Linux Audio

Check our new training course

Linux BSP development engineering services

Need help to port Linux and bootloaders to your hardware?
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}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0
   2
 
 
   3#include "misc.h"
   4#include "ctree.h"
   5#include "block-group.h"
   6#include "space-info.h"
   7#include "disk-io.h"
   8#include "free-space-cache.h"
   9#include "free-space-tree.h"
  10#include "volumes.h"
  11#include "transaction.h"
  12#include "ref-verify.h"
  13#include "sysfs.h"
  14#include "tree-log.h"
  15#include "delalloc-space.h"
  16#include "discard.h"
  17#include "raid56.h"
  18#include "zoned.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  19
  20/*
  21 * Return target flags in extended format or 0 if restripe for this chunk_type
  22 * is not in progress
  23 *
  24 * Should be called with balance_lock held
  25 */
  26static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
  27{
  28	struct btrfs_balance_control *bctl = fs_info->balance_ctl;
  29	u64 target = 0;
  30
  31	if (!bctl)
  32		return 0;
  33
  34	if (flags & BTRFS_BLOCK_GROUP_DATA &&
  35	    bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  36		target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
  37	} else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
  38		   bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  39		target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
  40	} else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
  41		   bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
  42		target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
  43	}
  44
  45	return target;
  46}
  47
  48/*
  49 * @flags: available profiles in extended format (see ctree.h)
  50 *
  51 * Return reduced profile in chunk format.  If profile changing is in progress
  52 * (either running or paused) picks the target profile (if it's already
  53 * available), otherwise falls back to plain reducing.
  54 */
  55static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
  56{
  57	u64 num_devices = fs_info->fs_devices->rw_devices;
  58	u64 target;
  59	u64 raid_type;
  60	u64 allowed = 0;
  61
  62	/*
  63	 * See if restripe for this chunk_type is in progress, if so try to
  64	 * reduce to the target profile
  65	 */
  66	spin_lock(&fs_info->balance_lock);
  67	target = get_restripe_target(fs_info, flags);
  68	if (target) {
  69		spin_unlock(&fs_info->balance_lock);
  70		return extended_to_chunk(target);
  71	}
  72	spin_unlock(&fs_info->balance_lock);
  73
  74	/* First, mask out the RAID levels which aren't possible */
  75	for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
  76		if (num_devices >= btrfs_raid_array[raid_type].devs_min)
  77			allowed |= btrfs_raid_array[raid_type].bg_flag;
  78	}
  79	allowed &= flags;
  80
  81	if (allowed & BTRFS_BLOCK_GROUP_RAID6)
 
 
 
  82		allowed = BTRFS_BLOCK_GROUP_RAID6;
 
 
  83	else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
  84		allowed = BTRFS_BLOCK_GROUP_RAID5;
  85	else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
  86		allowed = BTRFS_BLOCK_GROUP_RAID10;
  87	else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
  88		allowed = BTRFS_BLOCK_GROUP_RAID1;
 
 
  89	else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
  90		allowed = BTRFS_BLOCK_GROUP_RAID0;
  91
  92	flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
  93
  94	return extended_to_chunk(flags | allowed);
  95}
  96
  97u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
  98{
  99	unsigned seq;
 100	u64 flags;
 101
 102	do {
 103		flags = orig_flags;
 104		seq = read_seqbegin(&fs_info->profiles_lock);
 105
 106		if (flags & BTRFS_BLOCK_GROUP_DATA)
 107			flags |= fs_info->avail_data_alloc_bits;
 108		else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 109			flags |= fs_info->avail_system_alloc_bits;
 110		else if (flags & BTRFS_BLOCK_GROUP_METADATA)
 111			flags |= fs_info->avail_metadata_alloc_bits;
 112	} while (read_seqretry(&fs_info->profiles_lock, seq));
 113
 114	return btrfs_reduce_alloc_profile(fs_info, flags);
 115}
 116
 117void btrfs_get_block_group(struct btrfs_block_group *cache)
 118{
 119	refcount_inc(&cache->refs);
 120}
 121
 122void btrfs_put_block_group(struct btrfs_block_group *cache)
 123{
 124	if (refcount_dec_and_test(&cache->refs)) {
 125		WARN_ON(cache->pinned > 0);
 126		WARN_ON(cache->reserved > 0);
 
 
 
 
 
 
 
 
 
 127
 128		/*
 129		 * A block_group shouldn't be on the discard_list anymore.
 130		 * Remove the block_group from the discard_list to prevent us
 131		 * from causing a panic due to NULL pointer dereference.
 132		 */
 133		if (WARN_ON(!list_empty(&cache->discard_list)))
 134			btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
 135						  cache);
 136
 137		/*
 138		 * If not empty, someone is still holding mutex of
 139		 * full_stripe_lock, which can only be released by caller.
 140		 * And it will definitely cause use-after-free when caller
 141		 * tries to release full stripe lock.
 142		 *
 143		 * No better way to resolve, but only to warn.
 144		 */
 145		WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
 146		kfree(cache->free_space_ctl);
 
 147		kfree(cache);
 148	}
 149}
 150
 151/*
 152 * This adds the block group to the fs_info rb tree for the block group cache
 153 */
 154static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 155				       struct btrfs_block_group *block_group)
 156{
 157	struct rb_node **p;
 158	struct rb_node *parent = NULL;
 159	struct btrfs_block_group *cache;
 
 160
 161	ASSERT(block_group->length != 0);
 162
 163	spin_lock(&info->block_group_cache_lock);
 164	p = &info->block_group_cache_tree.rb_node;
 165
 166	while (*p) {
 167		parent = *p;
 168		cache = rb_entry(parent, struct btrfs_block_group, cache_node);
 169		if (block_group->start < cache->start) {
 170			p = &(*p)->rb_left;
 171		} else if (block_group->start > cache->start) {
 172			p = &(*p)->rb_right;
 
 173		} else {
 174			spin_unlock(&info->block_group_cache_lock);
 175			return -EEXIST;
 176		}
 177	}
 178
 179	rb_link_node(&block_group->cache_node, parent, p);
 180	rb_insert_color(&block_group->cache_node,
 181			&info->block_group_cache_tree);
 182
 183	if (info->first_logical_byte > block_group->start)
 184		info->first_logical_byte = block_group->start;
 185
 186	spin_unlock(&info->block_group_cache_lock);
 187
 188	return 0;
 189}
 190
 191/*
 192 * This will return the block group at or after bytenr if contains is 0, else
 193 * it will return the block group that contains the bytenr
 194 */
 195static struct btrfs_block_group *block_group_cache_tree_search(
 196		struct btrfs_fs_info *info, u64 bytenr, int contains)
 197{
 198	struct btrfs_block_group *cache, *ret = NULL;
 199	struct rb_node *n;
 200	u64 end, start;
 201
 202	spin_lock(&info->block_group_cache_lock);
 203	n = info->block_group_cache_tree.rb_node;
 204
 205	while (n) {
 206		cache = rb_entry(n, struct btrfs_block_group, cache_node);
 207		end = cache->start + cache->length - 1;
 208		start = cache->start;
 209
 210		if (bytenr < start) {
 211			if (!contains && (!ret || start < ret->start))
 212				ret = cache;
 213			n = n->rb_left;
 214		} else if (bytenr > start) {
 215			if (contains && bytenr <= end) {
 216				ret = cache;
 217				break;
 218			}
 219			n = n->rb_right;
 220		} else {
 221			ret = cache;
 222			break;
 223		}
 224	}
 225	if (ret) {
 226		btrfs_get_block_group(ret);
 227		if (bytenr == 0 && info->first_logical_byte > ret->start)
 228			info->first_logical_byte = ret->start;
 229	}
 230	spin_unlock(&info->block_group_cache_lock);
 231
 232	return ret;
 233}
 234
 235/*
 236 * Return the block group that starts at or after bytenr
 237 */
 238struct btrfs_block_group *btrfs_lookup_first_block_group(
 239		struct btrfs_fs_info *info, u64 bytenr)
 240{
 241	return block_group_cache_tree_search(info, bytenr, 0);
 242}
 243
 244/*
 245 * Return the block group that contains the given bytenr
 246 */
 247struct btrfs_block_group *btrfs_lookup_block_group(
 248		struct btrfs_fs_info *info, u64 bytenr)
 249{
 250	return block_group_cache_tree_search(info, bytenr, 1);
 251}
 252
 253struct btrfs_block_group *btrfs_next_block_group(
 254		struct btrfs_block_group *cache)
 255{
 256	struct btrfs_fs_info *fs_info = cache->fs_info;
 257	struct rb_node *node;
 258
 259	spin_lock(&fs_info->block_group_cache_lock);
 260
 261	/* If our block group was removed, we need a full search. */
 262	if (RB_EMPTY_NODE(&cache->cache_node)) {
 263		const u64 next_bytenr = cache->start + cache->length;
 264
 265		spin_unlock(&fs_info->block_group_cache_lock);
 266		btrfs_put_block_group(cache);
 267		cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
 268	}
 269	node = rb_next(&cache->cache_node);
 270	btrfs_put_block_group(cache);
 271	if (node) {
 272		cache = rb_entry(node, struct btrfs_block_group, cache_node);
 273		btrfs_get_block_group(cache);
 274	} else
 275		cache = NULL;
 276	spin_unlock(&fs_info->block_group_cache_lock);
 277	return cache;
 278}
 279
 280bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 281{
 282	struct btrfs_block_group *bg;
 283	bool ret = true;
 284
 285	bg = btrfs_lookup_block_group(fs_info, bytenr);
 286	if (!bg)
 287		return false;
 288
 289	spin_lock(&bg->lock);
 290	if (bg->ro)
 291		ret = false;
 292	else
 293		atomic_inc(&bg->nocow_writers);
 294	spin_unlock(&bg->lock);
 295
 296	/* No put on block group, done by btrfs_dec_nocow_writers */
 297	if (!ret)
 298		btrfs_put_block_group(bg);
 
 
 299
 300	return ret;
 
 301}
 302
 303void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
 
 
 
 
 
 
 
 
 
 
 
 304{
 305	struct btrfs_block_group *bg;
 306
 307	bg = btrfs_lookup_block_group(fs_info, bytenr);
 308	ASSERT(bg);
 309	if (atomic_dec_and_test(&bg->nocow_writers))
 310		wake_up_var(&bg->nocow_writers);
 311	/*
 312	 * Once for our lookup and once for the lookup done by a previous call
 313	 * to btrfs_inc_nocow_writers()
 314	 */
 315	btrfs_put_block_group(bg);
 316	btrfs_put_block_group(bg);
 317}
 318
 319void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
 320{
 321	wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
 322}
 323
 324void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
 325					const u64 start)
 326{
 327	struct btrfs_block_group *bg;
 328
 329	bg = btrfs_lookup_block_group(fs_info, start);
 330	ASSERT(bg);
 331	if (atomic_dec_and_test(&bg->reservations))
 332		wake_up_var(&bg->reservations);
 333	btrfs_put_block_group(bg);
 334}
 335
 336void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
 337{
 338	struct btrfs_space_info *space_info = bg->space_info;
 339
 340	ASSERT(bg->ro);
 341
 342	if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
 343		return;
 344
 345	/*
 346	 * Our block group is read only but before we set it to read only,
 347	 * some task might have had allocated an extent from it already, but it
 348	 * has not yet created a respective ordered extent (and added it to a
 349	 * root's list of ordered extents).
 350	 * Therefore wait for any task currently allocating extents, since the
 351	 * block group's reservations counter is incremented while a read lock
 352	 * on the groups' semaphore is held and decremented after releasing
 353	 * the read access on that semaphore and creating the ordered extent.
 354	 */
 355	down_write(&space_info->groups_sem);
 356	up_write(&space_info->groups_sem);
 357
 358	wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
 359}
 360
 361struct btrfs_caching_control *btrfs_get_caching_control(
 362		struct btrfs_block_group *cache)
 363{
 364	struct btrfs_caching_control *ctl;
 365
 366	spin_lock(&cache->lock);
 367	if (!cache->caching_ctl) {
 368		spin_unlock(&cache->lock);
 369		return NULL;
 370	}
 371
 372	ctl = cache->caching_ctl;
 373	refcount_inc(&ctl->count);
 374	spin_unlock(&cache->lock);
 375	return ctl;
 376}
 377
 378void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
 379{
 380	if (refcount_dec_and_test(&ctl->count))
 381		kfree(ctl);
 382}
 383
 384/*
 385 * When we wait for progress in the block group caching, its because our
 386 * allocation attempt failed at least once.  So, we must sleep and let some
 387 * progress happen before we try again.
 388 *
 389 * This function will sleep at least once waiting for new free space to show
 390 * up, and then it will check the block group free space numbers for our min
 391 * num_bytes.  Another option is to have it go ahead and look in the rbtree for
 392 * a free extent of a given size, but this is a good start.
 393 *
 394 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
 395 * any of the information in this block group.
 396 */
 397void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
 398					   u64 num_bytes)
 399{
 400	struct btrfs_caching_control *caching_ctl;
 
 401
 402	caching_ctl = btrfs_get_caching_control(cache);
 403	if (!caching_ctl)
 404		return;
 405
 
 
 
 
 
 
 
 
 406	wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
 407		   (cache->free_space_ctl->free_space >= num_bytes));
 
 408
 409	btrfs_put_caching_control(caching_ctl);
 410}
 411
 412int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
 
 
 
 
 
 
 
 413{
 414	struct btrfs_caching_control *caching_ctl;
 415	int ret = 0;
 416
 417	caching_ctl = btrfs_get_caching_control(cache);
 418	if (!caching_ctl)
 419		return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
 420
 421	wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
 422	if (cache->cached == BTRFS_CACHE_ERROR)
 423		ret = -EIO;
 424	btrfs_put_caching_control(caching_ctl);
 425	return ret;
 426}
 427
 428static bool space_cache_v1_done(struct btrfs_block_group *cache)
 429{
 430	bool ret;
 431
 432	spin_lock(&cache->lock);
 433	ret = cache->cached != BTRFS_CACHE_FAST;
 434	spin_unlock(&cache->lock);
 435
 436	return ret;
 437}
 438
 439void btrfs_wait_space_cache_v1_finished(struct btrfs_block_group *cache,
 440				struct btrfs_caching_control *caching_ctl)
 441{
 442	wait_event(caching_ctl->wait, space_cache_v1_done(cache));
 443}
 444
 445#ifdef CONFIG_BTRFS_DEBUG
 446static void fragment_free_space(struct btrfs_block_group *block_group)
 447{
 448	struct btrfs_fs_info *fs_info = block_group->fs_info;
 449	u64 start = block_group->start;
 450	u64 len = block_group->length;
 451	u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 452		fs_info->nodesize : fs_info->sectorsize;
 453	u64 step = chunk << 1;
 454
 455	while (len > chunk) {
 456		btrfs_remove_free_space(block_group, start, chunk);
 457		start += step;
 458		if (len < step)
 459			len = 0;
 460		else
 461			len -= step;
 462	}
 463}
 464#endif
 465
 466/*
 467 * This is only called by btrfs_cache_block_group, since we could have freed
 468 * extents we need to check the pinned_extents for any extents that can't be
 469 * used yet since their free space will be released as soon as the transaction
 470 * commits.
 
 
 
 
 
 
 
 471 */
 472u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
 
 473{
 474	struct btrfs_fs_info *info = block_group->fs_info;
 475	u64 extent_start, extent_end, size, total_added = 0;
 476	int ret;
 477
 
 
 
 478	while (start < end) {
 479		ret = find_first_extent_bit(&info->excluded_extents, start,
 480					    &extent_start, &extent_end,
 481					    EXTENT_DIRTY | EXTENT_UPTODATE,
 482					    NULL);
 483		if (ret)
 484			break;
 485
 486		if (extent_start <= start) {
 487			start = extent_end + 1;
 488		} else if (extent_start > start && extent_start < end) {
 489			size = extent_start - start;
 490			total_added += size;
 491			ret = btrfs_add_free_space_async_trimmed(block_group,
 492								 start, size);
 493			BUG_ON(ret); /* -ENOMEM or logic error */
 
 
 
 494			start = extent_end + 1;
 495		} else {
 496			break;
 497		}
 498	}
 499
 500	if (start < end) {
 501		size = end - start;
 502		total_added += size;
 503		ret = btrfs_add_free_space_async_trimmed(block_group, start,
 504							 size);
 505		BUG_ON(ret); /* -ENOMEM or logic error */
 
 
 
 506	}
 507
 508	return total_added;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 509}
 510
 511static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 512{
 513	struct btrfs_block_group *block_group = caching_ctl->block_group;
 514	struct btrfs_fs_info *fs_info = block_group->fs_info;
 515	struct btrfs_root *extent_root = fs_info->extent_root;
 516	struct btrfs_path *path;
 517	struct extent_buffer *leaf;
 518	struct btrfs_key key;
 519	u64 total_found = 0;
 520	u64 last = 0;
 521	u32 nritems;
 522	int ret;
 523	bool wakeup = true;
 524
 525	path = btrfs_alloc_path();
 526	if (!path)
 527		return -ENOMEM;
 528
 529	last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
 
 530
 531#ifdef CONFIG_BTRFS_DEBUG
 532	/*
 533	 * If we're fragmenting we don't want to make anybody think we can
 534	 * allocate from this block group until we've had a chance to fragment
 535	 * the free space.
 536	 */
 537	if (btrfs_should_fragment_free_space(block_group))
 538		wakeup = false;
 539#endif
 540	/*
 541	 * We don't want to deadlock with somebody trying to allocate a new
 542	 * extent for the extent root while also trying to search the extent
 543	 * root to add free space.  So we skip locking and search the commit
 544	 * root, since its read-only
 545	 */
 546	path->skip_locking = 1;
 547	path->search_commit_root = 1;
 548	path->reada = READA_FORWARD;
 549
 550	key.objectid = last;
 551	key.offset = 0;
 552	key.type = BTRFS_EXTENT_ITEM_KEY;
 553
 554next:
 555	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 556	if (ret < 0)
 557		goto out;
 558
 559	leaf = path->nodes[0];
 560	nritems = btrfs_header_nritems(leaf);
 561
 562	while (1) {
 563		if (btrfs_fs_closing(fs_info) > 1) {
 564			last = (u64)-1;
 565			break;
 566		}
 567
 568		if (path->slots[0] < nritems) {
 569			btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 570		} else {
 571			ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
 572			if (ret)
 573				break;
 574
 575			if (need_resched() ||
 576			    rwsem_is_contended(&fs_info->commit_root_sem)) {
 577				if (wakeup)
 578					caching_ctl->progress = last;
 579				btrfs_release_path(path);
 580				up_read(&fs_info->commit_root_sem);
 581				mutex_unlock(&caching_ctl->mutex);
 582				cond_resched();
 583				mutex_lock(&caching_ctl->mutex);
 584				down_read(&fs_info->commit_root_sem);
 585				goto next;
 586			}
 587
 588			ret = btrfs_next_leaf(extent_root, path);
 589			if (ret < 0)
 590				goto out;
 591			if (ret)
 592				break;
 593			leaf = path->nodes[0];
 594			nritems = btrfs_header_nritems(leaf);
 595			continue;
 596		}
 597
 598		if (key.objectid < last) {
 599			key.objectid = last;
 600			key.offset = 0;
 601			key.type = BTRFS_EXTENT_ITEM_KEY;
 602
 603			if (wakeup)
 604				caching_ctl->progress = last;
 605			btrfs_release_path(path);
 606			goto next;
 607		}
 608
 609		if (key.objectid < block_group->start) {
 610			path->slots[0]++;
 611			continue;
 612		}
 613
 614		if (key.objectid >= block_group->start + block_group->length)
 615			break;
 616
 617		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 618		    key.type == BTRFS_METADATA_ITEM_KEY) {
 619			total_found += add_new_free_space(block_group, last,
 620							  key.objectid);
 
 
 
 
 
 621			if (key.type == BTRFS_METADATA_ITEM_KEY)
 622				last = key.objectid +
 623					fs_info->nodesize;
 624			else
 625				last = key.objectid + key.offset;
 626
 627			if (total_found > CACHING_CTL_WAKE_UP) {
 628				total_found = 0;
 629				if (wakeup)
 
 630					wake_up(&caching_ctl->wait);
 
 631			}
 632		}
 633		path->slots[0]++;
 634	}
 635	ret = 0;
 636
 637	total_found += add_new_free_space(block_group, last,
 638				block_group->start + block_group->length);
 639	caching_ctl->progress = (u64)-1;
 640
 
 
 
 641out:
 642	btrfs_free_path(path);
 643	return ret;
 644}
 645
 
 
 
 
 
 
 646static noinline void caching_thread(struct btrfs_work *work)
 647{
 648	struct btrfs_block_group *block_group;
 649	struct btrfs_fs_info *fs_info;
 650	struct btrfs_caching_control *caching_ctl;
 651	int ret;
 652
 653	caching_ctl = container_of(work, struct btrfs_caching_control, work);
 654	block_group = caching_ctl->block_group;
 655	fs_info = block_group->fs_info;
 656
 657	mutex_lock(&caching_ctl->mutex);
 658	down_read(&fs_info->commit_root_sem);
 659
 
 660	if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
 661		ret = load_free_space_cache(block_group);
 662		if (ret == 1) {
 663			ret = 0;
 664			goto done;
 665		}
 666
 667		/*
 668		 * We failed to load the space cache, set ourselves to
 669		 * CACHE_STARTED and carry on.
 670		 */
 671		spin_lock(&block_group->lock);
 672		block_group->cached = BTRFS_CACHE_STARTED;
 673		spin_unlock(&block_group->lock);
 674		wake_up(&caching_ctl->wait);
 675	}
 676
 677	/*
 678	 * If we are in the transaction that populated the free space tree we
 679	 * can't actually cache from the free space tree as our commit root and
 680	 * real root are the same, so we could change the contents of the blocks
 681	 * while caching.  Instead do the slow caching in this case, and after
 682	 * the transaction has committed we will be safe.
 683	 */
 684	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
 685	    !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
 686		ret = load_free_space_tree(caching_ctl);
 687	else
 688		ret = load_extent_tree_free(caching_ctl);
 689done:
 690	spin_lock(&block_group->lock);
 691	block_group->caching_ctl = NULL;
 692	block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 693	spin_unlock(&block_group->lock);
 694
 695#ifdef CONFIG_BTRFS_DEBUG
 696	if (btrfs_should_fragment_free_space(block_group)) {
 697		u64 bytes_used;
 698
 699		spin_lock(&block_group->space_info->lock);
 700		spin_lock(&block_group->lock);
 701		bytes_used = block_group->length - block_group->used;
 702		block_group->space_info->bytes_used += bytes_used >> 1;
 703		spin_unlock(&block_group->lock);
 704		spin_unlock(&block_group->space_info->lock);
 705		fragment_free_space(block_group);
 706	}
 707#endif
 708
 709	caching_ctl->progress = (u64)-1;
 710
 711	up_read(&fs_info->commit_root_sem);
 712	btrfs_free_excluded_extents(block_group);
 713	mutex_unlock(&caching_ctl->mutex);
 714
 715	wake_up(&caching_ctl->wait);
 716
 717	btrfs_put_caching_control(caching_ctl);
 718	btrfs_put_block_group(block_group);
 719}
 720
 721int btrfs_cache_block_group(struct btrfs_block_group *cache, int load_cache_only)
 722{
 723	DEFINE_WAIT(wait);
 724	struct btrfs_fs_info *fs_info = cache->fs_info;
 725	struct btrfs_caching_control *caching_ctl = NULL;
 726	int ret = 0;
 727
 728	/* Allocator for zoned filesystems does not use the cache at all */
 729	if (btrfs_is_zoned(fs_info))
 730		return 0;
 731
 732	caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 733	if (!caching_ctl)
 734		return -ENOMEM;
 735
 736	INIT_LIST_HEAD(&caching_ctl->list);
 737	mutex_init(&caching_ctl->mutex);
 738	init_waitqueue_head(&caching_ctl->wait);
 739	caching_ctl->block_group = cache;
 740	caching_ctl->progress = cache->start;
 741	refcount_set(&caching_ctl->count, 2);
 742	btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 
 743
 744	spin_lock(&cache->lock);
 745	if (cache->cached != BTRFS_CACHE_NO) {
 746		kfree(caching_ctl);
 747
 748		caching_ctl = cache->caching_ctl;
 749		if (caching_ctl)
 750			refcount_inc(&caching_ctl->count);
 751		spin_unlock(&cache->lock);
 752		goto out;
 753	}
 754	WARN_ON(cache->caching_ctl);
 755	cache->caching_ctl = caching_ctl;
 756	if (btrfs_test_opt(fs_info, SPACE_CACHE))
 757		cache->cached = BTRFS_CACHE_FAST;
 758	else
 759		cache->cached = BTRFS_CACHE_STARTED;
 760	cache->has_caching_ctl = 1;
 761	spin_unlock(&cache->lock);
 762
 763	spin_lock(&fs_info->block_group_cache_lock);
 764	refcount_inc(&caching_ctl->count);
 765	list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 766	spin_unlock(&fs_info->block_group_cache_lock);
 767
 768	btrfs_get_block_group(cache);
 769
 770	btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 771out:
 772	if (load_cache_only && caching_ctl)
 773		btrfs_wait_space_cache_v1_finished(cache, caching_ctl);
 774	if (caching_ctl)
 775		btrfs_put_caching_control(caching_ctl);
 776
 777	return ret;
 778}
 779
 780static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
 781{
 782	u64 extra_flags = chunk_to_extended(flags) &
 783				BTRFS_EXTENDED_PROFILE_MASK;
 784
 785	write_seqlock(&fs_info->profiles_lock);
 786	if (flags & BTRFS_BLOCK_GROUP_DATA)
 787		fs_info->avail_data_alloc_bits &= ~extra_flags;
 788	if (flags & BTRFS_BLOCK_GROUP_METADATA)
 789		fs_info->avail_metadata_alloc_bits &= ~extra_flags;
 790	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
 791		fs_info->avail_system_alloc_bits &= ~extra_flags;
 792	write_sequnlock(&fs_info->profiles_lock);
 793}
 794
 795/*
 796 * Clear incompat bits for the following feature(s):
 797 *
 798 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
 799 *            in the whole filesystem
 800 *
 801 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
 802 */
 803static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
 804{
 805	bool found_raid56 = false;
 806	bool found_raid1c34 = false;
 807
 808	if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
 809	    (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
 810	    (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
 811		struct list_head *head = &fs_info->space_info;
 812		struct btrfs_space_info *sinfo;
 813
 814		list_for_each_entry_rcu(sinfo, head, list) {
 815			down_read(&sinfo->groups_sem);
 816			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
 817				found_raid56 = true;
 818			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
 819				found_raid56 = true;
 820			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
 821				found_raid1c34 = true;
 822			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
 823				found_raid1c34 = true;
 824			up_read(&sinfo->groups_sem);
 825		}
 826		if (!found_raid56)
 827			btrfs_clear_fs_incompat(fs_info, RAID56);
 828		if (!found_raid1c34)
 829			btrfs_clear_fs_incompat(fs_info, RAID1C34);
 830	}
 831}
 832
 833static int remove_block_group_item(struct btrfs_trans_handle *trans,
 834				   struct btrfs_path *path,
 835				   struct btrfs_block_group *block_group)
 836{
 837	struct btrfs_fs_info *fs_info = trans->fs_info;
 838	struct btrfs_root *root;
 839	struct btrfs_key key;
 840	int ret;
 841
 842	root = fs_info->extent_root;
 843	key.objectid = block_group->start;
 844	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
 845	key.offset = block_group->length;
 846
 847	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 848	if (ret > 0)
 849		ret = -ENOENT;
 850	if (ret < 0)
 851		return ret;
 852
 853	ret = btrfs_del_item(trans, root, path);
 854	return ret;
 855}
 856
 857int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 858			     u64 group_start, struct extent_map *em)
 859{
 860	struct btrfs_fs_info *fs_info = trans->fs_info;
 861	struct btrfs_path *path;
 862	struct btrfs_block_group *block_group;
 863	struct btrfs_free_cluster *cluster;
 864	struct inode *inode;
 865	struct kobject *kobj = NULL;
 866	int ret;
 867	int index;
 868	int factor;
 869	struct btrfs_caching_control *caching_ctl = NULL;
 870	bool remove_em;
 871	bool remove_rsv = false;
 872
 873	block_group = btrfs_lookup_block_group(fs_info, group_start);
 874	BUG_ON(!block_group);
 875	BUG_ON(!block_group->ro);
 876
 877	trace_btrfs_remove_block_group(block_group);
 878	/*
 879	 * Free the reserved super bytes from this block group before
 880	 * remove it.
 881	 */
 882	btrfs_free_excluded_extents(block_group);
 883	btrfs_free_ref_tree_range(fs_info, block_group->start,
 884				  block_group->length);
 885
 886	index = btrfs_bg_flags_to_raid_index(block_group->flags);
 887	factor = btrfs_bg_type_to_factor(block_group->flags);
 888
 889	/* make sure this block group isn't part of an allocation cluster */
 890	cluster = &fs_info->data_alloc_cluster;
 891	spin_lock(&cluster->refill_lock);
 892	btrfs_return_cluster_to_free_space(block_group, cluster);
 893	spin_unlock(&cluster->refill_lock);
 894
 895	/*
 896	 * make sure this block group isn't part of a metadata
 897	 * allocation cluster
 898	 */
 899	cluster = &fs_info->meta_alloc_cluster;
 900	spin_lock(&cluster->refill_lock);
 901	btrfs_return_cluster_to_free_space(block_group, cluster);
 902	spin_unlock(&cluster->refill_lock);
 903
 904	btrfs_clear_treelog_bg(block_group);
 
 905
 906	path = btrfs_alloc_path();
 907	if (!path) {
 908		ret = -ENOMEM;
 909		goto out;
 910	}
 911
 912	/*
 913	 * get the inode first so any iput calls done for the io_list
 914	 * aren't the final iput (no unlinks allowed now)
 915	 */
 916	inode = lookup_free_space_inode(block_group, path);
 917
 918	mutex_lock(&trans->transaction->cache_write_mutex);
 919	/*
 920	 * Make sure our free space cache IO is done before removing the
 921	 * free space inode
 922	 */
 923	spin_lock(&trans->transaction->dirty_bgs_lock);
 924	if (!list_empty(&block_group->io_list)) {
 925		list_del_init(&block_group->io_list);
 926
 927		WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
 928
 929		spin_unlock(&trans->transaction->dirty_bgs_lock);
 930		btrfs_wait_cache_io(trans, block_group, path);
 931		btrfs_put_block_group(block_group);
 932		spin_lock(&trans->transaction->dirty_bgs_lock);
 933	}
 934
 935	if (!list_empty(&block_group->dirty_list)) {
 936		list_del_init(&block_group->dirty_list);
 937		remove_rsv = true;
 938		btrfs_put_block_group(block_group);
 939	}
 940	spin_unlock(&trans->transaction->dirty_bgs_lock);
 941	mutex_unlock(&trans->transaction->cache_write_mutex);
 942
 943	ret = btrfs_remove_free_space_inode(trans, inode, block_group);
 944	if (ret)
 945		goto out;
 946
 947	spin_lock(&fs_info->block_group_cache_lock);
 948	rb_erase(&block_group->cache_node,
 949		 &fs_info->block_group_cache_tree);
 950	RB_CLEAR_NODE(&block_group->cache_node);
 951
 952	/* Once for the block groups rbtree */
 953	btrfs_put_block_group(block_group);
 954
 955	if (fs_info->first_logical_byte == block_group->start)
 956		fs_info->first_logical_byte = (u64)-1;
 957	spin_unlock(&fs_info->block_group_cache_lock);
 958
 959	down_write(&block_group->space_info->groups_sem);
 960	/*
 961	 * we must use list_del_init so people can check to see if they
 962	 * are still on the list after taking the semaphore
 963	 */
 964	list_del_init(&block_group->list);
 965	if (list_empty(&block_group->space_info->block_groups[index])) {
 966		kobj = block_group->space_info->block_group_kobjs[index];
 967		block_group->space_info->block_group_kobjs[index] = NULL;
 968		clear_avail_alloc_bits(fs_info, block_group->flags);
 969	}
 970	up_write(&block_group->space_info->groups_sem);
 971	clear_incompat_bg_bits(fs_info, block_group->flags);
 972	if (kobj) {
 973		kobject_del(kobj);
 974		kobject_put(kobj);
 975	}
 976
 977	if (block_group->has_caching_ctl)
 978		caching_ctl = btrfs_get_caching_control(block_group);
 979	if (block_group->cached == BTRFS_CACHE_STARTED)
 980		btrfs_wait_block_group_cache_done(block_group);
 981	if (block_group->has_caching_ctl) {
 982		spin_lock(&fs_info->block_group_cache_lock);
 983		if (!caching_ctl) {
 984			struct btrfs_caching_control *ctl;
 985
 986			list_for_each_entry(ctl,
 987				    &fs_info->caching_block_groups, list)
 988				if (ctl->block_group == block_group) {
 989					caching_ctl = ctl;
 990					refcount_inc(&caching_ctl->count);
 991					break;
 992				}
 993		}
 994		if (caching_ctl)
 995			list_del_init(&caching_ctl->list);
 996		spin_unlock(&fs_info->block_group_cache_lock);
 997		if (caching_ctl) {
 998			/* Once for the caching bgs list and once for us. */
 999			btrfs_put_caching_control(caching_ctl);
1000			btrfs_put_caching_control(caching_ctl);
1001		}
 
 
 
 
 
 
 
 
 
1002	}
1003
1004	spin_lock(&trans->transaction->dirty_bgs_lock);
1005	WARN_ON(!list_empty(&block_group->dirty_list));
1006	WARN_ON(!list_empty(&block_group->io_list));
1007	spin_unlock(&trans->transaction->dirty_bgs_lock);
1008
1009	btrfs_remove_free_space_cache(block_group);
1010
1011	spin_lock(&block_group->space_info->lock);
1012	list_del_init(&block_group->ro_list);
1013
1014	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1015		WARN_ON(block_group->space_info->total_bytes
1016			< block_group->length);
1017		WARN_ON(block_group->space_info->bytes_readonly
1018			< block_group->length - block_group->zone_unusable);
1019		WARN_ON(block_group->space_info->bytes_zone_unusable
1020			< block_group->zone_unusable);
1021		WARN_ON(block_group->space_info->disk_total
1022			< block_group->length * factor);
1023	}
1024	block_group->space_info->total_bytes -= block_group->length;
1025	block_group->space_info->bytes_readonly -=
1026		(block_group->length - block_group->zone_unusable);
1027	block_group->space_info->bytes_zone_unusable -=
1028		block_group->zone_unusable;
1029	block_group->space_info->disk_total -= block_group->length * factor;
1030
1031	spin_unlock(&block_group->space_info->lock);
1032
1033	/*
1034	 * Remove the free space for the block group from the free space tree
1035	 * and the block group's item from the extent tree before marking the
1036	 * block group as removed. This is to prevent races with tasks that
1037	 * freeze and unfreeze a block group, this task and another task
1038	 * allocating a new block group - the unfreeze task ends up removing
1039	 * the block group's extent map before the task calling this function
1040	 * deletes the block group item from the extent tree, allowing for
1041	 * another task to attempt to create another block group with the same
1042	 * item key (and failing with -EEXIST and a transaction abort).
1043	 */
1044	ret = remove_block_group_free_space(trans, block_group);
1045	if (ret)
1046		goto out;
1047
1048	ret = remove_block_group_item(trans, path, block_group);
1049	if (ret < 0)
1050		goto out;
1051
1052	spin_lock(&block_group->lock);
1053	block_group->removed = 1;
 
1054	/*
1055	 * At this point trimming or scrub can't start on this block group,
1056	 * because we removed the block group from the rbtree
1057	 * fs_info->block_group_cache_tree so no one can't find it anymore and
1058	 * even if someone already got this block group before we removed it
1059	 * from the rbtree, they have already incremented block_group->frozen -
1060	 * if they didn't, for the trimming case they won't find any free space
1061	 * entries because we already removed them all when we called
1062	 * btrfs_remove_free_space_cache().
1063	 *
1064	 * And we must not remove the extent map from the fs_info->mapping_tree
1065	 * to prevent the same logical address range and physical device space
1066	 * ranges from being reused for a new block group. This is needed to
1067	 * avoid races with trimming and scrub.
1068	 *
1069	 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1070	 * completely transactionless, so while it is trimming a range the
1071	 * currently running transaction might finish and a new one start,
1072	 * allowing for new block groups to be created that can reuse the same
1073	 * physical device locations unless we take this special care.
1074	 *
1075	 * There may also be an implicit trim operation if the file system
1076	 * is mounted with -odiscard. The same protections must remain
1077	 * in place until the extents have been discarded completely when
1078	 * the transaction commit has completed.
1079	 */
1080	remove_em = (atomic_read(&block_group->frozen) == 0);
1081	spin_unlock(&block_group->lock);
1082
1083	if (remove_em) {
1084		struct extent_map_tree *em_tree;
1085
1086		em_tree = &fs_info->mapping_tree;
1087		write_lock(&em_tree->lock);
1088		remove_extent_mapping(em_tree, em);
1089		write_unlock(&em_tree->lock);
1090		/* once for the tree */
1091		free_extent_map(em);
1092	}
1093
1094out:
1095	/* Once for the lookup reference */
1096	btrfs_put_block_group(block_group);
1097	if (remove_rsv)
1098		btrfs_delayed_refs_rsv_release(fs_info, 1);
1099	btrfs_free_path(path);
1100	return ret;
1101}
1102
1103struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1104		struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1105{
1106	struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1107	struct extent_map *em;
1108	struct map_lookup *map;
1109	unsigned int num_items;
1110
1111	read_lock(&em_tree->lock);
1112	em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1113	read_unlock(&em_tree->lock);
1114	ASSERT(em && em->start == chunk_offset);
1115
1116	/*
1117	 * We need to reserve 3 + N units from the metadata space info in order
1118	 * to remove a block group (done at btrfs_remove_chunk() and at
1119	 * btrfs_remove_block_group()), which are used for:
1120	 *
1121	 * 1 unit for adding the free space inode's orphan (located in the tree
1122	 * of tree roots).
1123	 * 1 unit for deleting the block group item (located in the extent
1124	 * tree).
1125	 * 1 unit for deleting the free space item (located in tree of tree
1126	 * roots).
1127	 * N units for deleting N device extent items corresponding to each
1128	 * stripe (located in the device tree).
1129	 *
1130	 * In order to remove a block group we also need to reserve units in the
1131	 * system space info in order to update the chunk tree (update one or
1132	 * more device items and remove one chunk item), but this is done at
1133	 * btrfs_remove_chunk() through a call to check_system_chunk().
1134	 */
1135	map = em->map_lookup;
1136	num_items = 3 + map->num_stripes;
1137	free_extent_map(em);
1138
1139	return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
1140							   num_items);
1141}
1142
1143/*
1144 * Mark block group @cache read-only, so later write won't happen to block
1145 * group @cache.
1146 *
1147 * If @force is not set, this function will only mark the block group readonly
1148 * if we have enough free space (1M) in other metadata/system block groups.
1149 * If @force is not set, this function will mark the block group readonly
1150 * without checking free space.
1151 *
1152 * NOTE: This function doesn't care if other block groups can contain all the
1153 * data in this block group. That check should be done by relocation routine,
1154 * not this function.
1155 */
1156static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1157{
1158	struct btrfs_space_info *sinfo = cache->space_info;
1159	u64 num_bytes;
1160	int ret = -ENOSPC;
1161
1162	spin_lock(&sinfo->lock);
1163	spin_lock(&cache->lock);
1164
1165	if (cache->swap_extents) {
1166		ret = -ETXTBSY;
1167		goto out;
1168	}
1169
1170	if (cache->ro) {
1171		cache->ro++;
1172		ret = 0;
1173		goto out;
1174	}
1175
1176	num_bytes = cache->length - cache->reserved - cache->pinned -
1177		    cache->bytes_super - cache->zone_unusable - cache->used;
1178
1179	/*
1180	 * Data never overcommits, even in mixed mode, so do just the straight
1181	 * check of left over space in how much we have allocated.
1182	 */
1183	if (force) {
1184		ret = 0;
1185	} else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1186		u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1187
1188		/*
1189		 * Here we make sure if we mark this bg RO, we still have enough
1190		 * free space as buffer.
1191		 */
1192		if (sinfo_used + num_bytes <= sinfo->total_bytes)
1193			ret = 0;
1194	} else {
1195		/*
1196		 * We overcommit metadata, so we need to do the
1197		 * btrfs_can_overcommit check here, and we need to pass in
1198		 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1199		 * leeway to allow us to mark this block group as read only.
1200		 */
1201		if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1202					 BTRFS_RESERVE_NO_FLUSH))
1203			ret = 0;
1204	}
1205
1206	if (!ret) {
1207		sinfo->bytes_readonly += num_bytes;
1208		if (btrfs_is_zoned(cache->fs_info)) {
1209			/* Migrate zone_unusable bytes to readonly */
1210			sinfo->bytes_readonly += cache->zone_unusable;
1211			sinfo->bytes_zone_unusable -= cache->zone_unusable;
1212			cache->zone_unusable = 0;
1213		}
1214		cache->ro++;
1215		list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1216	}
1217out:
1218	spin_unlock(&cache->lock);
1219	spin_unlock(&sinfo->lock);
1220	if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1221		btrfs_info(cache->fs_info,
1222			"unable to make block group %llu ro", cache->start);
1223		btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1224	}
1225	return ret;
1226}
1227
1228static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1229				 struct btrfs_block_group *bg)
1230{
1231	struct btrfs_fs_info *fs_info = bg->fs_info;
1232	struct btrfs_transaction *prev_trans = NULL;
1233	const u64 start = bg->start;
1234	const u64 end = start + bg->length - 1;
1235	int ret;
1236
1237	spin_lock(&fs_info->trans_lock);
1238	if (trans->transaction->list.prev != &fs_info->trans_list) {
1239		prev_trans = list_last_entry(&trans->transaction->list,
1240					     struct btrfs_transaction, list);
1241		refcount_inc(&prev_trans->use_count);
1242	}
1243	spin_unlock(&fs_info->trans_lock);
1244
1245	/*
1246	 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1247	 * btrfs_finish_extent_commit(). If we are at transaction N, another
1248	 * task might be running finish_extent_commit() for the previous
1249	 * transaction N - 1, and have seen a range belonging to the block
1250	 * group in pinned_extents before we were able to clear the whole block
1251	 * group range from pinned_extents. This means that task can lookup for
1252	 * the block group after we unpinned it from pinned_extents and removed
1253	 * it, leading to a BUG_ON() at unpin_extent_range().
1254	 */
1255	mutex_lock(&fs_info->unused_bg_unpin_mutex);
1256	if (prev_trans) {
1257		ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1258					EXTENT_DIRTY);
1259		if (ret)
1260			goto out;
1261	}
1262
1263	ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1264				EXTENT_DIRTY);
1265out:
1266	mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1267	if (prev_trans)
1268		btrfs_put_transaction(prev_trans);
1269
1270	return ret == 0;
1271}
1272
1273/*
1274 * Process the unused_bgs list and remove any that don't have any allocated
1275 * space inside of them.
1276 */
1277void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1278{
 
1279	struct btrfs_block_group *block_group;
1280	struct btrfs_space_info *space_info;
1281	struct btrfs_trans_handle *trans;
1282	const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1283	int ret = 0;
1284
1285	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1286		return;
1287
 
 
 
1288	/*
1289	 * Long running balances can keep us blocked here for eternity, so
1290	 * simply skip deletion if we're unable to get the mutex.
1291	 */
1292	if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1293		return;
1294
1295	spin_lock(&fs_info->unused_bgs_lock);
1296	while (!list_empty(&fs_info->unused_bgs)) {
 
1297		int trimming;
1298
1299		block_group = list_first_entry(&fs_info->unused_bgs,
1300					       struct btrfs_block_group,
1301					       bg_list);
1302		list_del_init(&block_group->bg_list);
1303
1304		space_info = block_group->space_info;
1305
1306		if (ret || btrfs_mixed_space_info(space_info)) {
1307			btrfs_put_block_group(block_group);
1308			continue;
1309		}
1310		spin_unlock(&fs_info->unused_bgs_lock);
1311
1312		btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1313
1314		/* Don't want to race with allocators so take the groups_sem */
1315		down_write(&space_info->groups_sem);
1316
1317		/*
1318		 * Async discard moves the final block group discard to be prior
1319		 * to the unused_bgs code path.  Therefore, if it's not fully
1320		 * trimmed, punt it back to the async discard lists.
1321		 */
1322		if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1323		    !btrfs_is_free_space_trimmed(block_group)) {
1324			trace_btrfs_skip_unused_block_group(block_group);
1325			up_write(&space_info->groups_sem);
1326			/* Requeue if we failed because of async discard */
1327			btrfs_discard_queue_work(&fs_info->discard_ctl,
1328						 block_group);
1329			goto next;
1330		}
1331
 
1332		spin_lock(&block_group->lock);
1333		if (block_group->reserved || block_group->pinned ||
1334		    block_group->used || block_group->ro ||
1335		    list_is_singular(&block_group->list)) {
1336			/*
1337			 * We want to bail if we made new allocations or have
1338			 * outstanding allocations in this block group.  We do
1339			 * the ro check in case balance is currently acting on
1340			 * this block group.
1341			 */
1342			trace_btrfs_skip_unused_block_group(block_group);
1343			spin_unlock(&block_group->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1344			up_write(&space_info->groups_sem);
1345			goto next;
1346		}
 
1347		spin_unlock(&block_group->lock);
 
1348
1349		/* We don't want to force the issue, only flip if it's ok. */
1350		ret = inc_block_group_ro(block_group, 0);
1351		up_write(&space_info->groups_sem);
1352		if (ret < 0) {
1353			ret = 0;
1354			goto next;
1355		}
1356
 
 
 
 
 
 
 
 
1357		/*
1358		 * Want to do this before we do anything else so we can recover
1359		 * properly if we fail to join the transaction.
1360		 */
1361		trans = btrfs_start_trans_remove_block_group(fs_info,
1362						     block_group->start);
1363		if (IS_ERR(trans)) {
1364			btrfs_dec_block_group_ro(block_group);
1365			ret = PTR_ERR(trans);
1366			goto next;
1367		}
1368
1369		/*
1370		 * We could have pending pinned extents for this block group,
1371		 * just delete them, we don't care about them anymore.
1372		 */
1373		if (!clean_pinned_extents(trans, block_group)) {
1374			btrfs_dec_block_group_ro(block_group);
1375			goto end_trans;
1376		}
1377
1378		/*
1379		 * At this point, the block_group is read only and should fail
1380		 * new allocations.  However, btrfs_finish_extent_commit() can
1381		 * cause this block_group to be placed back on the discard
1382		 * lists because now the block_group isn't fully discarded.
1383		 * Bail here and try again later after discarding everything.
1384		 */
1385		spin_lock(&fs_info->discard_ctl.lock);
1386		if (!list_empty(&block_group->discard_list)) {
1387			spin_unlock(&fs_info->discard_ctl.lock);
1388			btrfs_dec_block_group_ro(block_group);
1389			btrfs_discard_queue_work(&fs_info->discard_ctl,
1390						 block_group);
1391			goto end_trans;
1392		}
1393		spin_unlock(&fs_info->discard_ctl.lock);
1394
1395		/* Reset pinned so btrfs_put_block_group doesn't complain */
1396		spin_lock(&space_info->lock);
1397		spin_lock(&block_group->lock);
1398
1399		btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1400						     -block_group->pinned);
1401		space_info->bytes_readonly += block_group->pinned;
1402		block_group->pinned = 0;
1403
1404		spin_unlock(&block_group->lock);
1405		spin_unlock(&space_info->lock);
1406
1407		/*
1408		 * The normal path here is an unused block group is passed here,
1409		 * then trimming is handled in the transaction commit path.
1410		 * Async discard interposes before this to do the trimming
1411		 * before coming down the unused block group path as trimming
1412		 * will no longer be done later in the transaction commit path.
1413		 */
1414		if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1415			goto flip_async;
1416
1417		/*
1418		 * DISCARD can flip during remount. On zoned filesystems, we
1419		 * need to reset sequential-required zones.
1420		 */
1421		trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1422				btrfs_is_zoned(fs_info);
1423
1424		/* Implicit trim during transaction commit. */
1425		if (trimming)
1426			btrfs_freeze_block_group(block_group);
1427
1428		/*
1429		 * Btrfs_remove_chunk will abort the transaction if things go
1430		 * horribly wrong.
1431		 */
1432		ret = btrfs_remove_chunk(trans, block_group->start);
1433
1434		if (ret) {
1435			if (trimming)
1436				btrfs_unfreeze_block_group(block_group);
1437			goto end_trans;
1438		}
1439
1440		/*
1441		 * If we're not mounted with -odiscard, we can just forget
1442		 * about this block group. Otherwise we'll need to wait
1443		 * until transaction commit to do the actual discard.
1444		 */
1445		if (trimming) {
1446			spin_lock(&fs_info->unused_bgs_lock);
1447			/*
1448			 * A concurrent scrub might have added us to the list
1449			 * fs_info->unused_bgs, so use a list_move operation
1450			 * to add the block group to the deleted_bgs list.
1451			 */
1452			list_move(&block_group->bg_list,
1453				  &trans->transaction->deleted_bgs);
1454			spin_unlock(&fs_info->unused_bgs_lock);
1455			btrfs_get_block_group(block_group);
1456		}
1457end_trans:
1458		btrfs_end_transaction(trans);
1459next:
1460		btrfs_put_block_group(block_group);
1461		spin_lock(&fs_info->unused_bgs_lock);
1462	}
 
1463	spin_unlock(&fs_info->unused_bgs_lock);
1464	mutex_unlock(&fs_info->reclaim_bgs_lock);
1465	return;
1466
1467flip_async:
1468	btrfs_end_transaction(trans);
 
 
 
1469	mutex_unlock(&fs_info->reclaim_bgs_lock);
1470	btrfs_put_block_group(block_group);
1471	btrfs_discard_punt_unused_bgs_list(fs_info);
1472}
1473
1474void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1475{
1476	struct btrfs_fs_info *fs_info = bg->fs_info;
1477
1478	spin_lock(&fs_info->unused_bgs_lock);
1479	if (list_empty(&bg->bg_list)) {
1480		btrfs_get_block_group(bg);
1481		trace_btrfs_add_unused_block_group(bg);
1482		list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
 
 
 
 
1483	}
1484	spin_unlock(&fs_info->unused_bgs_lock);
1485}
1486
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1487void btrfs_reclaim_bgs_work(struct work_struct *work)
1488{
1489	struct btrfs_fs_info *fs_info =
1490		container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1491	struct btrfs_block_group *bg;
1492	struct btrfs_space_info *space_info;
1493	LIST_HEAD(again_list);
1494
1495	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1496		return;
1497
1498	if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE))
 
 
 
1499		return;
1500
 
 
 
 
 
 
 
1501	/*
1502	 * Long running balances can keep us blocked here for eternity, so
1503	 * simply skip reclaim if we're unable to get the mutex.
1504	 */
1505	if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1506		btrfs_exclop_finish(fs_info);
 
1507		return;
1508	}
1509
1510	spin_lock(&fs_info->unused_bgs_lock);
 
 
 
 
 
 
1511	while (!list_empty(&fs_info->reclaim_bgs)) {
1512		u64 zone_unusable;
1513		int ret = 0;
1514
1515		bg = list_first_entry(&fs_info->reclaim_bgs,
1516				      struct btrfs_block_group,
1517				      bg_list);
1518		list_del_init(&bg->bg_list);
1519
1520		space_info = bg->space_info;
1521		spin_unlock(&fs_info->unused_bgs_lock);
1522
1523		/* Don't race with allocators so take the groups_sem */
1524		down_write(&space_info->groups_sem);
1525
1526		spin_lock(&bg->lock);
1527		if (bg->reserved || bg->pinned || bg->ro) {
1528			/*
1529			 * We want to bail if we made new allocations or have
1530			 * outstanding allocations in this block group.  We do
1531			 * the ro check in case balance is currently acting on
1532			 * this block group.
1533			 */
1534			spin_unlock(&bg->lock);
1535			up_write(&space_info->groups_sem);
1536			goto next;
1537		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1538		spin_unlock(&bg->lock);
1539
1540		/* Get out fast, in case we're unmounting the filesystem */
1541		if (btrfs_fs_closing(fs_info)) {
 
 
 
 
 
 
 
1542			up_write(&space_info->groups_sem);
1543			goto next;
1544		}
1545
1546		/*
1547		 * Cache the zone_unusable value before turning the block group
1548		 * to read only. As soon as the blog group is read only it's
1549		 * zone_unusable value gets moved to the block group's read-only
1550		 * bytes and isn't available for calculations anymore.
1551		 */
1552		zone_unusable = bg->zone_unusable;
1553		ret = inc_block_group_ro(bg, 0);
1554		up_write(&space_info->groups_sem);
1555		if (ret < 0)
1556			goto next;
1557
1558		btrfs_info(fs_info,
1559			"reclaiming chunk %llu with %llu%% used %llu%% unusable",
1560				bg->start, div_u64(bg->used * 100, bg->length),
 
1561				div64_u64(zone_unusable * 100, bg->length));
1562		trace_btrfs_reclaim_block_group(bg);
1563		ret = btrfs_relocate_chunk(fs_info, bg->start);
1564		if (ret && ret != -EAGAIN)
 
1565			btrfs_err(fs_info, "error relocating chunk %llu",
1566				  bg->start);
 
1567
1568next:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1569		spin_lock(&fs_info->unused_bgs_lock);
1570		if (ret == -EAGAIN && list_empty(&bg->bg_list))
1571			list_add_tail(&bg->bg_list, &again_list);
1572		else
1573			btrfs_put_block_group(bg);
1574	}
1575	list_splice_tail(&again_list, &fs_info->reclaim_bgs);
1576	spin_unlock(&fs_info->unused_bgs_lock);
1577	mutex_unlock(&fs_info->reclaim_bgs_lock);
 
1578	btrfs_exclop_finish(fs_info);
 
1579}
1580
1581void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1582{
1583	spin_lock(&fs_info->unused_bgs_lock);
1584	if (!list_empty(&fs_info->reclaim_bgs))
1585		queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1586	spin_unlock(&fs_info->unused_bgs_lock);
1587}
1588
1589void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1590{
1591	struct btrfs_fs_info *fs_info = bg->fs_info;
1592
1593	spin_lock(&fs_info->unused_bgs_lock);
1594	if (list_empty(&bg->bg_list)) {
1595		btrfs_get_block_group(bg);
1596		trace_btrfs_add_reclaim_block_group(bg);
1597		list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1598	}
1599	spin_unlock(&fs_info->unused_bgs_lock);
1600}
1601
1602static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1603			   struct btrfs_path *path)
1604{
1605	struct extent_map_tree *em_tree;
1606	struct extent_map *em;
1607	struct btrfs_block_group_item bg;
1608	struct extent_buffer *leaf;
1609	int slot;
1610	u64 flags;
1611	int ret = 0;
1612
1613	slot = path->slots[0];
1614	leaf = path->nodes[0];
1615
1616	em_tree = &fs_info->mapping_tree;
1617	read_lock(&em_tree->lock);
1618	em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1619	read_unlock(&em_tree->lock);
1620	if (!em) {
1621		btrfs_err(fs_info,
1622			  "logical %llu len %llu found bg but no related chunk",
1623			  key->objectid, key->offset);
1624		return -ENOENT;
1625	}
1626
1627	if (em->start != key->objectid || em->len != key->offset) {
1628		btrfs_err(fs_info,
1629			"block group %llu len %llu mismatch with chunk %llu len %llu",
1630			key->objectid, key->offset, em->start, em->len);
1631		ret = -EUCLEAN;
1632		goto out_free_em;
1633	}
1634
1635	read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1636			   sizeof(bg));
1637	flags = btrfs_stack_block_group_flags(&bg) &
1638		BTRFS_BLOCK_GROUP_TYPE_MASK;
1639
1640	if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1641		btrfs_err(fs_info,
1642"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1643			  key->objectid, key->offset, flags,
1644			  (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1645		ret = -EUCLEAN;
1646	}
1647
1648out_free_em:
1649	free_extent_map(em);
1650	return ret;
1651}
1652
1653static int find_first_block_group(struct btrfs_fs_info *fs_info,
1654				  struct btrfs_path *path,
1655				  struct btrfs_key *key)
1656{
1657	struct btrfs_root *root = fs_info->extent_root;
1658	int ret;
1659	struct btrfs_key found_key;
1660	struct extent_buffer *leaf;
1661	int slot;
1662
1663	ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1664	if (ret < 0)
1665		return ret;
1666
1667	while (1) {
1668		slot = path->slots[0];
1669		leaf = path->nodes[0];
1670		if (slot >= btrfs_header_nritems(leaf)) {
1671			ret = btrfs_next_leaf(root, path);
1672			if (ret == 0)
1673				continue;
1674			if (ret < 0)
1675				goto out;
1676			break;
1677		}
1678		btrfs_item_key_to_cpu(leaf, &found_key, slot);
1679
 
1680		if (found_key.objectid >= key->objectid &&
1681		    found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1682			ret = read_bg_from_eb(fs_info, &found_key, path);
1683			break;
1684		}
1685
1686		path->slots[0]++;
1687	}
1688out:
1689	return ret;
1690}
1691
1692static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1693{
1694	u64 extra_flags = chunk_to_extended(flags) &
1695				BTRFS_EXTENDED_PROFILE_MASK;
1696
1697	write_seqlock(&fs_info->profiles_lock);
1698	if (flags & BTRFS_BLOCK_GROUP_DATA)
1699		fs_info->avail_data_alloc_bits |= extra_flags;
1700	if (flags & BTRFS_BLOCK_GROUP_METADATA)
1701		fs_info->avail_metadata_alloc_bits |= extra_flags;
1702	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1703		fs_info->avail_system_alloc_bits |= extra_flags;
1704	write_sequnlock(&fs_info->profiles_lock);
1705}
1706
1707/**
1708 * Map a physical disk address to a list of logical addresses
1709 *
1710 * @fs_info:       the filesystem
1711 * @chunk_start:   logical address of block group
1712 * @bdev:	   physical device to resolve, can be NULL to indicate any device
1713 * @physical:	   physical address to map to logical addresses
1714 * @logical:	   return array of logical addresses which map to @physical
1715 * @naddrs:	   length of @logical
1716 * @stripe_len:    size of IO stripe for the given block group
1717 *
1718 * Maps a particular @physical disk address to a list of @logical addresses.
1719 * Used primarily to exclude those portions of a block group that contain super
1720 * block copies.
1721 */
1722int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1723		     struct block_device *bdev, u64 physical, u64 **logical,
1724		     int *naddrs, int *stripe_len)
1725{
1726	struct extent_map *em;
1727	struct map_lookup *map;
1728	u64 *buf;
1729	u64 bytenr;
1730	u64 data_stripe_length;
1731	u64 io_stripe_size;
1732	int i, nr = 0;
1733	int ret = 0;
1734
1735	em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1736	if (IS_ERR(em))
1737		return -EIO;
1738
1739	map = em->map_lookup;
1740	data_stripe_length = em->orig_block_len;
1741	io_stripe_size = map->stripe_len;
1742	chunk_start = em->start;
1743
1744	/* For RAID5/6 adjust to a full IO stripe length */
1745	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1746		io_stripe_size = map->stripe_len * nr_data_stripes(map);
1747
1748	buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1749	if (!buf) {
1750		ret = -ENOMEM;
1751		goto out;
1752	}
1753
1754	for (i = 0; i < map->num_stripes; i++) {
1755		bool already_inserted = false;
1756		u64 stripe_nr;
1757		u64 offset;
1758		int j;
1759
1760		if (!in_range(physical, map->stripes[i].physical,
1761			      data_stripe_length))
1762			continue;
1763
1764		if (bdev && map->stripes[i].dev->bdev != bdev)
1765			continue;
1766
1767		stripe_nr = physical - map->stripes[i].physical;
1768		stripe_nr = div64_u64_rem(stripe_nr, map->stripe_len, &offset);
1769
1770		if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1771			stripe_nr = stripe_nr * map->num_stripes + i;
1772			stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1773		} else if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1774			stripe_nr = stripe_nr * map->num_stripes + i;
1775		}
1776		/*
1777		 * The remaining case would be for RAID56, multiply by
1778		 * nr_data_stripes().  Alternatively, just use rmap_len below
1779		 * instead of map->stripe_len
1780		 */
1781
1782		bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
1783
1784		/* Ensure we don't add duplicate addresses */
1785		for (j = 0; j < nr; j++) {
1786			if (buf[j] == bytenr) {
1787				already_inserted = true;
1788				break;
1789			}
1790		}
1791
1792		if (!already_inserted)
1793			buf[nr++] = bytenr;
1794	}
1795
1796	*logical = buf;
1797	*naddrs = nr;
1798	*stripe_len = io_stripe_size;
1799out:
1800	free_extent_map(em);
1801	return ret;
1802}
1803
1804static int exclude_super_stripes(struct btrfs_block_group *cache)
1805{
1806	struct btrfs_fs_info *fs_info = cache->fs_info;
1807	const bool zoned = btrfs_is_zoned(fs_info);
1808	u64 bytenr;
1809	u64 *logical;
1810	int stripe_len;
1811	int i, nr, ret;
1812
1813	if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1814		stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1815		cache->bytes_super += stripe_len;
1816		ret = btrfs_add_excluded_extent(fs_info, cache->start,
1817						stripe_len);
 
1818		if (ret)
1819			return ret;
1820	}
1821
1822	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1823		bytenr = btrfs_sb_offset(i);
1824		ret = btrfs_rmap_block(fs_info, cache->start, NULL,
1825				       bytenr, &logical, &nr, &stripe_len);
1826		if (ret)
1827			return ret;
1828
1829		/* Shouldn't have super stripes in sequential zones */
1830		if (zoned && nr) {
 
1831			btrfs_err(fs_info,
1832			"zoned: block group %llu must not contain super block",
1833				  cache->start);
1834			return -EUCLEAN;
1835		}
1836
1837		while (nr--) {
1838			u64 len = min_t(u64, stripe_len,
1839				cache->start + cache->length - logical[nr]);
1840
1841			cache->bytes_super += len;
1842			ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1843							len);
 
1844			if (ret) {
1845				kfree(logical);
1846				return ret;
1847			}
1848		}
1849
1850		kfree(logical);
1851	}
1852	return 0;
1853}
1854
1855static void link_block_group(struct btrfs_block_group *cache)
1856{
1857	struct btrfs_space_info *space_info = cache->space_info;
1858	int index = btrfs_bg_flags_to_raid_index(cache->flags);
1859
1860	down_write(&space_info->groups_sem);
1861	list_add_tail(&cache->list, &space_info->block_groups[index]);
1862	up_write(&space_info->groups_sem);
1863}
1864
1865static struct btrfs_block_group *btrfs_create_block_group_cache(
1866		struct btrfs_fs_info *fs_info, u64 start)
1867{
1868	struct btrfs_block_group *cache;
1869
1870	cache = kzalloc(sizeof(*cache), GFP_NOFS);
1871	if (!cache)
1872		return NULL;
1873
1874	cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1875					GFP_NOFS);
1876	if (!cache->free_space_ctl) {
1877		kfree(cache);
1878		return NULL;
1879	}
1880
1881	cache->start = start;
1882
1883	cache->fs_info = fs_info;
1884	cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1885
1886	cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1887
1888	refcount_set(&cache->refs, 1);
1889	spin_lock_init(&cache->lock);
1890	init_rwsem(&cache->data_rwsem);
1891	INIT_LIST_HEAD(&cache->list);
1892	INIT_LIST_HEAD(&cache->cluster_list);
1893	INIT_LIST_HEAD(&cache->bg_list);
1894	INIT_LIST_HEAD(&cache->ro_list);
1895	INIT_LIST_HEAD(&cache->discard_list);
1896	INIT_LIST_HEAD(&cache->dirty_list);
1897	INIT_LIST_HEAD(&cache->io_list);
 
1898	btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1899	atomic_set(&cache->frozen, 0);
1900	mutex_init(&cache->free_space_lock);
1901	btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
1902
1903	return cache;
1904}
1905
1906/*
1907 * Iterate all chunks and verify that each of them has the corresponding block
1908 * group
1909 */
1910static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
1911{
1912	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
1913	struct extent_map *em;
1914	struct btrfs_block_group *bg;
1915	u64 start = 0;
1916	int ret = 0;
1917
1918	while (1) {
1919		read_lock(&map_tree->lock);
 
 
1920		/*
1921		 * lookup_extent_mapping will return the first extent map
1922		 * intersecting the range, so setting @len to 1 is enough to
1923		 * get the first chunk.
1924		 */
1925		em = lookup_extent_mapping(map_tree, start, 1);
1926		read_unlock(&map_tree->lock);
1927		if (!em)
1928			break;
1929
1930		bg = btrfs_lookup_block_group(fs_info, em->start);
1931		if (!bg) {
1932			btrfs_err(fs_info,
1933	"chunk start=%llu len=%llu doesn't have corresponding block group",
1934				     em->start, em->len);
1935			ret = -EUCLEAN;
1936			free_extent_map(em);
1937			break;
1938		}
1939		if (bg->start != em->start || bg->length != em->len ||
1940		    (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
1941		    (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1942			btrfs_err(fs_info,
1943"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1944				em->start, em->len,
1945				em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
1946				bg->start, bg->length,
1947				bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
1948			ret = -EUCLEAN;
1949			free_extent_map(em);
1950			btrfs_put_block_group(bg);
1951			break;
1952		}
1953		start = em->start + em->len;
1954		free_extent_map(em);
1955		btrfs_put_block_group(bg);
1956	}
1957	return ret;
1958}
1959
1960static int read_one_block_group(struct btrfs_fs_info *info,
1961				struct btrfs_block_group_item *bgi,
1962				const struct btrfs_key *key,
1963				int need_clear)
1964{
1965	struct btrfs_block_group *cache;
1966	struct btrfs_space_info *space_info;
1967	const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
1968	int ret;
1969
1970	ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
1971
1972	cache = btrfs_create_block_group_cache(info, key->objectid);
1973	if (!cache)
1974		return -ENOMEM;
1975
1976	cache->length = key->offset;
1977	cache->used = btrfs_stack_block_group_used(bgi);
 
1978	cache->flags = btrfs_stack_block_group_flags(bgi);
 
1979
1980	set_free_space_tree_thresholds(cache);
1981
1982	if (need_clear) {
1983		/*
1984		 * When we mount with old space cache, we need to
1985		 * set BTRFS_DC_CLEAR and set dirty flag.
1986		 *
1987		 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
1988		 *    truncate the old free space cache inode and
1989		 *    setup a new one.
1990		 * b) Setting 'dirty flag' makes sure that we flush
1991		 *    the new space cache info onto disk.
1992		 */
1993		if (btrfs_test_opt(info, SPACE_CACHE))
1994			cache->disk_cache_state = BTRFS_DC_CLEAR;
1995	}
1996	if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
1997	    (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
1998			btrfs_err(info,
1999"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2000				  cache->start);
2001			ret = -EINVAL;
2002			goto error;
2003	}
2004
2005	ret = btrfs_load_block_group_zone_info(cache, false);
2006	if (ret) {
2007		btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2008			  cache->start);
2009		goto error;
2010	}
2011
2012	/*
2013	 * We need to exclude the super stripes now so that the space info has
2014	 * super bytes accounted for, otherwise we'll think we have more space
2015	 * than we actually do.
2016	 */
2017	ret = exclude_super_stripes(cache);
2018	if (ret) {
2019		/* We may have excluded something, so call this just in case. */
2020		btrfs_free_excluded_extents(cache);
2021		goto error;
2022	}
2023
2024	/*
2025	 * For zoned filesystem, space after the allocation offset is the only
2026	 * free space for a block group. So, we don't need any caching work.
2027	 * btrfs_calc_zone_unusable() will set the amount of free space and
2028	 * zone_unusable space.
2029	 *
2030	 * For regular filesystem, check for two cases, either we are full, and
2031	 * therefore don't need to bother with the caching work since we won't
2032	 * find any space, or we are empty, and we can just add all the space
2033	 * in and be done with it.  This saves us _a_lot_ of time, particularly
2034	 * in the full case.
2035	 */
2036	if (btrfs_is_zoned(info)) {
2037		btrfs_calc_zone_unusable(cache);
 
 
2038	} else if (cache->length == cache->used) {
2039		cache->last_byte_to_unpin = (u64)-1;
2040		cache->cached = BTRFS_CACHE_FINISHED;
2041		btrfs_free_excluded_extents(cache);
2042	} else if (cache->used == 0) {
2043		cache->last_byte_to_unpin = (u64)-1;
2044		cache->cached = BTRFS_CACHE_FINISHED;
2045		add_new_free_space(cache, cache->start,
2046				   cache->start + cache->length);
2047		btrfs_free_excluded_extents(cache);
 
 
2048	}
2049
2050	ret = btrfs_add_block_group_cache(info, cache);
2051	if (ret) {
2052		btrfs_remove_free_space_cache(cache);
2053		goto error;
2054	}
2055	trace_btrfs_add_block_group(info, cache, 0);
2056	btrfs_update_space_info(info, cache->flags, cache->length,
2057				cache->used, cache->bytes_super,
2058				cache->zone_unusable, &space_info);
2059
2060	cache->space_info = space_info;
2061
2062	link_block_group(cache);
2063
2064	set_avail_alloc_bits(info, cache->flags);
2065	if (btrfs_chunk_readonly(info, cache->start)) {
 
 
 
 
 
 
 
 
2066		inc_block_group_ro(cache, 1);
2067	} else if (cache->used == 0) {
2068		ASSERT(list_empty(&cache->bg_list));
2069		if (btrfs_test_opt(info, DISCARD_ASYNC))
2070			btrfs_discard_queue_work(&info->discard_ctl, cache);
2071		else
2072			btrfs_mark_bg_unused(cache);
2073	}
 
2074	return 0;
2075error:
2076	btrfs_put_block_group(cache);
2077	return ret;
2078}
2079
2080static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2081{
2082	struct extent_map_tree *em_tree = &fs_info->mapping_tree;
2083	struct btrfs_space_info *space_info;
2084	struct rb_node *node;
2085	int ret = 0;
2086
2087	for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
2088		struct extent_map *em;
2089		struct map_lookup *map;
2090		struct btrfs_block_group *bg;
2091
2092		em = rb_entry(node, struct extent_map, rb_node);
2093		map = em->map_lookup;
2094		bg = btrfs_create_block_group_cache(fs_info, em->start);
2095		if (!bg) {
2096			ret = -ENOMEM;
2097			break;
2098		}
2099
2100		/* Fill dummy cache as FULL */
2101		bg->length = em->len;
2102		bg->flags = map->type;
2103		bg->last_byte_to_unpin = (u64)-1;
2104		bg->cached = BTRFS_CACHE_FINISHED;
2105		bg->used = em->len;
2106		bg->flags = map->type;
2107		ret = btrfs_add_block_group_cache(fs_info, bg);
 
 
 
 
 
 
 
 
 
 
2108		if (ret) {
2109			btrfs_remove_free_space_cache(bg);
2110			btrfs_put_block_group(bg);
2111			break;
2112		}
2113		btrfs_update_space_info(fs_info, bg->flags, em->len, em->len,
2114					0, 0, &space_info);
2115		bg->space_info = space_info;
2116		link_block_group(bg);
2117
2118		set_avail_alloc_bits(fs_info, bg->flags);
2119	}
2120	if (!ret)
2121		btrfs_init_global_block_rsv(fs_info);
2122	return ret;
2123}
2124
2125int btrfs_read_block_groups(struct btrfs_fs_info *info)
2126{
 
2127	struct btrfs_path *path;
2128	int ret;
2129	struct btrfs_block_group *cache;
2130	struct btrfs_space_info *space_info;
2131	struct btrfs_key key;
2132	int need_clear = 0;
2133	u64 cache_gen;
2134
2135	if (!info->extent_root)
 
 
 
 
 
 
 
 
 
2136		return fill_dummy_bgs(info);
2137
2138	key.objectid = 0;
2139	key.offset = 0;
2140	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2141	path = btrfs_alloc_path();
2142	if (!path)
2143		return -ENOMEM;
2144
2145	cache_gen = btrfs_super_cache_generation(info->super_copy);
2146	if (btrfs_test_opt(info, SPACE_CACHE) &&
2147	    btrfs_super_generation(info->super_copy) != cache_gen)
2148		need_clear = 1;
2149	if (btrfs_test_opt(info, CLEAR_CACHE))
2150		need_clear = 1;
2151
2152	while (1) {
2153		struct btrfs_block_group_item bgi;
2154		struct extent_buffer *leaf;
2155		int slot;
2156
2157		ret = find_first_block_group(info, path, &key);
2158		if (ret > 0)
2159			break;
2160		if (ret != 0)
2161			goto error;
2162
2163		leaf = path->nodes[0];
2164		slot = path->slots[0];
2165
2166		read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2167				   sizeof(bgi));
2168
2169		btrfs_item_key_to_cpu(leaf, &key, slot);
2170		btrfs_release_path(path);
2171		ret = read_one_block_group(info, &bgi, &key, need_clear);
2172		if (ret < 0)
2173			goto error;
2174		key.objectid += key.offset;
2175		key.offset = 0;
2176	}
2177	btrfs_release_path(path);
2178
2179	list_for_each_entry(space_info, &info->space_info, list) {
2180		int i;
2181
2182		for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2183			if (list_empty(&space_info->block_groups[i]))
2184				continue;
2185			cache = list_first_entry(&space_info->block_groups[i],
2186						 struct btrfs_block_group,
2187						 list);
2188			btrfs_sysfs_add_block_group_type(cache);
2189		}
2190
2191		if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2192		      (BTRFS_BLOCK_GROUP_RAID10 |
2193		       BTRFS_BLOCK_GROUP_RAID1_MASK |
2194		       BTRFS_BLOCK_GROUP_RAID56_MASK |
2195		       BTRFS_BLOCK_GROUP_DUP)))
2196			continue;
2197		/*
2198		 * Avoid allocating from un-mirrored block group if there are
2199		 * mirrored block groups.
2200		 */
2201		list_for_each_entry(cache,
2202				&space_info->block_groups[BTRFS_RAID_RAID0],
2203				list)
2204			inc_block_group_ro(cache, 1);
2205		list_for_each_entry(cache,
2206				&space_info->block_groups[BTRFS_RAID_SINGLE],
2207				list)
2208			inc_block_group_ro(cache, 1);
2209	}
2210
2211	btrfs_init_global_block_rsv(info);
2212	ret = check_chunk_block_group_mappings(info);
2213error:
2214	btrfs_free_path(path);
 
 
 
 
 
 
 
 
2215	return ret;
2216}
2217
2218/*
2219 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2220 * allocation.
2221 *
2222 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2223 * phases.
2224 */
2225static int insert_block_group_item(struct btrfs_trans_handle *trans,
2226				   struct btrfs_block_group *block_group)
2227{
2228	struct btrfs_fs_info *fs_info = trans->fs_info;
2229	struct btrfs_block_group_item bgi;
2230	struct btrfs_root *root;
2231	struct btrfs_key key;
 
 
2232
2233	spin_lock(&block_group->lock);
2234	btrfs_set_stack_block_group_used(&bgi, block_group->used);
2235	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2236				BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2237	btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
 
 
2238	key.objectid = block_group->start;
2239	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2240	key.offset = block_group->length;
2241	spin_unlock(&block_group->lock);
2242
2243	root = fs_info->extent_root;
2244	return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2245}
2246
2247/*
2248 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2249 * chunk allocation.
2250 *
2251 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2252 * phases.
2253 */
2254void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2255{
2256	struct btrfs_fs_info *fs_info = trans->fs_info;
2257	struct btrfs_block_group *block_group;
2258	int ret = 0;
2259
2260	while (!list_empty(&trans->new_bgs)) {
2261		int index;
2262
2263		block_group = list_first_entry(&trans->new_bgs,
2264					       struct btrfs_block_group,
2265					       bg_list);
2266		if (ret)
2267			goto next;
2268
2269		index = btrfs_bg_flags_to_raid_index(block_group->flags);
2270
2271		ret = insert_block_group_item(trans, block_group);
2272		if (ret)
2273			btrfs_abort_transaction(trans, ret);
2274		if (!block_group->chunk_item_inserted) {
 
2275			mutex_lock(&fs_info->chunk_mutex);
2276			ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2277			mutex_unlock(&fs_info->chunk_mutex);
2278			if (ret)
2279				btrfs_abort_transaction(trans, ret);
2280		}
2281		ret = btrfs_finish_chunk_alloc(trans, block_group->start,
2282					block_group->length);
2283		if (ret)
2284			btrfs_abort_transaction(trans, ret);
2285		add_block_group_free_space(trans, block_group);
2286
2287		/*
2288		 * If we restriped during balance, we may have added a new raid
2289		 * type, so now add the sysfs entries when it is safe to do so.
2290		 * We don't have to worry about locking here as it's handled in
2291		 * btrfs_sysfs_add_block_group_type.
2292		 */
2293		if (block_group->space_info->block_group_kobjs[index] == NULL)
2294			btrfs_sysfs_add_block_group_type(block_group);
2295
2296		/* Already aborted the transaction if it failed. */
2297next:
2298		btrfs_delayed_refs_rsv_release(fs_info, 1);
2299		list_del_init(&block_group->bg_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2300	}
2301	btrfs_trans_release_chunk_metadata(trans);
2302}
2303
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2304struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2305						 u64 bytes_used, u64 type,
2306						 u64 chunk_offset, u64 size)
2307{
2308	struct btrfs_fs_info *fs_info = trans->fs_info;
2309	struct btrfs_block_group *cache;
2310	int ret;
2311
2312	btrfs_set_log_full_commit(trans);
2313
2314	cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2315	if (!cache)
2316		return ERR_PTR(-ENOMEM);
2317
 
 
 
 
 
 
 
2318	cache->length = size;
2319	set_free_space_tree_thresholds(cache);
2320	cache->used = bytes_used;
2321	cache->flags = type;
2322	cache->last_byte_to_unpin = (u64)-1;
2323	cache->cached = BTRFS_CACHE_FINISHED;
 
 
2324	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2325		cache->needs_free_space = 1;
2326
2327	ret = btrfs_load_block_group_zone_info(cache, true);
2328	if (ret) {
2329		btrfs_put_block_group(cache);
2330		return ERR_PTR(ret);
2331	}
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		btrfs_put_block_group(cache);
2338		return ERR_PTR(ret);
2339	}
2340
2341	add_new_free_space(cache, chunk_offset, chunk_offset + size);
2342
2343	btrfs_free_excluded_extents(cache);
 
 
 
 
2344
2345#ifdef CONFIG_BTRFS_DEBUG
2346	if (btrfs_should_fragment_free_space(cache)) {
2347		u64 new_bytes_used = size - bytes_used;
2348
2349		bytes_used += new_bytes_used >> 1;
2350		fragment_free_space(cache);
2351	}
2352#endif
2353	/*
2354	 * Ensure the corresponding space_info object is created and
2355	 * assigned to our block group. We want our bg to be added to the rbtree
2356	 * with its ->space_info set.
2357	 */
2358	cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2359	ASSERT(cache->space_info);
2360
2361	ret = btrfs_add_block_group_cache(fs_info, cache);
2362	if (ret) {
2363		btrfs_remove_free_space_cache(cache);
2364		btrfs_put_block_group(cache);
2365		return ERR_PTR(ret);
2366	}
2367
2368	/*
2369	 * Now that our block group has its ->space_info set and is inserted in
2370	 * the rbtree, update the space info's counters.
2371	 */
2372	trace_btrfs_add_block_group(fs_info, cache, 1);
2373	btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
2374				cache->bytes_super, 0, &cache->space_info);
2375	btrfs_update_global_block_rsv(fs_info);
2376
2377	link_block_group(cache);
 
 
 
 
 
2378
2379	list_add_tail(&cache->bg_list, &trans->new_bgs);
2380	trans->delayed_ref_updates++;
2381	btrfs_update_delayed_refs_rsv(trans);
2382
2383	set_avail_alloc_bits(fs_info, type);
2384	return cache;
2385}
2386
2387/*
2388 * Mark one block group RO, can be called several times for the same block
2389 * group.
2390 *
2391 * @cache:		the destination block group
2392 * @do_chunk_alloc:	whether need to do chunk pre-allocation, this is to
2393 * 			ensure we still have some free space after marking this
2394 * 			block group RO.
2395 */
2396int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2397			     bool do_chunk_alloc)
2398{
2399	struct btrfs_fs_info *fs_info = cache->fs_info;
2400	struct btrfs_trans_handle *trans;
 
2401	u64 alloc_flags;
2402	int ret;
2403	bool dirty_bg_running;
2404
 
 
 
 
 
 
 
 
 
 
 
 
 
2405	do {
2406		trans = btrfs_join_transaction(fs_info->extent_root);
2407		if (IS_ERR(trans))
2408			return PTR_ERR(trans);
2409
2410		dirty_bg_running = false;
2411
2412		/*
2413		 * We're not allowed to set block groups readonly after the dirty
2414		 * block group cache has started writing.  If it already started,
2415		 * back off and let this transaction commit.
2416		 */
2417		mutex_lock(&fs_info->ro_block_group_mutex);
2418		if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2419			u64 transid = trans->transid;
2420
2421			mutex_unlock(&fs_info->ro_block_group_mutex);
2422			btrfs_end_transaction(trans);
2423
2424			ret = btrfs_wait_for_commit(fs_info, transid);
2425			if (ret)
2426				return ret;
2427			dirty_bg_running = true;
2428		}
2429	} while (dirty_bg_running);
2430
2431	if (do_chunk_alloc) {
2432		/*
2433		 * If we are changing raid levels, try to allocate a
2434		 * corresponding block group with the new raid level.
2435		 */
2436		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2437		if (alloc_flags != cache->flags) {
2438			ret = btrfs_chunk_alloc(trans, alloc_flags,
2439						CHUNK_ALLOC_FORCE);
2440			/*
2441			 * ENOSPC is allowed here, we may have enough space
2442			 * already allocated at the new raid level to carry on
2443			 */
2444			if (ret == -ENOSPC)
2445				ret = 0;
2446			if (ret < 0)
2447				goto out;
2448		}
2449	}
2450
2451	ret = inc_block_group_ro(cache, 0);
2452	if (!do_chunk_alloc || ret == -ETXTBSY)
2453		goto unlock_out;
2454	if (!ret)
2455		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
2456	alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2457	ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2458	if (ret < 0)
2459		goto out;
 
 
 
 
 
 
 
 
2460	ret = inc_block_group_ro(cache, 0);
2461	if (ret == -ETXTBSY)
2462		goto unlock_out;
2463out:
2464	if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2465		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2466		mutex_lock(&fs_info->chunk_mutex);
2467		check_system_chunk(trans, alloc_flags);
2468		mutex_unlock(&fs_info->chunk_mutex);
2469	}
2470unlock_out:
2471	mutex_unlock(&fs_info->ro_block_group_mutex);
2472
2473	btrfs_end_transaction(trans);
2474	return ret;
2475}
2476
2477void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2478{
2479	struct btrfs_space_info *sinfo = cache->space_info;
2480	u64 num_bytes;
2481
2482	BUG_ON(!cache->ro);
2483
2484	spin_lock(&sinfo->lock);
2485	spin_lock(&cache->lock);
2486	if (!--cache->ro) {
2487		if (btrfs_is_zoned(cache->fs_info)) {
2488			/* Migrate zone_unusable bytes back */
2489			cache->zone_unusable = cache->alloc_offset - cache->used;
 
 
2490			sinfo->bytes_zone_unusable += cache->zone_unusable;
2491			sinfo->bytes_readonly -= cache->zone_unusable;
2492		}
2493		num_bytes = cache->length - cache->reserved -
2494			    cache->pinned - cache->bytes_super -
2495			    cache->zone_unusable - cache->used;
2496		sinfo->bytes_readonly -= num_bytes;
2497		list_del_init(&cache->ro_list);
2498	}
2499	spin_unlock(&cache->lock);
2500	spin_unlock(&sinfo->lock);
2501}
2502
2503static int update_block_group_item(struct btrfs_trans_handle *trans,
2504				   struct btrfs_path *path,
2505				   struct btrfs_block_group *cache)
2506{
2507	struct btrfs_fs_info *fs_info = trans->fs_info;
2508	int ret;
2509	struct btrfs_root *root = fs_info->extent_root;
2510	unsigned long bi;
2511	struct extent_buffer *leaf;
2512	struct btrfs_block_group_item bgi;
2513	struct btrfs_key key;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2514
2515	key.objectid = cache->start;
2516	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2517	key.offset = cache->length;
2518
2519	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2520	if (ret) {
2521		if (ret > 0)
2522			ret = -ENOENT;
2523		goto fail;
2524	}
2525
2526	leaf = path->nodes[0];
2527	bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2528	btrfs_set_stack_block_group_used(&bgi, cache->used);
2529	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2530			BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2531	btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2532	write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2533	btrfs_mark_buffer_dirty(leaf);
2534fail:
2535	btrfs_release_path(path);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2536	return ret;
2537
2538}
2539
2540static int cache_save_setup(struct btrfs_block_group *block_group,
2541			    struct btrfs_trans_handle *trans,
2542			    struct btrfs_path *path)
2543{
2544	struct btrfs_fs_info *fs_info = block_group->fs_info;
2545	struct btrfs_root *root = fs_info->tree_root;
2546	struct inode *inode = NULL;
2547	struct extent_changeset *data_reserved = NULL;
2548	u64 alloc_hint = 0;
2549	int dcs = BTRFS_DC_ERROR;
2550	u64 cache_size = 0;
2551	int retries = 0;
2552	int ret = 0;
2553
2554	if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2555		return 0;
2556
2557	/*
2558	 * If this block group is smaller than 100 megs don't bother caching the
2559	 * block group.
2560	 */
2561	if (block_group->length < (100 * SZ_1M)) {
2562		spin_lock(&block_group->lock);
2563		block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2564		spin_unlock(&block_group->lock);
2565		return 0;
2566	}
2567
2568	if (TRANS_ABORTED(trans))
2569		return 0;
2570again:
2571	inode = lookup_free_space_inode(block_group, path);
2572	if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2573		ret = PTR_ERR(inode);
2574		btrfs_release_path(path);
2575		goto out;
2576	}
2577
2578	if (IS_ERR(inode)) {
2579		BUG_ON(retries);
2580		retries++;
2581
2582		if (block_group->ro)
2583			goto out_free;
2584
2585		ret = create_free_space_inode(trans, block_group, path);
2586		if (ret)
2587			goto out_free;
2588		goto again;
2589	}
2590
2591	/*
2592	 * We want to set the generation to 0, that way if anything goes wrong
2593	 * from here on out we know not to trust this cache when we load up next
2594	 * time.
2595	 */
2596	BTRFS_I(inode)->generation = 0;
2597	ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2598	if (ret) {
2599		/*
2600		 * So theoretically we could recover from this, simply set the
2601		 * super cache generation to 0 so we know to invalidate the
2602		 * cache, but then we'd have to keep track of the block groups
2603		 * that fail this way so we know we _have_ to reset this cache
2604		 * before the next commit or risk reading stale cache.  So to
2605		 * limit our exposure to horrible edge cases lets just abort the
2606		 * transaction, this only happens in really bad situations
2607		 * anyway.
2608		 */
2609		btrfs_abort_transaction(trans, ret);
2610		goto out_put;
2611	}
2612	WARN_ON(ret);
2613
2614	/* We've already setup this transaction, go ahead and exit */
2615	if (block_group->cache_generation == trans->transid &&
2616	    i_size_read(inode)) {
2617		dcs = BTRFS_DC_SETUP;
2618		goto out_put;
2619	}
2620
2621	if (i_size_read(inode) > 0) {
2622		ret = btrfs_check_trunc_cache_free_space(fs_info,
2623					&fs_info->global_block_rsv);
2624		if (ret)
2625			goto out_put;
2626
2627		ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2628		if (ret)
2629			goto out_put;
2630	}
2631
2632	spin_lock(&block_group->lock);
2633	if (block_group->cached != BTRFS_CACHE_FINISHED ||
2634	    !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2635		/*
2636		 * don't bother trying to write stuff out _if_
2637		 * a) we're not cached,
2638		 * b) we're with nospace_cache mount option,
2639		 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2640		 */
2641		dcs = BTRFS_DC_WRITTEN;
2642		spin_unlock(&block_group->lock);
2643		goto out_put;
2644	}
2645	spin_unlock(&block_group->lock);
2646
2647	/*
2648	 * We hit an ENOSPC when setting up the cache in this transaction, just
2649	 * skip doing the setup, we've already cleared the cache so we're safe.
2650	 */
2651	if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2652		ret = -ENOSPC;
2653		goto out_put;
2654	}
2655
2656	/*
2657	 * Try to preallocate enough space based on how big the block group is.
2658	 * Keep in mind this has to include any pinned space which could end up
2659	 * taking up quite a bit since it's not folded into the other space
2660	 * cache.
2661	 */
2662	cache_size = div_u64(block_group->length, SZ_256M);
2663	if (!cache_size)
2664		cache_size = 1;
2665
2666	cache_size *= 16;
2667	cache_size *= fs_info->sectorsize;
2668
2669	ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2670					  cache_size);
2671	if (ret)
2672		goto out_put;
2673
2674	ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
2675					      cache_size, cache_size,
2676					      &alloc_hint);
2677	/*
2678	 * Our cache requires contiguous chunks so that we don't modify a bunch
2679	 * of metadata or split extents when writing the cache out, which means
2680	 * we can enospc if we are heavily fragmented in addition to just normal
2681	 * out of space conditions.  So if we hit this just skip setting up any
2682	 * other block groups for this transaction, maybe we'll unpin enough
2683	 * space the next time around.
2684	 */
2685	if (!ret)
2686		dcs = BTRFS_DC_SETUP;
2687	else if (ret == -ENOSPC)
2688		set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2689
2690out_put:
2691	iput(inode);
2692out_free:
2693	btrfs_release_path(path);
2694out:
2695	spin_lock(&block_group->lock);
2696	if (!ret && dcs == BTRFS_DC_SETUP)
2697		block_group->cache_generation = trans->transid;
2698	block_group->disk_cache_state = dcs;
2699	spin_unlock(&block_group->lock);
2700
2701	extent_changeset_free(data_reserved);
2702	return ret;
2703}
2704
2705int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2706{
2707	struct btrfs_fs_info *fs_info = trans->fs_info;
2708	struct btrfs_block_group *cache, *tmp;
2709	struct btrfs_transaction *cur_trans = trans->transaction;
2710	struct btrfs_path *path;
2711
2712	if (list_empty(&cur_trans->dirty_bgs) ||
2713	    !btrfs_test_opt(fs_info, SPACE_CACHE))
2714		return 0;
2715
2716	path = btrfs_alloc_path();
2717	if (!path)
2718		return -ENOMEM;
2719
2720	/* Could add new block groups, use _safe just in case */
2721	list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
2722				 dirty_list) {
2723		if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2724			cache_save_setup(cache, trans, path);
2725	}
2726
2727	btrfs_free_path(path);
2728	return 0;
2729}
2730
2731/*
2732 * Transaction commit does final block group cache writeback during a critical
2733 * section where nothing is allowed to change the FS.  This is required in
2734 * order for the cache to actually match the block group, but can introduce a
2735 * lot of latency into the commit.
2736 *
2737 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2738 * There's a chance we'll have to redo some of it if the block group changes
2739 * again during the commit, but it greatly reduces the commit latency by
2740 * getting rid of the easy block groups while we're still allowing others to
2741 * join the commit.
2742 */
2743int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
2744{
2745	struct btrfs_fs_info *fs_info = trans->fs_info;
2746	struct btrfs_block_group *cache;
2747	struct btrfs_transaction *cur_trans = trans->transaction;
2748	int ret = 0;
2749	int should_put;
2750	struct btrfs_path *path = NULL;
2751	LIST_HEAD(dirty);
2752	struct list_head *io = &cur_trans->io_bgs;
2753	int num_started = 0;
2754	int loops = 0;
2755
2756	spin_lock(&cur_trans->dirty_bgs_lock);
2757	if (list_empty(&cur_trans->dirty_bgs)) {
2758		spin_unlock(&cur_trans->dirty_bgs_lock);
2759		return 0;
2760	}
2761	list_splice_init(&cur_trans->dirty_bgs, &dirty);
2762	spin_unlock(&cur_trans->dirty_bgs_lock);
2763
2764again:
2765	/* Make sure all the block groups on our dirty list actually exist */
2766	btrfs_create_pending_block_groups(trans);
2767
2768	if (!path) {
2769		path = btrfs_alloc_path();
2770		if (!path) {
2771			ret = -ENOMEM;
2772			goto out;
2773		}
2774	}
2775
2776	/*
2777	 * cache_write_mutex is here only to save us from balance or automatic
2778	 * removal of empty block groups deleting this block group while we are
2779	 * writing out the cache
2780	 */
2781	mutex_lock(&trans->transaction->cache_write_mutex);
2782	while (!list_empty(&dirty)) {
2783		bool drop_reserve = true;
2784
2785		cache = list_first_entry(&dirty, struct btrfs_block_group,
2786					 dirty_list);
2787		/*
2788		 * This can happen if something re-dirties a block group that
2789		 * is already under IO.  Just wait for it to finish and then do
2790		 * it all again
2791		 */
2792		if (!list_empty(&cache->io_list)) {
2793			list_del_init(&cache->io_list);
2794			btrfs_wait_cache_io(trans, cache, path);
2795			btrfs_put_block_group(cache);
2796		}
2797
2798
2799		/*
2800		 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2801		 * it should update the cache_state.  Don't delete until after
2802		 * we wait.
2803		 *
2804		 * Since we're not running in the commit critical section
2805		 * we need the dirty_bgs_lock to protect from update_block_group
2806		 */
2807		spin_lock(&cur_trans->dirty_bgs_lock);
2808		list_del_init(&cache->dirty_list);
2809		spin_unlock(&cur_trans->dirty_bgs_lock);
2810
2811		should_put = 1;
2812
2813		cache_save_setup(cache, trans, path);
2814
2815		if (cache->disk_cache_state == BTRFS_DC_SETUP) {
2816			cache->io_ctl.inode = NULL;
2817			ret = btrfs_write_out_cache(trans, cache, path);
2818			if (ret == 0 && cache->io_ctl.inode) {
2819				num_started++;
2820				should_put = 0;
2821
2822				/*
2823				 * The cache_write_mutex is protecting the
2824				 * io_list, also refer to the definition of
2825				 * btrfs_transaction::io_bgs for more details
2826				 */
2827				list_add_tail(&cache->io_list, io);
2828			} else {
2829				/*
2830				 * If we failed to write the cache, the
2831				 * generation will be bad and life goes on
2832				 */
2833				ret = 0;
2834			}
2835		}
2836		if (!ret) {
2837			ret = update_block_group_item(trans, path, cache);
2838			/*
2839			 * Our block group might still be attached to the list
2840			 * of new block groups in the transaction handle of some
2841			 * other task (struct btrfs_trans_handle->new_bgs). This
2842			 * means its block group item isn't yet in the extent
2843			 * tree. If this happens ignore the error, as we will
2844			 * try again later in the critical section of the
2845			 * transaction commit.
2846			 */
2847			if (ret == -ENOENT) {
2848				ret = 0;
2849				spin_lock(&cur_trans->dirty_bgs_lock);
2850				if (list_empty(&cache->dirty_list)) {
2851					list_add_tail(&cache->dirty_list,
2852						      &cur_trans->dirty_bgs);
2853					btrfs_get_block_group(cache);
2854					drop_reserve = false;
2855				}
2856				spin_unlock(&cur_trans->dirty_bgs_lock);
2857			} else if (ret) {
2858				btrfs_abort_transaction(trans, ret);
2859			}
2860		}
2861
2862		/* If it's not on the io list, we need to put the block group */
2863		if (should_put)
2864			btrfs_put_block_group(cache);
2865		if (drop_reserve)
2866			btrfs_delayed_refs_rsv_release(fs_info, 1);
2867		/*
2868		 * Avoid blocking other tasks for too long. It might even save
2869		 * us from writing caches for block groups that are going to be
2870		 * removed.
2871		 */
2872		mutex_unlock(&trans->transaction->cache_write_mutex);
2873		if (ret)
2874			goto out;
2875		mutex_lock(&trans->transaction->cache_write_mutex);
2876	}
2877	mutex_unlock(&trans->transaction->cache_write_mutex);
2878
2879	/*
2880	 * Go through delayed refs for all the stuff we've just kicked off
2881	 * and then loop back (just once)
2882	 */
2883	if (!ret)
2884		ret = btrfs_run_delayed_refs(trans, 0);
2885	if (!ret && loops == 0) {
2886		loops++;
2887		spin_lock(&cur_trans->dirty_bgs_lock);
2888		list_splice_init(&cur_trans->dirty_bgs, &dirty);
2889		/*
2890		 * dirty_bgs_lock protects us from concurrent block group
2891		 * deletes too (not just cache_write_mutex).
2892		 */
2893		if (!list_empty(&dirty)) {
2894			spin_unlock(&cur_trans->dirty_bgs_lock);
2895			goto again;
2896		}
2897		spin_unlock(&cur_trans->dirty_bgs_lock);
2898	}
2899out:
2900	if (ret < 0) {
2901		spin_lock(&cur_trans->dirty_bgs_lock);
2902		list_splice_init(&dirty, &cur_trans->dirty_bgs);
2903		spin_unlock(&cur_trans->dirty_bgs_lock);
2904		btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
2905	}
2906
2907	btrfs_free_path(path);
2908	return ret;
2909}
2910
2911int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
2912{
2913	struct btrfs_fs_info *fs_info = trans->fs_info;
2914	struct btrfs_block_group *cache;
2915	struct btrfs_transaction *cur_trans = trans->transaction;
2916	int ret = 0;
2917	int should_put;
2918	struct btrfs_path *path;
2919	struct list_head *io = &cur_trans->io_bgs;
2920	int num_started = 0;
2921
2922	path = btrfs_alloc_path();
2923	if (!path)
2924		return -ENOMEM;
2925
2926	/*
2927	 * Even though we are in the critical section of the transaction commit,
2928	 * we can still have concurrent tasks adding elements to this
2929	 * transaction's list of dirty block groups. These tasks correspond to
2930	 * endio free space workers started when writeback finishes for a
2931	 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
2932	 * allocate new block groups as a result of COWing nodes of the root
2933	 * tree when updating the free space inode. The writeback for the space
2934	 * caches is triggered by an earlier call to
2935	 * btrfs_start_dirty_block_groups() and iterations of the following
2936	 * loop.
2937	 * Also we want to do the cache_save_setup first and then run the
2938	 * delayed refs to make sure we have the best chance at doing this all
2939	 * in one shot.
2940	 */
2941	spin_lock(&cur_trans->dirty_bgs_lock);
2942	while (!list_empty(&cur_trans->dirty_bgs)) {
2943		cache = list_first_entry(&cur_trans->dirty_bgs,
2944					 struct btrfs_block_group,
2945					 dirty_list);
2946
2947		/*
2948		 * This can happen if cache_save_setup re-dirties a block group
2949		 * that is already under IO.  Just wait for it to finish and
2950		 * then do it all again
2951		 */
2952		if (!list_empty(&cache->io_list)) {
2953			spin_unlock(&cur_trans->dirty_bgs_lock);
2954			list_del_init(&cache->io_list);
2955			btrfs_wait_cache_io(trans, cache, path);
2956			btrfs_put_block_group(cache);
2957			spin_lock(&cur_trans->dirty_bgs_lock);
2958		}
2959
2960		/*
2961		 * Don't remove from the dirty list until after we've waited on
2962		 * any pending IO
2963		 */
2964		list_del_init(&cache->dirty_list);
2965		spin_unlock(&cur_trans->dirty_bgs_lock);
2966		should_put = 1;
2967
2968		cache_save_setup(cache, trans, path);
2969
2970		if (!ret)
2971			ret = btrfs_run_delayed_refs(trans,
2972						     (unsigned long) -1);
2973
2974		if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
2975			cache->io_ctl.inode = NULL;
2976			ret = btrfs_write_out_cache(trans, cache, path);
2977			if (ret == 0 && cache->io_ctl.inode) {
2978				num_started++;
2979				should_put = 0;
2980				list_add_tail(&cache->io_list, io);
2981			} else {
2982				/*
2983				 * If we failed to write the cache, the
2984				 * generation will be bad and life goes on
2985				 */
2986				ret = 0;
2987			}
2988		}
2989		if (!ret) {
2990			ret = update_block_group_item(trans, path, cache);
2991			/*
2992			 * One of the free space endio workers might have
2993			 * created a new block group while updating a free space
2994			 * cache's inode (at inode.c:btrfs_finish_ordered_io())
2995			 * and hasn't released its transaction handle yet, in
2996			 * which case the new block group is still attached to
2997			 * its transaction handle and its creation has not
2998			 * finished yet (no block group item in the extent tree
2999			 * yet, etc). If this is the case, wait for all free
3000			 * space endio workers to finish and retry. This is a
3001			 * very rare case so no need for a more efficient and
3002			 * complex approach.
3003			 */
3004			if (ret == -ENOENT) {
3005				wait_event(cur_trans->writer_wait,
3006				   atomic_read(&cur_trans->num_writers) == 1);
3007				ret = update_block_group_item(trans, path, cache);
3008			}
3009			if (ret)
3010				btrfs_abort_transaction(trans, ret);
3011		}
3012
3013		/* If its not on the io list, we need to put the block group */
3014		if (should_put)
3015			btrfs_put_block_group(cache);
3016		btrfs_delayed_refs_rsv_release(fs_info, 1);
3017		spin_lock(&cur_trans->dirty_bgs_lock);
3018	}
3019	spin_unlock(&cur_trans->dirty_bgs_lock);
3020
3021	/*
3022	 * Refer to the definition of io_bgs member for details why it's safe
3023	 * to use it without any locking
3024	 */
3025	while (!list_empty(io)) {
3026		cache = list_first_entry(io, struct btrfs_block_group,
3027					 io_list);
3028		list_del_init(&cache->io_list);
3029		btrfs_wait_cache_io(trans, cache, path);
3030		btrfs_put_block_group(cache);
3031	}
3032
3033	btrfs_free_path(path);
3034	return ret;
3035}
3036
3037int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3038			     u64 bytenr, u64 num_bytes, int alloc)
3039{
3040	struct btrfs_fs_info *info = trans->fs_info;
3041	struct btrfs_block_group *cache = NULL;
3042	u64 total = num_bytes;
3043	u64 old_val;
3044	u64 byte_in_group;
 
3045	int factor;
3046	int ret = 0;
3047
3048	/* Block accounting for super block */
3049	spin_lock(&info->delalloc_root_lock);
3050	old_val = btrfs_super_bytes_used(info->super_copy);
3051	if (alloc)
3052		old_val += num_bytes;
3053	else
3054		old_val -= num_bytes;
3055	btrfs_set_super_bytes_used(info->super_copy, old_val);
3056	spin_unlock(&info->delalloc_root_lock);
3057
3058	while (total) {
3059		cache = btrfs_lookup_block_group(info, bytenr);
3060		if (!cache) {
3061			ret = -ENOENT;
3062			break;
3063		}
3064		factor = btrfs_bg_type_to_factor(cache->flags);
 
 
 
 
 
 
 
 
 
 
 
3065
3066		/*
3067		 * If this block group has free space cache written out, we
3068		 * need to make sure to load it if we are removing space.  This
3069		 * is because we need the unpinning stage to actually add the
3070		 * space back to the block group, otherwise we will leak space.
3071		 */
3072		if (!alloc && !btrfs_block_group_done(cache))
3073			btrfs_cache_block_group(cache, 1);
3074
3075		byte_in_group = bytenr - cache->start;
3076		WARN_ON(byte_in_group > cache->length);
 
3077
3078		spin_lock(&cache->space_info->lock);
3079		spin_lock(&cache->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3080
3081		if (btrfs_test_opt(info, SPACE_CACHE) &&
3082		    cache->disk_cache_state < BTRFS_DC_CLEAR)
3083			cache->disk_cache_state = BTRFS_DC_CLEAR;
3084
3085		old_val = cache->used;
3086		num_bytes = min(total, cache->length - byte_in_group);
3087		if (alloc) {
3088			old_val += num_bytes;
3089			cache->used = old_val;
3090			cache->reserved -= num_bytes;
3091			cache->space_info->bytes_reserved -= num_bytes;
3092			cache->space_info->bytes_used += num_bytes;
3093			cache->space_info->disk_used += num_bytes * factor;
3094			spin_unlock(&cache->lock);
3095			spin_unlock(&cache->space_info->lock);
3096		} else {
3097			old_val -= num_bytes;
3098			cache->used = old_val;
3099			cache->pinned += num_bytes;
3100			btrfs_space_info_update_bytes_pinned(info,
3101					cache->space_info, num_bytes);
3102			cache->space_info->bytes_used -= num_bytes;
3103			cache->space_info->disk_used -= num_bytes * factor;
3104			spin_unlock(&cache->lock);
3105			spin_unlock(&cache->space_info->lock);
3106
3107			set_extent_dirty(&trans->transaction->pinned_extents,
3108					 bytenr, bytenr + num_bytes - 1,
3109					 GFP_NOFS | __GFP_NOFAIL);
3110		}
3111
3112		spin_lock(&trans->transaction->dirty_bgs_lock);
3113		if (list_empty(&cache->dirty_list)) {
3114			list_add_tail(&cache->dirty_list,
3115				      &trans->transaction->dirty_bgs);
3116			trans->delayed_ref_updates++;
3117			btrfs_get_block_group(cache);
3118		}
3119		spin_unlock(&trans->transaction->dirty_bgs_lock);
3120
3121		/*
3122		 * No longer have used bytes in this block group, queue it for
3123		 * deletion. We do this after adding the block group to the
3124		 * dirty list to avoid races between cleaner kthread and space
3125		 * cache writeout.
3126		 */
3127		if (!alloc && old_val == 0) {
3128			if (!btrfs_test_opt(info, DISCARD_ASYNC))
3129				btrfs_mark_bg_unused(cache);
3130		}
3131
3132		btrfs_put_block_group(cache);
3133		total -= num_bytes;
3134		bytenr += num_bytes;
 
 
 
 
 
 
 
3135	}
3136
 
 
3137	/* Modified block groups are accounted for in the delayed_refs_rsv. */
3138	btrfs_update_delayed_refs_rsv(trans);
3139	return ret;
 
 
3140}
3141
3142/**
3143 * btrfs_add_reserved_bytes - update the block_group and space info counters
 
3144 * @cache:	The cache we are manipulating
3145 * @ram_bytes:  The number of bytes of file content, and will be same to
3146 *              @num_bytes except for the compress path.
3147 * @num_bytes:	The number of bytes in question
3148 * @delalloc:   The blocks are allocated for the delalloc write
3149 *
3150 * This is called by the allocator when it reserves space. If this is a
3151 * reservation and the block group has become read only we cannot make the
3152 * reservation and return -EAGAIN, otherwise this function always succeeds.
3153 */
3154int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3155			     u64 ram_bytes, u64 num_bytes, int delalloc)
 
3156{
3157	struct btrfs_space_info *space_info = cache->space_info;
 
3158	int ret = 0;
3159
3160	spin_lock(&space_info->lock);
3161	spin_lock(&cache->lock);
3162	if (cache->ro) {
3163		ret = -EAGAIN;
3164	} else {
3165		cache->reserved += num_bytes;
3166		space_info->bytes_reserved += num_bytes;
3167		trace_btrfs_space_reservation(cache->fs_info, "space_info",
3168					      space_info->flags, num_bytes, 1);
3169		btrfs_space_info_update_bytes_may_use(cache->fs_info,
3170						      space_info, -ram_bytes);
3171		if (delalloc)
3172			cache->delalloc_bytes += num_bytes;
3173
3174		/*
3175		 * Compression can use less space than we reserved, so wake
3176		 * tickets if that happens
3177		 */
3178		if (num_bytes < ram_bytes)
3179			btrfs_try_granting_tickets(cache->fs_info, space_info);
3180	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3181	spin_unlock(&cache->lock);
3182	spin_unlock(&space_info->lock);
3183	return ret;
3184}
3185
3186/**
3187 * btrfs_free_reserved_bytes - update the block_group and space info counters
 
3188 * @cache:      The cache we are manipulating
3189 * @num_bytes:  The number of bytes in question
3190 * @delalloc:   The blocks are allocated for the delalloc write
3191 *
3192 * This is called by somebody who is freeing space that was never actually used
3193 * on disk.  For example if you reserve some space for a new leaf in transaction
3194 * A and before transaction A commits you free that leaf, you call this with
3195 * reserve set to 0 in order to clear the reservation.
3196 */
3197void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3198			       u64 num_bytes, int delalloc)
3199{
3200	struct btrfs_space_info *space_info = cache->space_info;
3201
3202	spin_lock(&space_info->lock);
3203	spin_lock(&cache->lock);
3204	if (cache->ro)
3205		space_info->bytes_readonly += num_bytes;
3206	cache->reserved -= num_bytes;
3207	space_info->bytes_reserved -= num_bytes;
3208	space_info->max_extent_size = 0;
3209
3210	if (delalloc)
3211		cache->delalloc_bytes -= num_bytes;
3212	spin_unlock(&cache->lock);
3213
3214	btrfs_try_granting_tickets(cache->fs_info, space_info);
3215	spin_unlock(&space_info->lock);
3216}
3217
3218static void force_metadata_allocation(struct btrfs_fs_info *info)
3219{
3220	struct list_head *head = &info->space_info;
3221	struct btrfs_space_info *found;
3222
3223	list_for_each_entry(found, head, list) {
3224		if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3225			found->force_alloc = CHUNK_ALLOC_FORCE;
3226	}
3227}
3228
3229static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3230			      struct btrfs_space_info *sinfo, int force)
3231{
3232	u64 bytes_used = btrfs_space_info_used(sinfo, false);
3233	u64 thresh;
3234
3235	if (force == CHUNK_ALLOC_FORCE)
3236		return 1;
3237
3238	/*
3239	 * in limited mode, we want to have some free space up to
3240	 * about 1% of the FS size.
3241	 */
3242	if (force == CHUNK_ALLOC_LIMITED) {
3243		thresh = btrfs_super_total_bytes(fs_info->super_copy);
3244		thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3245
3246		if (sinfo->total_bytes - bytes_used < thresh)
3247			return 1;
3248	}
3249
3250	if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3251		return 0;
3252	return 1;
3253}
3254
3255int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3256{
3257	u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3258
3259	return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3260}
3261
3262static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3263{
3264	struct btrfs_block_group *bg;
3265	int ret;
3266
3267	/*
3268	 * Check if we have enough space in the system space info because we
3269	 * will need to update device items in the chunk btree and insert a new
3270	 * chunk item in the chunk btree as well. This will allocate a new
3271	 * system block group if needed.
3272	 */
3273	check_system_chunk(trans, flags);
3274
3275	bg = btrfs_alloc_chunk(trans, flags);
3276	if (IS_ERR(bg)) {
3277		ret = PTR_ERR(bg);
3278		goto out;
3279	}
3280
3281	/*
3282	 * If this is a system chunk allocation then stop right here and do not
3283	 * add the chunk item to the chunk btree. This is to prevent a deadlock
3284	 * because this system chunk allocation can be triggered while COWing
3285	 * some extent buffer of the chunk btree and while holding a lock on a
3286	 * parent extent buffer, in which case attempting to insert the chunk
3287	 * item (or update the device item) would result in a deadlock on that
3288	 * parent extent buffer. In this case defer the chunk btree updates to
3289	 * the second phase of chunk allocation and keep our reservation until
3290	 * the second phase completes.
3291	 *
3292	 * This is a rare case and can only be triggered by the very few cases
3293	 * we have where we need to touch the chunk btree outside chunk allocation
3294	 * and chunk removal. These cases are basically adding a device, removing
3295	 * a device or resizing a device.
3296	 */
3297	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3298		return 0;
3299
3300	ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3301	/*
3302	 * Normally we are not expected to fail with -ENOSPC here, since we have
3303	 * previously reserved space in the system space_info and allocated one
3304	 * new system chunk if necessary. However there are two exceptions:
3305	 *
3306	 * 1) We may have enough free space in the system space_info but all the
3307	 *    existing system block groups have a profile which can not be used
3308	 *    for extent allocation.
3309	 *
3310	 *    This happens when mounting in degraded mode. For example we have a
3311	 *    RAID1 filesystem with 2 devices, lose one device and mount the fs
3312	 *    using the other device in degraded mode. If we then allocate a chunk,
3313	 *    we may have enough free space in the existing system space_info, but
3314	 *    none of the block groups can be used for extent allocation since they
3315	 *    have a RAID1 profile, and because we are in degraded mode with a
3316	 *    single device, we are forced to allocate a new system chunk with a
3317	 *    SINGLE profile. Making check_system_chunk() iterate over all system
3318	 *    block groups and check if they have a usable profile and enough space
3319	 *    can be slow on very large filesystems, so we tolerate the -ENOSPC and
3320	 *    try again after forcing allocation of a new system chunk. Like this
3321	 *    we avoid paying the cost of that search in normal circumstances, when
3322	 *    we were not mounted in degraded mode;
3323	 *
3324	 * 2) We had enough free space info the system space_info, and one suitable
3325	 *    block group to allocate from when we called check_system_chunk()
3326	 *    above. However right after we called it, the only system block group
3327	 *    with enough free space got turned into RO mode by a running scrub,
3328	 *    and in this case we have to allocate a new one and retry. We only
3329	 *    need do this allocate and retry once, since we have a transaction
3330	 *    handle and scrub uses the commit root to search for block groups.
 
 
 
 
 
 
 
3331	 */
3332	if (ret == -ENOSPC) {
3333		const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3334		struct btrfs_block_group *sys_bg;
3335
3336		sys_bg = btrfs_alloc_chunk(trans, sys_flags);
3337		if (IS_ERR(sys_bg)) {
3338			ret = PTR_ERR(sys_bg);
3339			btrfs_abort_transaction(trans, ret);
3340			goto out;
3341		}
3342
3343		ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3344		if (ret) {
3345			btrfs_abort_transaction(trans, ret);
3346			goto out;
3347		}
3348
3349		ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3350		if (ret) {
3351			btrfs_abort_transaction(trans, ret);
3352			goto out;
3353		}
3354	} else if (ret) {
3355		btrfs_abort_transaction(trans, ret);
3356		goto out;
3357	}
3358out:
3359	btrfs_trans_release_chunk_metadata(trans);
3360
3361	return ret;
 
 
 
 
3362}
3363
3364/*
3365 * Chunk allocation is done in 2 phases:
3366 *
3367 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3368 *    the chunk, the chunk mapping, create its block group and add the items
3369 *    that belong in the chunk btree to it - more specifically, we need to
3370 *    update device items in the chunk btree and add a new chunk item to it.
3371 *
3372 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3373 *    group item to the extent btree and the device extent items to the devices
3374 *    btree.
3375 *
3376 * This is done to prevent deadlocks. For example when COWing a node from the
3377 * extent btree we are holding a write lock on the node's parent and if we
3378 * trigger chunk allocation and attempted to insert the new block group item
3379 * in the extent btree right way, we could deadlock because the path for the
3380 * insertion can include that parent node. At first glance it seems impossible
3381 * to trigger chunk allocation after starting a transaction since tasks should
3382 * reserve enough transaction units (metadata space), however while that is true
3383 * most of the time, chunk allocation may still be triggered for several reasons:
3384 *
3385 * 1) When reserving metadata, we check if there is enough free space in the
3386 *    metadata space_info and therefore don't trigger allocation of a new chunk.
3387 *    However later when the task actually tries to COW an extent buffer from
3388 *    the extent btree or from the device btree for example, it is forced to
3389 *    allocate a new block group (chunk) because the only one that had enough
3390 *    free space was just turned to RO mode by a running scrub for example (or
3391 *    device replace, block group reclaim thread, etc), so we can not use it
3392 *    for allocating an extent and end up being forced to allocate a new one;
3393 *
3394 * 2) Because we only check that the metadata space_info has enough free bytes,
3395 *    we end up not allocating a new metadata chunk in that case. However if
3396 *    the filesystem was mounted in degraded mode, none of the existing block
3397 *    groups might be suitable for extent allocation due to their incompatible
3398 *    profile (for e.g. mounting a 2 devices filesystem, where all block groups
3399 *    use a RAID1 profile, in degraded mode using a single device). In this case
3400 *    when the task attempts to COW some extent buffer of the extent btree for
3401 *    example, it will trigger allocation of a new metadata block group with a
3402 *    suitable profile (SINGLE profile in the example of the degraded mount of
3403 *    the RAID1 filesystem);
3404 *
3405 * 3) The task has reserved enough transaction units / metadata space, but when
3406 *    it attempts to COW an extent buffer from the extent or device btree for
3407 *    example, it does not find any free extent in any metadata block group,
3408 *    therefore forced to try to allocate a new metadata block group.
3409 *    This is because some other task allocated all available extents in the
3410 *    meanwhile - this typically happens with tasks that don't reserve space
3411 *    properly, either intentionally or as a bug. One example where this is
3412 *    done intentionally is fsync, as it does not reserve any transaction units
3413 *    and ends up allocating a variable number of metadata extents for log
3414 *    tree extent buffers.
 
 
 
 
 
 
 
 
3415 *
3416 * We also need this 2 phases setup when adding a device to a filesystem with
3417 * a seed device - we must create new metadata and system chunks without adding
3418 * any of the block group items to the chunk, extent and device btrees. If we
3419 * did not do it this way, we would get ENOSPC when attempting to update those
3420 * btrees, since all the chunks from the seed device are read-only.
3421 *
3422 * Phase 1 does the updates and insertions to the chunk btree because if we had
3423 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3424 * parallel, we risk having too many system chunks allocated by many tasks if
3425 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3426 * extreme case this leads to exhaustion of the system chunk array in the
3427 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3428 * and with RAID filesystems (so we have more device items in the chunk btree).
3429 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3430 * the system chunk array due to concurrent allocations") provides more details.
3431 *
3432 * For allocation of system chunks, we defer the updates and insertions into the
3433 * chunk btree to phase 2. This is to prevent deadlocks on extent buffers because
3434 * if the chunk allocation is triggered while COWing an extent buffer of the
3435 * chunk btree, we are holding a lock on the parent of that extent buffer and
3436 * doing the chunk btree updates and insertions can require locking that parent.
3437 * This is for the very few and rare cases where we update the chunk btree that
3438 * are not chunk allocation or chunk removal: adding a device, removing a device
3439 * or resizing a device.
3440 *
3441 * The reservation of system space, done through check_system_chunk(), as well
3442 * as all the updates and insertions into the chunk btree must be done while
3443 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
3444 * an extent buffer from the chunks btree we never trigger allocation of a new
3445 * system chunk, which would result in a deadlock (trying to lock twice an
3446 * extent buffer of the chunk btree, first time before triggering the chunk
3447 * allocation and the second time during chunk allocation while attempting to
3448 * update the chunks btree). The system chunk array is also updated while holding
3449 * that mutex. The same logic applies to removing chunks - we must reserve system
3450 * space, update the chunk btree and the system chunk array in the superblock
3451 * while holding fs_info->chunk_mutex.
3452 *
3453 * This function, btrfs_chunk_alloc(), belongs to phase 1.
3454 *
3455 * If @force is CHUNK_ALLOC_FORCE:
3456 *    - return 1 if it successfully allocates a chunk,
3457 *    - return errors including -ENOSPC otherwise.
3458 * If @force is NOT CHUNK_ALLOC_FORCE:
3459 *    - return 0 if it doesn't need to allocate a new chunk,
3460 *    - return 1 if it successfully allocates a chunk,
3461 *    - return errors including -ENOSPC otherwise.
3462 */
3463int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3464		      enum btrfs_chunk_alloc_enum force)
3465{
3466	struct btrfs_fs_info *fs_info = trans->fs_info;
3467	struct btrfs_space_info *space_info;
 
3468	bool wait_for_alloc = false;
3469	bool should_alloc = false;
 
3470	int ret = 0;
3471
 
 
 
 
 
3472	/* Don't re-enter if we're already allocating a chunk */
3473	if (trans->allocating_chunk)
3474		return -ENOSPC;
3475	/*
3476	 * If we are removing a chunk, don't re-enter or we would deadlock.
3477	 * System space reservation and system chunk allocation is done by the
3478	 * chunk remove operation (btrfs_remove_chunk()).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3479	 */
3480	if (trans->removing_chunk)
3481		return -ENOSPC;
3482
3483	space_info = btrfs_find_space_info(fs_info, flags);
3484	ASSERT(space_info);
3485
3486	do {
3487		spin_lock(&space_info->lock);
3488		if (force < space_info->force_alloc)
3489			force = space_info->force_alloc;
3490		should_alloc = should_alloc_chunk(fs_info, space_info, force);
3491		if (space_info->full) {
3492			/* No more free physical space */
3493			if (should_alloc)
3494				ret = -ENOSPC;
3495			else
3496				ret = 0;
3497			spin_unlock(&space_info->lock);
3498			return ret;
3499		} else if (!should_alloc) {
3500			spin_unlock(&space_info->lock);
3501			return 0;
3502		} else if (space_info->chunk_alloc) {
3503			/*
3504			 * Someone is already allocating, so we need to block
3505			 * until this someone is finished and then loop to
3506			 * recheck if we should continue with our allocation
3507			 * attempt.
3508			 */
3509			wait_for_alloc = true;
 
3510			spin_unlock(&space_info->lock);
3511			mutex_lock(&fs_info->chunk_mutex);
3512			mutex_unlock(&fs_info->chunk_mutex);
3513		} else {
3514			/* Proceed with allocation */
3515			space_info->chunk_alloc = 1;
3516			wait_for_alloc = false;
3517			spin_unlock(&space_info->lock);
3518		}
3519
3520		cond_resched();
3521	} while (wait_for_alloc);
3522
3523	mutex_lock(&fs_info->chunk_mutex);
3524	trans->allocating_chunk = true;
3525
3526	/*
3527	 * If we have mixed data/metadata chunks we want to make sure we keep
3528	 * allocating mixed chunks instead of individual chunks.
3529	 */
3530	if (btrfs_mixed_space_info(space_info))
3531		flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3532
3533	/*
3534	 * if we're doing a data chunk, go ahead and make sure that
3535	 * we keep a reasonable number of metadata chunks allocated in the
3536	 * FS as well.
3537	 */
3538	if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3539		fs_info->data_chunk_allocations++;
3540		if (!(fs_info->data_chunk_allocations %
3541		      fs_info->metadata_ratio))
3542			force_metadata_allocation(fs_info);
3543	}
3544
3545	ret = do_chunk_alloc(trans, flags);
3546	trans->allocating_chunk = false;
3547
 
 
 
 
 
 
 
 
 
 
 
 
 
3548	spin_lock(&space_info->lock);
3549	if (ret < 0) {
3550		if (ret == -ENOSPC)
3551			space_info->full = 1;
3552		else
3553			goto out;
3554	} else {
3555		ret = 1;
3556		space_info->max_extent_size = 0;
3557	}
3558
3559	space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3560out:
3561	space_info->chunk_alloc = 0;
3562	spin_unlock(&space_info->lock);
3563	mutex_unlock(&fs_info->chunk_mutex);
3564
3565	return ret;
3566}
3567
3568static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3569{
3570	u64 num_dev;
3571
3572	num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3573	if (!num_dev)
3574		num_dev = fs_info->fs_devices->rw_devices;
3575
3576	return num_dev;
3577}
3578
3579/*
3580 * Reserve space in the system space for allocating or removing a chunk
3581 */
3582void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3583{
3584	struct btrfs_fs_info *fs_info = trans->fs_info;
3585	struct btrfs_space_info *info;
3586	u64 left;
3587	u64 thresh;
3588	int ret = 0;
3589	u64 num_devs;
3590
3591	/*
3592	 * Needed because we can end up allocating a system chunk and for an
3593	 * atomic and race free space reservation in the chunk block reserve.
3594	 */
3595	lockdep_assert_held(&fs_info->chunk_mutex);
3596
3597	info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3598	spin_lock(&info->lock);
3599	left = info->total_bytes - btrfs_space_info_used(info, true);
3600	spin_unlock(&info->lock);
3601
3602	num_devs = get_profile_num_devs(fs_info, type);
3603
3604	/* num_devs device items to update and 1 chunk item to add or remove */
3605	thresh = btrfs_calc_metadata_size(fs_info, num_devs) +
3606		btrfs_calc_insert_metadata_size(fs_info, 1);
3607
3608	if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3609		btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3610			   left, thresh, type);
3611		btrfs_dump_space_info(fs_info, info, 0, 0);
3612	}
3613
3614	if (left < thresh) {
3615		u64 flags = btrfs_system_alloc_profile(fs_info);
3616		struct btrfs_block_group *bg;
3617
3618		/*
3619		 * Ignore failure to create system chunk. We might end up not
3620		 * needing it, as we might not need to COW all nodes/leafs from
3621		 * the paths we visit in the chunk tree (they were already COWed
3622		 * or created in the current transaction for example).
3623		 *
3624		 * Also, if our caller is allocating a system chunk, do not
3625		 * attempt to insert the chunk item in the chunk btree, as we
3626		 * could deadlock on an extent buffer since our caller may be
3627		 * COWing an extent buffer from the chunk btree.
3628		 */
3629		bg = btrfs_alloc_chunk(trans, flags);
3630		if (IS_ERR(bg)) {
3631			ret = PTR_ERR(bg);
3632		} else if (!(type & BTRFS_BLOCK_GROUP_SYSTEM)) {
 
 
 
 
 
 
 
 
3633			/*
3634			 * If we fail to add the chunk item here, we end up
3635			 * trying again at phase 2 of chunk allocation, at
3636			 * btrfs_create_pending_block_groups(). So ignore
3637			 * any error here.
 
 
 
 
3638			 */
3639			btrfs_chunk_alloc_add_chunk_item(trans, bg);
3640		}
3641	}
3642
3643	if (!ret) {
3644		ret = btrfs_block_rsv_add(fs_info->chunk_root,
3645					  &fs_info->chunk_block_rsv,
3646					  thresh, BTRFS_RESERVE_NO_FLUSH);
3647		if (!ret)
3648			trans->chunk_bytes_reserved += thresh;
3649	}
3650}
3651
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3652void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
3653{
3654	struct btrfs_block_group *block_group;
3655	u64 last = 0;
3656
3657	while (1) {
3658		struct inode *inode;
 
 
 
 
 
 
 
 
3659
3660		block_group = btrfs_lookup_first_block_group(info, last);
3661		while (block_group) {
3662			btrfs_wait_block_group_cache_done(block_group);
3663			spin_lock(&block_group->lock);
3664			if (block_group->iref)
3665				break;
3666			spin_unlock(&block_group->lock);
3667			block_group = btrfs_next_block_group(block_group);
3668		}
3669		if (!block_group) {
3670			if (last == 0)
3671				break;
3672			last = 0;
3673			continue;
3674		}
3675
3676		inode = block_group->inode;
3677		block_group->iref = 0;
3678		block_group->inode = NULL;
3679		spin_unlock(&block_group->lock);
3680		ASSERT(block_group->io_ctl.inode == NULL);
3681		iput(inode);
3682		last = block_group->start + block_group->length;
3683		btrfs_put_block_group(block_group);
3684	}
3685}
3686
3687/*
3688 * Must be called only after stopping all workers, since we could have block
3689 * group caching kthreads running, and therefore they could race with us if we
3690 * freed the block groups before stopping them.
3691 */
3692int btrfs_free_block_groups(struct btrfs_fs_info *info)
3693{
3694	struct btrfs_block_group *block_group;
3695	struct btrfs_space_info *space_info;
3696	struct btrfs_caching_control *caching_ctl;
3697	struct rb_node *n;
3698
3699	spin_lock(&info->block_group_cache_lock);
 
 
 
 
 
 
 
 
 
 
 
3700	while (!list_empty(&info->caching_block_groups)) {
3701		caching_ctl = list_entry(info->caching_block_groups.next,
3702					 struct btrfs_caching_control, list);
3703		list_del(&caching_ctl->list);
3704		btrfs_put_caching_control(caching_ctl);
3705	}
3706	spin_unlock(&info->block_group_cache_lock);
3707
3708	spin_lock(&info->unused_bgs_lock);
3709	while (!list_empty(&info->unused_bgs)) {
3710		block_group = list_first_entry(&info->unused_bgs,
3711					       struct btrfs_block_group,
3712					       bg_list);
3713		list_del_init(&block_group->bg_list);
3714		btrfs_put_block_group(block_group);
3715	}
3716	spin_unlock(&info->unused_bgs_lock);
3717
3718	spin_lock(&info->unused_bgs_lock);
3719	while (!list_empty(&info->reclaim_bgs)) {
3720		block_group = list_first_entry(&info->reclaim_bgs,
3721					       struct btrfs_block_group,
3722					       bg_list);
3723		list_del_init(&block_group->bg_list);
3724		btrfs_put_block_group(block_group);
3725	}
3726	spin_unlock(&info->unused_bgs_lock);
3727
3728	spin_lock(&info->block_group_cache_lock);
3729	while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
 
 
 
 
 
 
 
 
 
 
3730		block_group = rb_entry(n, struct btrfs_block_group,
3731				       cache_node);
3732		rb_erase(&block_group->cache_node,
3733			 &info->block_group_cache_tree);
3734		RB_CLEAR_NODE(&block_group->cache_node);
3735		spin_unlock(&info->block_group_cache_lock);
3736
3737		down_write(&block_group->space_info->groups_sem);
3738		list_del(&block_group->list);
3739		up_write(&block_group->space_info->groups_sem);
3740
3741		/*
3742		 * We haven't cached this block group, which means we could
3743		 * possibly have excluded extents on this block group.
3744		 */
3745		if (block_group->cached == BTRFS_CACHE_NO ||
3746		    block_group->cached == BTRFS_CACHE_ERROR)
3747			btrfs_free_excluded_extents(block_group);
3748
3749		btrfs_remove_free_space_cache(block_group);
3750		ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
3751		ASSERT(list_empty(&block_group->dirty_list));
3752		ASSERT(list_empty(&block_group->io_list));
3753		ASSERT(list_empty(&block_group->bg_list));
3754		ASSERT(refcount_read(&block_group->refs) == 1);
3755		ASSERT(block_group->swap_extents == 0);
3756		btrfs_put_block_group(block_group);
3757
3758		spin_lock(&info->block_group_cache_lock);
3759	}
3760	spin_unlock(&info->block_group_cache_lock);
3761
3762	btrfs_release_global_block_rsv(info);
3763
3764	while (!list_empty(&info->space_info)) {
3765		space_info = list_entry(info->space_info.next,
3766					struct btrfs_space_info,
3767					list);
3768
3769		/*
3770		 * Do not hide this behind enospc_debug, this is actually
3771		 * important and indicates a real bug if this happens.
3772		 */
3773		if (WARN_ON(space_info->bytes_pinned > 0 ||
3774			    space_info->bytes_reserved > 0 ||
3775			    space_info->bytes_may_use > 0))
3776			btrfs_dump_space_info(info, space_info, 0, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3777		WARN_ON(space_info->reclaim_size > 0);
3778		list_del(&space_info->list);
3779		btrfs_sysfs_remove_space_info(space_info);
3780	}
3781	return 0;
3782}
3783
3784void btrfs_freeze_block_group(struct btrfs_block_group *cache)
3785{
3786	atomic_inc(&cache->frozen);
3787}
3788
3789void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
3790{
3791	struct btrfs_fs_info *fs_info = block_group->fs_info;
3792	struct extent_map_tree *em_tree;
3793	struct extent_map *em;
3794	bool cleanup;
3795
3796	spin_lock(&block_group->lock);
3797	cleanup = (atomic_dec_and_test(&block_group->frozen) &&
3798		   block_group->removed);
3799	spin_unlock(&block_group->lock);
3800
3801	if (cleanup) {
3802		em_tree = &fs_info->mapping_tree;
3803		write_lock(&em_tree->lock);
3804		em = lookup_extent_mapping(em_tree, block_group->start,
3805					   1);
3806		BUG_ON(!em); /* logic error, can't happen */
3807		remove_extent_mapping(em_tree, em);
3808		write_unlock(&em_tree->lock);
3809
3810		/* once for us and once for the tree */
3811		free_extent_map(em);
3812		free_extent_map(em);
3813
3814		/*
3815		 * We may have left one free space entry and other possible
3816		 * tasks trimming this block group have left 1 entry each one.
3817		 * Free them if any.
3818		 */
3819		__btrfs_remove_free_space_cache(block_group->free_space_ctl);
3820	}
3821}
3822
3823bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
3824{
3825	bool ret = true;
3826
3827	spin_lock(&bg->lock);
3828	if (bg->ro)
3829		ret = false;
3830	else
3831		bg->swap_extents++;
3832	spin_unlock(&bg->lock);
3833
3834	return ret;
3835}
3836
3837void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
3838{
3839	spin_lock(&bg->lock);
3840	ASSERT(!bg->ro);
3841	ASSERT(bg->swap_extents >= amount);
3842	bg->swap_extents -= amount;
3843	spin_unlock(&bg->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3844}