROUP_DATA)
2031		fs_info->avail_data_alloc_bits |= extra_flags;
2032	if (flags & BTRFS_BLOCK_GROUP_METADATA)
2033		fs_info->avail_metadata_alloc_bits |= extra_flags;
2034	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2035		fs_info->avail_system_alloc_bits |= extra_flags;
2036	write_sequnlock(&fs_info->profiles_lock);
2037}
2038
2039/*
2040 * Map a physical disk address to a list of logical addresses.
2041 *
2042 * @fs_info:       the filesystem
2043 * @chunk_start:   logical address of block group
2044 * @physical:	   physical address to map to logical addresses
2045 * @logical:	   return array of logical addresses which map to @physical
2046 * @naddrs:	   length of @logical
2047 * @stripe_len:    size of IO stripe for the given block group
2048 *
2049 * Maps a particular @physical disk address to a list of @logical addresses.
2050 * Used primarily to exclude those portions of a block group that contain super
2051 * block copies.
2052 */
2053int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2054		     u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2055{
2056	struct btrfs_chunk_map *map;
2057	u64 *buf;
2058	u64 bytenr;
2059	u64 data_stripe_length;
2060	u64 io_stripe_size;
2061	int i, nr = 0;
2062	int ret = 0;
2063
2064	map = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2065	if (IS_ERR(map))
2066		return -EIO;
2067
2068	data_stripe_length = map->stripe_size;
2069	io_stripe_size = BTRFS_STRIPE_LEN;
2070	chunk_start = map->start;
2071
2072	/* For RAID5/6 adjust to a full IO stripe length */
2073	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2074		io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2075
2076	buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2077	if (!buf) {
2078		ret = -ENOMEM;
2079		goto out;
2080	}
2081
2082	for (i = 0; i < map->num_stripes; i++) {
2083		bool already_inserted = false;
2084		u32 stripe_nr;
2085		u32 offset;
2086		int j;
2087
2088		if (!in_range(physical, map->stripes[i].physical,
2089			      data_stripe_length))
2090			continue;
2091
2092		stripe_nr = (physical - map->stripes[i].physical) >>
2093			    BTRFS_STRIPE_LEN_SHIFT;
2094		offset = (physical - map->stripes[i].physical) &
2095			 BTRFS_STRIPE_LEN_MASK;
2096
2097		if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2098				 BTRFS_BLOCK_GROUP_RAID10))
2099			stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2100					    map->sub_stripes);
2101		/*
2102		 * The remaining case would be for RAID56, multiply by
2103		 * nr_data_stripes().  Alternatively, just use rmap_len below
2104		 * instead of map->stripe_len
2105		 */
2106		bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2107
2108		/* Ensure we don't add duplicate addresses */
2109		for (j = 0; j < nr; j++) {
2110			if (buf[j] == bytenr) {
2111				already_inserted = true;
2112				break;
2113			}
2114		}
2115
2116		if (!already_inserted)
2117			buf[nr++] = bytenr;
2118	}
2119
2120	*logical = buf;
2121	*naddrs = nr;
2122	*stripe_len = io_stripe_size;
2123out:
2124	btrfs_free_chunk_map(map);
2125	return ret;
2126}
2127
2128static int exclude_super_stripes(struct btrfs_block_group *cache)
2129{
2130	struct btrfs_fs_info *fs_info = cache->fs_info;
2131	const bool zoned = btrfs_is_zoned(fs_info);
2132	u64 bytenr;
2133	u64 *logical;
2134	int stripe_len;
2135	int i, nr, ret;
2136
2137	if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2138		stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2139		cache->bytes_super += stripe_len;
2140		ret = set_extent_bit(&fs_info->excluded_extents, cache->start,
2141				     cache->start + stripe_len - 1,
2142				     EXTENT_UPTODATE, NULL);
2143		if (ret)
2144			return ret;
2145	}
2146
2147	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2148		bytenr = btrfs_sb_offset(i);
2149		ret = btrfs_rmap_block(fs_info, cache->start,
2150				       bytenr, &logical, &nr, &stripe_len);
2151		if (ret)
2152			return ret;
2153
2154		/* Shouldn't have super stripes in sequential zones */
2155		if (zoned && nr) {
2156			kfree(logical);
2157			btrfs_err(fs_info,
2158			"zoned: block group %llu must not contain super block",
2159				  cache->start);
2160			return -EUCLEAN;
2161		}
2162
2163		while (nr--) {
2164			u64 len = min_t(u64, stripe_len,
2165				cache->start + cache->length - logical[nr]);
2166
2167			cache->bytes_super += len;
2168			ret = set_extent_bit(&fs_info->excluded_extents, logical[nr],
2169					     logical[nr] + len - 1,
2170					     EXTENT_UPTODATE, NULL);
2171			if (ret) {
2172				kfree(logical);
2173				return ret;
2174			}
2175		}
2176
2177		kfree(logical);
2178	}
2179	return 0;
2180}
2181
2182static struct btrfs_block_group *btrfs_create_block_group_cache(
2183		struct btrfs_fs_info *fs_info, u64 start)
2184{
2185	struct btrfs_block_group *cache;
2186
2187	cache = kzalloc(sizeof(*cache), GFP_NOFS);
2188	if (!cache)
2189		return NULL;
2190
2191	cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2192					GFP_NOFS);
2193	if (!cache->free_space_ctl) {
2194		kfree(cache);
2195		return NULL;
2196	}
2197
2198	cache->start = start;
2199
2200	cache->fs_info = fs_info;
2201	cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2202
2203	cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2204
2205	refcount_set(&cache->refs, 1);
2206	spin_lock_init(&cache->lock);
2207	init_rwsem(&cache->data_rwsem);
2208	INIT_LIST_HEAD(&cache->list);
2209	INIT_LIST_HEAD(&cache->cluster_list);
2210	INIT_LIST_HEAD(&cache->bg_list);
2211	INIT_LIST_HEAD(&cache->ro_list);
2212	INIT_LIST_HEAD(&cache->discard_list);
2213	INIT_LIST_HEAD(&cache->dirty_list);
2214	INIT_LIST_HEAD(&cache->io_list);
2215	INIT_LIST_HEAD(&cache->active_bg_list);
2216	btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2217	atomic_set(&cache->frozen, 0);
2218	mutex_init(&cache->free_space_lock);
2219
2220	return cache;
2221}
2222
2223/*
2224 * Iterate all chunks and verify that each of them has the corresponding block
2225 * group
2226 */
2227static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2228{
2229	u64 start = 0;
2230	int ret = 0;
2231
2232	while (1) {
2233		struct btrfs_chunk_map *map;
2234		struct btrfs_block_group *bg;
2235
2236		/*
2237		 * btrfs_find_chunk_map() will return the first chunk map
2238		 * intersecting the range, so setting @length to 1 is enough to
2239		 * get the first chunk.
2240		 */
2241		map = btrfs_find_chunk_map(fs_info, start, 1);
2242		if (!map)
2243			break;
2244
2245		bg = btrfs_lookup_block_group(fs_info, map->start);
2246		if (!bg) {
2247			btrfs_err(fs_info,
2248	"chunk start=%llu len=%llu doesn't have corresponding block group",
2249				     map->start, map->chunk_len);
2250			ret = -EUCLEAN;
2251			btrfs_free_chunk_map(map);
2252			break;
2253		}
2254		if (bg->start != map->start || bg->length != map->chunk_len ||
2255		    (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2256		    (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2257			btrfs_err(fs_info,
2258"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2259				map->start, map->chunk_len,
2260				map->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2261				bg->start, bg->length,
2262				bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2263			ret = -EUCLEAN;
2264			btrfs_free_chunk_map(map);
2265			btrfs_put_block_group(bg);
2266			break;
2267		}
2268		start = map->start + map->chunk_len;
2269		btrfs_free_chunk_map(map);
2270		btrfs_put_block_group(bg);
2271	}
2272	return ret;
2273}
2274
2275static int read_one_block_group(struct btrfs_fs_info *info,
2276				struct btrfs_block_group_item *bgi,
2277				const struct btrfs_key *key,
2278				int need_clear)
2279{
2280	struct btrfs_block_group *cache;
2281	const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2282	int ret;
2283
2284	ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2285
2286	cache = btrfs_create_block_group_cache(info, key->objectid);
2287	if (!cache)
2288		return -ENOMEM;
2289
2290	cache->length = key->offset;
2291	cache->used = btrfs_stack_block_group_used(bgi);
2292	cache->commit_used = cache->used;
2293	cache->flags = btrfs_stack_block_group_flags(bgi);
2294	cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2295
2296	set_free_space_tree_thresholds(cache);
2297
2298	if (need_clear) {
2299		/*
2300		 * When we mount with old space cache, we need to
2301		 * set BTRFS_DC_CLEAR and set dirty flag.
2302		 *
2303		 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2304		 *    truncate the old free space cache inode and
2305		 *    setup a new one.
2306		 * b) Setting 'dirty flag' makes sure that we flush
2307		 *    the new space cache info onto disk.
2308		 */
2309		if (btrfs_test_opt(info, SPACE_CACHE))
2310			cache->disk_cache_state = BTRFS_DC_CLEAR;
2311	}
2312	if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2313	    (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2314			btrfs_err(info,
2315"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2316				  cache->start);
2317			ret = -EINVAL;
2318			goto error;
2319	}
2320
2321	ret = btrfs_load_block_group_zone_info(cache, false);
2322	if (ret) {
2323		btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2324			  cache->start);
2325		goto error;
2326	}
2327
2328	/*
2329	 * We need to exclude the super stripes now so that the space info has
2330	 * super bytes accounted for, otherwise we'll think we have more space
2331	 * than we actually do.
2332	 */
2333	ret = exclude_super_stripes(cache);
2334	if (ret) {
2335		/* We may have excluded something, so call this just in case. */
2336		btrfs_free_excluded_extents(cache);
2337		goto error;
2338	}
2339
2340	/*
2341	 * For zoned filesystem, space after the allocation offset is the only
2342	 * free space for a block group. So, we don't need any caching work.
2343	 * btrfs_calc_zone_unusable() will set the amount of free space and
2344	 * zone_unusable space.
2345	 *
2346	 * For regular filesystem, check for two cases, either we are full, and
2347	 * therefore don't need to bother with the caching work since we won't
2348	 * find any space, or we are empty, and we can just add all the space
2349	 * in and be done with it.  This saves us _a_lot_ of time, particularly
2350	 * in the full case.
2351	 */
2352	if (btrfs_is_zoned(info)) {
2353		btrfs_calc_zone_unusable(cache);
2354		/* Should not have any excluded extents. Just in case, though. */
2355		btrfs_free_excluded_extents(cache);
2356	} else if (cache->length == cache->used) {
2357		cache->cached = BTRFS_CACHE_FINISHED;
2358		btrfs_free_excluded_extents(cache);
2359	} else if (cache->used == 0) {
2360		cache->cached = BTRFS_CACHE_FINISHED;
2361		ret = btrfs_add_new_free_space(cache, cache->start,
2362					       cache->start + cache->length, NULL);
2363		btrfs_free_excluded_extents(cache);
2364		if (ret)
2365			goto error;
2366	}
2367
2368	ret = btrfs_add_block_group_cache(info, cache);
2369	if (ret) {
2370		btrfs_remove_free_space_cache(cache);
2371		goto error;
2372	}
2373	trace_btrfs_add_block_group(info, cache, 0);
2374	btrfs_add_bg_to_space_info(info, cache);
2375
2376	set_avail_alloc_bits(info, cache->flags);
2377	if (btrfs_chunk_writeable(info, cache->start)) {
2378		if (cache->used == 0) {
2379			ASSERT(list_empty(&cache->bg_list));
2380			if (btrfs_test_opt(info, DISCARD_ASYNC))
2381				btrfs_discard_queue_work(&info->discard_ctl, cache);
2382			else
2383				btrfs_mark_bg_unused(cache);
2384		}
2385	} else {
2386		inc_block_group_ro(cache, 1);
2387	}
2388
2389	return 0;
2390error:
2391	btrfs_put_block_group(cache);
2392	return ret;
2393}
2394
2395static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2396{
2397	struct rb_node *node;
2398	int ret = 0;
2399
2400	for (node = rb_first_cached(&fs_info->mapping_tree); node; node = rb_next(node)) {
2401		struct btrfs_chunk_map *map;
2402		struct btrfs_block_group *bg;
2403
2404		map = rb_entry(node, struct btrfs_chunk_map, rb_node);
2405		bg = btrfs_create_block_group_cache(fs_info, map->start);
2406		if (!bg) {
2407			ret = -ENOMEM;
2408			break;
2409		}
2410
2411		/* Fill dummy cache as FULL */
2412		bg->length = map->chunk_len;
2413		bg->flags = map->type;
2414		bg->cached = BTRFS_CACHE_FINISHED;
2415		bg->used = map->chunk_len;
2416		bg->flags = map->type;
2417		ret = btrfs_add_block_group_cache(fs_info, bg);
2418		/*
2419		 * We may have some valid block group cache added already, in
2420		 * that case we skip to the next one.
2421		 */
2422		if (ret == -EEXIST) {
2423			ret = 0;
2424			btrfs_put_block_group(bg);
2425			continue;
2426		}
2427
2428		if (ret) {
2429			btrfs_remove_free_space_cache(bg);
2430			btrfs_put_block_group(bg);
2431			break;
2432		}
2433
2434		btrfs_add_bg_to_space_info(fs_info, bg);
2435
2436		set_avail_alloc_bits(fs_info, bg->flags);
2437	}
2438	if (!ret)
2439		btrfs_init_global_block_rsv(fs_info);
2440	return ret;
2441}
2442
2443int btrfs_read_block_groups(struct btrfs_fs_info *info)
2444{
2445	struct btrfs_root *root = btrfs_block_group_root(info);
2446	struct btrfs_path *path;
2447	int ret;
2448	struct btrfs_block_group *cache;
2449	struct btrfs_space_info *space_info;
2450	struct btrfs_key key;
2451	int need_clear = 0;
2452	u64 cache_gen;
2453
2454	/*
2455	 * Either no extent root (with ibadroots rescue option) or we have
2456	 * unsupported RO options. The fs can never be mounted read-write, so no
2457	 * need to waste time searching block group items.
2458	 *
2459	 * This also allows new extent tree related changes to be RO compat,
2460	 * no need for a full incompat flag.
2461	 */
2462	if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2463		      ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2464		return fill_dummy_bgs(info);
2465
2466	key.objectid = 0;
2467	key.offset = 0;
2468	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2469	path = btrfs_alloc_path();
2470	if (!path)
2471		return -ENOMEM;
2472
2473	cache_gen = btrfs_super_cache_generation(info->super_copy);
2474	if (btrfs_test_opt(info, SPACE_CACHE) &&
2475	    btrfs_super_generation(info->super_copy) != cache_gen)
2476		need_clear = 1;
2477	if (btrfs_test_opt(info, CLEAR_CACHE))
2478		need_clear = 1;
2479
2480	while (1) {
2481		struct btrfs_block_group_item bgi;
2482		struct extent_buffer *leaf;
2483		int slot;
2484
2485		ret = find_first_block_group(info, path, &key);
2486		if (ret > 0)
2487			break;
2488		if (ret != 0)
2489			goto error;
2490
2491		leaf = path->nodes[0];
2492		slot = path->slots[0];
2493
2494		read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2495				   sizeof(bgi));
2496
2497		btrfs_item_key_to_cpu(leaf, &key, slot);
2498		btrfs_release_path(path);
2499		ret = read_one_block_group(info, &bgi, &key, need_clear);
2500		if (ret < 0)
2501			goto error;
2502		key.objectid += key.offset;
2503		key.offset = 0;
2504	}
2505	btrfs_release_path(path);
2506
2507	list_for_each_entry(space_info, &info->space_info, list) {
2508		int i;
2509
2510		for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2511			if (list_empty(&space_info->block_groups[i]))
2512				continue;
2513			cache = list_first_entry(&space_info->block_groups[i],
2514						 struct btrfs_block_group,
2515						 list);
2516			btrfs_sysfs_add_block_group_type(cache);
2517		}
2518
2519		if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2520		      (BTRFS_BLOCK_GROUP_RAID10 |
2521		       BTRFS_BLOCK_GROUP_RAID1_MASK |
2522		       BTRFS_BLOCK_GROUP_RAID56_MASK |
2523		       BTRFS_BLOCK_GROUP_DUP)))
2524			continue;
2525		/*
2526		 * Avoid allocating from un-mirrored block group if there are
2527		 * mirrored block groups.
2528		 */
2529		list_for_each_entry(cache,
2530				&space_info->block_groups[BTRFS_RAID_RAID0],
2531				list)
2532			inc_block_group_ro(cache, 1);
2533		list_for_each_entry(cache,
2534				&space_info->block_groups[BTRFS_RAID_SINGLE],
2535				list)
2536			inc_block_group_ro(cache, 1);
2537	}
2538
2539	btrfs_init_global_block_rsv(info);
2540	ret = check_chunk_block_group_mappings(info);
2541error:
2542	btrfs_free_path(path);
2543	/*
2544	 * We've hit some error while reading the extent tree, and have
2545	 * rescue=ibadroots mount option.
2546	 * Try to fill the tree using dummy block groups so that the user can
2547	 * continue to mount and grab their data.
2548	 */
2549	if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2550		ret = fill_dummy_bgs(info);
2551	return ret;
2552}
2553
2554/*
2555 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2556 * allocation.
2557 *
2558 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2559 * phases.
2560 */
2561static int insert_block_group_item(struct btrfs_trans_handle *trans,
2562				   struct btrfs_block_group *block_group)
2563{
2564	struct btrfs_fs_info *fs_info = trans->fs_info;
2565	struct btrfs_block_group_item bgi;
2566	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2567	struct btrfs_key key;
2568	u64 old_commit_used;
2569	int ret;
2570
2571	spin_lock(&block_group->lock);
2572	btrfs_set_stack_block_group_used(&bgi, block_group->used);
2573	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2574						   block_group->global_root_id);
2575	btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2576	old_commit_used = block_group->commit_used;
2577	block_group->commit_used = block_group->used;
2578	key.objectid = block_group->start;
2579	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2580	key.offset = block_group->length;
2581	spin_unlock(&block_group->lock);
2582
2583	ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2584	if (ret < 0) {
2585		spin_lock(&block_group->lock);
2586		block_group->commit_used = old_commit_used;
2587		spin_unlock(&block_group->lock);
2588	}
2589
2590	return ret;
2591}
2592
2593static int insert_dev_extent(struct btrfs_trans_handle *trans,
2594			    struct btrfs_device *device, u64 chunk_offset,
2595			    u64 start, u64 num_bytes)
2596{
2597	struct btrfs_fs_info *fs_info = device->fs_info;
2598	struct btrfs_root *root = fs_info->dev_root;
2599	struct btrfs_path *path;
2600	struct btrfs_dev_extent *extent;
2601	struct extent_buffer *leaf;
2602	struct btrfs_key key;
2603	int ret;
2604
2605	WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2606	WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2607	path = btrfs_alloc_path();
2608	if (!path)
2609		return -ENOMEM;
2610
2611	key.objectid = device->devid;
2612	key.type = BTRFS_DEV_EXTENT_KEY;
2613	key.offset = start;
2614	ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2615	if (ret)
2616		goto out;
2617
2618	leaf = path->nodes[0];
2619	extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2620	btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2621	btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2622					    BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2623	btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2624
2625	btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2626	btrfs_mark_buffer_dirty(trans, leaf);
2627out:
2628	btrfs_free_path(path);
2629	return ret;
2630}
2631
2632/*
2633 * This function belongs to phase 2.
2634 *
2635 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2636 * phases.
2637 */
2638static int insert_dev_extents(struct btrfs_trans_handle *trans,
2639				   u64 chunk_offset, u64 chunk_size)
2640{
2641	struct btrfs_fs_info *fs_info = trans->fs_info;
2642	struct btrfs_device *device;
2643	struct btrfs_chunk_map *map;
2644	u64 dev_offset;
2645	int i;
2646	int ret = 0;
2647
2648	map = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2649	if (IS_ERR(map))
2650		return PTR_ERR(map);
2651
2652	/*
2653	 * Take the device list mutex to prevent races with the final phase of
2654	 * a device replace operation that replaces the device object associated
2655	 * with the map's stripes, because the device object's id can change
2656	 * at any time during that final phase of the device replace operation
2657	 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2658	 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2659	 * resulting in persisting a device extent item with such ID.
2660	 */
2661	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2662	for (i = 0; i < map->num_stripes; i++) {
2663		device = map->stripes[i].dev;
2664		dev_offset = map->stripes[i].physical;
2665
2666		ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2667					map->stripe_size);
2668		if (ret)
2669			break;
2670	}
2671	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2672
2673	btrfs_free_chunk_map(map);
2674	return ret;
2675}
2676
2677/*
2678 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2679 * chunk allocation.
2680 *
2681 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2682 * phases.
2683 */
2684void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2685{
2686	struct btrfs_fs_info *fs_info = trans->fs_info;
2687	struct btrfs_block_group *block_group;
2688	int ret = 0;
2689
2690	while (!list_empty(&trans->new_bgs)) {
2691		int index;
2692
2693		block_group = list_first_entry(&trans->new_bgs,
2694					       struct btrfs_block_group,
2695					       bg_list);
2696		if (ret)
2697			goto next;
2698
2699		index = btrfs_bg_flags_to_raid_index(block_group->flags);
2700
2701		ret = insert_block_group_item(trans, block_group);
2702		if (ret)
2703			btrfs_abort_transaction(trans, ret);
2704		if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2705			      &block_group->runtime_flags)) {
2706			mutex_lock(&fs_info->chunk_mutex);
2707			ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2708			mutex_unlock(&fs_info->chunk_mutex);
2709			if (ret)
2710				btrfs_abort_transaction(trans, ret);
2711		}
2712		ret = insert_dev_extents(trans, block_group->start,
2713					 block_group->length);
2714		if (ret)
2715			btrfs_abort_transaction(trans, ret);
2716		add_block_group_free_space(trans, block_group);
2717
2718		/*
2719		 * If we restriped during balance, we may have added a new raid
2720		 * type, so now add the sysfs entries when it is safe to do so.
2721		 * We don't have to worry about locking here as it's handled in
2722		 * btrfs_sysfs_add_block_group_type.
2723		 */
2724		if (block_group->space_info->block_group_kobjs[index] == NULL)
2725			btrfs_sysfs_add_block_group_type(block_group);
2726
2727		/* Already aborted the transaction if it failed. */
2728next:
2729		btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2730		list_del_init(&block_group->bg_list);
2731		clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2732
2733		/*
2734		 * If the block group is still unused, add it to the list of
2735		 * unused block groups. The block group may have been created in
2736		 * order to satisfy a space reservation, in which case the
2737		 * extent allocation only happens later. But often we don't
2738		 * actually need to allocate space that we previously reserved,
2739		 * so the block group may become unused for a long time. For
2740		 * example for metadata we generally reserve space for a worst
2741		 * possible scenario, but then don't end up allocating all that
2742		 * space or none at all (due to no need to COW, extent buffers
2743		 * were already COWed in the current transaction and still
2744		 * unwritten, tree heights lower than the maximum possible
2745		 * height, etc). For data we generally reserve the axact amount
2746		 * of space we are going to allocate later, the exception is
2747		 * when using compression, as we must reserve space based on the
2748		 * uncompressed data size, because the compression is only done
2749		 * when writeback triggered and we don't know how much space we
2750		 * are actually going to need, so we reserve the uncompressed
2751		 * size because the data may be uncompressible in the worst case.
2752		 */
2753		if (ret == 0) {
2754			bool used;
2755
2756			spin_lock(&block_group->lock);
2757			used = btrfs_is_block_group_used(block_group);
2758			spin_unlock(&block_group->lock);
2759
2760			if (!used)
2761				btrfs_mark_bg_unused(block_group);
2762		}
2763	}
2764	btrfs_trans_release_chunk_metadata(trans);
2765}
2766
2767/*
2768 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2769 * global root id.  For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2770 */
2771static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2772{
2773	u64 div = SZ_1G;
2774	u64 index;
2775
2776	if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2777		return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2778
2779	/* If we have a smaller fs index based on 128MiB. */
2780	if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2781		div = SZ_128M;
2782
2783	offset = div64_u64(offset, div);
2784	div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2785	return index;
2786}
2787
2788struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2789						 u64 type,
2790						 u64 chunk_offset, u64 size)
2791{
2792	struct btrfs_fs_info *fs_info = trans->fs_info;
2793	struct btrfs_block_group *cache;
2794	int ret;
2795
2796	btrfs_set_log_full_commit(trans);
2797
2798	cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2799	if (!cache)
2800		return ERR_PTR(-ENOMEM);
2801
2802	/*
2803	 * Mark it as new before adding it to the rbtree of block groups or any
2804	 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2805	 * before the new flag is set.
2806	 */
2807	set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2808
2809	cache->length = size;
2810	set_free_space_tree_thresholds(cache);
2811	cache->flags = type;
2812	cache->cached = BTRFS_CACHE_FINISHED;
2813	cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2814
2815	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2816		set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2817
2818	ret = btrfs_load_block_group_zone_info(cache, true);
2819	if (ret) {
2820		btrfs_put_block_group(cache);
2821		return ERR_PTR(ret);
2822	}
2823
2824	ret = exclude_super_stripes(cache);
2825	if (ret) {
2826		/* We may have excluded something, so call this just in case */
2827		btrfs_free_excluded_extents(cache);
2828		btrfs_put_block_group(cache);
2829		return ERR_PTR(ret);
2830	}
2831
2832	ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2833	btrfs_free_excluded_extents(cache);
2834	if (ret) {
2835		btrfs_put_block_group(cache);
2836		return ERR_PTR(ret);
2837	}
2838
2839	/*
2840	 * Ensure the corresponding space_info object is created and
2841	 * assigned to our block group. We want our bg to be added to the rbtree
2842	 * with its ->space_info set.
2843	 */
2844	cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2845	ASSERT(cache->space_info);
2846
2847	ret = btrfs_add_block_group_cache(fs_info, cache);
2848	if (ret) {
2849		btrfs_remove_free_space_cache(cache);
2850		btrfs_put_block_group(cache);
2851		return ERR_PTR(ret);
2852	}
2853
2854	/*
2855	 * Now that our block group has its ->space_info set and is inserted in
2856	 * the rbtree, update the space info's counters.
2857	 */
2858	trace_btrfs_add_block_group(fs_info, cache, 1);
2859	btrfs_add_bg_to_space_info(fs_info, cache);
2860	btrfs_update_global_block_rsv(fs_info);
2861
2862#ifdef CONFIG_BTRFS_DEBUG
2863	if (btrfs_should_fragment_free_space(cache)) {
2864		cache->space_info->bytes_used += size >> 1;
2865		fragment_free_space(cache);
2866	}
2867#endif
2868
2869	list_add_tail(&cache->bg_list, &trans->new_bgs);
2870	btrfs_inc_delayed_refs_rsv_bg_inserts(fs_info);
2871
2872	set_avail_alloc_bits(fs_info, type);
2873	return cache;
2874}
2875
2876/*
2877 * Mark one block group RO, can be called several times for the same block
2878 * group.
2879 *
2880 * @cache:		the destination block group
2881 * @do_chunk_alloc:	whether need to do chunk pre-allocation, this is to
2882 * 			ensure we still have some free space after marking this
2883 * 			block group RO.
2884 */
2885int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2886			     bool do_chunk_alloc)
2887{
2888	struct btrfs_fs_info *fs_info = cache->fs_info;
2889	struct btrfs_trans_handle *trans;
2890	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2891	u64 alloc_flags;
2892	int ret;
2893	bool dirty_bg_running;
2894
2895	/*
2896	 * This can only happen when we are doing read-only scrub on read-only
2897	 * mount.
2898	 * In that case we should not start a new transaction on read-only fs.
2899	 * Thus here we skip all chunk allocations.
2900	 */
2901	if (sb_rdonly(fs_info->sb)) {
2902		mutex_lock(&fs_info->ro_block_group_mutex);
2903		ret = inc_block_group_ro(cache, 0);
2904		mutex_unlock(&fs_info->ro_block_group_mutex);
2905		return ret;
2906	}
2907
2908	do {
2909		trans = btrfs_join_transaction(root);
2910		if (IS_ERR(trans))
2911			return PTR_ERR(trans);
2912
2913		dirty_bg_running = false;
2914
2915		/*
2916		 * We're not allowed to set block groups readonly after the dirty
2917		 * block group cache has started writing.  If it already started,
2918		 * back off and let this transaction commit.
2919		 */
2920		mutex_lock(&fs_info->ro_block_group_mutex);
2921		if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2922			u64 transid = trans->transid;
2923
2924			mutex_unlock(&fs_info->ro_block_group_mutex);
2925			btrfs_end_transaction(trans);
2926
2927			ret = btrfs_wait_for_commit(fs_info, transid);
2928			if (ret)
2929				return ret;
2930			dirty_bg_running = true;
2931		}
2932	} while (dirty_bg_running);
2933
2934	if (do_chunk_alloc) {
2935		/*
2936		 * If we are changing raid levels, try to allocate a
2937		 * corresponding block group with the new raid level.
2938		 */
2939		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2940		if (alloc_flags != cache->flags) {
2941			ret = btrfs_chunk_alloc(trans, alloc_flags,
2942						CHUNK_ALLOC_FORCE);
2943			/*
2944			 * ENOSPC is allowed here, we may have enough space
2945			 * already allocated at the new raid level to carry on
2946			 */
2947			if (ret == -ENOSPC)
2948				ret = 0;
2949			if (ret < 0)
2950				goto out;
2951		}
2952	}
2953
2954	ret = inc_block_group_ro(cache, 0);
2955	if (!ret)
2956		goto out;
2957	if (ret == -ETXTBSY)
2958		goto unlock_out;
2959
2960	/*
2961	 * Skip chunk allocation if the bg is SYSTEM, this is to avoid system
2962	 * chunk allocation storm to exhaust the system chunk array.  Otherwise
2963	 * we still want to try our best to mark the block group read-only.
2964	 */
2965	if (!do_chunk_alloc && ret == -ENOSPC &&
2966	    (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
2967		goto unlock_out;
2968
2969	alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2970	ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2971	if (ret < 0)
2972		goto out;
2973	/*
2974	 * We have allocated a new chunk. We also need to activate that chunk to
2975	 * grant metadata tickets for zoned filesystem.
2976	 */
2977	ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2978	if (ret < 0)
2979		goto out;
2980
2981	ret = inc_block_group_ro(cache, 0);
2982	if (ret == -ETXTBSY)
2983		goto unlock_out;
2984out:
2985	if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2986		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2987		mutex_lock(&fs_info->chunk_mutex);
2988		check_system_chunk(trans, alloc_flags);
2989		mutex_unlock(&fs_info->chunk_mutex);
2990	}
2991unlock_out:
2992	mutex_unlock(&fs_info->ro_block_group_mutex);
2993
2994	btrfs_end_transaction(trans);
2995	return ret;
2996}
2997
2998void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2999{
3000	struct btrfs_space_info *sinfo = cache->space_info;
3001	u64 num_bytes;
3002
3003	BUG_ON(!cache->ro);
3004
3005	spin_lock(&sinfo->lock);
3006	spin_lock(&cache->lock);
3007	if (!--cache->ro) {
3008		if (btrfs_is_zoned(cache->fs_info)) {
3009			/* Migrate zone_unusable bytes back */
3010			cache->zone_unusable =
3011				(cache->alloc_offset - cache->used) +
3012				(cache->length - cache->zone_capacity);
3013			sinfo->bytes_zone_unusable += cache->zone_unusable;
3014			sinfo->bytes_readonly -= cache->zone_unusable;
3015		}
3016		num_bytes = cache->length - cache->reserved -
3017			    cache->pinned - cache->bytes_super -
3018			    cache->zone_unusable - cache->used;
3019		sinfo->bytes_readonly -= num_bytes;
3020		list_del_init(&cache->ro_list);
3021	}
3022	spin_unlock(&cache->lock);
3023	spin_unlock(&sinfo->lock);
3024}
3025
3026static int update_block_group_item(struct btrfs_trans_handle *trans,
3027				   struct btrfs_path *path,
3028				   struct btrfs_block_group *cache)
3029{
3030	struct btrfs_fs_info *fs_info = trans->fs_info;
3031	int ret;
3032	struct btrfs_root *root = btrfs_block_group_root(fs_info);
3033	unsigned long bi;
3034	struct extent_buffer *leaf;
3035	struct btrfs_block_group_item bgi;
3036	struct btrfs_key key;
3037	u64 old_commit_used;
3038	u64 used;
3039
3040	/*
3041	 * Block group items update can be triggered out of commit transaction
3042	 * critical section, thus we need a consistent view of used bytes.
3043	 * We cannot use cache->used directly outside of the spin lock, as it
3044	 * may be changed.
3045	 */
3046	spin_lock(&cache->lock);
3047	old_commit_used = cache->commit_used;
3048	used = cache->used;
3049	/* No change in used bytes, can safely skip it. */
3050	if (cache->commit_used == used) {
3051		spin_unlock(&cache->lock);
3052		return 0;
3053	}
3054	cache->commit_used = used;
3055	spin_unlock(&cache->lock);
3056
3057	key.objectid = cache->start;
3058	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3059	key.offset = cache->length;
3060
3061	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3062	if (ret) {
3063		if (ret > 0)
3064			ret = -ENOENT;
3065		goto fail;
3066	}
3067
3068	leaf = path->nodes[0];
3069	bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3070	btrfs_set_stack_block_group_used(&bgi, used);
3071	btrfs_set_stack_block_group_chunk_objectid(&bgi,
3072						   cache->global_root_id);
3073	btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3074	write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3075	btrfs_mark_buffer_dirty(trans, leaf);
3076fail:
3077	btrfs_release_path(path);
3078	/*
3079	 * We didn't update the block group item, need to revert commit_used
3080	 * unless the block group item didn't exist yet - this is to prevent a
3081	 * race with a concurrent insertion of the block group item, with
3082	 * insert_block_group_item(), that happened just after we attempted to
3083	 * update. In that case we would reset commit_used to 0 just after the
3084	 * insertion set it to a value greater than 0 - if the block group later
3085	 * becomes with 0 used bytes, we would incorrectly skip its update.
3086	 */
3087	if (ret < 0 && ret != -ENOENT) {
3088		spin_lock(&cache->lock);
3089		cache->commit_used = old_commit_used;
3090		spin_unlock(&cache->lock);
3091	}
3092	return ret;
3093
3094}
3095
3096static int cache_save_setup(struct btrfs_block_group *block_group,
3097			    struct btrfs_trans_handle *trans,
3098			    struct btrfs_path *path)
3099{
3100	struct btrfs_fs_info *fs_info = block_group->fs_info;
3101	struct inode *inode = NULL;
3102	struct extent_changeset *data_reserved = NULL;
3103	u64 alloc_hint = 0;
3104	int dcs = BTRFS_DC_ERROR;
3105	u64 cache_size = 0;
3106	int retries = 0;
3107	int ret = 0;
3108
3109	if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3110		return 0;
3111
3112	/*
3113	 * If this block group is smaller than 100 megs don't bother caching the
3114	 * block group.
3115	 */
3116	if (block_group->length < (100 * SZ_1M)) {
3117		spin_lock(&block_group->lock);
3118		block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3119		spin_unlock(&block_group->lock);
3120		return 0;
3121	}
3122
3123	if (TRANS_ABORTED(trans))
3124		return 0;
3125again:
3126	inode = lookup_free_space_inode(block_group, path);
3127	if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3128		ret = PTR_ERR(inode);
3129		btrfs_release_path(path);
3130		goto out;
3131	}
3132
3133	if (IS_ERR(inode)) {
3134		BUG_ON(retries);
3135		retries++;
3136
3137		if (block_group->ro)
3138			goto out_free;
3139
3140		ret = create_free_space_inode(trans, block_group, path);
3141		if (ret)
3142			goto out_free;
3143		goto again;
3144	}
3145
3146	/*
3147	 * We want to set the generation to 0, that way if anything goes wrong
3148	 * from here on out we know not to trust this cache when we load up next
3149	 * time.
3150	 */
3151	BTRFS_I(inode)->generation = 0;
3152	ret = btrfs_update_inode(trans, BTRFS_I(inode));
3153	if (ret) {
3154		/*
3155		 * So theoretically we could recover from this, simply set the
3156		 * super cache generation to 0 so we know to invalidate the
3157		 * cache, but then we'd have to keep track of the block groups
3158		 * that fail this way so we know we _have_ to reset this cache
3159		 * before the next commit or risk reading stale cache.  So to
3160		 * limit our exposure to horrible edge cases lets just abort the
3161		 * transaction, this only happens in really bad situations
3162		 * anyway.
3163		 */
3164		btrfs_abort_transaction(trans, ret);
3165		goto out_put;
3166	}
3167	WARN_ON(ret);
3168
3169	/* We've already setup this transaction, go ahead and exit */
3170	if (block_group->cache_generation == trans->transid &&
3171	    i_size_read(inode)) {
3172		dcs = BTRFS_DC_SETUP;
3173		goto out_put;
3174	}
3175
3176	if (i_size_read(inode) > 0) {
3177		ret = btrfs_check_trunc_cache_free_space(fs_info,
3178					&fs_info->global_block_rsv);
3179		if (ret)
3180			goto out_put;
3181
3182		ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3183		if (ret)
3184			goto out_put;
3185	}
3186
3187	spin_lock(&block_group->lock);
3188	if (block_group->cached != BTRFS_CACHE_FINISHED ||
3189	    !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3190		/*
3191		 * don't bother trying to write stuff out _if_
3192		 * a) we're not cached,
3193		 * b) we're with nospace_cache mount option,
3194		 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3195		 */
3196		dcs = BTRFS_DC_WRITTEN;
3197		spin_unlock(&block_group->lock);
3198		goto out_put;
3199	}
3200	spin_unlock(&block_group->lock);
3201
3202	/*
3203	 * We hit an ENOSPC when setting up the cache in this transaction, just
3204	 * skip doing the setup, we've already cleared the cache so we're safe.
3205	 */
3206	if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3207		ret = -ENOSPC;
3208		goto out_put;
3209	}
3210
3211	/*
3212	 * Try to preallocate enough space based on how big the block group is.
3213	 * Keep in mind this has to include any pinned space which could end up
3214	 * taking up quite a bit since it's not folded into the other space
3215	 * cache.
3216	 */
3217	cache_size = div_u64(block_group->length, SZ_256M);
3218	if (!cache_size)
3219		cache_size = 1;
3220
3221	cache_size *= 16;
3222	cache_size *= fs_info->sectorsize;
3223
3224	ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3225					  cache_size, false);
3226	if (ret)
3227		goto out_put;
3228
3229	ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3230					      cache_size, cache_size,
3231					      &alloc_hint);
3232	/*
3233	 * Our cache requires contiguous chunks so that we don't modify a bunch
3234	 * of metadata or split extents when writing the cache out, which means
3235	 * we can enospc if we are heavily fragmented in addition to just normal
3236	 * out of space conditions.  So if we hit this just skip setting up any
3237	 * other block groups for this transaction, maybe we'll unpin enough
3238	 * space the next time around.
3239	 */
3240	if (!ret)
3241		dcs = BTRFS_DC_SETUP;
3242	else if (ret == -ENOSPC)
3243		set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3244
3245out_put:
3246	iput(inode);
3247out_free:
3248	btrfs_release_path(path);
3249out:
3250	spin_lock(&block_group->lock);
3251	if (!ret && dcs == BTRFS_DC_SETUP)
3252		block_group->cache_generation = trans->transid;
3253	block_group->disk_cache_state = dcs;
3254	spin_unlock(&block_group->lock);
3255
3256	extent_changeset_free(data_reserved);
3257	return ret;
3258}
3259
3260int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3261{
3262	struct btrfs_fs_info *fs_info = trans->fs_info;
3263	struct btrfs_block_group *cache, *tmp;
3264	struct btrfs_transaction *cur_trans = trans->transaction;
3265	struct btrfs_path *path;
3266
3267	if (list_empty(&cur_trans->dirty_bgs) ||
3268	    !btrfs_test_opt(fs_info, SPACE_CACHE))
3269		return 0;
3270
3271	path = btrfs_alloc_path();
3272	if (!path)
3273		return -ENOMEM;
3274
3275	/* Could add new block groups, use _safe just in case */
3276	list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3277				 dirty_list) {
3278		if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3279			cache_save_setup(cache, trans, path);
3280	}
3281
3282	btrfs_free_path(path);
3283	return 0;
3284}
3285
3286/*
3287 * Transaction commit does final block group cache writeback during a critical
3288 * section where nothing is allowed to change the FS.  This is required in
3289 * order for the cache to actually match the block group, but can introduce a
3290 * lot of latency into the commit.
3291 *
3292 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3293 * There's a chance we'll have to redo some of it if the block group changes
3294 * again during the commit, but it greatly reduces the commit latency by
3295 * getting rid of the easy block groups while we're still allowing others to
3296 * join the commit.
3297 */
3298int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3299{
3300	struct btrfs_fs_info *fs_info = trans->fs_info;
3301	struct btrfs_block_group *cache;
3302	struct btrfs_transaction *cur_trans = trans->transaction;
3303	int ret = 0;
3304	int should_put;
3305	struct btrfs_path *path = NULL;
3306	LIST_HEAD(dirty);
3307	struct list_head *io = &cur_trans->io_bgs;
3308	int loops = 0;
3309
3310	spin_lock(&cur_trans->dirty_bgs_lock);
3311	if (list_empty(&cur_trans->dirty_bgs)) {
3312		spin_unlock(&cur_trans->dirty_bgs_lock);
3313		return 0;
3314	}
3315	list_splice_init(&cur_trans->dirty_bgs, &dirty);
3316	spin_unlock(&cur_trans->dirty_bgs_lock);
3317
3318again:
3319	/* Make sure all the block groups on our dirty list actually exist */
3320	btrfs_create_pending_block_groups(trans);
3321
3322	if (!path) {
3323		path = btrfs_alloc_path();
3324		if (!path) {
3325			ret = -ENOMEM;
3326			goto out;
3327		}
3328	}
3329
3330	/*
3331	 * cache_write_mutex is here only to save us from balance or automatic
3332	 * removal of empty block groups deleting this block group while we are
3333	 * writing out the cache
3334	 */
3335	mutex_lock(&trans->transaction->cache_write_mutex);
3336	while (!list_empty(&dirty)) {
3337		bool drop_reserve = true;
3338
3339		cache = list_first_entry(&dirty, struct btrfs_block_group,
3340					 dirty_list);
3341		/*
3342		 * This can happen if something re-dirties a block group that
3343		 * is already under IO.  Just wait for it to finish and then do
3344		 * it all again
3345		 */
3346		if (!list_empty(&cache->io_list)) {
3347			list_del_init(&cache->io_list);
3348			btrfs_wait_cache_io(trans, cache, path);
3349			btrfs_put_block_group(cache);
3350		}
3351
3352
3353		/*
3354		 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3355		 * it should update the cache_state.  Don't delete until after
3356		 * we wait.
3357		 *
3358		 * Since we're not running in the commit critical section
3359		 * we need the dirty_bgs_lock to protect from update_block_group
3360		 */
3361		spin_lock(&cur_trans->dirty_bgs_lock);
3362		list_del_init(&cache->dirty_list);
3363		spin_unlock(&cur_trans->dirty_bgs_lock);
3364
3365		should_put = 1;
3366
3367		cache_save_setup(cache, trans, path);
3368
3369		if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3370			cache->io_ctl.inode = NULL;
3371			ret = btrfs_write_out_cache(trans, cache, path);
3372			if (ret == 0 && cache->io_ctl.inode) {
3373				should_put = 0;
3374
3375				/*
3376				 * The cache_write_mutex is protecting the
3377				 * io_list, also refer to the definition of
3378				 * btrfs_transaction::io_bgs for more details
3379				 */
3380				list_add_tail(&cache->io_list, io);
3381			} else {
3382				/*
3383				 * If we failed to write the cache, the
3384				 * generation will be bad and life goes on
3385				 */
3386				ret = 0;
3387			}
3388		}
3389		if (!ret) {
3390			ret = update_block_group_item(trans, path, cache);
3391			/*
3392			 * Our block group might still be attached to the list
3393			 * of new block groups in the transaction handle of some
3394			 * other task (struct btrfs_trans_handle->new_bgs). This
3395			 * means its block group item isn't yet in the extent
3396			 * tree. If this happens ignore the error, as we will
3397			 * try again later in the critical section of the
3398			 * transaction commit.
3399			 */
3400			if (ret == -ENOENT) {
3401				ret = 0;
3402				spin_lock(&cur_trans->dirty_bgs_lock);
3403				if (list_empty(&cache->dirty_list)) {
3404					list_add_tail(&cache->dirty_list,
3405						      &cur_trans->dirty_bgs);
3406					btrfs_get_block_group(cache);
3407					drop_reserve = false;
3408				}
3409				spin_unlock(&cur_trans->dirty_bgs_lock);
3410			} else if (ret) {
3411				btrfs_abort_transaction(trans, ret);
3412			}
3413		}
3414
3415		/* If it's not on the io list, we need to put the block group */
3416		if (should_put)
3417			btrfs_put_block_group(cache);
3418		if (drop_reserve)
3419			btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3420		/*
3421		 * Avoid blocking other tasks for too long. It might even save
3422		 * us from writing caches for block groups that are going to be
3423		 * removed.
3424		 */
3425		mutex_unlock(&trans->transaction->cache_write_mutex);
3426		if (ret)
3427			goto out;
3428		mutex_lock(&trans->transaction->cache_write_mutex);
3429	}
3430	mutex_unlock(&trans->transaction->cache_write_mutex);
3431
3432	/*
3433	 * Go through delayed refs for all the stuff we've just kicked off
3434	 * and then loop back (just once)
3435	 */
3436	if (!ret)
3437		ret = btrfs_run_delayed_refs(trans, 0);
3438	if (!ret && loops == 0) {
3439		loops++;
3440		spin_lock(&cur_trans->dirty_bgs_lock);
3441		list_splice_init(&cur_trans->dirty_bgs, &dirty);
3442		/*
3443		 * dirty_bgs_lock protects us from concurrent block group
3444		 * deletes too (not just cache_write_mutex).
3445		 */
3446		if (!list_empty(&dirty)) {
3447			spin_unlock(&cur_trans->dirty_bgs_lock);
3448			goto again;
3449		}
3450		spin_unlock(&cur_trans->dirty_bgs_lock);
3451	}
3452out:
3453	if (ret < 0) {
3454		spin_lock(&cur_trans->dirty_bgs_lock);
3455		list_splice_init(&dirty, &cur_trans->dirty_bgs);
3456		spin_unlock(&cur_trans->dirty_bgs_lock);
3457		btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3458	}
3459
3460	btrfs_free_path(path);
3461	return ret;
3462}
3463
3464int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3465{
3466	struct btrfs_fs_info *fs_info = trans->fs_info;
3467	struct btrfs_block_group *cache;
3468	struct btrfs_transaction *cur_trans = trans->transaction;
3469	int ret = 0;
3470	int should_put;
3471	struct btrfs_path *path;
3472	struct list_head *io = &cur_trans->io_bgs;
3473
3474	path = btrfs_alloc_path();
3475	if (!path)
3476		return -ENOMEM;
3477
3478	/*
3479	 * Even though we are in the critical section of the transaction commit,
3480	 * we can still have concurrent tasks adding elements to this
3481	 * transaction's list of dirty block groups. These tasks correspond to
3482	 * endio free space workers started when writeback finishes for a
3483	 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3484	 * allocate new block groups as a result of COWing nodes of the root
3485	 * tree when updating the free space inode. The writeback for the space
3486	 * caches is triggered by an earlier call to
3487	 * btrfs_start_dirty_block_groups() and iterations of the following
3488	 * loop.
3489	 * Also we want to do the cache_save_setup first and then run the
3490	 * delayed refs to make sure we have the best chance at doing this all
3491	 * in one shot.
3492	 */
3493	spin_lock(&cur_trans->dirty_bgs_lock);
3494	while (!list_empty(&cur_trans->dirty_bgs)) {
3495		cache = list_first_entry(&cur_trans->dirty_bgs,
3496					 struct btrfs_block_group,
3497					 dirty_list);
3498
3499		/*
3500		 * This can happen if cache_save_setup re-dirties a block group
3501		 * that is already under IO.  Just wait for it to finish and
3502		 * then do it all again
3503		 */
3504		if (!list_empty(&cache->io_list)) {
3505			spin_unlock(&cur_trans->dirty_bgs_lock);
3506			list_del_init(&cache->io_list);
3507			btrfs_wait_cache_io(trans, cache, path);
3508			btrfs_put_block_group(cache);
3509			spin_lock(&cur_trans->dirty_bgs_lock);
3510		}
3511
3512		/*
3513		 * Don't remove from the dirty list until after we've waited on
3514		 * any pending IO
3515		 */
3516		list_del_init(&cache->dirty_list);
3517		spin_unlock(&cur_trans->dirty_bgs_lock);
3518		should_put = 1;
3519
3520		cache_save_setup(cache, trans, path);
3521
3522		if (!ret)
3523			ret = btrfs_run_delayed_refs(trans, U64_MAX);
3524
3525		if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3526			cache->io_ctl.inode = NULL;
3527			ret = btrfs_write_out_cache(trans, cache, path);
3528			if (ret == 0 && cache->io_ctl.inode) {
3529				should_put = 0;
3530				list_add_tail(&cache->io_list, io);
3531			} else {
3532				/*
3533				 * If we failed to write the cache, the
3534				 * generation will be bad and life goes on
3535				 */
3536				ret = 0;
3537			}
3538		}
3539		if (!ret) {
3540			ret = update_block_group_item(trans, path, cache);
3541			/*
3542			 * One of the free space endio workers might have
3543			 * created a new block group while updating a free space
3544			 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3545			 * and hasn't released its transaction handle yet, in
3546			 * which case the new block group is still attached to
3547			 * its transaction handle and its creation has not
3548			 * finished yet (no block group item in the extent tree
3549			 * yet, etc). If this is the case, wait for all free
3550			 * space endio workers to finish and retry. This is a
3551			 * very rare case so no need for a more efficient and
3552			 * complex approach.
3553			 */
3554			if (ret == -ENOENT) {
3555				wait_event(cur_trans->writer_wait,
3556				   atomic_read(&cur_trans->num_writers) == 1);
3557				ret = update_block_group_item(trans, path, cache);
3558			}
3559			if (ret)
3560				btrfs_abort_transaction(trans, ret);
3561		}
3562
3563		/* If its not on the io list, we need to put the block group */
3564		if (should_put)
3565			btrfs_put_block_group(cache);
3566		btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3567		spin_lock(&cur_trans->dirty_bgs_lock);
3568	}
3569	spin_unlock(&cur_trans->dirty_bgs_lock);
3570
3571	/*
3572	 * Refer to the definition of io_bgs member for details why it's safe
3573	 * to use it without any locking
3574	 */
3575	while (!list_empty(io)) {
3576		cache = list_first_entry(io, struct btrfs_block_group,
3577					 io_list);
3578		list_del_init(&cache->io_list);
3579		btrfs_wait_cache_io(trans, cache, path);
3580		btrfs_put_block_group(cache);
3581	}
3582
3583	btrfs_free_path(path);
3584	return ret;
3585}
3586
3587int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3588			     u64 bytenr, u64 num_bytes, bool alloc)
3589{
3590	struct btrfs_fs_info *info = trans->fs_info;
3591	struct btrfs_space_info *space_info;
3592	struct btrfs_block_group *cache;
3593	u64 old_val;
3594	bool reclaim = false;
3595	bool bg_already_dirty = true;
3596	int factor;
3597
3598	/* Block accounting for super block */
3599	spin_lock(&info->delalloc_root_lock);
3600	old_val = btrfs_super_bytes_used(info->super_copy);
3601	if (alloc)
3602		old_val += num_bytes;
3603	else
3604		old_val -= num_bytes;
3605	btrfs_set_super_bytes_used(info->super_copy, old_val);
3606	spin_unlock(&info->delalloc_root_lock);
3607
3608	cache = btrfs_lookup_block_group(info, bytenr);
3609	if (!cache)
3610		return -ENOENT;
3611
3612	/* An extent can not span multiple block groups. */
3613	ASSERT(bytenr + num_bytes <= cache->start + cache->length);
3614
3615	space_info = cache->space_info;
3616	factor = btrfs_bg_type_to_factor(cache->flags);
3617
3618	/*
3619	 * If this block group has free space cache written out, we need to make
3620	 * sure to load it if we are removing space.  This is because we need
3621	 * the unpinning stage to actually add the space back to the block group,
3622	 * otherwise we will leak space.
3623	 */
3624	if (!alloc && !btrfs_block_group_done(cache))
3625		btrfs_cache_block_group(cache, true);
3626
3627	spin_lock(&space_info->lock);
3628	spin_lock(&cache->lock);
3629
3630	if (btrfs_test_opt(info, SPACE_CACHE) &&
3631	    cache->disk_cache_state < BTRFS_DC_CLEAR)
3632		cache->disk_cache_state = BTRFS_DC_CLEAR;
3633
3634	old_val = cache->used;
3635	if (alloc) {
3636		old_val += num_bytes;
3637		cache->used = old_val;
3638		cache->reserved -= num_bytes;
3639		space_info->bytes_reserved -= num_bytes;
3640		space_info->bytes_used += num_bytes;
3641		space_info->disk_used += num_bytes * factor;
3642		spin_unlock(&cache->lock);
3643		spin_unlock(&space_info->lock);
3644	} else {
3645		old_val -= num_bytes;
3646		cache->used = old_val;
3647		cache->pinned += num_bytes;
3648		btrfs_space_info_update_bytes_pinned(info, space_info, num_bytes);
3649		space_info->bytes_used -= num_bytes;
3650		space_info->disk_used -= num_bytes * factor;
3651
3652		reclaim = should_reclaim_block_group(cache, num_bytes);
3653
3654		spin_unlock(&cache->lock);
3655		spin_unlock(&space_info->lock);
3656
3657		set_extent_bit(&trans->transaction->pinned_extents, bytenr,
3658			       bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
3659	}
3660
3661	spin_lock(&trans->transaction->dirty_bgs_lock);
3662	if (list_empty(&cache->dirty_list)) {
3663		list_add_tail(&cache->dirty_list, &trans->transaction->dirty_bgs);
3664		bg_already_dirty = false;
3665		btrfs_get_block_group(cache);
3666	}
3667	spin_unlock(&trans->transaction->dirty_bgs_lock);
3668
3669	/*
3670	 * No longer have used bytes in this block group, queue it for deletion.
3671	 * We do this after adding the block group to the dirty list to avoid
3672	 * races between cleaner kthread and space cache writeout.
3673	 */
3674	if (!alloc && old_val == 0) {
3675		if (!btrfs_test_opt(info, DISCARD_ASYNC))
3676			btrfs_mark_bg_unused(cache);
3677	} else if (!alloc && reclaim) {
3678		btrfs_mark_bg_to_reclaim(cache);
3679	}
3680
3681	btrfs_put_block_group(cache);
3682
3683	/* Modified block groups are accounted for in the delayed_refs_rsv. */
3684	if (!bg_already_dirty)
3685		btrfs_inc_delayed_refs_rsv_bg_updates(info);
3686
3687	return 0;
3688}
3689
3690/*
3691 * Update the block_group and space info counters.
3692 *
3693 * @cache:	The cache we are manipulating
3694 * @ram_bytes:  The number of bytes of file content, and will be same to
3695 *              @num_bytes except for the compress path.
3696 * @num_bytes:	The number of bytes in question
3697 * @delalloc:   The blocks are allocated for the delalloc write
3698 *
3699 * This is called by the allocator when it reserves space. If this is a
3700 * reservation and the block group has become read only we cannot make the
3701 * reservation and return -EAGAIN, otherwise this function always succeeds.
3702 */
3703int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3704			     u64 ram_bytes, u64 num_bytes, int delalloc,
3705			     bool force_wrong_size_class)
3706{
3707	struct btrfs_space_info *space_info = cache->space_info;
3708	enum btrfs_block_group_size_class size_class;
3709	int ret = 0;
3710
3711	spin_lock(&space_info->lock);
3712	spin_lock(&cache->lock);
3713	if (cache->ro) {
3714		ret = -EAGAIN;
3715		goto out;
3716	}
3717
3718	if (btrfs_block_group_should_use_size_class(cache)) {
3719		size_class = btrfs_calc_block_group_size_class(num_bytes);
3720		ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3721		if (ret)
3722			goto out;
3723	}
3724	cache->reserved += num_bytes;
3725	space_info->bytes_reserved += num_bytes;
3726	trace_btrfs_space_reservation(cache->fs_info, "space_info",
3727				      space_info->flags, num_bytes, 1);
3728	btrfs_space_info_update_bytes_may_use(cache->fs_info,
3729					      space_info, -ram_bytes);
3730	if (delalloc)
3731		cache->delalloc_bytes += num_bytes;
3732
3733	/*
3734	 * Compression can use less space than we reserved, so wake tickets if
3735	 * that happens.
3736	 */
3737	if (num_bytes < ram_bytes)
3738		btrfs_try_granting_tickets(cache->fs_info, space_info);
3739out:
3740	spin_unlock(&cache->lock);
3741	spin_unlock(&space_info->lock);
3742	return ret;
3743}
3744
3745/*
3746 * Update the block_group and space info counters.
3747 *
3748 * @cache:      The cache we are manipulating
3749 * @num_bytes:  The number of bytes in question
3750 * @delalloc:   The blocks are allocated for the delalloc write
3751 *
3752 * This is called by somebody who is freeing space that was never actually used
3753 * on disk.  For example if you reserve some space for a new leaf in transaction
3754 * A and before transaction A commits you free that leaf, you call this with
3755 * reserve set to 0 in order to clear the reservation.
3756 */
3757void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3758			       u64 num_bytes, int delalloc)
3759{
3760	struct btrfs_space_info *space_info = cache->space_info;
3761
3762	spin_lock(&space_info->lock);
3763	spin_lock(&cache->lock);
3764	if (cache->ro)
3765		space_info->bytes_readonly += num_bytes;
3766	cache->reserved -= num_bytes;
3767	space_info->bytes_reserved -= num_bytes;
3768	space_info->max_extent_size = 0;
3769
3770	if (delalloc)
3771		cache->delalloc_bytes -= num_bytes;
3772	spin_unlock(&cache->lock);
3773
3774	btrfs_try_granting_tickets(cache->fs_info, space_info);
3775	spin_unlock(&space_info->lock);
3776}
3777
3778static void force_metadata_allocation(struct btrfs_fs_info *info)
3779{
3780	struct list_head *head = &info->space_info;
3781	struct btrfs_space_info *found;
3782
3783	list_for_each_entry(found, head, list) {
3784		if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3785			found->force_alloc = CHUNK_ALLOC_FORCE;
3786	}
3787}
3788
3789static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3790			      struct btrfs_space_info *sinfo, int force)
3791{
3792	u64 bytes_used = btrfs_space_info_used(sinfo, false);
3793	u64 thresh;
3794
3795	if (force == CHUNK_ALLOC_FORCE)
3796		return 1;
3797
3798	/*
3799	 * in limited mode, we want to have some free space up to
3800	 * about 1% of the FS size.
3801	 */
3802	if (force == CHUNK_ALLOC_LIMITED) {
3803		thresh = btrfs_super_total_bytes(fs_info->super_copy);
3804		thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3805
3806		if (sinfo->total_bytes - bytes_used < thresh)
3807			return 1;
3808	}
3809
3810	if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3811		return 0;
3812	return 1;
3813}
3814
3815int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3816{
3817	u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3818
3819	return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3820}
3821
3822static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3823{
3824	struct btrfs_block_group *bg;
3825	int ret;
3826
3827	/*
3828	 * Check if we have enough space in the system space info because we
3829	 * will need to update device items in the chunk btree and insert a new
3830	 * chunk item in the chunk btree as well. This will allocate a new
3831	 * system block group if needed.
3832	 */
3833	check_system_chunk(trans, flags);
3834
3835	bg = btrfs_create_chunk(trans, flags);
3836	if (IS_ERR(bg)) {
3837		ret = PTR_ERR(bg);
3838		goto out;
3839	}
3840
3841	ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3842	/*
3843	 * Normally we are not expected to fail with -ENOSPC here, since we have
3844	 * previously reserved space in the system space_info and allocated one
3845	 * new system chunk if necessary. However there are three exceptions:
3846	 *
3847	 * 1) We may have enough free space in the system space_info but all the
3848	 *    existing system block groups have a profile which can not be used
3849	 *    for extent allocation.
3850	 *
3851	 *    This happens when mounting in degraded mode. For example we have a
3852	 *    RAID1 filesystem with 2 devices, lose one device and mount the fs
3853	 *    using the other device in degraded mode. If we then allocate a chunk,
3854	 *    we may have enough free space in the existing system space_info, but
3855	 *    none of the block groups can be used for extent allocation since they
3856	 *    have a RAID1 profile, and because we are in degraded mode with a
3857	 *    single device, we are forced to allocate a new system chunk with a
3858	 *    SINGLE profile. Making check_system_chunk() iterate over all system
3859	 *    block groups and check if they have a usable profile and enough space
3860	 *    can be slow on very large filesystems, so we tolerate the -ENOSPC and
3861	 *    try again after forcing allocation of a new system chunk. Like this
3862	 *    we avoid paying the cost of that search in normal circumstances, when
3863	 *    we were not mounted in degraded mode;
3864	 *
3865	 * 2) We had enough free space info the system space_info, and one suitable
3866	 *    block group to allocate from when we called check_system_chunk()
3867	 *    above. However right after we called it, the only system block group
3868	 *    with enough free space got turned into RO mode by a running scrub,
3869	 *    and in this case we have to allocate a new one and retry. We only
3870	 *    need do this allocate and retry once, since we have a transaction
3871	 *    handle and scrub uses the commit root to search for block groups;
3872	 *
3873	 * 3) We had one system block group with enough free space when we called
3874	 *    check_system_chunk(), but after that, right before we tried to
3875	 *    allocate the last extent buffer we needed, a discard operation came
3876	 *    in and it temporarily removed the last free space entry from the
3877	 *    block group (discard removes a free space entry, discards it, and
3878	 *    then adds back the entry to the block group cache).
3879	 */
3880	if (ret == -ENOSPC) {
3881		const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3882		struct btrfs_block_group *sys_bg;
3883
3884		sys_bg = btrfs_create_chunk(trans, sys_flags);
3885		if (IS_ERR(sys_bg)) {
3886			ret = PTR_ERR(sys_bg);
3887			btrfs_abort_transaction(trans, ret);
3888			goto out;
3889		}
3890
3891		ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3892		if (ret) {
3893			btrfs_abort_transaction(trans, ret);
3894			goto out;
3895		}
3896
3897		ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3898		if (ret) {
3899			btrfs_abort_transaction(trans, ret);
3900			goto out;
3901		}
3902	} else if (ret) {
3903		btrfs_abort_transaction(trans, ret);
3904		goto out;
3905	}
3906out:
3907	btrfs_trans_release_chunk_metadata(trans);
3908
3909	if (ret)
3910		return ERR_PTR(ret);
3911
3912	btrfs_get_block_group(bg);
3913	return bg;
3914}
3915
3916/*
3917 * Chunk allocation is done in 2 phases:
3918 *
3919 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3920 *    the chunk, the chunk mapping, create its block group and add the items
3921 *    that belong in the chunk btree to it - more specifically, we need to
3922 *    update device items in the chunk btree and add a new chunk item to it.
3923 *
3924 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3925 *    group item to the extent btree and the device extent items to the devices
3926 *    btree.
3927 *
3928 * This is done to prevent deadlocks. For example when COWing a node from the
3929 * extent btree we are holding a write lock on the node's parent and if we
3930 * trigger chunk allocation and attempted to insert the new block group item
3931 * in the extent btree right way, we could deadlock because the path for the
3932 * insertion can include that parent node. At first glance it seems impossible
3933 * to trigger chunk allocation after starting a transaction since tasks should
3934 * reserve enough transaction units (metadata space), however while that is true
3935 * most of the time, chunk allocation may still be triggered for several reasons:
3936 *
3937 * 1) When reserving metadata, we check if there is enough free space in the
3938 *    metadata space_info and therefore don't trigger allocation of a new chunk.
3939 *    However later when the task actually tries to COW an extent buffer from
3940 *    the extent btree or from the device btree for example, it is forced to
3941 *    allocate a new block group (chunk) because the only one that had enough
3942 *    free space was just turned to RO mode by a running scrub for example (or
3943 *    device replace, block group reclaim thread, etc), so we can not use it
3944 *    for allocating an extent and end up being forced to allocate a new one;
3945 *
3946 * 2) Because we only check that the metadata space_info has enough free bytes,
3947 *    we end up not allocating a new metadata chunk in that case. However if
3948 *    the filesystem was mounted in degraded mode, none of the existing block
3949 *    groups might be suitable for extent allocation due to their incompatible
3950 *    profile (for e.g. mounting a 2 devices filesystem, where all block groups
3951 *    use a RAID1 profile, in degraded mode using a single device). In this case
3952 *    when the task attempts to COW some extent buffer of the extent btree for
3953 *    example, it will trigger allocation of a new metadata block group with a
3954 *    suitable profile (SINGLE profile in the example of the degraded mount of
3955 *    the RAID1 filesystem);
3956 *
3957 * 3) The task has reserved enough transaction units / metadata space, but when
3958 *    it attempts to COW an extent buffer from the extent or device btree for
3959 *    example, it does not find any free extent in any metadata block group,
3960 *    therefore forced to try to allocate a new metadata block group.
3961 *    This is because some other task allocated all available extents in the
3962 *    meanwhile - this typically happens with tasks that don't reserve space
3963 *    properly, either intentionally or as a bug. One example where this is
3964 *    done intentionally is fsync, as it does not reserve any transaction units
3965 *    and ends up allocating a variable number of metadata extents for log
3966 *    tree extent buffers;
3967 *
3968 * 4) The task has reserved enough transaction units / metadata space, but right
3969 *    before it tries to allocate the last extent buffer it needs, a discard
3970 *    operation comes in and, temporarily, removes the last free space entry from
3971 *    the only metadata block group that had free space (discard starts by
3972 *    removing a free space entry from a block group, then does the discard
3973 *    operation and, once it's done, it adds back the free space entry to the
3974 *    block group).
3975 *
3976 * We also need this 2 phases setup when adding a device to a filesystem with
3977 * a seed device - we must create new metadata and system chunks without adding
3978 * any of the block group items to the chunk, extent and device btrees. If we
3979 * did not do it this way, we would get ENOSPC when attempting to update those
3980 * btrees, since all the chunks from the seed device are read-only.
3981 *
3982 * Phase 1 does the updates and insertions to the chunk btree because if we had
3983 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3984 * parallel, we risk having too many system chunks allocated by many tasks if
3985 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3986 * extreme case this leads to exhaustion of the system chunk array in the
3987 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3988 * and with RAID filesystems (so we have more device items in the chunk btree).
3989 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3990 * the system chunk array due to concurrent allocations") provides more details.
3991 *
3992 * Allocation of system chunks does not happen through this function. A task that
3993 * needs to update the chunk btree (the only btree that uses system chunks), must
3994 * preallocate chunk space by calling either check_system_chunk() or
3995 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
3996 * metadata chunk or when removing a chunk, while the later is used before doing
3997 * a modification to the chunk btree - use cases for the later are adding,
3998 * removing and resizing a device as well as relocation of a system chunk.
3999 * See the comment below for more details.
4000 *
4001 * The reservation of system space, done through check_system_chunk(), as well
4002 * as all the updates and insertions into the chunk btree must be done while
4003 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4004 * an extent buffer from the chunks btree we never trigger allocation of a new
4005 * system chunk, which would result in a deadlock (trying to lock twice an
4006 * extent buffer of the chunk btree, first time before triggering the chunk
4007 * allocation and the second time during chunk allocation while attempting to
4008 * update the chunks btree). The system chunk array is also updated while holding
4009 * that mutex. The same logic applies to removing chunks - we must reserve system
4010 * space, update the chunk btree and the system chunk array in the superblock
4011 * while holding fs_info->chunk_mutex.
4012 *
4013 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4014 *
4015 * If @force is CHUNK_ALLOC_FORCE:
4016 *    - return 1 if it successfully allocates a chunk,
4017 *    - return errors including -ENOSPC otherwise.
4018 * If @force is NOT CHUNK_ALLOC_FORCE:
4019 *    - return 0 if it doesn't need to allocate a new chunk,
4020 *    - return 1 if it successfully allocates a chunk,
4021 *    - return errors including -ENOSPC otherwise.
4022 */
4023int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4024		      enum btrfs_chunk_alloc_enum force)
4025{
4026	struct btrfs_fs_info *fs_info = trans->fs_info;
4027	struct btrfs_space_info *space_info;
4028	struct btrfs_block_group *ret_bg;
4029	bool wait_for_alloc = false;
4030	bool should_alloc = false;
4031	bool from_extent_allocation = false;
4032	int ret = 0;
4033
4034	if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4035		from_extent_allocation = true;
4036		force = CHUNK_ALLOC_FORCE;
4037	}
4038
4039	/* Don't re-enter if we're already allocating a chunk */
4040	if (trans->allocating_chunk)
4041		return -ENOSPC;
4042	/*
4043	 * Allocation of system chunks can not happen through this path, as we
4044	 * could end up in a deadlock if we are allocating a data or metadata
4045	 * chunk and there is another task modifying the chunk btree.
4046	 *
4047	 * This is because while we are holding the chunk mutex, we will attempt
4048	 * to add the new chunk item to the chunk btree or update an existing
4049	 * device item in the chunk btree, while the other task that is modifying
4050	 * the chunk btree is attempting to COW an extent buffer while holding a
4051	 * lock on it and on its parent - if the COW operation triggers a system
4052	 * chunk allocation, then we can deadlock because we are holding the
4053	 * chunk mutex and we may need to access that extent buffer or its parent
4054	 * in order to add the chunk item or update a device item.
4055	 *
4056	 * Tasks that want to modify the chunk tree should reserve system space
4057	 * before updating the chunk btree, by calling either
4058	 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4059	 * It's possible that after a task reserves the space, it still ends up
4060	 * here - this happens in the cases described above at do_chunk_alloc().
4061	 * The task will have to either retry or fail.
4062	 */
4063	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4064		return -ENOSPC;
4065
4066	space_info = btrfs_find_space_info(fs_info, flags);
4067	ASSERT(space_info);
4068
4069	do {
4070		spin_lock(&space_info->lock);
4071		if (force < space_info->force_alloc)
4072			force = space_info->force_alloc;
4073		should_alloc = should_alloc_chunk(fs_info, space_info, force);
4074		if (space_info->full) {
4075			/* No more free physical space */
4076			if (should_alloc)
4077				ret = -ENOSPC;
4078			else
4079				ret = 0;
4080			spin_unlock(&space_info->lock);
4081			return ret;
4082		} else if (!should_alloc) {
4083			spin_unlock(&space_info->lock);
4084			return 0;
4085		} else if (space_info->chunk_alloc) {
4086			/*
4087			 * Someone is already allocating, so we need to block
4088			 * until this someone is finished and then loop to
4089			 * recheck if we should continue with our allocation
4090			 * attempt.
4091			 */
4092			wait_for_alloc = true;
4093			force = CHUNK_ALLOC_NO_FORCE;
4094			spin_unlock(&space_info->lock);
4095			mutex_lock(&fs_info->chunk_mutex);
4096			mutex_unlock(&fs_info->chunk_mutex);
4097		} else {
4098			/* Proceed with allocation */
4099			space_info->chunk_alloc = 1;
4100			wait_for_alloc = false;
4101			spin_unlock(&space_info->lock);
4102		}
4103
4104		cond_resched();
4105	} while (wait_for_alloc);
4106
4107	mutex_lock(&fs_info->chunk_mutex);
4108	trans->allocating_chunk = true;
4109
4110	/*
4111	 * If we have mixed data/metadata chunks we want to make sure we keep
4112	 * allocating mixed chunks instead of individual chunks.
4113	 */
4114	if (btrfs_mixed_space_info(space_info))
4115		flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4116
4117	/*
4118	 * if we're doing a data chunk, go ahead and make sure that
4119	 * we keep a reasonable number of metadata chunks allocated in the
4120	 * FS as well.
4121	 */
4122	if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4123		fs_info->data_chunk_allocations++;
4124		if (!(fs_info->data_chunk_allocations %
4125		      fs_info->metadata_ratio))
4126			force_metadata_allocation(fs_info);
4127	}
4128
4129	ret_bg = do_chunk_alloc(trans, flags);
4130	trans->allocating_chunk = false;
4131
4132	if (IS_ERR(ret_bg)) {
4133		ret = PTR_ERR(ret_bg);
4134	} else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4135		/*
4136		 * New block group is likely to be used soon. Try to activate
4137		 * it now. Failure is OK for now.
4138		 */
4139		btrfs_zone_activate(ret_bg);
4140	}
4141
4142	if (!ret)
4143		btrfs_put_block_group(ret_bg);
4144
4145	spin_lock(&space_info->lock);
4146	if (ret < 0) {
4147		if (ret == -ENOSPC)
4148			space_info->full = 1;
4149		else
4150			goto out;
4151	} else {
4152		ret = 1;
4153		space_info->max_extent_size = 0;
4154	}
4155
4156	space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4157out:
4158	space_info->chunk_alloc = 0;
4159	spin_unlock(&space_info->lock);
4160	mutex_unlock(&fs_info->chunk_mutex);
4161
4162	return ret;
4163}
4164
4165static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4166{
4167	u64 num_dev;
4168
4169	num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4170	if (!num_dev)
4171		num_dev = fs_info->fs_devices->rw_devices;
4172
4173	return num_dev;
4174}
4175
4176static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4177				u64 bytes,
4178				u64 type)
4179{
4180	struct btrfs_fs_info *fs_info = trans->fs_info;
4181	struct btrfs_space_info *info;
4182	u64 left;
4183	int ret = 0;
4184
4185	/*
4186	 * Needed because we can end up allocating a system chunk and for an
4187	 * atomic and race free space reservation in the chunk block reserve.
4188	 */
4189	lockdep_assert_held(&fs_info->chunk_mutex);
4190
4191	info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4192	spin_lock(&info->lock);
4193	left = info->total_bytes - btrfs_space_info_used(info, true);
4194	spin_unlock(&info->lock);
4195
4196	if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4197		btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4198			   left, bytes, type);
4199		btrfs_dump_space_info(fs_info, info, 0, 0);
4200	}
4201
4202	if (left < bytes) {
4203		u64 flags = btrfs_system_alloc_profile(fs_info);
4204		struct btrfs_block_group *bg;
4205
4206		/*
4207		 * Ignore failure to create system chunk. We might end up not
4208		 * needing it, as we might not need to COW all nodes/leafs from
4209		 * the paths we visit in the chunk tree (they were already COWed
4210		 * or created in the current transaction for example).
4211		 */
4212		bg = btrfs_create_chunk(trans, flags);
4213		if (IS_ERR(bg)) {
4214			ret = PTR_ERR(bg);
4215		} else {
4216			/*
4217			 * We have a new chunk. We also need to activate it for
4218			 * zoned filesystem.
4219			 */
4220			ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
4221			if (ret < 0)
4222				return;
4223
4224			/*
4225			 * If we fail to add the chunk item here, we end up
4226			 * trying again at phase 2 of chunk allocation, at
4227			 * btrfs_create_pending_block_groups(). So ignore
4228			 * any error here. An ENOSPC here could happen, due to
4229			 * the cases described at do_chunk_alloc() - the system
4230			 * block group we just created was just turned into RO
4231			 * mode by a scrub for example, or a running discard
4232			 * temporarily removed its free space entries, etc.
4233			 */
4234			btrfs_chunk_alloc_add_chunk_item(trans, bg);
4235		}
4236	}
4237
4238	if (!ret) {
4239		ret = btrfs_block_rsv_add(fs_info,
4240					  &fs_info->chunk_block_rsv,
4241					  bytes, BTRFS_RESERVE_NO_FLUSH);
4242		if (!ret)
4243			trans->chunk_bytes_reserved += bytes;
4244	}
4245}
4246
4247/*
4248 * Reserve space in the system space for allocating or removing a chunk.
4249 * The caller must be holding fs_info->chunk_mutex.
4250 */
4251void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4252{
4253	struct btrfs_fs_info *fs_info = trans->fs_info;
4254	const u64 num_devs = get_profile_num_devs(fs_info, type);
4255	u64 bytes;
4256
4257	/* num_devs device items to update and 1 chunk item to add or remove. */
4258	bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4259		btrfs_calc_insert_metadata_size(fs_info, 1);
4260
4261	reserve_chunk_space(trans, bytes, type);
4262}
4263
4264/*
4265 * Reserve space in the system space, if needed, for doing a modification to the
4266 * chunk btree.
4267 *
4268 * @trans:		A transaction handle.
4269 * @is_item_insertion:	Indicate if the modification is for inserting a new item
4270 *			in the chunk btree or if it's for the deletion or update
4271 *			of an existing item.
4272 *
4273 * This is used in a context where we need to update the chunk btree outside
4274 * block group allocation and removal, to avoid a deadlock with a concurrent
4275 * task that is allocating a metadata or data block group and therefore needs to
4276 * update the chunk btree while holding the chunk mutex. After the update to the
4277 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4278 *
4279 */
4280void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4281				  bool is_item_insertion)
4282{
4283	struct btrfs_fs_info *fs_info = trans->fs_info;
4284	u64 bytes;
4285
4286	if (is_item_insertion)
4287		bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4288	else
4289		bytes = btrfs_calc_metadata_size(fs_info, 1);
4290
4291	mutex_lock(&fs_info->chunk_mutex);
4292	reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4293	mutex_unlock(&fs_info->chunk_mutex);
4294}
4295
4296void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4297{
4298	struct btrfs_block_group *block_group;
4299
4300	block_group = btrfs_lookup_first_block_group(info, 0);
4301	while (block_group) {
4302		btrfs_wait_block_group_cache_done(block_group);
4303		spin_lock(&block_group->lock);
4304		if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4305				       &block_group->runtime_flags)) {
4306			struct inode *inode = block_group->inode;
4307
4308			block_group->inode = NULL;
4309			spin_unlock(&block_group->lock);
4310
4311			ASSERT(block_group->io_ctl.inode == NULL);
4312			iput(inode);
4313		} else {
4314			spin_unlock(&block_group->lock);
4315		}
4316		block_group = btrfs_next_block_group(block_group);
4317	}
4318}
4319
4320/*
4321 * Must be called only after stopping all workers, since we could have block
4322 * group caching kthreads running, and therefore they could race with us if we
4323 * freed the block groups before stopping them.
4324 */
4325int btrfs_free_block_groups(struct btrfs_fs_info *info)
4326{
4327	struct btrfs_block_group *block_group;
4328	struct btrfs_space_info *space_info;
4329	struct btrfs_caching_control *caching_ctl;
4330	struct rb_node *n;
4331
4332	if (btrfs_is_zoned(info)) {
4333		if (info->active_meta_bg) {
4334			btrfs_put_block_group(info->active_meta_bg);
4335			info->active_meta_bg = NULL;
4336		}
4337		if (info->active_system_bg) {
4338			btrfs_put_block_group(info->active_system_bg);
4339			info->active_system_bg = NULL;
4340		}
4341	}
4342
4343	write_lock(&info->block_group_cache_lock);
4344	while (!list_empty(&info->caching_block_groups)) {
4345		caching_ctl = list_entry(info->caching_block_groups.next,
4346					 struct btrfs_caching_control, list);
4347		list_del(&caching_ctl->list);
4348		btrfs_put_caching_control(caching_ctl);
4349	}
4350	write_unlock(&info->block_group_cache_lock);
4351
4352	spin_lock(&info->unused_bgs_lock);
4353	while (!list_empty(&info->unused_bgs)) {
4354		block_group = list_first_entry(&info->unused_bgs,
4355					       struct btrfs_block_group,
4356					       bg_list);
4357		list_del_init(&block_group->bg_list);
4358		btrfs_put_block_group(block_group);
4359	}
4360
4361	while (!list_empty(&info->reclaim_bgs)) {
4362		block_group = list_first_entry(&info->reclaim_bgs,
4363					       struct btrfs_block_group,
4364					       bg_list);
4365		list_del_init(&block_group->bg_list);
4366		btrfs_put_block_group(block_group);
4367	}
4368	spin_unlock(&info->unused_bgs_lock);
4369
4370	spin_lock(&info->zone_active_bgs_lock);
4371	while (!list_empty(&info->zone_active_bgs)) {
4372		block_group = list_first_entry(&info->zone_active_bgs,
4373					       struct btrfs_block_group,
4374					       active_bg_list);
4375		list_del_init(&block_group->active_bg_list);
4376		btrfs_put_block_group(block_group);
4377	}
4378	spin_unlock(&info->zone_active_bgs_lock);
4379
4380	write_lock(&info->block_group_cache_lock);
4381	while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4382		block_group = rb_entry(n, struct btrfs_block_group,
4383				       cache_node);
4384		rb_erase_cached(&block_group->cache_node,
4385				&info->block_group_cache_tree);
4386		RB_CLEAR_NODE(&block_group->cache_node);
4387		write_unlock(&info->block_group_cache_lock);
4388
4389		down_write(&block_group->space_info->groups_sem);
4390		list_del(&block_group->list);
4391		up_write(&block_group->space_info->groups_sem);
4392
4393		/*
4394		 * We haven't cached this block group, which means we could
4395		 * possibly have excluded extents on this block group.
4396		 */
4397		if (block_group->cached == BTRFS_CACHE_NO ||
4398		    block_group->cached == BTRFS_CACHE_ERROR)
4399			btrfs_free_excluded_extents(block_group);
4400
4401		btrfs_remove_free_space_cache(block_group);
4402		ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4403		ASSERT(list_empty(&block_group->dirty_list));
4404		ASSERT(list_empty(&block_group->io_list));
4405		ASSERT(list_empty(&block_group->bg_list));
4406		ASSERT(refcount_read(&block_group->refs) == 1);
4407		ASSERT(block_group->swap_extents == 0);
4408		btrfs_put_block_group(block_group);
4409
4410		write_lock(&info->block_group_cache_lock);
4411	}
4412	write_unlock(&info->block_group_cache_lock);
4413
4414	btrfs_release_global_block_rsv(info);
4415
4416	while (!list_empty(&info->space_info)) {
4417		space_info = list_entry(info->space_info.next,
4418					struct btrfs_space_info,
4419					list);
4420
4421		/*
4422		 * Do not hide this behind enospc_debug, this is actually
4423		 * important and indicates a real bug if this happens.
4424		 */
4425		if (WARN_ON(space_info->bytes_pinned > 0 ||
4426			    space_info->bytes_may_use > 0))
4427			btrfs_dump_space_info(info, space_info, 0, 0);
4428
4429		/*
4430		 * If there was a failure to cleanup a log tree, very likely due
4431		 * to an IO failure on a writeback attempt of one or more of its
4432		 * extent buffers, we could not do proper (and cheap) unaccounting
4433		 * of their reserved space, so don't warn on bytes_reserved > 0 in
4434		 * that case.
4435		 */
4436		if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4437		    !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4438			if (WARN_ON(space_info->bytes_reserved > 0))
4439				btrfs_dump_space_info(info, space_info, 0, 0);
4440		}
4441
4442		WARN_ON(space_info->reclaim_size > 0);
4443		list_del(&space_info->list);
4444		btrfs_sysfs_remove_space_info(space_info);
4445	}
4446	return 0;
4447}
4448
4449void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4450{
4451	atomic_inc(&cache->frozen);
4452}
4453
4454void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4455{
4456	struct btrfs_fs_info *fs_info = block_group->fs_info;
4457	bool cleanup;
4458
4459	spin_lock(&block_group->lock);
4460	cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4461		   test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4462	spin_unlock(&block_group->lock);
4463
4464	if (cleanup) {
4465		struct btrfs_chunk_map *map;
4466
4467		map = btrfs_find_chunk_map(fs_info, block_group->start, 1);
4468		/* Logic error, can't happen. */
4469		ASSERT(map);
4470
4471		btrfs_remove_chunk_map(fs_info, map);
4472
4473		/* Once for our lookup reference. */
4474		btrfs_free_chunk_map(map);
4475
4476		/*
4477		 * We may have left one free space entry and other possible
4478		 * tasks trimming this block group have left 1 entry each one.
4479		 * Free them if any.
4480		 */
4481		btrfs_remove_free_space_cache(block_group);
4482	}
4483}
4484
4485bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4486{
4487	bool ret = true;
4488
4489	spin_lock(&bg->lock);
4490	if (bg->ro)
4491		ret = false;
4492	else
4493		bg->swap_extents++;
4494	spin_unlock(&bg->lock);
4495
4496	return ret;
4497}
4498
4499void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4500{
4501	spin_lock(&bg->lock);
4502	ASSERT(!bg->ro);
4503	ASSERT(bg->swap_extents >= amount);
4504	bg->swap_extents -= amount;
4505	spin_unlock(&bg->lock);
4506}
4507
4508enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4509{
4510	if (size <= SZ_128K)
4511		return BTRFS_BG_SZ_SMALL;
4512	if (size <= SZ_8M)
4513		return BTRFS_BG_SZ_MEDIUM;
4514	return BTRFS_BG_SZ_LARGE;
4515}
4516
4517/*
4518 * Handle a block group allocating an extent in a size class
4519 *
4520 * @bg:				The block group we allocated in.
4521 * @size_class:			The size class of the allocation.
4522 * @force_wrong_size_class:	Whether we are desperate enough to allow
4523 *				mismatched size classes.
4524 *
4525 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4526 * case of a race that leads to the wrong size class without
4527 * force_wrong_size_class set.
4528 *
4529 * find_free_extent will skip block groups with a mismatched size class until
4530 * it really needs to avoid ENOSPC. In that case it will set
4531 * force_wrong_size_class. However, if a block group is newly allocated and
4532 * doesn't yet have a size class, then it is possible for two allocations of
4533 * different sizes to race and both try to use it. The loser is caught here and
4534 * has to retry.
4535 */
4536int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4537				     enum btrfs_block_group_size_class size_class,
4538				     bool force_wrong_size_class)
4539{
4540	ASSERT(size_class != BTRFS_BG_SZ_NONE);
4541
4542	/* The new allocation is in the right size class, do nothing */
4543	if (bg->size_class == size_class)
4544		return 0;
4545	/*
4546	 * The new allocation is in a mismatched size class.
4547	 * This means one of two things:
4548	 *
4549	 * 1. Two tasks in find_free_extent for different size_classes raced
4550	 *    and hit the same empty block_group. Make the loser try again.
4551	 * 2. A call to find_free_extent got desperate enough to set
4552	 *    'force_wrong_slab'. Don't change the size_class, but allow the
4553	 *    allocation.
4554	 */
4555	if (bg->size_class != BTRFS_BG_SZ_NONE) {
4556		if (force_wrong_size_class)
4557			return 0;
4558		return -EAGAIN;
4559	}
4560	/*
4561	 * The happy new block group case: the new allocation is the first
4562	 * one in the block_group so we set size_class.
4563	 */
4564	bg->size_class = size_class;
4565
4566	return 0;
4567}
4568
4569bool btrfs_block_group_should_use_size_class(struct btrfs_block_group *bg)
4570{
4571	if (btrfs_is_zoned(bg->fs_info))
4572		return false;
4573	if (!btrfs_is_block_group_data_only(bg))
4574		return false;
4575	return true;
4576}
v6.9.4
   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(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
  37/*
  38 * Return target flags in extended format or 0 if restripe for this chunk_type
  39 * is not in progress
  40 *
  41 * Should be called with balance_lock held
  42 */
  43static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
  44{
  45	struct btrfs_balance_control *bctl = fs_info->balance_ctl;
  46	u64 target = 0;
  47
  48	if (!bctl)
  49		return 0;
  50
  51	if (flags & BTRFS_BLOCK_GROUP_DATA &&
  52	    bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  53		target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
  54	} else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
  55		   bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  56		target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
  57	} else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
  58		   bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  59		target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
  60	}
  61
  62	return target;
  63}
  64
  65/*
  66 * @flags: available profiles in extended format (see ctree.h)
  67 *
  68 * Return reduced profile in chunk format.  If profile changing is in progress
  69 * (either running or paused) picks the target profile (if it's already
  70 * available), otherwise falls back to plain reducing.
  71 */
  72static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
  73{
  74	u64 num_devices = fs_info->fs_devices->rw_devices;
  75	u64 target;
  76	u64 raid_type;
  77	u64 allowed = 0;
  78
  79	/*
  80	 * See if restripe for this chunk_type is in progress, if so try to
  81	 * reduce to the target profile
  82	 */
  83	spin_lock(&fs_info->balance_lock);
  84	target = get_restripe_target(fs_info, flags);
  85	if (target) {
  86		spin_unlock(&fs_info->balance_lock);
  87		return extended_to_chunk(target);
  88	}
  89	spin_unlock(&fs_info->balance_lock);
  90
  91	/* First, mask out the RAID levels which aren't possible */
  92	for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
  93		if (num_devices >= btrfs_raid_array[raid_type].devs_min)
  94			allowed |= btrfs_raid_array[raid_type].bg_flag;
  95	}
  96	allowed &= flags;
  97
  98	/* Select the highest-redundancy RAID level. */
  99	if (allowed & BTRFS_BLOCK_GROUP_RAID1C4)
 100		allowed = BTRFS_BLOCK_GROUP_RAID1C4;
 101	else if (allowed & BTRFS_BLOCK_GROUP_RAID6)
 102		allowed = BTRFS_BLOCK_GROUP_RAID6;
 103	else if (allowed & BTRFS_BLOCK_GROUP_RAID1C3)
 104		allowed = BTRFS_BLOCK_GROUP_RAID1C3;
 105	else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
 106		allowed = BTRFS_BLOCK_GROUP_RAID5;
 107	else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
 108		allowed = BTRFS_BLOCK_GROUP_RAID10;
 109	else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
 110		allowed = BTRFS_BLOCK_GROUP_RAID1;
 111	else if (allowed & BTRFS_BLOCK_GROUP_DUP)
 112		allowed = BTRFS_BLOCK_GROUP_DUP;
 113	else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
 114		allowed = BTRFS_BLOCK_GROUP_RAID0;
 115
 116	flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
 117
 118	return extended_to_chunk(flags | allowed);
 119}
 120
 121u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
 122{
 123	unsigned seq;
 124	u64 flags;
 125
 126	do {
 127		flags = orig_flags;
 128		seq = read_seqbegin(&fs_info->profiles_lock);
 129
 130		if (flags & BTRFS_BLOCK_GROUP_DATA)
 131			flags |= fs_info->avail_data_alloc_bits;
 132		else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 133			flags |= fs_info->avail_system_alloc_bits;
 134		else if (flags & BTRFS_BLOCK_GROUP_METADATA)
 135			flags |= fs_info->avail_metadata_alloc_bits;
 136	} while (read_seqretry(&fs_info->profiles_lock, seq));
 137
 138	return btrfs_reduce_alloc_profile(fs_info, flags);
 139}
 140
 141void btrfs_get_block_group(struct btrfs_block_group *cache)
 142{
 143	refcount_inc(&cache->refs);
 144}
 145
 146void btrfs_put_block_group(struct btrfs_block_group *cache)
 147{
 148	if (refcount_dec_and_test(&cache->refs)) {
 149		WARN_ON(cache->pinned > 0);
 150		/*
 151		 * If there was a failure to cleanup a log tree, very likely due
 152		 * to an IO failure on a writeback attempt of one or more of its
 153		 * extent buffers, we could not do proper (and cheap) unaccounting
 154		 * of their reserved space, so don't warn on reserved > 0 in that
 155		 * case.
 156		 */
 157		if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
 158		    !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
 159			WARN_ON(cache->reserved > 0);
 160
 161		/*
 162		 * A block_group shouldn't be on the discard_list anymore.
 163		 * Remove the block_group from the discard_list to prevent us
 164		 * from causing a panic due to NULL pointer dereference.
 165		 */
 166		if (WARN_ON(!list_empty(&cache->discard_list)))
 167			btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
 168						  cache);
 169
 170		kfree(cache->free_space_ctl);
 171		btrfs_free_chunk_map(cache->physical_map);
 172		kfree(cache);
 173	}
 174}
 175
 176/*
 177 * This adds the block group to the fs_info rb tree for the block group cache
 178 */
 179static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 180				       struct btrfs_block_group *block_group)
 181{
 182	struct rb_node **p;
 183	struct rb_node *parent = NULL;
 184	struct btrfs_block_group *cache;
 185	bool leftmost = true;
 186
 187	ASSERT(block_group->length != 0);
 188
 189	write_lock(&info->block_group_cache_lock);
 190	p = &info->block_group_cache_tree.rb_root.rb_node;
 191
 192	while (*p) {
 193		parent = *p;
 194		cache = rb_entry(parent, struct btrfs_block_group, cache_node);
 195		if (block_group->start < cache->start) {
 196			p = &(*p)->rb_left;
 197		} else if (block_group->start > cache->start) {
 198			p = &(*p)->rb_right;
 199			leftmost = false;
 200		} else {
 201			write_unlock(&info->block_group_cache_lock);
 202			return -EEXIST;
 203		}
 204	}
 205
 206	rb_link_node(&block_group->cache_node, parent, p);
 207	rb_insert_color_cached(&block_group->cache_node,
 208			       &info->block_group_cache_tree, leftmost);
 209
 210	write_unlock(&info->block_group_cache_lock);
 211
 212	return 0;
 213}
 214
 215/*
 216 * This will return the block group at or after bytenr if contains is 0, else
 217 * it will return the block group that contains the bytenr
 218 */
 219static struct btrfs_block_group *block_group_cache_tree_search(
 220		struct btrfs_fs_info *info, u64 bytenr, int contains)
 221{
 222	struct btrfs_block_group *cache, *ret = NULL;
 223	struct rb_node *n;
 224	u64 end, start;
 225
 226	read_lock(&info->block_group_cache_lock);
 227	n = info->block_group_cache_tree.rb_root.rb_node;
 228
 229	while (n) {
 230		cache = rb_entry(n, struct btrfs_block_group, cache_node);
 231		end = cache->start + cache->length - 1;
 232		start = cache->start;
 233
 234		if (bytenr < start) {
 235			if (!contains && (!ret || start < ret->start))
 236				ret = cache;
 237			n = n->rb_left;
 238		} else if (bytenr > start) {
 239			if (contains && bytenr <= end) {
 240				ret = cache;
 241				break;
 242			}
 243			n = n->rb_right;
 244		} else {
 245			ret = cache;
 246			break;
 247		}
 248	}
 249	if (ret)
 250		btrfs_get_block_group(ret);
 251	read_unlock(&info->block_group_cache_lock);
 252
 253	return ret;
 254}
 255
 256/*
 257 * Return the block group that starts at or after bytenr
 258 */
 259struct btrfs_block_group *btrfs_lookup_first_block_group(
 260		struct btrfs_fs_info *info, u64 bytenr)
 261{
 262	return block_group_cache_tree_search(info, bytenr, 0);
 263}
 264
 265/*
 266 * Return the block group that contains the given bytenr
 267 */
 268struct btrfs_block_group *btrfs_lookup_block_group(
 269		struct btrfs_fs_info *info, u64 bytenr)
 270{
 271	return block_group_cache_tree_search(info, bytenr, 1);
 272}
 273
 274struct btrfs_block_group *btrfs_next_block_group(
 275		struct btrfs_block_group *cache)
 276{
 277	struct btrfs_fs_info *fs_info = cache->fs_info;
 278	struct rb_node *node;
 279
 280	read_lock(&fs_info->block_group_cache_lock);
 281
 282	/* If our block group was removed, we need a full search. */
 283	if (RB_EMPTY_NODE(&cache->cache_node)) {
 284		const u64 next_bytenr = cache->start + cache->length;
 285
 286		read_unlock(&fs_info->block_group_cache_lock);
 287		btrfs_put_block_group(cache);
 288		return btrfs_lookup_first_block_group(fs_info, next_bytenr);
 289	}
 290	node = rb_next(&cache->cache_node);
 291	btrfs_put_block_group(cache);
 292	if (node) {
 293		cache = rb_entry(node, struct btrfs_block_group, cache_node);
 294		btrfs_get_block_group(cache);
 295	} else
 296		cache = NULL;
 297	read_unlock(&fs_info->block_group_cache_lock);
 298	return cache;
 299}
 300
 301/*
 302 * Check if we can do a NOCOW write for a given extent.
 303 *
 304 * @fs_info:       The filesystem information object.
 305 * @bytenr:        Logical start address of the extent.
 306 *
 307 * Check if we can do a NOCOW write for the given extent, and increments the
 308 * number of NOCOW writers in the block group that contains the extent, as long
 309 * as the block group exists and it's currently not in read-only mode.
 310 *
 311 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
 312 *          is responsible for calling btrfs_dec_nocow_writers() later.
 313 *
 314 *          Or NULL if we can not do a NOCOW write
 315 */
 316struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
 317						  u64 bytenr)
 318{
 319	struct btrfs_block_group *bg;
 320	bool can_nocow = true;
 321
 322	bg = btrfs_lookup_block_group(fs_info, bytenr);
 323	if (!bg)
 324		return NULL;
 325
 326	spin_lock(&bg->lock);
 327	if (bg->ro)
 328		can_nocow = false;
 329	else
 330		atomic_inc(&bg->nocow_writers);
 331	spin_unlock(&bg->lock);
 332
 333	if (!can_nocow) {
 334		btrfs_put_block_group(bg);
 335		return NULL;
 336	}
 337
 338	/* No put on block group, done by btrfs_dec_nocow_writers(). */
 339	return bg;
 340}
 341
 342/*
 343 * Decrement the number of NOCOW writers in a block group.
 344 *
 345 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
 346 * and on the block group returned by that call. Typically this is called after
 347 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
 348 * relocation.
 349 *
 350 * After this call, the caller should not use the block group anymore. It it wants
 351 * to use it, then it should get a reference on it before calling this function.
 352 */
 353void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
 354{
 355	if (atomic_dec_and_test(&bg->nocow_writers))
 356		wake_up_var(&bg->nocow_writers);
 357
 358	/* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
 359	btrfs_put_block_group(bg);
 360}
 361
 362void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
 363{
 364	wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
 365}
 366
 367void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
 368					const u64 start)
 369{
 370	struct btrfs_block_group *bg;
 371
 372	bg = btrfs_lookup_block_group(fs_info, start);
 373	ASSERT(bg);
 374	if (atomic_dec_and_test(&bg->reservations))
 375		wake_up_var(&bg->reservations);
 376	btrfs_put_block_group(bg);
 377}
 378
 379void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
 380{
 381	struct btrfs_space_info *space_info = bg->space_info;
 382
 383	ASSERT(bg->ro);
 384
 385	if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
 386		return;
 387
 388	/*
 389	 * Our block group is read only but before we set it to read only,
 390	 * some task might have had allocated an extent from it already, but it
 391	 * has not yet created a respective ordered extent (and added it to a
 392	 * root's list of ordered extents).
 393	 * Therefore wait for any task currently allocating extents, since the
 394	 * block group's reservations counter is incremented while a read lock
 395	 * on the groups' semaphore is held and decremented after releasing
 396	 * the read access on that semaphore and creating the ordered extent.
 397	 */
 398	down_write(&space_info->groups_sem);
 399	up_write(&space_info->groups_sem);
 400
 401	wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
 402}
 403
 404struct btrfs_caching_control *btrfs_get_caching_control(
 405		struct btrfs_block_group *cache)
 406{
 407	struct btrfs_caching_control *ctl;
 408
 409	spin_lock(&cache->lock);
 410	if (!cache->caching_ctl) {
 411		spin_unlock(&cache->lock);
 412		return NULL;
 413	}
 414
 415	ctl = cache->caching_ctl;
 416	refcount_inc(&ctl->count);
 417	spin_unlock(&cache->lock);
 418	return ctl;
 419}
 420
 421static void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
 422{
 423	if (refcount_dec_and_test(&ctl->count))
 424		kfree(ctl);
 425}
 426
 427/*
 428 * When we wait for progress in the block group caching, its because our
 429 * allocation attempt failed at least once.  So, we must sleep and let some
 430 * progress happen before we try again.
 431 *
 432 * This function will sleep at least once waiting for new free space to show
 433 * up, and then it will check the block group free space numbers for our min
 434 * num_bytes.  Another option is to have it go ahead and look in the rbtree for
 435 * a free extent of a given size, but this is a good start.
 436 *
 437 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
 438 * any of the information in this block group.
 439 */
 440void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
 441					   u64 num_bytes)
 442{
 443	struct btrfs_caching_control *caching_ctl;
 444	int progress;
 445
 446	caching_ctl = btrfs_get_caching_control(cache);
 447	if (!caching_ctl)
 448		return;
 449
 450	/*
 451	 * We've already failed to allocate from this block group, so even if
 452	 * there's enough space in the block group it isn't contiguous enough to
 453	 * allow for an allocation, so wait for at least the next wakeup tick,
 454	 * or for the thing to be done.
 455	 */
 456	progress = atomic_read(&caching_ctl->progress);
 457
 458	wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
 459		   (progress != atomic_read(&caching_ctl->progress) &&
 460		    (cache->free_space_ctl->free_space >= num_bytes)));
 461
 462	btrfs_put_caching_control(caching_ctl);
 463}
 464
 465static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
 466				       struct btrfs_caching_control *caching_ctl)
 467{
 468	wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
 469	return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
 470}
 471
 472static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
 473{
 474	struct btrfs_caching_control *caching_ctl;
 475	int ret;
 476
 477	caching_ctl = btrfs_get_caching_control(cache);
 478	if (!caching_ctl)
 479		return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
 480	ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
 481	btrfs_put_caching_control(caching_ctl);
 482	return ret;
 483}
 484
 485#ifdef CONFIG_BTRFS_DEBUG
 486static void fragment_free_space(struct btrfs_block_group *block_group)
 487{
 488	struct btrfs_fs_info *fs_info = block_group->fs_info;
 489	u64 start = block_group->start;
 490	u64 len = block_group->length;
 491	u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 492		fs_info->nodesize : fs_info->sectorsize;
 493	u64 step = chunk << 1;
 494
 495	while (len > chunk) {
 496		btrfs_remove_free_space(block_group, start, chunk);
 497		start += step;
 498		if (len < step)
 499			len = 0;
 500		else
 501			len -= step;
 502	}
 503}
 504#endif
 505
 506/*
 507 * Add a free space range to the in memory free space cache of a block group.
 508 * This checks if the range contains super block locations and any such
 509 * locations are not added to the free space cache.
 510 *
 511 * @block_group:      The target block group.
 512 * @start:            Start offset of the range.
 513 * @end:              End offset of the range (exclusive).
 514 * @total_added_ret:  Optional pointer to return the total amount of space
 515 *                    added to the block group's free space cache.
 516 *
 517 * Returns 0 on success or < 0 on error.
 518 */
 519int btrfs_add_new_free_space(struct btrfs_block_group *block_group, u64 start,
 520			     u64 end, u64 *total_added_ret)
 521{
 522	struct btrfs_fs_info *info = block_group->fs_info;
 523	u64 extent_start, extent_end, size;
 524	int ret;
 525
 526	if (total_added_ret)
 527		*total_added_ret = 0;
 528
 529	while (start < end) {
 530		if (!find_first_extent_bit(&info->excluded_extents, start,
 531					   &extent_start, &extent_end,
 532					   EXTENT_DIRTY | EXTENT_UPTODATE,
 533					   NULL))
 534			break;
 535
 536		if (extent_start <= start) {
 537			start = extent_end + 1;
 538		} else if (extent_start > start && extent_start < end) {
 539			size = extent_start - start;
 540			ret = btrfs_add_free_space_async_trimmed(block_group,
 541								 start, size);
 542			if (ret)
 543				return ret;
 544			if (total_added_ret)
 545				*total_added_ret += size;
 546			start = extent_end + 1;
 547		} else {
 548			break;
 549		}
 550	}
 551
 552	if (start < end) {
 553		size = end - start;
 554		ret = btrfs_add_free_space_async_trimmed(block_group, start,
 555							 size);
 556		if (ret)
 557			return ret;
 558		if (total_added_ret)
 559			*total_added_ret += size;
 560	}
 561
 562	return 0;
 563}
 564
 565/*
 566 * Get an arbitrary extent item index / max_index through the block group
 567 *
 568 * @block_group   the block group to sample from
 569 * @index:        the integral step through the block group to grab from
 570 * @max_index:    the granularity of the sampling
 571 * @key:          return value parameter for the item we find
 572 *
 573 * Pre-conditions on indices:
 574 * 0 <= index <= max_index
 575 * 0 < max_index
 576 *
 577 * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
 578 * error code on error.
 579 */
 580static int sample_block_group_extent_item(struct btrfs_caching_control *caching_ctl,
 581					  struct btrfs_block_group *block_group,
 582					  int index, int max_index,
 583					  struct btrfs_key *found_key)
 584{
 585	struct btrfs_fs_info *fs_info = block_group->fs_info;
 586	struct btrfs_root *extent_root;
 587	u64 search_offset;
 588	u64 search_end = block_group->start + block_group->length;
 589	struct btrfs_path *path;
 590	struct btrfs_key search_key;
 591	int ret = 0;
 592
 593	ASSERT(index >= 0);
 594	ASSERT(index <= max_index);
 595	ASSERT(max_index > 0);
 596	lockdep_assert_held(&caching_ctl->mutex);
 597	lockdep_assert_held_read(&fs_info->commit_root_sem);
 598
 599	path = btrfs_alloc_path();
 600	if (!path)
 601		return -ENOMEM;
 602
 603	extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
 604						       BTRFS_SUPER_INFO_OFFSET));
 605
 606	path->skip_locking = 1;
 607	path->search_commit_root = 1;
 608	path->reada = READA_FORWARD;
 609
 610	search_offset = index * div_u64(block_group->length, max_index);
 611	search_key.objectid = block_group->start + search_offset;
 612	search_key.type = BTRFS_EXTENT_ITEM_KEY;
 613	search_key.offset = 0;
 614
 615	btrfs_for_each_slot(extent_root, &search_key, found_key, path, ret) {
 616		/* Success; sampled an extent item in the block group */
 617		if (found_key->type == BTRFS_EXTENT_ITEM_KEY &&
 618		    found_key->objectid >= block_group->start &&
 619		    found_key->objectid + found_key->offset <= search_end)
 620			break;
 621
 622		/* We can't possibly find a valid extent item anymore */
 623		if (found_key->objectid >= search_end) {
 624			ret = 1;
 625			break;
 626		}
 627	}
 628
 629	lockdep_assert_held(&caching_ctl->mutex);
 630	lockdep_assert_held_read(&fs_info->commit_root_sem);
 631	btrfs_free_path(path);
 632	return ret;
 633}
 634
 635/*
 636 * Best effort attempt to compute a block group's size class while caching it.
 637 *
 638 * @block_group: the block group we are caching
 639 *
 640 * We cannot infer the size class while adding free space extents, because that
 641 * logic doesn't care about contiguous file extents (it doesn't differentiate
 642 * between a 100M extent and 100 contiguous 1M extents). So we need to read the
 643 * file extent items. Reading all of them is quite wasteful, because usually
 644 * only a handful are enough to give a good answer. Therefore, we just grab 5 of
 645 * them at even steps through the block group and pick the smallest size class
 646 * we see. Since size class is best effort, and not guaranteed in general,
 647 * inaccuracy is acceptable.
 648 *
 649 * To be more explicit about why this algorithm makes sense:
 650 *
 651 * If we are caching in a block group from disk, then there are three major cases
 652 * to consider:
 653 * 1. the block group is well behaved and all extents in it are the same size
 654 *    class.
 655 * 2. the block group is mostly one size class with rare exceptions for last
 656 *    ditch allocations
 657 * 3. the block group was populated before size classes and can have a totally
 658 *    arbitrary mix of size classes.
 659 *
 660 * In case 1, looking at any extent in the block group will yield the correct
 661 * result. For the mixed cases, taking the minimum size class seems like a good
 662 * approximation, since gaps from frees will be usable to the size class. For
 663 * 2., a small handful of file extents is likely to yield the right answer. For
 664 * 3, we can either read every file extent, or admit that this is best effort
 665 * anyway and try to stay fast.
 666 *
 667 * Returns: 0 on success, negative error code on error.
 668 */
 669static int load_block_group_size_class(struct btrfs_caching_control *caching_ctl,
 670				       struct btrfs_block_group *block_group)
 671{
 672	struct btrfs_fs_info *fs_info = block_group->fs_info;
 673	struct btrfs_key key;
 674	int i;
 675	u64 min_size = block_group->length;
 676	enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
 677	int ret;
 678
 679	if (!btrfs_block_group_should_use_size_class(block_group))
 680		return 0;
 681
 682	lockdep_assert_held(&caching_ctl->mutex);
 683	lockdep_assert_held_read(&fs_info->commit_root_sem);
 684	for (i = 0; i < 5; ++i) {
 685		ret = sample_block_group_extent_item(caching_ctl, block_group, i, 5, &key);
 686		if (ret < 0)
 687			goto out;
 688		if (ret > 0)
 689			continue;
 690		min_size = min_t(u64, min_size, key.offset);
 691		size_class = btrfs_calc_block_group_size_class(min_size);
 692	}
 693	if (size_class != BTRFS_BG_SZ_NONE) {
 694		spin_lock(&block_group->lock);
 695		block_group->size_class = size_class;
 696		spin_unlock(&block_group->lock);
 697	}
 698out:
 699	return ret;
 700}
 701
 702static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 703{
 704	struct btrfs_block_group *block_group = caching_ctl->block_group;
 705	struct btrfs_fs_info *fs_info = block_group->fs_info;
 706	struct btrfs_root *extent_root;
 707	struct btrfs_path *path;
 708	struct extent_buffer *leaf;
 709	struct btrfs_key key;
 710	u64 total_found = 0;
 711	u64 last = 0;
 712	u32 nritems;
 713	int ret;
 714	bool wakeup = true;
 715
 716	path = btrfs_alloc_path();
 717	if (!path)
 718		return -ENOMEM;
 719
 720	last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
 721	extent_root = btrfs_extent_root(fs_info, last);
 722
 723#ifdef CONFIG_BTRFS_DEBUG
 724	/*
 725	 * If we're fragmenting we don't want to make anybody think we can
 726	 * allocate from this block group until we've had a chance to fragment
 727	 * the free space.
 728	 */
 729	if (btrfs_should_fragment_free_space(block_group))
 730		wakeup = false;
 731#endif
 732	/*
 733	 * We don't want to deadlock with somebody trying to allocate a new
 734	 * extent for the extent root while also trying to search the extent
 735	 * root to add free space.  So we skip locking and search the commit
 736	 * root, since its read-only
 737	 */
 738	path->skip_locking = 1;
 739	path->search_commit_root = 1;
 740	path->reada = READA_FORWARD;
 741
 742	key.objectid = last;
 743	key.offset = 0;
 744	key.type = BTRFS_EXTENT_ITEM_KEY;
 745
 746next:
 747	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 748	if (ret < 0)
 749		goto out;
 750
 751	leaf = path->nodes[0];
 752	nritems = btrfs_header_nritems(leaf);
 753
 754	while (1) {
 755		if (btrfs_fs_closing(fs_info) > 1) {
 756			last = (u64)-1;
 757			break;
 758		}
 759
 760		if (path->slots[0] < nritems) {
 761			btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 762		} else {
 763			ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
 764			if (ret)
 765				break;
 766
 767			if (need_resched() ||
 768			    rwsem_is_contended(&fs_info->commit_root_sem)) {
 769				btrfs_release_path(path);
 770				up_read(&fs_info->commit_root_sem);
 771				mutex_unlock(&caching_ctl->mutex);
 772				cond_resched();
 773				mutex_lock(&caching_ctl->mutex);
 774				down_read(&fs_info->commit_root_sem);
 775				goto next;
 776			}
 777
 778			ret = btrfs_next_leaf(extent_root, path);
 779			if (ret < 0)
 780				goto out;
 781			if (ret)
 782				break;
 783			leaf = path->nodes[0];
 784			nritems = btrfs_header_nritems(leaf);
 785			continue;
 786		}
 787
 788		if (key.objectid < last) {
 789			key.objectid = last;
 790			key.offset = 0;
 791			key.type = BTRFS_EXTENT_ITEM_KEY;
 792			btrfs_release_path(path);
 793			goto next;
 794		}
 795
 796		if (key.objectid < block_group->start) {
 797			path->slots[0]++;
 798			continue;
 799		}
 800
 801		if (key.objectid >= block_group->start + block_group->length)
 802			break;
 803
 804		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 805		    key.type == BTRFS_METADATA_ITEM_KEY) {
 806			u64 space_added;
 807
 808			ret = btrfs_add_new_free_space(block_group, last,
 809						       key.objectid, &space_added);
 810			if (ret)
 811				goto out;
 812			total_found += space_added;
 813			if (key.type == BTRFS_METADATA_ITEM_KEY)
 814				last = key.objectid +
 815					fs_info->nodesize;
 816			else
 817				last = key.objectid + key.offset;
 818
 819			if (total_found > CACHING_CTL_WAKE_UP) {
 820				total_found = 0;
 821				if (wakeup) {
 822					atomic_inc(&caching_ctl->progress);
 823					wake_up(&caching_ctl->wait);
 824				}
 825			}
 826		}
 827		path->slots[0]++;
 828	}
 829
 830	ret = btrfs_add_new_free_space(block_group, last,
 831				       block_group->start + block_group->length,
 832				       NULL);
 833out:
 834	btrfs_free_path(path);
 835	return ret;
 836}
 837
 838static inline void btrfs_free_excluded_extents(const struct btrfs_block_group *bg)
 839{
 840	clear_extent_bits(&bg->fs_info->excluded_extents, bg->start,
 841			  bg->start + bg->length - 1, EXTENT_UPTODATE);
 842}
 843
 844static noinline void caching_thread(struct btrfs_work *work)
 845{
 846	struct btrfs_block_group *block_group;
 847	struct btrfs_fs_info *fs_info;
 848	struct btrfs_caching_control *caching_ctl;
 849	int ret;
 850
 851	caching_ctl = container_of(work, struct btrfs_caching_control, work);
 852	block_group = caching_ctl->block_group;
 853	fs_info = block_group->fs_info;
 854
 855	mutex_lock(&caching_ctl->mutex);
 856	down_read(&fs_info->commit_root_sem);
 857
 858	load_block_group_size_class(caching_ctl, block_group);
 859	if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 860		ret = load_free_space_cache(block_group);
 861		if (ret == 1) {
 862			ret = 0;
 863			goto done;
 864		}
 865
 866		/*
 867		 * We failed to load the space cache, set ourselves to
 868		 * CACHE_STARTED and carry on.
 869		 */
 870		spin_lock(&block_group->lock);
 871		block_group->cached = BTRFS_CACHE_STARTED;
 872		spin_unlock(&block_group->lock);
 873		wake_up(&caching_ctl->wait);
 874	}
 875
 876	/*
 877	 * If we are in the transaction that populated the free space tree we
 878	 * can't actually cache from the free space tree as our commit root and
 879	 * real root are the same, so we could change the contents of the blocks
 880	 * while caching.  Instead do the slow caching in this case, and after
 881	 * the transaction has committed we will be safe.
 882	 */
 883	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
 884	    !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
 885		ret = load_free_space_tree(caching_ctl);
 886	else
 887		ret = load_extent_tree_free(caching_ctl);
 888done:
 889	spin_lock(&block_group->lock);
 890	block_group->caching_ctl = NULL;
 891	block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 892	spin_unlock(&block_group->lock);
 893
 894#ifdef CONFIG_BTRFS_DEBUG
 895	if (btrfs_should_fragment_free_space(block_group)) {
 896		u64 bytes_used;
 897
 898		spin_lock(&block_group->space_info->lock);
 899		spin_lock(&block_group->lock);
 900		bytes_used = block_group->length - block_group->used;
 901		block_group->space_info->bytes_used += bytes_used >> 1;
 902		spin_unlock(&block_group->lock);
 903		spin_unlock(&block_group->space_info->lock);
 904		fragment_free_space(block_group);
 905	}
 906#endif
 907
 908	up_read(&fs_info->commit_root_sem);
 909	btrfs_free_excluded_extents(block_group);
 910	mutex_unlock(&caching_ctl->mutex);
 911
 912	wake_up(&caching_ctl->wait);
 913
 914	btrfs_put_caching_control(caching_ctl);
 915	btrfs_put_block_group(block_group);
 916}
 917
 918int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
 919{
 920	struct btrfs_fs_info *fs_info = cache->fs_info;
 921	struct btrfs_caching_control *caching_ctl = NULL;
 922	int ret = 0;
 923
 924	/* Allocator for zoned filesystems does not use the cache at all */
 925	if (btrfs_is_zoned(fs_info))
 926		return 0;
 927
 928	caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 929	if (!caching_ctl)
 930		return -ENOMEM;
 931
 932	INIT_LIST_HEAD(&caching_ctl->list);
 933	mutex_init(&caching_ctl->mutex);
 934	init_waitqueue_head(&caching_ctl->wait);
 935	caching_ctl->block_group = cache;
 936	refcount_set(&caching_ctl->count, 2);
 937	atomic_set(&caching_ctl->progress, 0);
 938	btrfs_init_work(&caching_ctl->work, caching_thread, NULL);
 939
 940	spin_lock(&cache->lock);
 941	if (cache->cached != BTRFS_CACHE_NO) {
 942		kfree(caching_ctl);
 943
 944		caching_ctl = cache->caching_ctl;
 945		if (caching_ctl)
 946			refcount_inc(&caching_ctl->count);
 947		spin_unlock(&cache->lock);
 948		goto out;
 949	}
 950	WARN_ON(cache->caching_ctl);
 951	cache->caching_ctl = caching_ctl;
 952	cache->cached = BTRFS_CACHE_STARTED;
 953	spin_unlock(&cache->lock);
 954
 955	write_lock(&fs_info->block_group_cache_lock);
 956	refcount_inc(&caching_ctl->count);
 957	list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 958	write_unlock(&fs_info->block_group_cache_lock);
 959
 960	btrfs_get_block_group(cache);
 961
 962	btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 963out:
 964	if (wait && caching_ctl)
 965		ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
 966	if (caching_ctl)
 967		btrfs_put_caching_control(caching_ctl);
 968
 969	return ret;
 970}
 971
 972static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
 973{
 974	u64 extra_flags = chunk_to_extended(flags) &
 975				BTRFS_EXTENDED_PROFILE_MASK;
 976
 977	write_seqlock(&fs_info->profiles_lock);
 978	if (flags & BTRFS_BLOCK_GROUP_DATA)
 979		fs_info->avail_data_alloc_bits &= ~extra_flags;
 980	if (flags & BTRFS_BLOCK_GROUP_METADATA)
 981		fs_info->avail_metadata_alloc_bits &= ~extra_flags;
 982	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 983		fs_info->avail_system_alloc_bits &= ~extra_flags;
 984	write_sequnlock(&fs_info->profiles_lock);
 985}
 986
 987/*
 988 * Clear incompat bits for the following feature(s):
 989 *
 990 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
 991 *            in the whole filesystem
 992 *
 993 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
 994 */
 995static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
 996{
 997	bool found_raid56 = false;
 998	bool found_raid1c34 = false;
 999
1000	if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
1001	    (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
1002	    (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
1003		struct list_head *head = &fs_info->space_info;
1004		struct btrfs_space_info *sinfo;
1005
1006		list_for_each_entry_rcu(sinfo, head, list) {
1007			down_read(&sinfo->groups_sem);
1008			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
1009				found_raid56 = true;
1010			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
1011				found_raid56 = true;
1012			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
1013				found_raid1c34 = true;
1014			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
1015				found_raid1c34 = true;
1016			up_read(&sinfo->groups_sem);
1017		}
1018		if (!found_raid56)
1019			btrfs_clear_fs_incompat(fs_info, RAID56);
1020		if (!found_raid1c34)
1021			btrfs_clear_fs_incompat(fs_info, RAID1C34);
1022	}
1023}
1024
1025static int remove_block_group_item(struct btrfs_trans_handle *trans,
1026				   struct btrfs_path *path,
1027				   struct btrfs_block_group *block_group)
1028{
1029	struct btrfs_fs_info *fs_info = trans->fs_info;
1030	struct btrfs_root *root;
1031	struct btrfs_key key;
1032	int ret;
1033
1034	root = btrfs_block_group_root(fs_info);
1035	key.objectid = block_group->start;
1036	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1037	key.offset = block_group->length;
1038
1039	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1040	if (ret > 0)
1041		ret = -ENOENT;
1042	if (ret < 0)
1043		return ret;
1044
1045	ret = btrfs_del_item(trans, root, path);
1046	return ret;
1047}
1048
1049int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1050			     struct btrfs_chunk_map *map)
1051{
1052	struct btrfs_fs_info *fs_info = trans->fs_info;
1053	struct btrfs_path *path;
1054	struct btrfs_block_group *block_group;
1055	struct btrfs_free_cluster *cluster;
1056	struct inode *inode;
1057	struct kobject *kobj = NULL;
1058	int ret;
1059	int index;
1060	int factor;
1061	struct btrfs_caching_control *caching_ctl = NULL;
1062	bool remove_map;
1063	bool remove_rsv = false;
1064
1065	block_group = btrfs_lookup_block_group(fs_info, map->start);
1066	if (!block_group)
1067		return -ENOENT;
1068
1069	BUG_ON(!block_group->ro);
1070
1071	trace_btrfs_remove_block_group(block_group);
1072	/*
1073	 * Free the reserved super bytes from this block group before
1074	 * remove it.
1075	 */
1076	btrfs_free_excluded_extents(block_group);
1077	btrfs_free_ref_tree_range(fs_info, block_group->start,
1078				  block_group->length);
1079
1080	index = btrfs_bg_flags_to_raid_index(block_group->flags);
1081	factor = btrfs_bg_type_to_factor(block_group->flags);
1082
1083	/* make sure this block group isn't part of an allocation cluster */
1084	cluster = &fs_info->data_alloc_cluster;
1085	spin_lock(&cluster->refill_lock);
1086	btrfs_return_cluster_to_free_space(block_group, cluster);
1087	spin_unlock(&cluster->refill_lock);
1088
1089	/*
1090	 * make sure this block group isn't part of a metadata
1091	 * allocation cluster
1092	 */
1093	cluster = &fs_info->meta_alloc_cluster;
1094	spin_lock(&cluster->refill_lock);
1095	btrfs_return_cluster_to_free_space(block_group, cluster);
1096	spin_unlock(&cluster->refill_lock);
1097
1098	btrfs_clear_treelog_bg(block_group);
1099	btrfs_clear_data_reloc_bg(block_group);
1100
1101	path = btrfs_alloc_path();
1102	if (!path) {
1103		ret = -ENOMEM;
1104		goto out;
1105	}
1106
1107	/*
1108	 * get the inode first so any iput calls done for the io_list
1109	 * aren't the final iput (no unlinks allowed now)
1110	 */
1111	inode = lookup_free_space_inode(block_group, path);
1112
1113	mutex_lock(&trans->transaction->cache_write_mutex);
1114	/*
1115	 * Make sure our free space cache IO is done before removing the
1116	 * free space inode
1117	 */
1118	spin_lock(&trans->transaction->dirty_bgs_lock);
1119	if (!list_empty(&block_group->io_list)) {
1120		list_del_init(&block_group->io_list);
1121
1122		WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1123
1124		spin_unlock(&trans->transaction->dirty_bgs_lock);
1125		btrfs_wait_cache_io(trans, block_group, path);
1126		btrfs_put_block_group(block_group);
1127		spin_lock(&trans->transaction->dirty_bgs_lock);
1128	}
1129
1130	if (!list_empty(&block_group->dirty_list)) {
1131		list_del_init(&block_group->dirty_list);
1132		remove_rsv = true;
1133		btrfs_put_block_group(block_group);
1134	}
1135	spin_unlock(&trans->transaction->dirty_bgs_lock);
1136	mutex_unlock(&trans->transaction->cache_write_mutex);
1137
1138	ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1139	if (ret)
1140		goto out;
1141
1142	write_lock(&fs_info->block_group_cache_lock);
1143	rb_erase_cached(&block_group->cache_node,
1144			&fs_info->block_group_cache_tree);
1145	RB_CLEAR_NODE(&block_group->cache_node);
1146
1147	/* Once for the block groups rbtree */
1148	btrfs_put_block_group(block_group);
1149
1150	write_unlock(&fs_info->block_group_cache_lock);
1151
1152	down_write(&block_group->space_info->groups_sem);
1153	/*
1154	 * we must use list_del_init so people can check to see if they
1155	 * are still on the list after taking the semaphore
1156	 */
1157	list_del_init(&block_group->list);
1158	if (list_empty(&block_group->space_info->block_groups[index])) {
1159		kobj = block_group->space_info->block_group_kobjs[index];
1160		block_group->space_info->block_group_kobjs[index] = NULL;
1161		clear_avail_alloc_bits(fs_info, block_group->flags);
1162	}
1163	up_write(&block_group->space_info->groups_sem);
1164	clear_incompat_bg_bits(fs_info, block_group->flags);
1165	if (kobj) {
1166		kobject_del(kobj);
1167		kobject_put(kobj);
1168	}
1169
1170	if (block_group->cached == BTRFS_CACHE_STARTED)
1171		btrfs_wait_block_group_cache_done(block_group);
1172
1173	write_lock(&fs_info->block_group_cache_lock);
1174	caching_ctl = btrfs_get_caching_control(block_group);
1175	if (!caching_ctl) {
1176		struct btrfs_caching_control *ctl;
1177
1178		list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1179			if (ctl->block_group == block_group) {
1180				caching_ctl = ctl;
1181				refcount_inc(&caching_ctl->count);
1182				break;
1183			}
1184		}
1185	}
1186	if (caching_ctl)
1187		list_del_init(&caching_ctl->list);
1188	write_unlock(&fs_info->block_group_cache_lock);
1189
1190	if (caching_ctl) {
1191		/* Once for the caching bgs list and once for us. */
1192		btrfs_put_caching_control(caching_ctl);
1193		btrfs_put_caching_control(caching_ctl);
1194	}
1195
1196	spin_lock(&trans->transaction->dirty_bgs_lock);
1197	WARN_ON(!list_empty(&block_group->dirty_list));
1198	WARN_ON(!list_empty(&block_group->io_list));
1199	spin_unlock(&trans->transaction->dirty_bgs_lock);
1200
1201	btrfs_remove_free_space_cache(block_group);
1202
1203	spin_lock(&block_group->space_info->lock);
1204	list_del_init(&block_group->ro_list);
1205
1206	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1207		WARN_ON(block_group->space_info->total_bytes
1208			< block_group->length);
1209		WARN_ON(block_group->space_info->bytes_readonly
1210			< block_group->length - block_group->zone_unusable);
1211		WARN_ON(block_group->space_info->bytes_zone_unusable
1212			< block_group->zone_unusable);
1213		WARN_ON(block_group->space_info->disk_total
1214			< block_group->length * factor);
1215	}
1216	block_group->space_info->total_bytes -= block_group->length;
1217	block_group->space_info->bytes_readonly -=
1218		(block_group->length - block_group->zone_unusable);
1219	block_group->space_info->bytes_zone_unusable -=
1220		block_group->zone_unusable;
1221	block_group->space_info->disk_total -= block_group->length * factor;
1222
1223	spin_unlock(&block_group->space_info->lock);
1224
1225	/*
1226	 * Remove the free space for the block group from the free space tree
1227	 * and the block group's item from the extent tree before marking the
1228	 * block group as removed. This is to prevent races with tasks that
1229	 * freeze and unfreeze a block group, this task and another task
1230	 * allocating a new block group - the unfreeze task ends up removing
1231	 * the block group's extent map before the task calling this function
1232	 * deletes the block group item from the extent tree, allowing for
1233	 * another task to attempt to create another block group with the same
1234	 * item key (and failing with -EEXIST and a transaction abort).
1235	 */
1236	ret = remove_block_group_free_space(trans, block_group);
1237	if (ret)
1238		goto out;
1239
1240	ret = remove_block_group_item(trans, path, block_group);
1241	if (ret < 0)
1242		goto out;
1243
1244	spin_lock(&block_group->lock);
1245	set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1246
1247	/*
1248	 * At this point trimming or scrub can't start on this block group,
1249	 * because we removed the block group from the rbtree
1250	 * fs_info->block_group_cache_tree so no one can't find it anymore and
1251	 * even if someone already got this block group before we removed it
1252	 * from the rbtree, they have already incremented block_group->frozen -
1253	 * if they didn't, for the trimming case they won't find any free space
1254	 * entries because we already removed them all when we called
1255	 * btrfs_remove_free_space_cache().
1256	 *
1257	 * And we must not remove the chunk map from the fs_info->mapping_tree
1258	 * to prevent the same logical address range and physical device space
1259	 * ranges from being reused for a new block group. This is needed to
1260	 * avoid races with trimming and scrub.
1261	 *
1262	 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1263	 * completely transactionless, so while it is trimming a range the
1264	 * currently running transaction might finish and a new one start,
1265	 * allowing for new block groups to be created that can reuse the same
1266	 * physical device locations unless we take this special care.
1267	 *
1268	 * There may also be an implicit trim operation if the file system
1269	 * is mounted with -odiscard. The same protections must remain
1270	 * in place until the extents have been discarded completely when
1271	 * the transaction commit has completed.
1272	 */
1273	remove_map = (atomic_read(&block_group->frozen) == 0);
1274	spin_unlock(&block_group->lock);
1275
1276	if (remove_map)
1277		btrfs_remove_chunk_map(fs_info, map);
1278
1279out:
1280	/* Once for the lookup reference */
1281	btrfs_put_block_group(block_group);
1282	if (remove_rsv)
1283		btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
1284	btrfs_free_path(path);
1285	return ret;
1286}
1287
1288struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1289		struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1290{
1291	struct btrfs_root *root = btrfs_block_group_root(fs_info);
1292	struct btrfs_chunk_map *map;
1293	unsigned int num_items;
1294
1295	map = btrfs_find_chunk_map(fs_info, chunk_offset, 1);
1296	ASSERT(map != NULL);
1297	ASSERT(map->start == chunk_offset);
1298
1299	/*
1300	 * We need to reserve 3 + N units from the metadata space info in order
1301	 * to remove a block group (done at btrfs_remove_chunk() and at
1302	 * btrfs_remove_block_group()), which are used for:
1303	 *
1304	 * 1 unit for adding the free space inode's orphan (located in the tree
1305	 * of tree roots).
1306	 * 1 unit for deleting the block group item (located in the extent
1307	 * tree).
1308	 * 1 unit for deleting the free space item (located in tree of tree
1309	 * roots).
1310	 * N units for deleting N device extent items corresponding to each
1311	 * stripe (located in the device tree).
1312	 *
1313	 * In order to remove a block group we also need to reserve units in the
1314	 * system space info in order to update the chunk tree (update one or
1315	 * more device items and remove one chunk item), but this is done at
1316	 * btrfs_remove_chunk() through a call to check_system_chunk().
1317	 */
1318	num_items = 3 + map->num_stripes;
1319	btrfs_free_chunk_map(map);
1320
1321	return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1322}
1323
1324/*
1325 * Mark block group @cache read-only, so later write won't happen to block
1326 * group @cache.
1327 *
1328 * If @force is not set, this function will only mark the block group readonly
1329 * if we have enough free space (1M) in other metadata/system block groups.
1330 * If @force is not set, this function will mark the block group readonly
1331 * without checking free space.
1332 *
1333 * NOTE: This function doesn't care if other block groups can contain all the
1334 * data in this block group. That check should be done by relocation routine,
1335 * not this function.
1336 */
1337static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1338{
1339	struct btrfs_space_info *sinfo = cache->space_info;
1340	u64 num_bytes;
1341	int ret = -ENOSPC;
1342
1343	spin_lock(&sinfo->lock);
1344	spin_lock(&cache->lock);
1345
1346	if (cache->swap_extents) {
1347		ret = -ETXTBSY;
1348		goto out;
1349	}
1350
1351	if (cache->ro) {
1352		cache->ro++;
1353		ret = 0;
1354		goto out;
1355	}
1356
1357	num_bytes = cache->length - cache->reserved - cache->pinned -
1358		    cache->bytes_super - cache->zone_unusable - cache->used;
1359
1360	/*
1361	 * Data never overcommits, even in mixed mode, so do just the straight
1362	 * check of left over space in how much we have allocated.
1363	 */
1364	if (force) {
1365		ret = 0;
1366	} else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1367		u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1368
1369		/*
1370		 * Here we make sure if we mark this bg RO, we still have enough
1371		 * free space as buffer.
1372		 */
1373		if (sinfo_used + num_bytes <= sinfo->total_bytes)
1374			ret = 0;
1375	} else {
1376		/*
1377		 * We overcommit metadata, so we need to do the
1378		 * btrfs_can_overcommit check here, and we need to pass in
1379		 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1380		 * leeway to allow us to mark this block group as read only.
1381		 */
1382		if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1383					 BTRFS_RESERVE_NO_FLUSH))
1384			ret = 0;
1385	}
1386
1387	if (!ret) {
1388		sinfo->bytes_readonly += num_bytes;
1389		if (btrfs_is_zoned(cache->fs_info)) {
1390			/* Migrate zone_unusable bytes to readonly */
1391			sinfo->bytes_readonly += cache->zone_unusable;
1392			sinfo->bytes_zone_unusable -= cache->zone_unusable;
1393			cache->zone_unusable = 0;
1394		}
1395		cache->ro++;
1396		list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1397	}
1398out:
1399	spin_unlock(&cache->lock);
1400	spin_unlock(&sinfo->lock);
1401	if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1402		btrfs_info(cache->fs_info,
1403			"unable to make block group %llu ro", cache->start);
1404		btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1405	}
1406	return ret;
1407}
1408
1409static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1410				 struct btrfs_block_group *bg)
1411{
1412	struct btrfs_fs_info *fs_info = bg->fs_info;
1413	struct btrfs_transaction *prev_trans = NULL;
1414	const u64 start = bg->start;
1415	const u64 end = start + bg->length - 1;
1416	int ret;
1417
1418	spin_lock(&fs_info->trans_lock);
1419	if (trans->transaction->list.prev != &fs_info->trans_list) {
1420		prev_trans = list_last_entry(&trans->transaction->list,
1421					     struct btrfs_transaction, list);
1422		refcount_inc(&prev_trans->use_count);
1423	}
1424	spin_unlock(&fs_info->trans_lock);
1425
1426	/*
1427	 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1428	 * btrfs_finish_extent_commit(). If we are at transaction N, another
1429	 * task might be running finish_extent_commit() for the previous
1430	 * transaction N - 1, and have seen a range belonging to the block
1431	 * group in pinned_extents before we were able to clear the whole block
1432	 * group range from pinned_extents. This means that task can lookup for
1433	 * the block group after we unpinned it from pinned_extents and removed
1434	 * it, leading to an error at unpin_extent_range().
1435	 */
1436	mutex_lock(&fs_info->unused_bg_unpin_mutex);
1437	if (prev_trans) {
1438		ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1439					EXTENT_DIRTY);
1440		if (ret)
1441			goto out;
1442	}
1443
1444	ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1445				EXTENT_DIRTY);
1446out:
1447	mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1448	if (prev_trans)
1449		btrfs_put_transaction(prev_trans);
1450
1451	return ret == 0;
1452}
1453
1454/*
1455 * Process the unused_bgs list and remove any that don't have any allocated
1456 * space inside of them.
1457 */
1458void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1459{
1460	LIST_HEAD(retry_list);
1461	struct btrfs_block_group *block_group;
1462	struct btrfs_space_info *space_info;
1463	struct btrfs_trans_handle *trans;
1464	const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1465	int ret = 0;
1466
1467	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1468		return;
1469
1470	if (btrfs_fs_closing(fs_info))
1471		return;
1472
1473	/*
1474	 * Long running balances can keep us blocked here for eternity, so
1475	 * simply skip deletion if we're unable to get the mutex.
1476	 */
1477	if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1478		return;
1479
1480	spin_lock(&fs_info->unused_bgs_lock);
1481	while (!list_empty(&fs_info->unused_bgs)) {
1482		u64 used;
1483		int trimming;
1484
1485		block_group = list_first_entry(&fs_info->unused_bgs,
1486					       struct btrfs_block_group,
1487					       bg_list);
1488		list_del_init(&block_group->bg_list);
1489
1490		space_info = block_group->space_info;
1491
1492		if (ret || btrfs_mixed_space_info(space_info)) {
1493			btrfs_put_block_group(block_group);
1494			continue;
1495		}
1496		spin_unlock(&fs_info->unused_bgs_lock);
1497
1498		btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1499
1500		/* Don't want to race with allocators so take the groups_sem */
1501		down_write(&space_info->groups_sem);
1502
1503		/*
1504		 * Async discard moves the final block group discard to be prior
1505		 * to the unused_bgs code path.  Therefore, if it's not fully
1506		 * trimmed, punt it back to the async discard lists.
1507		 */
1508		if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1509		    !btrfs_is_free_space_trimmed(block_group)) {
1510			trace_btrfs_skip_unused_block_group(block_group);
1511			up_write(&space_info->groups_sem);
1512			/* Requeue if we failed because of async discard */
1513			btrfs_discard_queue_work(&fs_info->discard_ctl,
1514						 block_group);
1515			goto next;
1516		}
1517
1518		spin_lock(&space_info->lock);
1519		spin_lock(&block_group->lock);
1520		if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1521		    list_is_singular(&block_group->list)) {
1522			/*
1523			 * We want to bail if we made new allocations or have
1524			 * outstanding allocations in this block group.  We do
1525			 * the ro check in case balance is currently acting on
1526			 * this block group.
1527			 *
1528			 * Also bail out if this is the only block group for its
1529			 * type, because otherwise we would lose profile
1530			 * information from fs_info->avail_*_alloc_bits and the
1531			 * next block group of this type would be created with a
1532			 * "single" profile (even if we're in a raid fs) because
1533			 * fs_info->avail_*_alloc_bits would be 0.
1534			 */
1535			trace_btrfs_skip_unused_block_group(block_group);
1536			spin_unlock(&block_group->lock);
1537			spin_unlock(&space_info->lock);
1538			up_write(&space_info->groups_sem);
1539			goto next;
1540		}
1541
1542		/*
1543		 * The block group may be unused but there may be space reserved
1544		 * accounting with the existence of that block group, that is,
1545		 * space_info->bytes_may_use was incremented by a task but no
1546		 * space was yet allocated from the block group by the task.
1547		 * That space may or may not be allocated, as we are generally
1548		 * pessimistic about space reservation for metadata as well as
1549		 * for data when using compression (as we reserve space based on
1550		 * the worst case, when data can't be compressed, and before
1551		 * actually attempting compression, before starting writeback).
1552		 *
1553		 * So check if the total space of the space_info minus the size
1554		 * of this block group is less than the used space of the
1555		 * space_info - if that's the case, then it means we have tasks
1556		 * that might be relying on the block group in order to allocate
1557		 * extents, and add back the block group to the unused list when
1558		 * we finish, so that we retry later in case no tasks ended up
1559		 * needing to allocate extents from the block group.
1560		 */
1561		used = btrfs_space_info_used(space_info, true);
1562		if (space_info->total_bytes - block_group->length < used &&
1563		    block_group->zone_unusable < block_group->length) {
1564			/*
1565			 * Add a reference for the list, compensate for the ref
1566			 * drop under the "next" label for the
1567			 * fs_info->unused_bgs list.
1568			 */
1569			btrfs_get_block_group(block_group);
1570			list_add_tail(&block_group->bg_list, &retry_list);
1571
1572			trace_btrfs_skip_unused_block_group(block_group);
1573			spin_unlock(&block_group->lock);
1574			spin_unlock(&space_info->lock);
1575			up_write(&space_info->groups_sem);
1576			goto next;
1577		}
1578
1579		spin_unlock(&block_group->lock);
1580		spin_unlock(&space_info->lock);
1581
1582		/* We don't want to force the issue, only flip if it's ok. */
1583		ret = inc_block_group_ro(block_group, 0);
1584		up_write(&space_info->groups_sem);
1585		if (ret < 0) {
1586			ret = 0;
1587			goto next;
1588		}
1589
1590		ret = btrfs_zone_finish(block_group);
1591		if (ret < 0) {
1592			btrfs_dec_block_group_ro(block_group);
1593			if (ret == -EAGAIN)
1594				ret = 0;
1595			goto next;
1596		}
1597
1598		/*
1599		 * Want to do this before we do anything else so we can recover
1600		 * properly if we fail to join the transaction.
1601		 */
1602		trans = btrfs_start_trans_remove_block_group(fs_info,
1603						     block_group->start);
1604		if (IS_ERR(trans)) {
1605			btrfs_dec_block_group_ro(block_group);
1606			ret = PTR_ERR(trans);
1607			goto next;
1608		}
1609
1610		/*
1611		 * We could have pending pinned extents for this block group,
1612		 * just delete them, we don't care about them anymore.
1613		 */
1614		if (!clean_pinned_extents(trans, block_group)) {
1615			btrfs_dec_block_group_ro(block_group);
1616			goto end_trans;
1617		}
1618
1619		/*
1620		 * At this point, the block_group is read only and should fail
1621		 * new allocations.  However, btrfs_finish_extent_commit() can
1622		 * cause this block_group to be placed back on the discard
1623		 * lists because now the block_group isn't fully discarded.
1624		 * Bail here and try again later after discarding everything.
1625		 */
1626		spin_lock(&fs_info->discard_ctl.lock);
1627		if (!list_empty(&block_group->discard_list)) {
1628			spin_unlock(&fs_info->discard_ctl.lock);
1629			btrfs_dec_block_group_ro(block_group);
1630			btrfs_discard_queue_work(&fs_info->discard_ctl,
1631						 block_group);
1632			goto end_trans;
1633		}
1634		spin_unlock(&fs_info->discard_ctl.lock);
1635
1636		/* Reset pinned so btrfs_put_block_group doesn't complain */
1637		spin_lock(&space_info->lock);
1638		spin_lock(&block_group->lock);
1639
1640		btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1641						     -block_group->pinned);
1642		space_info->bytes_readonly += block_group->pinned;
1643		block_group->pinned = 0;
1644
1645		spin_unlock(&block_group->lock);
1646		spin_unlock(&space_info->lock);
1647
1648		/*
1649		 * The normal path here is an unused block group is passed here,
1650		 * then trimming is handled in the transaction commit path.
1651		 * Async discard interposes before this to do the trimming
1652		 * before coming down the unused block group path as trimming
1653		 * will no longer be done later in the transaction commit path.
1654		 */
1655		if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1656			goto flip_async;
1657
1658		/*
1659		 * DISCARD can flip during remount. On zoned filesystems, we
1660		 * need to reset sequential-required zones.
1661		 */
1662		trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1663				btrfs_is_zoned(fs_info);
1664
1665		/* Implicit trim during transaction commit. */
1666		if (trimming)
1667			btrfs_freeze_block_group(block_group);
1668
1669		/*
1670		 * Btrfs_remove_chunk will abort the transaction if things go
1671		 * horribly wrong.
1672		 */
1673		ret = btrfs_remove_chunk(trans, block_group->start);
1674
1675		if (ret) {
1676			if (trimming)
1677				btrfs_unfreeze_block_group(block_group);
1678			goto end_trans;
1679		}
1680
1681		/*
1682		 * If we're not mounted with -odiscard, we can just forget
1683		 * about this block group. Otherwise we'll need to wait
1684		 * until transaction commit to do the actual discard.
1685		 */
1686		if (trimming) {
1687			spin_lock(&fs_info->unused_bgs_lock);
1688			/*
1689			 * A concurrent scrub might have added us to the list
1690			 * fs_info->unused_bgs, so use a list_move operation
1691			 * to add the block group to the deleted_bgs list.
1692			 */
1693			list_move(&block_group->bg_list,
1694				  &trans->transaction->deleted_bgs);
1695			spin_unlock(&fs_info->unused_bgs_lock);
1696			btrfs_get_block_group(block_group);
1697		}
1698end_trans:
1699		btrfs_end_transaction(trans);
1700next:
1701		btrfs_put_block_group(block_group);
1702		spin_lock(&fs_info->unused_bgs_lock);
1703	}
1704	list_splice_tail(&retry_list, &fs_info->unused_bgs);
1705	spin_unlock(&fs_info->unused_bgs_lock);
1706	mutex_unlock(&fs_info->reclaim_bgs_lock);
1707	return;
1708
1709flip_async:
1710	btrfs_end_transaction(trans);
1711	spin_lock(&fs_info->unused_bgs_lock);
1712	list_splice_tail(&retry_list, &fs_info->unused_bgs);
1713	spin_unlock(&fs_info->unused_bgs_lock);
1714	mutex_unlock(&fs_info->reclaim_bgs_lock);
1715	btrfs_put_block_group(block_group);
1716	btrfs_discard_punt_unused_bgs_list(fs_info);
1717}
1718
1719void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1720{
1721	struct btrfs_fs_info *fs_info = bg->fs_info;
1722
1723	spin_lock(&fs_info->unused_bgs_lock);
1724	if (list_empty(&bg->bg_list)) {
1725		btrfs_get_block_group(bg);
1726		trace_btrfs_add_unused_block_group(bg);
1727		list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1728	} else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1729		/* Pull out the block group from the reclaim_bgs list. */
1730		trace_btrfs_add_unused_block_group(bg);
1731		list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1732	}
1733	spin_unlock(&fs_info->unused_bgs_lock);
1734}
1735
1736/*
1737 * We want block groups with a low number of used bytes to be in the beginning
1738 * of the list, so they will get reclaimed first.
1739 */
1740static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1741			   const struct list_head *b)
1742{
1743	const struct btrfs_block_group *bg1, *bg2;
1744
1745	bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1746	bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1747
1748	return bg1->used > bg2->used;
1749}
1750
1751static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1752{
1753	if (btrfs_is_zoned(fs_info))
1754		return btrfs_zoned_should_reclaim(fs_info);
1755	return true;
1756}
1757
1758static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1759{
1760	const struct btrfs_space_info *space_info = bg->space_info;
1761	const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1762	const u64 new_val = bg->used;
1763	const u64 old_val = new_val + bytes_freed;
1764	u64 thresh;
1765
1766	if (reclaim_thresh == 0)
1767		return false;
1768
1769	thresh = mult_perc(bg->length, reclaim_thresh);
1770
1771	/*
1772	 * If we were below the threshold before don't reclaim, we are likely a
1773	 * brand new block group and we don't want to relocate new block groups.
1774	 */
1775	if (old_val < thresh)
1776		return false;
1777	if (new_val >= thresh)
1778		return false;
1779	return true;
1780}
1781
1782void btrfs_reclaim_bgs_work(struct work_struct *work)
1783{
1784	struct btrfs_fs_info *fs_info =
1785		container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1786	struct btrfs_block_group *bg;
1787	struct btrfs_space_info *space_info;
1788
1789	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1790		return;
1791
1792	if (btrfs_fs_closing(fs_info))
1793		return;
1794
1795	if (!btrfs_should_reclaim(fs_info))
1796		return;
1797
1798	sb_start_write(fs_info->sb);
1799
1800	if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1801		sb_end_write(fs_info->sb);
1802		return;
1803	}
1804
1805	/*
1806	 * Long running balances can keep us blocked here for eternity, so
1807	 * simply skip reclaim if we're unable to get the mutex.
1808	 */
1809	if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1810		btrfs_exclop_finish(fs_info);
1811		sb_end_write(fs_info->sb);
1812		return;
1813	}
1814
1815	spin_lock(&fs_info->unused_bgs_lock);
1816	/*
1817	 * Sort happens under lock because we can't simply splice it and sort.
1818	 * The block groups might still be in use and reachable via bg_list,
1819	 * and their presence in the reclaim_bgs list must be preserved.
1820	 */
1821	list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1822	while (!list_empty(&fs_info->reclaim_bgs)) {
1823		u64 zone_unusable;
1824		int ret = 0;
1825
1826		bg = list_first_entry(&fs_info->reclaim_bgs,
1827				      struct btrfs_block_group,
1828				      bg_list);
1829		list_del_init(&bg->bg_list);
1830
1831		space_info = bg->space_info;
1832		spin_unlock(&fs_info->unused_bgs_lock);
1833
1834		/* Don't race with allocators so take the groups_sem */
1835		down_write(&space_info->groups_sem);
1836
1837		spin_lock(&bg->lock);
1838		if (bg->reserved || bg->pinned || bg->ro) {
1839			/*
1840			 * We want to bail if we made new allocations or have
1841			 * outstanding allocations in this block group.  We do
1842			 * the ro check in case balance is currently acting on
1843			 * this block group.
1844			 */
1845			spin_unlock(&bg->lock);
1846			up_write(&space_info->groups_sem);
1847			goto next;
1848		}
1849		if (bg->used == 0) {
1850			/*
1851			 * It is possible that we trigger relocation on a block
1852			 * group as its extents are deleted and it first goes
1853			 * below the threshold, then shortly after goes empty.
1854			 *
1855			 * In this case, relocating it does delete it, but has
1856			 * some overhead in relocation specific metadata, looking
1857			 * for the non-existent extents and running some extra
1858			 * transactions, which we can avoid by using one of the
1859			 * other mechanisms for dealing with empty block groups.
1860			 */
1861			if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1862				btrfs_mark_bg_unused(bg);
1863			spin_unlock(&bg->lock);
1864			up_write(&space_info->groups_sem);
1865			goto next;
1866
1867		}
1868		/*
1869		 * The block group might no longer meet the reclaim condition by
1870		 * the time we get around to reclaiming it, so to avoid
1871		 * reclaiming overly full block_groups, skip reclaiming them.
1872		 *
1873		 * Since the decision making process also depends on the amount
1874		 * being freed, pass in a fake giant value to skip that extra
1875		 * check, which is more meaningful when adding to the list in
1876		 * the first place.
1877		 */
1878		if (!should_reclaim_block_group(bg, bg->length)) {
1879			spin_unlock(&bg->lock);
1880			up_write(&space_info->groups_sem);
1881			goto next;
1882		}
1883		spin_unlock(&bg->lock);
1884
1885		/*
1886		 * Get out fast, in case we're read-only or unmounting the
1887		 * filesystem. It is OK to drop block groups from the list even
1888		 * for the read-only case. As we did sb_start_write(),
1889		 * "mount -o remount,ro" won't happen and read-only filesystem
1890		 * means it is forced read-only due to a fatal error. So, it
1891		 * never gets back to read-write to let us reclaim again.
1892		 */
1893		if (btrfs_need_cleaner_sleep(fs_info)) {
1894			up_write(&space_info->groups_sem);
1895			goto next;
1896		}
1897
1898		/*
1899		 * Cache the zone_unusable value before turning the block group
1900		 * to read only. As soon as the blog group is read only it's
1901		 * zone_unusable value gets moved to the block group's read-only
1902		 * bytes and isn't available for calculations anymore.
1903		 */
1904		zone_unusable = bg->zone_unusable;
1905		ret = inc_block_group_ro(bg, 0);
1906		up_write(&space_info->groups_sem);
1907		if (ret < 0)
1908			goto next;
1909
1910		btrfs_info(fs_info,
1911			"reclaiming chunk %llu with %llu%% used %llu%% unusable",
1912				bg->start,
1913				div64_u64(bg->used * 100, bg->length),
1914				div64_u64(zone_unusable * 100, bg->length));
1915		trace_btrfs_reclaim_block_group(bg);
1916		ret = btrfs_relocate_chunk(fs_info, bg->start);
1917		if (ret) {
1918			btrfs_dec_block_group_ro(bg);
1919			btrfs_err(fs_info, "error relocating chunk %llu",
1920				  bg->start);
1921		}
1922
1923next:
1924		if (ret)
1925			btrfs_mark_bg_to_reclaim(bg);
1926		btrfs_put_block_group(bg);
1927
1928		mutex_unlock(&fs_info->reclaim_bgs_lock);
1929		/*
1930		 * Reclaiming all the block groups in the list can take really
1931		 * long.  Prioritize cleaning up unused block groups.
1932		 */
1933		btrfs_delete_unused_bgs(fs_info);
1934		/*
1935		 * If we are interrupted by a balance, we can just bail out. The
1936		 * cleaner thread restart again if necessary.
1937		 */
1938		if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1939			goto end;
1940		spin_lock(&fs_info->unused_bgs_lock);
1941	}
1942	spin_unlock(&fs_info->unused_bgs_lock);
1943	mutex_unlock(&fs_info->reclaim_bgs_lock);
1944end:
1945	btrfs_exclop_finish(fs_info);
1946	sb_end_write(fs_info->sb);
1947}
1948
1949void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1950{
1951	spin_lock(&fs_info->unused_bgs_lock);
1952	if (!list_empty(&fs_info->reclaim_bgs))
1953		queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1954	spin_unlock(&fs_info->unused_bgs_lock);
1955}
1956
1957void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1958{
1959	struct btrfs_fs_info *fs_info = bg->fs_info;
1960
1961	spin_lock(&fs_info->unused_bgs_lock);
1962	if (list_empty(&bg->bg_list)) {
1963		btrfs_get_block_group(bg);
1964		trace_btrfs_add_reclaim_block_group(bg);
1965		list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1966	}
1967	spin_unlock(&fs_info->unused_bgs_lock);
1968}
1969
1970static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1971			   struct btrfs_path *path)
1972{
1973	struct btrfs_chunk_map *map;
1974	struct btrfs_block_group_item bg;
1975	struct extent_buffer *leaf;
1976	int slot;
1977	u64 flags;
1978	int ret = 0;
1979
1980	slot = path->slots[0];
1981	leaf = path->nodes[0];
1982
1983	map = btrfs_find_chunk_map(fs_info, key->objectid, key->offset);
1984	if (!map) {
1985		btrfs_err(fs_info,
1986			  "logical %llu len %llu found bg but no related chunk",
1987			  key->objectid, key->offset);
1988		return -ENOENT;
1989	}
1990
1991	if (map->start != key->objectid || map->chunk_len != key->offset) {
1992		btrfs_err(fs_info,
1993			"block group %llu len %llu mismatch with chunk %llu len %llu",
1994			  key->objectid, key->offset, map->start, map->chunk_len);
1995		ret = -EUCLEAN;
1996		goto out_free_map;
1997	}
1998
1999	read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
2000			   sizeof(bg));
2001	flags = btrfs_stack_block_group_flags(&bg) &
2002		BTRFS_BLOCK_GROUP_TYPE_MASK;
2003
2004	if (flags != (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2005		btrfs_err(fs_info,
2006"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
2007			  key->objectid, key->offset, flags,
2008			  (BTRFS_BLOCK_GROUP_TYPE_MASK & map->type));
2009		ret = -EUCLEAN;
2010	}
2011
2012out_free_map:
2013	btrfs_free_chunk_map(map);
2014	return ret;
2015}
2016
2017static int find_first_block_group(struct btrfs_fs_info *fs_info,
2018				  struct btrfs_path *path,
2019				  struct btrfs_key *key)
2020{
2021	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2022	int ret;
2023	struct btrfs_key found_key;
2024
2025	btrfs_for_each_slot(root, key, &found_key, path, ret) {
2026		if (found_key.objectid >= key->objectid &&
2027		    found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2028			return read_bg_from_eb(fs_info, &found_key, path);
2029		}
2030	}
2031	return ret;
2032}
2033
2034static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2035{
2036	u64 extra_flags = chunk_to_extended(flags) &
2037				BTRFS_EXTENDED_PROFILE_MASK;
2038
2039	write_seqlock(&fs_info->profiles_lock);
2040	if (flags & BTRFS_BLOCK_GROUP_DATA)
2041		fs_info->avail_data_alloc_bits |= extra_flags;
2042	if (flags & BTRFS_BLOCK_GROUP_METADATA)
2043		fs_info->avail_metadata_alloc_bits |= extra_flags;
2044	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2045		fs_info->avail_system_alloc_bits |= extra_flags;
2046	write_sequnlock(&fs_info->profiles_lock);
2047}
2048
2049/*
2050 * Map a physical disk address to a list of logical addresses.
2051 *
2052 * @fs_info:       the filesystem
2053 * @chunk_start:   logical address of block group
2054 * @physical:	   physical address to map to logical addresses
2055 * @logical:	   return array of logical addresses which map to @physical
2056 * @naddrs:	   length of @logical
2057 * @stripe_len:    size of IO stripe for the given block group
2058 *
2059 * Maps a particular @physical disk address to a list of @logical addresses.
2060 * Used primarily to exclude those portions of a block group that contain super
2061 * block copies.
2062 */
2063int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2064		     u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2065{
2066	struct btrfs_chunk_map *map;
2067	u64 *buf;
2068	u64 bytenr;
2069	u64 data_stripe_length;
2070	u64 io_stripe_size;
2071	int i, nr = 0;
2072	int ret = 0;
2073
2074	map = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2075	if (IS_ERR(map))
2076		return -EIO;
2077
2078	data_stripe_length = map->stripe_size;
2079	io_stripe_size = BTRFS_STRIPE_LEN;
2080	chunk_start = map->start;
2081
2082	/* For RAID5/6 adjust to a full IO stripe length */
2083	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2084		io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2085
2086	buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2087	if (!buf) {
2088		ret = -ENOMEM;
2089		goto out;
2090	}
2091
2092	for (i = 0; i < map->num_stripes; i++) {
2093		bool already_inserted = false;
2094		u32 stripe_nr;
2095		u32 offset;
2096		int j;
2097
2098		if (!in_range(physical, map->stripes[i].physical,
2099			      data_stripe_length))
2100			continue;
2101
2102		stripe_nr = (physical - map->stripes[i].physical) >>
2103			    BTRFS_STRIPE_LEN_SHIFT;
2104		offset = (physical - map->stripes[i].physical) &
2105			 BTRFS_STRIPE_LEN_MASK;
2106
2107		if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2108				 BTRFS_BLOCK_GROUP_RAID10))
2109			stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2110					    map->sub_stripes);
2111		/*
2112		 * The remaining case would be for RAID56, multiply by
2113		 * nr_data_stripes().  Alternatively, just use rmap_len below
2114		 * instead of map->stripe_len
2115		 */
2116		bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2117
2118		/* Ensure we don't add duplicate addresses */
2119		for (j = 0; j < nr; j++) {
2120			if (buf[j] == bytenr) {
2121				already_inserted = true;
2122				break;
2123			}
2124		}
2125
2126		if (!already_inserted)
2127			buf[nr++] = bytenr;
2128	}
2129
2130	*logical = buf;
2131	*naddrs = nr;
2132	*stripe_len = io_stripe_size;
2133out:
2134	btrfs_free_chunk_map(map);
2135	return ret;
2136}
2137
2138static int exclude_super_stripes(struct btrfs_block_group *cache)
2139{
2140	struct btrfs_fs_info *fs_info = cache->fs_info;
2141	const bool zoned = btrfs_is_zoned(fs_info);
2142	u64 bytenr;
2143	u64 *logical;
2144	int stripe_len;
2145	int i, nr, ret;
2146
2147	if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2148		stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2149		cache->bytes_super += stripe_len;
2150		ret = set_extent_bit(&fs_info->excluded_extents, cache->start,
2151				     cache->start + stripe_len - 1,
2152				     EXTENT_UPTODATE, NULL);
2153		if (ret)
2154			return ret;
2155	}
2156
2157	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2158		bytenr = btrfs_sb_offset(i);
2159		ret = btrfs_rmap_block(fs_info, cache->start,
2160				       bytenr, &logical, &nr, &stripe_len);
2161		if (ret)
2162			return ret;
2163
2164		/* Shouldn't have super stripes in sequential zones */
2165		if (zoned && nr) {
2166			kfree(logical);
2167			btrfs_err(fs_info,
2168			"zoned: block group %llu must not contain super block",
2169				  cache->start);
2170			return -EUCLEAN;
2171		}
2172
2173		while (nr--) {
2174			u64 len = min_t(u64, stripe_len,
2175				cache->start + cache->length - logical[nr]);
2176
2177			cache->bytes_super += len;
2178			ret = set_extent_bit(&fs_info->excluded_extents, logical[nr],
2179					     logical[nr] + len - 1,
2180					     EXTENT_UPTODATE, NULL);
2181			if (ret) {
2182				kfree(logical);
2183				return ret;
2184			}
2185		}
2186
2187		kfree(logical);
2188	}
2189	return 0;
2190}
2191
2192static struct btrfs_block_group *btrfs_create_block_group_cache(
2193		struct btrfs_fs_info *fs_info, u64 start)
2194{
2195	struct btrfs_block_group *cache;
2196
2197	cache = kzalloc(sizeof(*cache), GFP_NOFS);
2198	if (!cache)
2199		return NULL;
2200
2201	cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2202					GFP_NOFS);
2203	if (!cache->free_space_ctl) {
2204		kfree(cache);
2205		return NULL;
2206	}
2207
2208	cache->start = start;
2209
2210	cache->fs_info = fs_info;
2211	cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2212
2213	cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2214
2215	refcount_set(&cache->refs, 1);
2216	spin_lock_init(&cache->lock);
2217	init_rwsem(&cache->data_rwsem);
2218	INIT_LIST_HEAD(&cache->list);
2219	INIT_LIST_HEAD(&cache->cluster_list);
2220	INIT_LIST_HEAD(&cache->bg_list);
2221	INIT_LIST_HEAD(&cache->ro_list);
2222	INIT_LIST_HEAD(&cache->discard_list);
2223	INIT_LIST_HEAD(&cache->dirty_list);
2224	INIT_LIST_HEAD(&cache->io_list);
2225	INIT_LIST_HEAD(&cache->active_bg_list);
2226	btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2227	atomic_set(&cache->frozen, 0);
2228	mutex_init(&cache->free_space_lock);
2229
2230	return cache;
2231}
2232
2233/*
2234 * Iterate all chunks and verify that each of them has the corresponding block
2235 * group
2236 */
2237static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2238{
2239	u64 start = 0;
2240	int ret = 0;
2241
2242	while (1) {
2243		struct btrfs_chunk_map *map;
2244		struct btrfs_block_group *bg;
2245
2246		/*
2247		 * btrfs_find_chunk_map() will return the first chunk map
2248		 * intersecting the range, so setting @length to 1 is enough to
2249		 * get the first chunk.
2250		 */
2251		map = btrfs_find_chunk_map(fs_info, start, 1);
2252		if (!map)
2253			break;
2254
2255		bg = btrfs_lookup_block_group(fs_info, map->start);
2256		if (!bg) {
2257			btrfs_err(fs_info,
2258	"chunk start=%llu len=%llu doesn't have corresponding block group",
2259				     map->start, map->chunk_len);
2260			ret = -EUCLEAN;
2261			btrfs_free_chunk_map(map);
2262			break;
2263		}
2264		if (bg->start != map->start || bg->length != map->chunk_len ||
2265		    (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2266		    (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2267			btrfs_err(fs_info,
2268"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2269				map->start, map->chunk_len,
2270				map->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2271				bg->start, bg->length,
2272				bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2273			ret = -EUCLEAN;
2274			btrfs_free_chunk_map(map);
2275			btrfs_put_block_group(bg);
2276			break;
2277		}
2278		start = map->start + map->chunk_len;
2279		btrfs_free_chunk_map(map);
2280		btrfs_put_block_group(bg);
2281	}
2282	return ret;
2283}
2284
2285static int read_one_block_group(struct btrfs_fs_info *info,
2286				struct btrfs_block_group_item *bgi,
2287				const struct btrfs_key *key,
2288				int need_clear)
2289{
2290	struct btrfs_block_group *cache;
2291	const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2292	int ret;
2293
2294	ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2295
2296	cache = btrfs_create_block_group_cache(info, key->objectid);
2297	if (!cache)
2298		return -ENOMEM;
2299
2300	cache->length = key->offset;
2301	cache->used = btrfs_stack_block_group_used(bgi);
2302	cache->commit_used = cache->used;
2303	cache->flags = btrfs_stack_block_group_flags(bgi);
2304	cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2305
2306	set_free_space_tree_thresholds(cache);
2307
2308	if (need_clear) {
2309		/*
2310		 * When we mount with old space cache, we need to
2311		 * set BTRFS_DC_CLEAR and set dirty flag.
2312		 *
2313		 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2314		 *    truncate the old free space cache inode and
2315		 *    setup a new one.
2316		 * b) Setting 'dirty flag' makes sure that we flush
2317		 *    the new space cache info onto disk.
2318		 */
2319		if (btrfs_test_opt(info, SPACE_CACHE))
2320			cache->disk_cache_state = BTRFS_DC_CLEAR;
2321	}
2322	if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2323	    (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2324			btrfs_err(info,
2325"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2326				  cache->start);
2327			ret = -EINVAL;
2328			goto error;
2329	}
2330
2331	ret = btrfs_load_block_group_zone_info(cache, false);
2332	if (ret) {
2333		btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2334			  cache->start);
2335		goto error;
2336	}
2337
2338	/*
2339	 * We need to exclude the super stripes now so that the space info has
2340	 * super bytes accounted for, otherwise we'll think we have more space
2341	 * than we actually do.
2342	 */
2343	ret = exclude_super_stripes(cache);
2344	if (ret) {
2345		/* We may have excluded something, so call this just in case. */
2346		btrfs_free_excluded_extents(cache);
2347		goto error;
2348	}
2349
2350	/*
2351	 * For zoned filesystem, space after the allocation offset is the only
2352	 * free space for a block group. So, we don't need any caching work.
2353	 * btrfs_calc_zone_unusable() will set the amount of free space and
2354	 * zone_unusable space.
2355	 *
2356	 * For regular filesystem, check for two cases, either we are full, and
2357	 * therefore don't need to bother with the caching work since we won't
2358	 * find any space, or we are empty, and we can just add all the space
2359	 * in and be done with it.  This saves us _a_lot_ of time, particularly
2360	 * in the full case.
2361	 */
2362	if (btrfs_is_zoned(info)) {
2363		btrfs_calc_zone_unusable(cache);
2364		/* Should not have any excluded extents. Just in case, though. */
2365		btrfs_free_excluded_extents(cache);
2366	} else if (cache->length == cache->used) {
2367		cache->cached = BTRFS_CACHE_FINISHED;
2368		btrfs_free_excluded_extents(cache);
2369	} else if (cache->used == 0) {
2370		cache->cached = BTRFS_CACHE_FINISHED;
2371		ret = btrfs_add_new_free_space(cache, cache->start,
2372					       cache->start + cache->length, NULL);
2373		btrfs_free_excluded_extents(cache);
2374		if (ret)
2375			goto error;
2376	}
2377
2378	ret = btrfs_add_block_group_cache(info, cache);
2379	if (ret) {
2380		btrfs_remove_free_space_cache(cache);
2381		goto error;
2382	}
2383	trace_btrfs_add_block_group(info, cache, 0);
2384	btrfs_add_bg_to_space_info(info, cache);
2385
2386	set_avail_alloc_bits(info, cache->flags);
2387	if (btrfs_chunk_writeable(info, cache->start)) {
2388		if (cache->used == 0) {
2389			ASSERT(list_empty(&cache->bg_list));
2390			if (btrfs_test_opt(info, DISCARD_ASYNC))
2391				btrfs_discard_queue_work(&info->discard_ctl, cache);
2392			else
2393				btrfs_mark_bg_unused(cache);
2394		}
2395	} else {
2396		inc_block_group_ro(cache, 1);
2397	}
2398
2399	return 0;
2400error:
2401	btrfs_put_block_group(cache);
2402	return ret;
2403}
2404
2405static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2406{
2407	struct rb_node *node;
2408	int ret = 0;
2409
2410	for (node = rb_first_cached(&fs_info->mapping_tree); node; node = rb_next(node)) {
2411		struct btrfs_chunk_map *map;
2412		struct btrfs_block_group *bg;
2413
2414		map = rb_entry(node, struct btrfs_chunk_map, rb_node);
2415		bg = btrfs_create_block_group_cache(fs_info, map->start);
2416		if (!bg) {
2417			ret = -ENOMEM;
2418			break;
2419		}
2420
2421		/* Fill dummy cache as FULL */
2422		bg->length = map->chunk_len;
2423		bg->flags = map->type;
2424		bg->cached = BTRFS_CACHE_FINISHED;
2425		bg->used = map->chunk_len;
2426		bg->flags = map->type;
2427		ret = btrfs_add_block_group_cache(fs_info, bg);
2428		/*
2429		 * We may have some valid block group cache added already, in
2430		 * that case we skip to the next one.
2431		 */
2432		if (ret == -EEXIST) {
2433			ret = 0;
2434			btrfs_put_block_group(bg);
2435			continue;
2436		}
2437
2438		if (ret) {
2439			btrfs_remove_free_space_cache(bg);
2440			btrfs_put_block_group(bg);
2441			break;
2442		}
2443
2444		btrfs_add_bg_to_space_info(fs_info, bg);
2445
2446		set_avail_alloc_bits(fs_info, bg->flags);
2447	}
2448	if (!ret)
2449		btrfs_init_global_block_rsv(fs_info);
2450	return ret;
2451}
2452
2453int btrfs_read_block_groups(struct btrfs_fs_info *info)
2454{
2455	struct btrfs_root *root = btrfs_block_group_root(info);
2456	struct btrfs_path *path;
2457	int ret;
2458	struct btrfs_block_group *cache;
2459	struct btrfs_space_info *space_info;
2460	struct btrfs_key key;
2461	int need_clear = 0;
2462	u64 cache_gen;
2463
2464	/*
2465	 * Either no extent root (with ibadroots rescue option) or we have
2466	 * unsupported RO options. The fs can never be mounted read-write, so no
2467	 * need to waste time searching block group items.
2468	 *
2469	 * This also allows new extent tree related changes to be RO compat,
2470	 * no need for a full incompat flag.
2471	 */
2472	if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2473		      ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2474		return fill_dummy_bgs(info);
2475
2476	key.objectid = 0;
2477	key.offset = 0;
2478	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2479	path = btrfs_alloc_path();
2480	if (!path)
2481		return -ENOMEM;
2482
2483	cache_gen = btrfs_super_cache_generation(info->super_copy);
2484	if (btrfs_test_opt(info, SPACE_CACHE) &&
2485	    btrfs_super_generation(info->super_copy) != cache_gen)
2486		need_clear = 1;
2487	if (btrfs_test_opt(info, CLEAR_CACHE))
2488		need_clear = 1;
2489
2490	while (1) {
2491		struct btrfs_block_group_item bgi;
2492		struct extent_buffer *leaf;
2493		int slot;
2494
2495		ret = find_first_block_group(info, path, &key);
2496		if (ret > 0)
2497			break;
2498		if (ret != 0)
2499			goto error;
2500
2501		leaf = path->nodes[0];
2502		slot = path->slots[0];
2503
2504		read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2505				   sizeof(bgi));
2506
2507		btrfs_item_key_to_cpu(leaf, &key, slot);
2508		btrfs_release_path(path);
2509		ret = read_one_block_group(info, &bgi, &key, need_clear);
2510		if (ret < 0)
2511			goto error;
2512		key.objectid += key.offset;
2513		key.offset = 0;
2514	}
2515	btrfs_release_path(path);
2516
2517	list_for_each_entry(space_info, &info->space_info, list) {
2518		int i;
2519
2520		for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2521			if (list_empty(&space_info->block_groups[i]))
2522				continue;
2523			cache = list_first_entry(&space_info->block_groups[i],
2524						 struct btrfs_block_group,
2525						 list);
2526			btrfs_sysfs_add_block_group_type(cache);
2527		}
2528
2529		if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2530		      (BTRFS_BLOCK_GROUP_RAID10 |
2531		       BTRFS_BLOCK_GROUP_RAID1_MASK |
2532		       BTRFS_BLOCK_GROUP_RAID56_MASK |
2533		       BTRFS_BLOCK_GROUP_DUP)))
2534			continue;
2535		/*
2536		 * Avoid allocating from un-mirrored block group if there are
2537		 * mirrored block groups.
2538		 */
2539		list_for_each_entry(cache,
2540				&space_info->block_groups[BTRFS_RAID_RAID0],
2541				list)
2542			inc_block_group_ro(cache, 1);
2543		list_for_each_entry(cache,
2544				&space_info->block_groups[BTRFS_RAID_SINGLE],
2545				list)
2546			inc_block_group_ro(cache, 1);
2547	}
2548
2549	btrfs_init_global_block_rsv(info);
2550	ret = check_chunk_block_group_mappings(info);
2551error:
2552	btrfs_free_path(path);
2553	/*
2554	 * We've hit some error while reading the extent tree, and have
2555	 * rescue=ibadroots mount option.
2556	 * Try to fill the tree using dummy block groups so that the user can
2557	 * continue to mount and grab their data.
2558	 */
2559	if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2560		ret = fill_dummy_bgs(info);
2561	return ret;
2562}
2563
2564/*
2565 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2566 * allocation.
2567 *
2568 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2569 * phases.
2570 */
2571static int insert_block_group_item(struct btrfs_trans_handle *trans,
2572				   struct btrfs_block_group *block_group)
2573{
2574	struct btrfs_fs_info *fs_info = trans->fs_info;
2575	struct btrfs_block_group_item bgi;
2576	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2577	struct btrfs_key key;
2578	u64 old_commit_used;
2579	int ret;
2580
2581	spin_lock(&block_group->lock);
2582	btrfs_set_stack_block_group_used(&bgi, block_group->used);
2583	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2584						   block_group->global_root_id);
2585	btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2586	old_commit_used = block_group->commit_used;
2587	block_group->commit_used = block_group->used;
2588	key.objectid = block_group->start;
2589	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2590	key.offset = block_group->length;
2591	spin_unlock(&block_group->lock);
2592
2593	ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2594	if (ret < 0) {
2595		spin_lock(&block_group->lock);
2596		block_group->commit_used = old_commit_used;
2597		spin_unlock(&block_group->lock);
2598	}
2599
2600	return ret;
2601}
2602
2603static int insert_dev_extent(struct btrfs_trans_handle *trans,
2604			    struct btrfs_device *device, u64 chunk_offset,
2605			    u64 start, u64 num_bytes)
2606{
2607	struct btrfs_fs_info *fs_info = device->fs_info;
2608	struct btrfs_root *root = fs_info->dev_root;
2609	struct btrfs_path *path;
2610	struct btrfs_dev_extent *extent;
2611	struct extent_buffer *leaf;
2612	struct btrfs_key key;
2613	int ret;
2614
2615	WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2616	WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2617	path = btrfs_alloc_path();
2618	if (!path)
2619		return -ENOMEM;
2620
2621	key.objectid = device->devid;
2622	key.type = BTRFS_DEV_EXTENT_KEY;
2623	key.offset = start;
2624	ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2625	if (ret)
2626		goto out;
2627
2628	leaf = path->nodes[0];
2629	extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2630	btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2631	btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2632					    BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2633	btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2634
2635	btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2636	btrfs_mark_buffer_dirty(trans, leaf);
2637out:
2638	btrfs_free_path(path);
2639	return ret;
2640}
2641
2642/*
2643 * This function belongs to phase 2.
2644 *
2645 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2646 * phases.
2647 */
2648static int insert_dev_extents(struct btrfs_trans_handle *trans,
2649				   u64 chunk_offset, u64 chunk_size)
2650{
2651	struct btrfs_fs_info *fs_info = trans->fs_info;
2652	struct btrfs_device *device;
2653	struct btrfs_chunk_map *map;
2654	u64 dev_offset;
2655	int i;
2656	int ret = 0;
2657
2658	map = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2659	if (IS_ERR(map))
2660		return PTR_ERR(map);
2661
2662	/*
2663	 * Take the device list mutex to prevent races with the final phase of
2664	 * a device replace operation that replaces the device object associated
2665	 * with the map's stripes, because the device object's id can change
2666	 * at any time during that final phase of the device replace operation
2667	 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2668	 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2669	 * resulting in persisting a device extent item with such ID.
2670	 */
2671	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2672	for (i = 0; i < map->num_stripes; i++) {
2673		device = map->stripes[i].dev;
2674		dev_offset = map->stripes[i].physical;
2675
2676		ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2677					map->stripe_size);
2678		if (ret)
2679			break;
2680	}
2681	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2682
2683	btrfs_free_chunk_map(map);
2684	return ret;
2685}
2686
2687/*
2688 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2689 * chunk allocation.
2690 *
2691 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2692 * phases.
2693 */
2694void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2695{
2696	struct btrfs_fs_info *fs_info = trans->fs_info;
2697	struct btrfs_block_group *block_group;
2698	int ret = 0;
2699
2700	while (!list_empty(&trans->new_bgs)) {
2701		int index;
2702
2703		block_group = list_first_entry(&trans->new_bgs,
2704					       struct btrfs_block_group,
2705					       bg_list);
2706		if (ret)
2707			goto next;
2708
2709		index = btrfs_bg_flags_to_raid_index(block_group->flags);
2710
2711		ret = insert_block_group_item(trans, block_group);
2712		if (ret)
2713			btrfs_abort_transaction(trans, ret);
2714		if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2715			      &block_group->runtime_flags)) {
2716			mutex_lock(&fs_info->chunk_mutex);
2717			ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2718			mutex_unlock(&fs_info->chunk_mutex);
2719			if (ret)
2720				btrfs_abort_transaction(trans, ret);
2721		}
2722		ret = insert_dev_extents(trans, block_group->start,
2723					 block_group->length);
2724		if (ret)
2725			btrfs_abort_transaction(trans, ret);
2726		add_block_group_free_space(trans, block_group);
2727
2728		/*
2729		 * If we restriped during balance, we may have added a new raid
2730		 * type, so now add the sysfs entries when it is safe to do so.
2731		 * We don't have to worry about locking here as it's handled in
2732		 * btrfs_sysfs_add_block_group_type.
2733		 */
2734		if (block_group->space_info->block_group_kobjs[index] == NULL)
2735			btrfs_sysfs_add_block_group_type(block_group);
2736
2737		/* Already aborted the transaction if it failed. */
2738next:
2739		btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2740		list_del_init(&block_group->bg_list);
2741		clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2742
2743		/*
2744		 * If the block group is still unused, add it to the list of
2745		 * unused block groups. The block group may have been created in
2746		 * order to satisfy a space reservation, in which case the
2747		 * extent allocation only happens later. But often we don't
2748		 * actually need to allocate space that we previously reserved,
2749		 * so the block group may become unused for a long time. For
2750		 * example for metadata we generally reserve space for a worst
2751		 * possible scenario, but then don't end up allocating all that
2752		 * space or none at all (due to no need to COW, extent buffers
2753		 * were already COWed in the current transaction and still
2754		 * unwritten, tree heights lower than the maximum possible
2755		 * height, etc). For data we generally reserve the axact amount
2756		 * of space we are going to allocate later, the exception is
2757		 * when using compression, as we must reserve space based on the
2758		 * uncompressed data size, because the compression is only done
2759		 * when writeback triggered and we don't know how much space we
2760		 * are actually going to need, so we reserve the uncompressed
2761		 * size because the data may be uncompressible in the worst case.
2762		 */
2763		if (ret == 0) {
2764			bool used;
2765
2766			spin_lock(&block_group->lock);
2767			used = btrfs_is_block_group_used(block_group);
2768			spin_unlock(&block_group->lock);
2769
2770			if (!used)
2771				btrfs_mark_bg_unused(block_group);
2772		}
2773	}
2774	btrfs_trans_release_chunk_metadata(trans);
2775}
2776
2777/*
2778 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2779 * global root id.  For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2780 */
2781static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2782{
2783	u64 div = SZ_1G;
2784	u64 index;
2785
2786	if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2787		return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2788
2789	/* If we have a smaller fs index based on 128MiB. */
2790	if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2791		div = SZ_128M;
2792
2793	offset = div64_u64(offset, div);
2794	div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2795	return index;
2796}
2797
2798struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2799						 u64 type,
2800						 u64 chunk_offset, u64 size)
2801{
2802	struct btrfs_fs_info *fs_info = trans->fs_info;
2803	struct btrfs_block_group *cache;
2804	int ret;
2805
2806	btrfs_set_log_full_commit(trans);
2807
2808	cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2809	if (!cache)
2810		return ERR_PTR(-ENOMEM);
2811
2812	/*
2813	 * Mark it as new before adding it to the rbtree of block groups or any
2814	 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2815	 * before the new flag is set.
2816	 */
2817	set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2818
2819	cache->length = size;
2820	set_free_space_tree_thresholds(cache);
2821	cache->flags = type;
2822	cache->cached = BTRFS_CACHE_FINISHED;
2823	cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2824
2825	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2826		set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2827
2828	ret = btrfs_load_block_group_zone_info(cache, true);
2829	if (ret) {
2830		btrfs_put_block_group(cache);
2831		return ERR_PTR(ret);
2832	}
2833
2834	ret = exclude_super_stripes(cache);
2835	if (ret) {
2836		/* We may have excluded something, so call this just in case */
2837		btrfs_free_excluded_extents(cache);
2838		btrfs_put_block_group(cache);
2839		return ERR_PTR(ret);
2840	}
2841
2842	ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2843	btrfs_free_excluded_extents(cache);
2844	if (ret) {
2845		btrfs_put_block_group(cache);
2846		return ERR_PTR(ret);
2847	}
2848
2849	/*
2850	 * Ensure the corresponding space_info object is created and
2851	 * assigned to our block group. We want our bg to be added to the rbtree
2852	 * with its ->space_info set.
2853	 */
2854	cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2855	ASSERT(cache->space_info);
2856
2857	ret = btrfs_add_block_group_cache(fs_info, cache);
2858	if (ret) {
2859		btrfs_remove_free_space_cache(cache);
2860		btrfs_put_block_group(cache);
2861		return ERR_PTR(ret);
2862	}
2863
2864	/*
2865	 * Now that our block group has its ->space_info set and is inserted in
2866	 * the rbtree, update the space info's counters.
2867	 */
2868	trace_btrfs_add_block_group(fs_info, cache, 1);
2869	btrfs_add_bg_to_space_info(fs_info, cache);
2870	btrfs_update_global_block_rsv(fs_info);
2871
2872#ifdef CONFIG_BTRFS_DEBUG
2873	if (btrfs_should_fragment_free_space(cache)) {
2874		cache->space_info->bytes_used += size >> 1;
2875		fragment_free_space(cache);
2876	}
2877#endif
2878
2879	list_add_tail(&cache->bg_list, &trans->new_bgs);
2880	btrfs_inc_delayed_refs_rsv_bg_inserts(fs_info);
2881
2882	set_avail_alloc_bits(fs_info, type);
2883	return cache;
2884}
2885
2886/*
2887 * Mark one block group RO, can be called several times for the same block
2888 * group.
2889 *
2890 * @cache:		the destination block group
2891 * @do_chunk_alloc:	whether need to do chunk pre-allocation, this is to
2892 * 			ensure we still have some free space after marking this
2893 * 			block group RO.
2894 */
2895int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2896			     bool do_chunk_alloc)
2897{
2898	struct btrfs_fs_info *fs_info = cache->fs_info;
2899	struct btrfs_trans_handle *trans;
2900	struct btrfs_root *root = btrfs_block_group_root(fs_info);
2901	u64 alloc_flags;
2902	int ret;
2903	bool dirty_bg_running;
2904
2905	/*
2906	 * This can only happen when we are doing read-only scrub on read-only
2907	 * mount.
2908	 * In that case we should not start a new transaction on read-only fs.
2909	 * Thus here we skip all chunk allocations.
2910	 */
2911	if (sb_rdonly(fs_info->sb)) {
2912		mutex_lock(&fs_info->ro_block_group_mutex);
2913		ret = inc_block_group_ro(cache, 0);
2914		mutex_unlock(&fs_info->ro_block_group_mutex);
2915		return ret;
2916	}
2917
2918	do {
2919		trans = btrfs_join_transaction(root);
2920		if (IS_ERR(trans))
2921			return PTR_ERR(trans);
2922
2923		dirty_bg_running = false;
2924
2925		/*
2926		 * We're not allowed to set block groups readonly after the dirty
2927		 * block group cache has started writing.  If it already started,
2928		 * back off and let this transaction commit.
2929		 */
2930		mutex_lock(&fs_info->ro_block_group_mutex);
2931		if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2932			u64 transid = trans->transid;
2933
2934			mutex_unlock(&fs_info->ro_block_group_mutex);
2935			btrfs_end_transaction(trans);
2936
2937			ret = btrfs_wait_for_commit(fs_info, transid);
2938			if (ret)
2939				return ret;
2940			dirty_bg_running = true;
2941		}
2942	} while (dirty_bg_running);
2943
2944	if (do_chunk_alloc) {
2945		/*
2946		 * If we are changing raid levels, try to allocate a
2947		 * corresponding block group with the new raid level.
2948		 */
2949		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2950		if (alloc_flags != cache->flags) {
2951			ret = btrfs_chunk_alloc(trans, alloc_flags,
2952						CHUNK_ALLOC_FORCE);
2953			/*
2954			 * ENOSPC is allowed here, we may have enough space
2955			 * already allocated at the new raid level to carry on
2956			 */
2957			if (ret == -ENOSPC)
2958				ret = 0;
2959			if (ret < 0)
2960				goto out;
2961		}
2962	}
2963
2964	ret = inc_block_group_ro(cache, 0);
2965	if (!ret)
2966		goto out;
2967	if (ret == -ETXTBSY)
2968		goto unlock_out;
2969
2970	/*
2971	 * Skip chunk allocation if the bg is SYSTEM, this is to avoid system
2972	 * chunk allocation storm to exhaust the system chunk array.  Otherwise
2973	 * we still want to try our best to mark the block group read-only.
2974	 */
2975	if (!do_chunk_alloc && ret == -ENOSPC &&
2976	    (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
2977		goto unlock_out;
2978
2979	alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2980	ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2981	if (ret < 0)
2982		goto out;
2983	/*
2984	 * We have allocated a new chunk. We also need to activate that chunk to
2985	 * grant metadata tickets for zoned filesystem.
2986	 */
2987	ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2988	if (ret < 0)
2989		goto out;
2990
2991	ret = inc_block_group_ro(cache, 0);
2992	if (ret == -ETXTBSY)
2993		goto unlock_out;
2994out:
2995	if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2996		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2997		mutex_lock(&fs_info->chunk_mutex);
2998		check_system_chunk(trans, alloc_flags);
2999		mutex_unlock(&fs_info->chunk_mutex);
3000	}
3001unlock_out:
3002	mutex_unlock(&fs_info->ro_block_group_mutex);
3003
3004	btrfs_end_transaction(trans);
3005	return ret;
3006}
3007
3008void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
3009{
3010	struct btrfs_space_info *sinfo = cache->space_info;
3011	u64 num_bytes;
3012
3013	BUG_ON(!cache->ro);
3014
3015	spin_lock(&sinfo->lock);
3016	spin_lock(&cache->lock);
3017	if (!--cache->ro) {
3018		if (btrfs_is_zoned(cache->fs_info)) {
3019			/* Migrate zone_unusable bytes back */
3020			cache->zone_unusable =
3021				(cache->alloc_offset - cache->used) +
3022				(cache->length - cache->zone_capacity);
3023			sinfo->bytes_zone_unusable += cache->zone_unusable;
3024			sinfo->bytes_readonly -= cache->zone_unusable;
3025		}
3026		num_bytes = cache->length - cache->reserved -
3027			    cache->pinned - cache->bytes_super -
3028			    cache->zone_unusable - cache->used;
3029		sinfo->bytes_readonly -= num_bytes;
3030		list_del_init(&cache->ro_list);
3031	}
3032	spin_unlock(&cache->lock);
3033	spin_unlock(&sinfo->lock);
3034}
3035
3036static int update_block_group_item(struct btrfs_trans_handle *trans,
3037				   struct btrfs_path *path,
3038				   struct btrfs_block_group *cache)
3039{
3040	struct btrfs_fs_info *fs_info = trans->fs_info;
3041	int ret;
3042	struct btrfs_root *root = btrfs_block_group_root(fs_info);
3043	unsigned long bi;
3044	struct extent_buffer *leaf;
3045	struct btrfs_block_group_item bgi;
3046	struct btrfs_key key;
3047	u64 old_commit_used;
3048	u64 used;
3049
3050	/*
3051	 * Block group items update can be triggered out of commit transaction
3052	 * critical section, thus we need a consistent view of used bytes.
3053	 * We cannot use cache->used directly outside of the spin lock, as it
3054	 * may be changed.
3055	 */
3056	spin_lock(&cache->lock);
3057	old_commit_used = cache->commit_used;
3058	used = cache->used;
3059	/* No change in used bytes, can safely skip it. */
3060	if (cache->commit_used == used) {
3061		spin_unlock(&cache->lock);
3062		return 0;
3063	}
3064	cache->commit_used = used;
3065	spin_unlock(&cache->lock);
3066
3067	key.objectid = cache->start;
3068	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3069	key.offset = cache->length;
3070
3071	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3072	if (ret) {
3073		if (ret > 0)
3074			ret = -ENOENT;
3075		goto fail;
3076	}
3077
3078	leaf = path->nodes[0];
3079	bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3080	btrfs_set_stack_block_group_used(&bgi, used);
3081	btrfs_set_stack_block_group_chunk_objectid(&bgi,
3082						   cache->global_root_id);
3083	btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3084	write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3085	btrfs_mark_buffer_dirty(trans, leaf);
3086fail:
3087	btrfs_release_path(path);
3088	/*
3089	 * We didn't update the block group item, need to revert commit_used
3090	 * unless the block group item didn't exist yet - this is to prevent a
3091	 * race with a concurrent insertion of the block group item, with
3092	 * insert_block_group_item(), that happened just after we attempted to
3093	 * update. In that case we would reset commit_used to 0 just after the
3094	 * insertion set it to a value greater than 0 - if the block group later
3095	 * becomes with 0 used bytes, we would incorrectly skip its update.
3096	 */
3097	if (ret < 0 && ret != -ENOENT) {
3098		spin_lock(&cache->lock);
3099		cache->commit_used = old_commit_used;
3100		spin_unlock(&cache->lock);
3101	}
3102	return ret;
3103
3104}
3105
3106static int cache_save_setup(struct btrfs_block_group *block_group,
3107			    struct btrfs_trans_handle *trans,
3108			    struct btrfs_path *path)
3109{
3110	struct btrfs_fs_info *fs_info = block_group->fs_info;
3111	struct inode *inode = NULL;
3112	struct extent_changeset *data_reserved = NULL;
3113	u64 alloc_hint = 0;
3114	int dcs = BTRFS_DC_ERROR;
3115	u64 cache_size = 0;
3116	int retries = 0;
3117	int ret = 0;
3118
3119	if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3120		return 0;
3121
3122	/*
3123	 * If this block group is smaller than 100 megs don't bother caching the
3124	 * block group.
3125	 */
3126	if (block_group->length < (100 * SZ_1M)) {
3127		spin_lock(&block_group->lock);
3128		block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3129		spin_unlock(&block_group->lock);
3130		return 0;
3131	}
3132
3133	if (TRANS_ABORTED(trans))
3134		return 0;
3135again:
3136	inode = lookup_free_space_inode(block_group, path);
3137	if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3138		ret = PTR_ERR(inode);
3139		btrfs_release_path(path);
3140		goto out;
3141	}
3142
3143	if (IS_ERR(inode)) {
3144		BUG_ON(retries);
3145		retries++;
3146
3147		if (block_group->ro)
3148			goto out_free;
3149
3150		ret = create_free_space_inode(trans, block_group, path);
3151		if (ret)
3152			goto out_free;
3153		goto again;
3154	}
3155
3156	/*
3157	 * We want to set the generation to 0, that way if anything goes wrong
3158	 * from here on out we know not to trust this cache when we load up next
3159	 * time.
3160	 */
3161	BTRFS_I(inode)->generation = 0;
3162	ret = btrfs_update_inode(trans, BTRFS_I(inode));
3163	if (ret) {
3164		/*
3165		 * So theoretically we could recover from this, simply set the
3166		 * super cache generation to 0 so we know to invalidate the
3167		 * cache, but then we'd have to keep track of the block groups
3168		 * that fail this way so we know we _have_ to reset this cache
3169		 * before the next commit or risk reading stale cache.  So to
3170		 * limit our exposure to horrible edge cases lets just abort the
3171		 * transaction, this only happens in really bad situations
3172		 * anyway.
3173		 */
3174		btrfs_abort_transaction(trans, ret);
3175		goto out_put;
3176	}
3177	WARN_ON(ret);
3178
3179	/* We've already setup this transaction, go ahead and exit */
3180	if (block_group->cache_generation == trans->transid &&
3181	    i_size_read(inode)) {
3182		dcs = BTRFS_DC_SETUP;
3183		goto out_put;
3184	}
3185
3186	if (i_size_read(inode) > 0) {
3187		ret = btrfs_check_trunc_cache_free_space(fs_info,
3188					&fs_info->global_block_rsv);
3189		if (ret)
3190			goto out_put;
3191
3192		ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3193		if (ret)
3194			goto out_put;
3195	}
3196
3197	spin_lock(&block_group->lock);
3198	if (block_group->cached != BTRFS_CACHE_FINISHED ||
3199	    !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3200		/*
3201		 * don't bother trying to write stuff out _if_
3202		 * a) we're not cached,
3203		 * b) we're with nospace_cache mount option,
3204		 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3205		 */
3206		dcs = BTRFS_DC_WRITTEN;
3207		spin_unlock(&block_group->lock);
3208		goto out_put;
3209	}
3210	spin_unlock(&block_group->lock);
3211
3212	/*
3213	 * We hit an ENOSPC when setting up the cache in this transaction, just
3214	 * skip doing the setup, we've already cleared the cache so we're safe.
3215	 */
3216	if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3217		ret = -ENOSPC;
3218		goto out_put;
3219	}
3220
3221	/*
3222	 * Try to preallocate enough space based on how big the block group is.
3223	 * Keep in mind this has to include any pinned space which could end up
3224	 * taking up quite a bit since it's not folded into the other space
3225	 * cache.
3226	 */
3227	cache_size = div_u64(block_group->length, SZ_256M);
3228	if (!cache_size)
3229		cache_size = 1;
3230
3231	cache_size *= 16;
3232	cache_size *= fs_info->sectorsize;
3233
3234	ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3235					  cache_size, false);
3236	if (ret)
3237		goto out_put;
3238
3239	ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3240					      cache_size, cache_size,
3241					      &alloc_hint);
3242	/*
3243	 * Our cache requires contiguous chunks so that we don't modify a bunch
3244	 * of metadata or split extents when writing the cache out, which means
3245	 * we can enospc if we are heavily fragmented in addition to just normal
3246	 * out of space conditions.  So if we hit this just skip setting up any
3247	 * other block groups for this transaction, maybe we'll unpin enough
3248	 * space the next time around.
3249	 */
3250	if (!ret)
3251		dcs = BTRFS_DC_SETUP;
3252	else if (ret == -ENOSPC)
3253		set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3254
3255out_put:
3256	iput(inode);
3257out_free:
3258	btrfs_release_path(path);
3259out:
3260	spin_lock(&block_group->lock);
3261	if (!ret && dcs == BTRFS_DC_SETUP)
3262		block_group->cache_generation = trans->transid;
3263	block_group->disk_cache_state = dcs;
3264	spin_unlock(&block_group->lock);
3265
3266	extent_changeset_free(data_reserved);
3267	return ret;
3268}
3269
3270int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3271{
3272	struct btrfs_fs_info *fs_info = trans->fs_info;
3273	struct btrfs_block_group *cache, *tmp;
3274	struct btrfs_transaction *cur_trans = trans->transaction;
3275	struct btrfs_path *path;
3276
3277	if (list_empty(&cur_trans->dirty_bgs) ||
3278	    !btrfs_test_opt(fs_info, SPACE_CACHE))
3279		return 0;
3280
3281	path = btrfs_alloc_path();
3282	if (!path)
3283		return -ENOMEM;
3284
3285	/* Could add new block groups, use _safe just in case */
3286	list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3287				 dirty_list) {
3288		if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3289			cache_save_setup(cache, trans, path);
3290	}
3291
3292	btrfs_free_path(path);
3293	return 0;
3294}
3295
3296/*
3297 * Transaction commit does final block group cache writeback during a critical
3298 * section where nothing is allowed to change the FS.  This is required in
3299 * order for the cache to actually match the block group, but can introduce a
3300 * lot of latency into the commit.
3301 *
3302 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3303 * There's a chance we'll have to redo some of it if the block group changes
3304 * again during the commit, but it greatly reduces the commit latency by
3305 * getting rid of the easy block groups while we're still allowing others to
3306 * join the commit.
3307 */
3308int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3309{
3310	struct btrfs_fs_info *fs_info = trans->fs_info;
3311	struct btrfs_block_group *cache;
3312	struct btrfs_transaction *cur_trans = trans->transaction;
3313	int ret = 0;
3314	int should_put;
3315	struct btrfs_path *path = NULL;
3316	LIST_HEAD(dirty);
3317	struct list_head *io = &cur_trans->io_bgs;
3318	int loops = 0;
3319
3320	spin_lock(&cur_trans->dirty_bgs_lock);
3321	if (list_empty(&cur_trans->dirty_bgs)) {
3322		spin_unlock(&cur_trans->dirty_bgs_lock);
3323		return 0;
3324	}
3325	list_splice_init(&cur_trans->dirty_bgs, &dirty);
3326	spin_unlock(&cur_trans->dirty_bgs_lock);
3327
3328again:
3329	/* Make sure all the block groups on our dirty list actually exist */
3330	btrfs_create_pending_block_groups(trans);
3331
3332	if (!path) {
3333		path = btrfs_alloc_path();
3334		if (!path) {
3335			ret = -ENOMEM;
3336			goto out;
3337		}
3338	}
3339
3340	/*
3341	 * cache_write_mutex is here only to save us from balance or automatic
3342	 * removal of empty block groups deleting this block group while we are
3343	 * writing out the cache
3344	 */
3345	mutex_lock(&trans->transaction->cache_write_mutex);
3346	while (!list_empty(&dirty)) {
3347		bool drop_reserve = true;
3348
3349		cache = list_first_entry(&dirty, struct btrfs_block_group,
3350					 dirty_list);
3351		/*
3352		 * This can happen if something re-dirties a block group that
3353		 * is already under IO.  Just wait for it to finish and then do
3354		 * it all again
3355		 */
3356		if (!list_empty(&cache->io_list)) {
3357			list_del_init(&cache->io_list);
3358			btrfs_wait_cache_io(trans, cache, path);
3359			btrfs_put_block_group(cache);
3360		}
3361
3362
3363		/*
3364		 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3365		 * it should update the cache_state.  Don't delete until after
3366		 * we wait.
3367		 *
3368		 * Since we're not running in the commit critical section
3369		 * we need the dirty_bgs_lock to protect from update_block_group
3370		 */
3371		spin_lock(&cur_trans->dirty_bgs_lock);
3372		list_del_init(&cache->dirty_list);
3373		spin_unlock(&cur_trans->dirty_bgs_lock);
3374
3375		should_put = 1;
3376
3377		cache_save_setup(cache, trans, path);
3378
3379		if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3380			cache->io_ctl.inode = NULL;
3381			ret = btrfs_write_out_cache(trans, cache, path);
3382			if (ret == 0 && cache->io_ctl.inode) {
3383				should_put = 0;
3384
3385				/*
3386				 * The cache_write_mutex is protecting the
3387				 * io_list, also refer to the definition of
3388				 * btrfs_transaction::io_bgs for more details
3389				 */
3390				list_add_tail(&cache->io_list, io);
3391			} else {
3392				/*
3393				 * If we failed to write the cache, the
3394				 * generation will be bad and life goes on
3395				 */
3396				ret = 0;
3397			}
3398		}
3399		if (!ret) {
3400			ret = update_block_group_item(trans, path, cache);
3401			/*
3402			 * Our block group might still be attached to the list
3403			 * of new block groups in the transaction handle of some
3404			 * other task (struct btrfs_trans_handle->new_bgs). This
3405			 * means its block group item isn't yet in the extent
3406			 * tree. If this happens ignore the error, as we will
3407			 * try again later in the critical section of the
3408			 * transaction commit.
3409			 */
3410			if (ret == -ENOENT) {
3411				ret = 0;
3412				spin_lock(&cur_trans->dirty_bgs_lock);
3413				if (list_empty(&cache->dirty_list)) {
3414					list_add_tail(&cache->dirty_list,
3415						      &cur_trans->dirty_bgs);
3416					btrfs_get_block_group(cache);
3417					drop_reserve = false;
3418				}
3419				spin_unlock(&cur_trans->dirty_bgs_lock);
3420			} else if (ret) {
3421				btrfs_abort_transaction(trans, ret);
3422			}
3423		}
3424
3425		/* If it's not on the io list, we need to put the block group */
3426		if (should_put)
3427			btrfs_put_block_group(cache);
3428		if (drop_reserve)
3429			btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3430		/*
3431		 * Avoid blocking other tasks for too long. It might even save
3432		 * us from writing caches for block groups that are going to be
3433		 * removed.
3434		 */
3435		mutex_unlock(&trans->transaction->cache_write_mutex);
3436		if (ret)
3437			goto out;
3438		mutex_lock(&trans->transaction->cache_write_mutex);
3439	}
3440	mutex_unlock(&trans->transaction->cache_write_mutex);
3441
3442	/*
3443	 * Go through delayed refs for all the stuff we've just kicked off
3444	 * and then loop back (just once)
3445	 */
3446	if (!ret)
3447		ret = btrfs_run_delayed_refs(trans, 0);
3448	if (!ret && loops == 0) {
3449		loops++;
3450		spin_lock(&cur_trans->dirty_bgs_lock);
3451		list_splice_init(&cur_trans->dirty_bgs, &dirty);
3452		/*
3453		 * dirty_bgs_lock protects us from concurrent block group
3454		 * deletes too (not just cache_write_mutex).
3455		 */
3456		if (!list_empty(&dirty)) {
3457			spin_unlock(&cur_trans->dirty_bgs_lock);
3458			goto again;
3459		}
3460		spin_unlock(&cur_trans->dirty_bgs_lock);
3461	}
3462out:
3463	if (ret < 0) {
3464		spin_lock(&cur_trans->dirty_bgs_lock);
3465		list_splice_init(&dirty, &cur_trans->dirty_bgs);
3466		spin_unlock(&cur_trans->dirty_bgs_lock);
3467		btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3468	}
3469
3470	btrfs_free_path(path);
3471	return ret;
3472}
3473
3474int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3475{
3476	struct btrfs_fs_info *fs_info = trans->fs_info;
3477	struct btrfs_block_group *cache;
3478	struct btrfs_transaction *cur_trans = trans->transaction;
3479	int ret = 0;
3480	int should_put;
3481	struct btrfs_path *path;
3482	struct list_head *io = &cur_trans->io_bgs;
3483
3484	path = btrfs_alloc_path();
3485	if (!path)
3486		return -ENOMEM;
3487
3488	/*
3489	 * Even though we are in the critical section of the transaction commit,
3490	 * we can still have concurrent tasks adding elements to this
3491	 * transaction's list of dirty block groups. These tasks correspond to
3492	 * endio free space workers started when writeback finishes for a
3493	 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3494	 * allocate new block groups as a result of COWing nodes of the root
3495	 * tree when updating the free space inode. The writeback for the space
3496	 * caches is triggered by an earlier call to
3497	 * btrfs_start_dirty_block_groups() and iterations of the following
3498	 * loop.
3499	 * Also we want to do the cache_save_setup first and then run the
3500	 * delayed refs to make sure we have the best chance at doing this all
3501	 * in one shot.
3502	 */
3503	spin_lock(&cur_trans->dirty_bgs_lock);
3504	while (!list_empty(&cur_trans->dirty_bgs)) {
3505		cache = list_first_entry(&cur_trans->dirty_bgs,
3506					 struct btrfs_block_group,
3507					 dirty_list);
3508
3509		/*
3510		 * This can happen if cache_save_setup re-dirties a block group
3511		 * that is already under IO.  Just wait for it to finish and
3512		 * then do it all again
3513		 */
3514		if (!list_empty(&cache->io_list)) {
3515			spin_unlock(&cur_trans->dirty_bgs_lock);
3516			list_del_init(&cache->io_list);
3517			btrfs_wait_cache_io(trans, cache, path);
3518			btrfs_put_block_group(cache);
3519			spin_lock(&cur_trans->dirty_bgs_lock);
3520		}
3521
3522		/*
3523		 * Don't remove from the dirty list until after we've waited on
3524		 * any pending IO
3525		 */
3526		list_del_init(&cache->dirty_list);
3527		spin_unlock(&cur_trans->dirty_bgs_lock);
3528		should_put = 1;
3529
3530		cache_save_setup(cache, trans, path);
3531
3532		if (!ret)
3533			ret = btrfs_run_delayed_refs(trans, U64_MAX);
3534
3535		if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3536			cache->io_ctl.inode = NULL;
3537			ret = btrfs_write_out_cache(trans, cache, path);
3538			if (ret == 0 && cache->io_ctl.inode) {
3539				should_put = 0;
3540				list_add_tail(&cache->io_list, io);
3541			} else {
3542				/*
3543				 * If we failed to write the cache, the
3544				 * generation will be bad and life goes on
3545				 */
3546				ret = 0;
3547			}
3548		}
3549		if (!ret) {
3550			ret = update_block_group_item(trans, path, cache);
3551			/*
3552			 * One of the free space endio workers might have
3553			 * created a new block group while updating a free space
3554			 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3555			 * and hasn't released its transaction handle yet, in
3556			 * which case the new block group is still attached to
3557			 * its transaction handle and its creation has not
3558			 * finished yet (no block group item in the extent tree
3559			 * yet, etc). If this is the case, wait for all free
3560			 * space endio workers to finish and retry. This is a
3561			 * very rare case so no need for a more efficient and
3562			 * complex approach.
3563			 */
3564			if (ret == -ENOENT) {
3565				wait_event(cur_trans->writer_wait,
3566				   atomic_read(&cur_trans->num_writers) == 1);
3567				ret = update_block_group_item(trans, path, cache);
3568			}
3569			if (ret)
3570				btrfs_abort_transaction(trans, ret);
3571		}
3572
3573		/* If its not on the io list, we need to put the block group */
3574		if (should_put)
3575			btrfs_put_block_group(cache);
3576		btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3577		spin_lock(&cur_trans->dirty_bgs_lock);
3578	}
3579	spin_unlock(&cur_trans->dirty_bgs_lock);
3580
3581	/*
3582	 * Refer to the definition of io_bgs member for details why it's safe
3583	 * to use it without any locking
3584	 */
3585	while (!list_empty(io)) {
3586		cache = list_first_entry(io, struct btrfs_block_group,
3587					 io_list);
3588		list_del_init(&cache->io_list);
3589		btrfs_wait_cache_io(trans, cache, path);
3590		btrfs_put_block_group(cache);
3591	}
3592
3593	btrfs_free_path(path);
3594	return ret;
3595}
3596
3597int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3598			     u64 bytenr, u64 num_bytes, bool alloc)
3599{
3600	struct btrfs_fs_info *info = trans->fs_info;
3601	struct btrfs_space_info *space_info;
3602	struct btrfs_block_group *cache;
3603	u64 old_val;
3604	bool reclaim = false;
3605	bool bg_already_dirty = true;
3606	int factor;
3607
3608	/* Block accounting for super block */
3609	spin_lock(&info->delalloc_root_lock);
3610	old_val = btrfs_super_bytes_used(info->super_copy);
3611	if (alloc)
3612		old_val += num_bytes;
3613	else
3614		old_val -= num_bytes;
3615	btrfs_set_super_bytes_used(info->super_copy, old_val);
3616	spin_unlock(&info->delalloc_root_lock);
3617
3618	cache = btrfs_lookup_block_group(info, bytenr);
3619	if (!cache)
3620		return -ENOENT;
3621
3622	/* An extent can not span multiple block groups. */
3623	ASSERT(bytenr + num_bytes <= cache->start + cache->length);
3624
3625	space_info = cache->space_info;
3626	factor = btrfs_bg_type_to_factor(cache->flags);
3627
3628	/*
3629	 * If this block group has free space cache written out, we need to make
3630	 * sure to load it if we are removing space.  This is because we need
3631	 * the unpinning stage to actually add the space back to the block group,
3632	 * otherwise we will leak space.
3633	 */
3634	if (!alloc && !btrfs_block_group_done(cache))
3635		btrfs_cache_block_group(cache, true);
3636
3637	spin_lock(&space_info->lock);
3638	spin_lock(&cache->lock);
3639
3640	if (btrfs_test_opt(info, SPACE_CACHE) &&
3641	    cache->disk_cache_state < BTRFS_DC_CLEAR)
3642		cache->disk_cache_state = BTRFS_DC_CLEAR;
3643
3644	old_val = cache->used;
3645	if (alloc) {
3646		old_val += num_bytes;
3647		cache->used = old_val;
3648		cache->reserved -= num_bytes;
3649		space_info->bytes_reserved -= num_bytes;
3650		space_info->bytes_used += num_bytes;
3651		space_info->disk_used += num_bytes * factor;
3652		spin_unlock(&cache->lock);
3653		spin_unlock(&space_info->lock);
3654	} else {
3655		old_val -= num_bytes;
3656		cache->used = old_val;
3657		cache->pinned += num_bytes;
3658		btrfs_space_info_update_bytes_pinned(info, space_info, num_bytes);
3659		space_info->bytes_used -= num_bytes;
3660		space_info->disk_used -= num_bytes * factor;
3661
3662		reclaim = should_reclaim_block_group(cache, num_bytes);
3663
3664		spin_unlock(&cache->lock);
3665		spin_unlock(&space_info->lock);
3666
3667		set_extent_bit(&trans->transaction->pinned_extents, bytenr,
3668			       bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
3669	}
3670
3671	spin_lock(&trans->transaction->dirty_bgs_lock);
3672	if (list_empty(&cache->dirty_list)) {
3673		list_add_tail(&cache->dirty_list, &trans->transaction->dirty_bgs);
3674		bg_already_dirty = false;
3675		btrfs_get_block_group(cache);
3676	}
3677	spin_unlock(&trans->transaction->dirty_bgs_lock);
3678
3679	/*
3680	 * No longer have used bytes in this block group, queue it for deletion.
3681	 * We do this after adding the block group to the dirty list to avoid
3682	 * races between cleaner kthread and space cache writeout.
3683	 */
3684	if (!alloc && old_val == 0) {
3685		if (!btrfs_test_opt(info, DISCARD_ASYNC))
3686			btrfs_mark_bg_unused(cache);
3687	} else if (!alloc && reclaim) {
3688		btrfs_mark_bg_to_reclaim(cache);
3689	}
3690
3691	btrfs_put_block_group(cache);
3692
3693	/* Modified block groups are accounted for in the delayed_refs_rsv. */
3694	if (!bg_already_dirty)
3695		btrfs_inc_delayed_refs_rsv_bg_updates(info);
3696
3697	return 0;
3698}
3699
3700/*
3701 * Update the block_group and space info counters.
3702 *
3703 * @cache:	The cache we are manipulating
3704 * @ram_bytes:  The number of bytes of file content, and will be same to
3705 *              @num_bytes except for the compress path.
3706 * @num_bytes:	The number of bytes in question
3707 * @delalloc:   The blocks are allocated for the delalloc write
3708 *
3709 * This is called by the allocator when it reserves space. If this is a
3710 * reservation and the block group has become read only we cannot make the
3711 * reservation and return -EAGAIN, otherwise this function always succeeds.
3712 */
3713int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3714			     u64 ram_bytes, u64 num_bytes, int delalloc,
3715			     bool force_wrong_size_class)
3716{
3717	struct btrfs_space_info *space_info = cache->space_info;
3718	enum btrfs_block_group_size_class size_class;
3719	int ret = 0;
3720
3721	spin_lock(&space_info->lock);
3722	spin_lock(&cache->lock);
3723	if (cache->ro) {
3724		ret = -EAGAIN;
3725		goto out;
3726	}
3727
3728	if (btrfs_block_group_should_use_size_class(cache)) {
3729		size_class = btrfs_calc_block_group_size_class(num_bytes);
3730		ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3731		if (ret)
3732			goto out;
3733	}
3734	cache->reserved += num_bytes;
3735	space_info->bytes_reserved += num_bytes;
3736	trace_btrfs_space_reservation(cache->fs_info, "space_info",
3737				      space_info->flags, num_bytes, 1);
3738	btrfs_space_info_update_bytes_may_use(cache->fs_info,
3739					      space_info, -ram_bytes);
3740	if (delalloc)
3741		cache->delalloc_bytes += num_bytes;
3742
3743	/*
3744	 * Compression can use less space than we reserved, so wake tickets if
3745	 * that happens.
3746	 */
3747	if (num_bytes < ram_bytes)
3748		btrfs_try_granting_tickets(cache->fs_info, space_info);
3749out:
3750	spin_unlock(&cache->lock);
3751	spin_unlock(&space_info->lock);
3752	return ret;
3753}
3754
3755/*
3756 * Update the block_group and space info counters.
3757 *
3758 * @cache:      The cache we are manipulating
3759 * @num_bytes:  The number of bytes in question
3760 * @delalloc:   The blocks are allocated for the delalloc write
3761 *
3762 * This is called by somebody who is freeing space that was never actually used
3763 * on disk.  For example if you reserve some space for a new leaf in transaction
3764 * A and before transaction A commits you free that leaf, you call this with
3765 * reserve set to 0 in order to clear the reservation.
3766 */
3767void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3768			       u64 num_bytes, int delalloc)
3769{
3770	struct btrfs_space_info *space_info = cache->space_info;
3771
3772	spin_lock(&space_info->lock);
3773	spin_lock(&cache->lock);
3774	if (cache->ro)
3775		space_info->bytes_readonly += num_bytes;
3776	cache->reserved -= num_bytes;
3777	space_info->bytes_reserved -= num_bytes;
3778	space_info->max_extent_size = 0;
3779
3780	if (delalloc)
3781		cache->delalloc_bytes -= num_bytes;
3782	spin_unlock(&cache->lock);
3783
3784	btrfs_try_granting_tickets(cache->fs_info, space_info);
3785	spin_unlock(&space_info->lock);
3786}
3787
3788static void force_metadata_allocation(struct btrfs_fs_info *info)
3789{
3790	struct list_head *head = &info->space_info;
3791	struct btrfs_space_info *found;
3792
3793	list_for_each_entry(found, head, list) {
3794		if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3795			found->force_alloc = CHUNK_ALLOC_FORCE;
3796	}
3797}
3798
3799static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3800			      struct btrfs_space_info *sinfo, int force)
3801{
3802	u64 bytes_used = btrfs_space_info_used(sinfo, false);
3803	u64 thresh;
3804
3805	if (force == CHUNK_ALLOC_FORCE)
3806		return 1;
3807
3808	/*
3809	 * in limited mode, we want to have some free space up to
3810	 * about 1% of the FS size.
3811	 */
3812	if (force == CHUNK_ALLOC_LIMITED) {
3813		thresh = btrfs_super_total_bytes(fs_info->super_copy);
3814		thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3815
3816		if (sinfo->total_bytes - bytes_used < thresh)
3817			return 1;
3818	}
3819
3820	if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3821		return 0;
3822	return 1;
3823}
3824
3825int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3826{
3827	u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3828
3829	return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3830}
3831
3832static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3833{
3834	struct btrfs_block_group *bg;
3835	int ret;
3836
3837	/*
3838	 * Check if we have enough space in the system space info because we
3839	 * will need to update device items in the chunk btree and insert a new
3840	 * chunk item in the chunk btree as well. This will allocate a new
3841	 * system block group if needed.
3842	 */
3843	check_system_chunk(trans, flags);
3844
3845	bg = btrfs_create_chunk(trans, flags);
3846	if (IS_ERR(bg)) {
3847		ret = PTR_ERR(bg);
3848		goto out;
3849	}
3850
3851	ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3852	/*
3853	 * Normally we are not expected to fail with -ENOSPC here, since we have
3854	 * previously reserved space in the system space_info and allocated one
3855	 * new system chunk if necessary. However there are three exceptions:
3856	 *
3857	 * 1) We may have enough free space in the system space_info but all the
3858	 *    existing system block groups have a profile which can not be used
3859	 *    for extent allocation.
3860	 *
3861	 *    This happens when mounting in degraded mode. For example we have a
3862	 *    RAID1 filesystem with 2 devices, lose one device and mount the fs
3863	 *    using the other device in degraded mode. If we then allocate a chunk,
3864	 *    we may have enough free space in the existing system space_info, but
3865	 *    none of the block groups can be used for extent allocation since they
3866	 *    have a RAID1 profile, and because we are in degraded mode with a
3867	 *    single device, we are forced to allocate a new system chunk with a
3868	 *    SINGLE profile. Making check_system_chunk() iterate over all system
3869	 *    block groups and check if they have a usable profile and enough space
3870	 *    can be slow on very large filesystems, so we tolerate the -ENOSPC and
3871	 *    try again after forcing allocation of a new system chunk. Like this
3872	 *    we avoid paying the cost of that search in normal circumstances, when
3873	 *    we were not mounted in degraded mode;
3874	 *
3875	 * 2) We had enough free space info the system space_info, and one suitable
3876	 *    block group to allocate from when we called check_system_chunk()
3877	 *    above. However right after we called it, the only system block group
3878	 *    with enough free space got turned into RO mode by a running scrub,
3879	 *    and in this case we have to allocate a new one and retry. We only
3880	 *    need do this allocate and retry once, since we have a transaction
3881	 *    handle and scrub uses the commit root to search for block groups;
3882	 *
3883	 * 3) We had one system block group with enough free space when we called
3884	 *    check_system_chunk(), but after that, right before we tried to
3885	 *    allocate the last extent buffer we needed, a discard operation came
3886	 *    in and it temporarily removed the last free space entry from the
3887	 *    block group (discard removes a free space entry, discards it, and
3888	 *    then adds back the entry to the block group cache).
3889	 */
3890	if (ret == -ENOSPC) {
3891		const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3892		struct btrfs_block_group *sys_bg;
3893
3894		sys_bg = btrfs_create_chunk(trans, sys_flags);
3895		if (IS_ERR(sys_bg)) {
3896			ret = PTR_ERR(sys_bg);
3897			btrfs_abort_transaction(trans, ret);
3898			goto out;
3899		}
3900
3901		ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3902		if (ret) {
3903			btrfs_abort_transaction(trans, ret);
3904			goto out;
3905		}
3906
3907		ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3908		if (ret) {
3909			btrfs_abort_transaction(trans, ret);
3910			goto out;
3911		}
3912	} else if (ret) {
3913		btrfs_abort_transaction(trans, ret);
3914		goto out;
3915	}
3916out:
3917	btrfs_trans_release_chunk_metadata(trans);
3918
3919	if (ret)
3920		return ERR_PTR(ret);
3921
3922	btrfs_get_block_group(bg);
3923	return bg;
3924}
3925
3926/*
3927 * Chunk allocation is done in 2 phases:
3928 *
3929 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3930 *    the chunk, the chunk mapping, create its block group and add the items
3931 *    that belong in the chunk btree to it - more specifically, we need to
3932 *    update device items in the chunk btree and add a new chunk item to it.
3933 *
3934 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3935 *    group item to the extent btree and the device extent items to the devices
3936 *    btree.
3937 *
3938 * This is done to prevent deadlocks. For example when COWing a node from the
3939 * extent btree we are holding a write lock on the node's parent and if we
3940 * trigger chunk allocation and attempted to insert the new block group item
3941 * in the extent btree right way, we could deadlock because the path for the
3942 * insertion can include that parent node. At first glance it seems impossible
3943 * to trigger chunk allocation after starting a transaction since tasks should
3944 * reserve enough transaction units (metadata space), however while that is true
3945 * most of the time, chunk allocation may still be triggered for several reasons:
3946 *
3947 * 1) When reserving metadata, we check if there is enough free space in the
3948 *    metadata space_info and therefore don't trigger allocation of a new chunk.
3949 *    However later when the task actually tries to COW an extent buffer from
3950 *    the extent btree or from the device btree for example, it is forced to
3951 *    allocate a new block group (chunk) because the only one that had enough
3952 *    free space was just turned to RO mode by a running scrub for example (or
3953 *    device replace, block group reclaim thread, etc), so we can not use it
3954 *    for allocating an extent and end up being forced to allocate a new one;
3955 *
3956 * 2) Because we only check that the metadata space_info has enough free bytes,
3957 *    we end up not allocating a new metadata chunk in that case. However if
3958 *    the filesystem was mounted in degraded mode, none of the existing block
3959 *    groups might be suitable for extent allocation due to their incompatible
3960 *    profile (for e.g. mounting a 2 devices filesystem, where all block groups
3961 *    use a RAID1 profile, in degraded mode using a single device). In this case
3962 *    when the task attempts to COW some extent buffer of the extent btree for
3963 *    example, it will trigger allocation of a new metadata block group with a
3964 *    suitable profile (SINGLE profile in the example of the degraded mount of
3965 *    the RAID1 filesystem);
3966 *
3967 * 3) The task has reserved enough transaction units / metadata space, but when
3968 *    it attempts to COW an extent buffer from the extent or device btree for
3969 *    example, it does not find any free extent in any metadata block group,
3970 *    therefore forced to try to allocate a new metadata block group.
3971 *    This is because some other task allocated all available extents in the
3972 *    meanwhile - this typically happens with tasks that don't reserve space
3973 *    properly, either intentionally or as a bug. One example where this is
3974 *    done intentionally is fsync, as it does not reserve any transaction units
3975 *    and ends up allocating a variable number of metadata extents for log
3976 *    tree extent buffers;
3977 *
3978 * 4) The task has reserved enough transaction units / metadata space, but right
3979 *    before it tries to allocate the last extent buffer it needs, a discard
3980 *    operation comes in and, temporarily, removes the last free space entry from
3981 *    the only metadata block group that had free space (discard starts by
3982 *    removing a free space entry from a block group, then does the discard
3983 *    operation and, once it's done, it adds back the free space entry to the
3984 *    block group).
3985 *
3986 * We also need this 2 phases setup when adding a device to a filesystem with
3987 * a seed device - we must create new metadata and system chunks without adding
3988 * any of the block group items to the chunk, extent and device btrees. If we
3989 * did not do it this way, we would get ENOSPC when attempting to update those
3990 * btrees, since all the chunks from the seed device are read-only.
3991 *
3992 * Phase 1 does the updates and insertions to the chunk btree because if we had
3993 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3994 * parallel, we risk having too many system chunks allocated by many tasks if
3995 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3996 * extreme case this leads to exhaustion of the system chunk array in the
3997 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3998 * and with RAID filesystems (so we have more device items in the chunk btree).
3999 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
4000 * the system chunk array due to concurrent allocations") provides more details.
4001 *
4002 * Allocation of system chunks does not happen through this function. A task that
4003 * needs to update the chunk btree (the only btree that uses system chunks), must
4004 * preallocate chunk space by calling either check_system_chunk() or
4005 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
4006 * metadata chunk or when removing a chunk, while the later is used before doing
4007 * a modification to the chunk btree - use cases for the later are adding,
4008 * removing and resizing a device as well as relocation of a system chunk.
4009 * See the comment below for more details.
4010 *
4011 * The reservation of system space, done through check_system_chunk(), as well
4012 * as all the updates and insertions into the chunk btree must be done while
4013 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4014 * an extent buffer from the chunks btree we never trigger allocation of a new
4015 * system chunk, which would result in a deadlock (trying to lock twice an
4016 * extent buffer of the chunk btree, first time before triggering the chunk
4017 * allocation and the second time during chunk allocation while attempting to
4018 * update the chunks btree). The system chunk array is also updated while holding
4019 * that mutex. The same logic applies to removing chunks - we must reserve system
4020 * space, update the chunk btree and the system chunk array in the superblock
4021 * while holding fs_info->chunk_mutex.
4022 *
4023 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4024 *
4025 * If @force is CHUNK_ALLOC_FORCE:
4026 *    - return 1 if it successfully allocates a chunk,
4027 *    - return errors including -ENOSPC otherwise.
4028 * If @force is NOT CHUNK_ALLOC_FORCE:
4029 *    - return 0 if it doesn't need to allocate a new chunk,
4030 *    - return 1 if it successfully allocates a chunk,
4031 *    - return errors including -ENOSPC otherwise.
4032 */
4033int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4034		      enum btrfs_chunk_alloc_enum force)
4035{
4036	struct btrfs_fs_info *fs_info = trans->fs_info;
4037	struct btrfs_space_info *space_info;
4038	struct btrfs_block_group *ret_bg;
4039	bool wait_for_alloc = false;
4040	bool should_alloc = false;
4041	bool from_extent_allocation = false;
4042	int ret = 0;
4043
4044	if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4045		from_extent_allocation = true;
4046		force = CHUNK_ALLOC_FORCE;
4047	}
4048
4049	/* Don't re-enter if we're already allocating a chunk */
4050	if (trans->allocating_chunk)
4051		return -ENOSPC;
4052	/*
4053	 * Allocation of system chunks can not happen through this path, as we
4054	 * could end up in a deadlock if we are allocating a data or metadata
4055	 * chunk and there is another task modifying the chunk btree.
4056	 *
4057	 * This is because while we are holding the chunk mutex, we will attempt
4058	 * to add the new chunk item to the chunk btree or update an existing
4059	 * device item in the chunk btree, while the other task that is modifying
4060	 * the chunk btree is attempting to COW an extent buffer while holding a
4061	 * lock on it and on its parent - if the COW operation triggers a system
4062	 * chunk allocation, then we can deadlock because we are holding the
4063	 * chunk mutex and we may need to access that extent buffer or its parent
4064	 * in order to add the chunk item or update a device item.
4065	 *
4066	 * Tasks that want to modify the chunk tree should reserve system space
4067	 * before updating the chunk btree, by calling either
4068	 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4069	 * It's possible that after a task reserves the space, it still ends up
4070	 * here - this happens in the cases described above at do_chunk_alloc().
4071	 * The task will have to either retry or fail.
4072	 */
4073	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4074		return -ENOSPC;
4075
4076	space_info = btrfs_find_space_info(fs_info, flags);
4077	ASSERT(space_info);
4078
4079	do {
4080		spin_lock(&space_info->lock);
4081		if (force < space_info->force_alloc)
4082			force = space_info->force_alloc;
4083		should_alloc = should_alloc_chunk(fs_info, space_info, force);
4084		if (space_info->full) {
4085			/* No more free physical space */
4086			if (should_alloc)
4087				ret = -ENOSPC;
4088			else
4089				ret = 0;
4090			spin_unlock(&space_info->lock);
4091			return ret;
4092		} else if (!should_alloc) {
4093			spin_unlock(&space_info->lock);
4094			return 0;
4095		} else if (space_info->chunk_alloc) {
4096			/*
4097			 * Someone is already allocating, so we need to block
4098			 * until this someone is finished and then loop to
4099			 * recheck if we should continue with our allocation
4100			 * attempt.
4101			 */
4102			wait_for_alloc = true;
4103			force = CHUNK_ALLOC_NO_FORCE;
4104			spin_unlock(&space_info->lock);
4105			mutex_lock(&fs_info->chunk_mutex);
4106			mutex_unlock(&fs_info->chunk_mutex);
4107		} else {
4108			/* Proceed with allocation */
4109			space_info->chunk_alloc = 1;
4110			wait_for_alloc = false;
4111			spin_unlock(&space_info->lock);
4112		}
4113
4114		cond_resched();
4115	} while (wait_for_alloc);
4116
4117	mutex_lock(&fs_info->chunk_mutex);
4118	trans->allocating_chunk = true;
4119
4120	/*
4121	 * If we have mixed data/metadata chunks we want to make sure we keep
4122	 * allocating mixed chunks instead of individual chunks.
4123	 */
4124	if (btrfs_mixed_space_info(space_info))
4125		flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4126
4127	/*
4128	 * if we're doing a data chunk, go ahead and make sure that
4129	 * we keep a reasonable number of metadata chunks allocated in the
4130	 * FS as well.
4131	 */
4132	if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4133		fs_info->data_chunk_allocations++;
4134		if (!(fs_info->data_chunk_allocations %
4135		      fs_info->metadata_ratio))
4136			force_metadata_allocation(fs_info);
4137	}
4138
4139	ret_bg = do_chunk_alloc(trans, flags);
4140	trans->allocating_chunk = false;
4141
4142	if (IS_ERR(ret_bg)) {
4143		ret = PTR_ERR(ret_bg);
4144	} else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4145		/*
4146		 * New block group is likely to be used soon. Try to activate
4147		 * it now. Failure is OK for now.
4148		 */
4149		btrfs_zone_activate(ret_bg);
4150	}
4151
4152	if (!ret)
4153		btrfs_put_block_group(ret_bg);
4154
4155	spin_lock(&space_info->lock);
4156	if (ret < 0) {
4157		if (ret == -ENOSPC)
4158			space_info->full = 1;
4159		else
4160			goto out;
4161	} else {
4162		ret = 1;
4163		space_info->max_extent_size = 0;
4164	}
4165
4166	space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4167out:
4168	space_info->chunk_alloc = 0;
4169	spin_unlock(&space_info->lock);
4170	mutex_unlock(&fs_info->chunk_mutex);
4171
4172	return ret;
4173}
4174
4175static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4176{
4177	u64 num_dev;
4178
4179	num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4180	if (!num_dev)
4181		num_dev = fs_info->fs_devices->rw_devices;
4182
4183	return num_dev;
4184}
4185
4186static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4187				u64 bytes,
4188				u64 type)
4189{
4190	struct btrfs_fs_info *fs_info = trans->fs_info;
4191	struct btrfs_space_info *info;
4192	u64 left;
4193	int ret = 0;
4194
4195	/*
4196	 * Needed because we can end up allocating a system chunk and for an
4197	 * atomic and race free space reservation in the chunk block reserve.
4198	 */
4199	lockdep_assert_held(&fs_info->chunk_mutex);
4200
4201	info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4202	spin_lock(&info->lock);
4203	left = info->total_bytes - btrfs_space_info_used(info, true);
4204	spin_unlock(&info->lock);
4205
4206	if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4207		btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4208			   left, bytes, type);
4209		btrfs_dump_space_info(fs_info, info, 0, 0);
4210	}
4211
4212	if (left < bytes) {
4213		u64 flags = btrfs_system_alloc_profile(fs_info);
4214		struct btrfs_block_group *bg;
4215
4216		/*
4217		 * Ignore failure to create system chunk. We might end up not
4218		 * needing it, as we might not need to COW all nodes/leafs from
4219		 * the paths we visit in the chunk tree (they were already COWed
4220		 * or created in the current transaction for example).
4221		 */
4222		bg = btrfs_create_chunk(trans, flags);
4223		if (IS_ERR(bg)) {
4224			ret = PTR_ERR(bg);
4225		} else {
4226			/*
4227			 * We have a new chunk. We also need to activate it for
4228			 * zoned filesystem.
4229			 */
4230			ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
4231			if (ret < 0)
4232				return;
4233
4234			/*
4235			 * If we fail to add the chunk item here, we end up
4236			 * trying again at phase 2 of chunk allocation, at
4237			 * btrfs_create_pending_block_groups(). So ignore
4238			 * any error here. An ENOSPC here could happen, due to
4239			 * the cases described at do_chunk_alloc() - the system
4240			 * block group we just created was just turned into RO
4241			 * mode by a scrub for example, or a running discard
4242			 * temporarily removed its free space entries, etc.
4243			 */
4244			btrfs_chunk_alloc_add_chunk_item(trans, bg);
4245		}
4246	}
4247
4248	if (!ret) {
4249		ret = btrfs_block_rsv_add(fs_info,
4250					  &fs_info->chunk_block_rsv,
4251					  bytes, BTRFS_RESERVE_NO_FLUSH);
4252		if (!ret)
4253			trans->chunk_bytes_reserved += bytes;
4254	}
4255}
4256
4257/*
4258 * Reserve space in the system space for allocating or removing a chunk.
4259 * The caller must be holding fs_info->chunk_mutex.
4260 */
4261void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4262{
4263	struct btrfs_fs_info *fs_info = trans->fs_info;
4264	const u64 num_devs = get_profile_num_devs(fs_info, type);
4265	u64 bytes;
4266
4267	/* num_devs device items to update and 1 chunk item to add or remove. */
4268	bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4269		btrfs_calc_insert_metadata_size(fs_info, 1);
4270
4271	reserve_chunk_space(trans, bytes, type);
4272}
4273
4274/*
4275 * Reserve space in the system space, if needed, for doing a modification to the
4276 * chunk btree.
4277 *
4278 * @trans:		A transaction handle.
4279 * @is_item_insertion:	Indicate if the modification is for inserting a new item
4280 *			in the chunk btree or if it's for the deletion or update
4281 *			of an existing item.
4282 *
4283 * This is used in a context where we need to update the chunk btree outside
4284 * block group allocation and removal, to avoid a deadlock with a concurrent
4285 * task that is allocating a metadata or data block group and therefore needs to
4286 * update the chunk btree while holding the chunk mutex. After the update to the
4287 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4288 *
4289 */
4290void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4291				  bool is_item_insertion)
4292{
4293	struct btrfs_fs_info *fs_info = trans->fs_info;
4294	u64 bytes;
4295
4296	if (is_item_insertion)
4297		bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4298	else
4299		bytes = btrfs_calc_metadata_size(fs_info, 1);
4300
4301	mutex_lock(&fs_info->chunk_mutex);
4302	reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4303	mutex_unlock(&fs_info->chunk_mutex);
4304}
4305
4306void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4307{
4308	struct btrfs_block_group *block_group;
4309
4310	block_group = btrfs_lookup_first_block_group(info, 0);
4311	while (block_group) {
4312		btrfs_wait_block_group_cache_done(block_group);
4313		spin_lock(&block_group->lock);
4314		if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4315				       &block_group->runtime_flags)) {
4316			struct inode *inode = block_group->inode;
4317
4318			block_group->inode = NULL;
4319			spin_unlock(&block_group->lock);
4320
4321			ASSERT(block_group->io_ctl.inode == NULL);
4322			iput(inode);
4323		} else {
4324			spin_unlock(&block_group->lock);
4325		}
4326		block_group = btrfs_next_block_group(block_group);
4327	}
4328}
4329
4330/*
4331 * Must be called only after stopping all workers, since we could have block
4332 * group caching kthreads running, and therefore they could race with us if we
4333 * freed the block groups before stopping them.
4334 */
4335int btrfs_free_block_groups(struct btrfs_fs_info *info)
4336{
4337	struct btrfs_block_group *block_group;
4338	struct btrfs_space_info *space_info;
4339	struct btrfs_caching_control *caching_ctl;
4340	struct rb_node *n;
4341
4342	if (btrfs_is_zoned(info)) {
4343		if (info->active_meta_bg) {
4344			btrfs_put_block_group(info->active_meta_bg);
4345			info->active_meta_bg = NULL;
4346		}
4347		if (info->active_system_bg) {
4348			btrfs_put_block_group(info->active_system_bg);
4349			info->active_system_bg = NULL;
4350		}
4351	}
4352
4353	write_lock(&info->block_group_cache_lock);
4354	while (!list_empty(&info->caching_block_groups)) {
4355		caching_ctl = list_entry(info->caching_block_groups.next,
4356					 struct btrfs_caching_control, list);
4357		list_del(&caching_ctl->list);
4358		btrfs_put_caching_control(caching_ctl);
4359	}
4360	write_unlock(&info->block_group_cache_lock);
4361
4362	spin_lock(&info->unused_bgs_lock);
4363	while (!list_empty(&info->unused_bgs)) {
4364		block_group = list_first_entry(&info->unused_bgs,
4365					       struct btrfs_block_group,
4366					       bg_list);
4367		list_del_init(&block_group->bg_list);
4368		btrfs_put_block_group(block_group);
4369	}
4370
4371	while (!list_empty(&info->reclaim_bgs)) {
4372		block_group = list_first_entry(&info->reclaim_bgs,
4373					       struct btrfs_block_group,
4374					       bg_list);
4375		list_del_init(&block_group->bg_list);
4376		btrfs_put_block_group(block_group);
4377	}
4378	spin_unlock(&info->unused_bgs_lock);
4379
4380	spin_lock(&info->zone_active_bgs_lock);
4381	while (!list_empty(&info->zone_active_bgs)) {
4382		block_group = list_first_entry(&info->zone_active_bgs,
4383					       struct btrfs_block_group,
4384					       active_bg_list);
4385		list_del_init(&block_group->active_bg_list);
4386		btrfs_put_block_group(block_group);
4387	}
4388	spin_unlock(&info->zone_active_bgs_lock);
4389
4390	write_lock(&info->block_group_cache_lock);
4391	while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4392		block_group = rb_entry(n, struct btrfs_block_group,
4393				       cache_node);
4394		rb_erase_cached(&block_group->cache_node,
4395				&info->block_group_cache_tree);
4396		RB_CLEAR_NODE(&block_group->cache_node);
4397		write_unlock(&info->block_group_cache_lock);
4398
4399		down_write(&block_group->space_info->groups_sem);
4400		list_del(&block_group->list);
4401		up_write(&block_group->space_info->groups_sem);
4402
4403		/*
4404		 * We haven't cached this block group, which means we could
4405		 * possibly have excluded extents on this block group.
4406		 */
4407		if (block_group->cached == BTRFS_CACHE_NO ||
4408		    block_group->cached == BTRFS_CACHE_ERROR)
4409			btrfs_free_excluded_extents(block_group);
4410
4411		btrfs_remove_free_space_cache(block_group);
4412		ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4413		ASSERT(list_empty(&block_group->dirty_list));
4414		ASSERT(list_empty(&block_group->io_list));
4415		ASSERT(list_empty(&block_group->bg_list));
4416		ASSERT(refcount_read(&block_group->refs) == 1);
4417		ASSERT(block_group->swap_extents == 0);
4418		btrfs_put_block_group(block_group);
4419
4420		write_lock(&info->block_group_cache_lock);
4421	}
4422	write_unlock(&info->block_group_cache_lock);
4423
4424	btrfs_release_global_block_rsv(info);
4425
4426	while (!list_empty(&info->space_info)) {
4427		space_info = list_entry(info->space_info.next,
4428					struct btrfs_space_info,
4429					list);
4430
4431		/*
4432		 * Do not hide this behind enospc_debug, this is actually
4433		 * important and indicates a real bug if this happens.
4434		 */
4435		if (WARN_ON(space_info->bytes_pinned > 0 ||
4436			    space_info->bytes_may_use > 0))
4437			btrfs_dump_space_info(info, space_info, 0, 0);
4438
4439		/*
4440		 * If there was a failure to cleanup a log tree, very likely due
4441		 * to an IO failure on a writeback attempt of one or more of its
4442		 * extent buffers, we could not do proper (and cheap) unaccounting
4443		 * of their reserved space, so don't warn on bytes_reserved > 0 in
4444		 * that case.
4445		 */
4446		if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4447		    !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4448			if (WARN_ON(space_info->bytes_reserved > 0))
4449				btrfs_dump_space_info(info, space_info, 0, 0);
4450		}
4451
4452		WARN_ON(space_info->reclaim_size > 0);
4453		list_del(&space_info->list);
4454		btrfs_sysfs_remove_space_info(space_info);
4455	}
4456	return 0;
4457}
4458
4459void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4460{
4461	atomic_inc(&cache->frozen);
4462}
4463
4464void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4465{
4466	struct btrfs_fs_info *fs_info = block_group->fs_info;
4467	bool cleanup;
4468
4469	spin_lock(&block_group->lock);
4470	cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4471		   test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4472	spin_unlock(&block_group->lock);
4473
4474	if (cleanup) {
4475		struct btrfs_chunk_map *map;
4476
4477		map = btrfs_find_chunk_map(fs_info, block_group->start, 1);
4478		/* Logic error, can't happen. */
4479		ASSERT(map);
4480
4481		btrfs_remove_chunk_map(fs_info, map);
4482
4483		/* Once for our lookup reference. */
4484		btrfs_free_chunk_map(map);
4485
4486		/*
4487		 * We may have left one free space entry and other possible
4488		 * tasks trimming this block group have left 1 entry each one.
4489		 * Free them if any.
4490		 */
4491		btrfs_remove_free_space_cache(block_group);
4492	}
4493}
4494
4495bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4496{
4497	bool ret = true;
4498
4499	spin_lock(&bg->lock);
4500	if (bg->ro)
4501		ret = false;
4502	else
4503		bg->swap_extents++;
4504	spin_unlock(&bg->lock);
4505
4506	return ret;
4507}
4508
4509void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4510{
4511	spin_lock(&bg->lock);
4512	ASSERT(!bg->ro);
4513	ASSERT(bg->swap_extents >= amount);
4514	bg->swap_extents -= amount;
4515	spin_unlock(&bg->lock);
4516}
4517
4518enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4519{
4520	if (size <= SZ_128K)
4521		return BTRFS_BG_SZ_SMALL;
4522	if (size <= SZ_8M)
4523		return BTRFS_BG_SZ_MEDIUM;
4524	return BTRFS_BG_SZ_LARGE;
4525}
4526
4527/*
4528 * Handle a block group allocating an extent in a size class
4529 *
4530 * @bg:				The block group we allocated in.
4531 * @size_class:			The size class of the allocation.
4532 * @force_wrong_size_class:	Whether we are desperate enough to allow
4533 *				mismatched size classes.
4534 *
4535 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4536 * case of a race that leads to the wrong size class without
4537 * force_wrong_size_class set.
4538 *
4539 * find_free_extent will skip block groups with a mismatched size class until
4540 * it really needs to avoid ENOSPC. In that case it will set
4541 * force_wrong_size_class. However, if a block group is newly allocated and
4542 * doesn't yet have a size class, then it is possible for two allocations of
4543 * different sizes to race and both try to use it. The loser is caught here and
4544 * has to retry.
4545 */
4546int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4547				     enum btrfs_block_group_size_class size_class,
4548				     bool force_wrong_size_class)
4549{
4550	ASSERT(size_class != BTRFS_BG_SZ_NONE);
4551
4552	/* The new allocation is in the right size class, do nothing */
4553	if (bg->size_class == size_class)
4554		return 0;
4555	/*
4556	 * The new allocation is in a mismatched size class.
4557	 * This means one of two things:
4558	 *
4559	 * 1. Two tasks in find_free_extent for different size_classes raced
4560	 *    and hit the same empty block_group. Make the loser try again.
4561	 * 2. A call to find_free_extent got desperate enough to set
4562	 *    'force_wrong_slab'. Don't change the size_class, but allow the
4563	 *    allocation.
4564	 */
4565	if (bg->size_class != BTRFS_BG_SZ_NONE) {
4566		if (force_wrong_size_class)
4567			return 0;
4568		return -EAGAIN;
4569	}
4570	/*
4571	 * The happy new block group case: the new allocation is the first
4572	 * one in the block_group so we set size_class.
4573	 */
4574	bg->size_class = size_class;
4575
4576	return 0;
4577}
4578
4579bool btrfs_block_group_should_use_size_class(struct btrfs_block_group *bg)
4580{
4581	if (btrfs_is_zoned(bg->fs_info))
4582		return false;
4583	if (!btrfs_is_block_group_data_only(bg))
4584		return false;
4585	return true;
4586}