free-space-cache.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2008 Red Hat.  All rights reserved.
   4 */
   5
   6#include <linux/pagemap.h>
   7#include <linux/sched.h>
   8#include <linux/sched/signal.h>
   9#include <linux/slab.h>
  10#include <linux/math64.h>
  11#include <linux/ratelimit.h>
  12#include <linux/error-injection.h>
  13#include <linux/sched/mm.h>
  14#include <linux/string_choices.h>
  15#include "extent-tree.h"
  16#include "fs.h"
  17#include "messages.h"
  18#include "misc.h"
  19#include "free-space-cache.h"
  20#include "transaction.h"
  21#include "disk-io.h"
  22#include "extent_io.h"
  23#include "space-info.h"
  24#include "block-group.h"
  25#include "discard.h"
  26#include "subpage.h"
  27#include "inode-item.h"
  28#include "accessors.h"
  29#include "file-item.h"
  30#include "file.h"
  31#include "super.h"
  32
  33#define BITS_PER_BITMAP		(PAGE_SIZE * 8UL)
  34#define MAX_CACHE_BYTES_PER_GIG	SZ_64K
  35#define FORCE_EXTENT_THRESHOLD	SZ_1M
  36
  37static struct kmem_cache *btrfs_free_space_cachep;
  38static struct kmem_cache *btrfs_free_space_bitmap_cachep;
  39
  40struct btrfs_trim_range {
  41	u64 start;
  42	u64 bytes;
  43	struct list_head list;
  44};
  45
  46static int link_free_space(struct btrfs_free_space_ctl *ctl,
  47			   struct btrfs_free_space *info);
  48static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  49			      struct btrfs_free_space *info, bool update_stat);
  50static int search_bitmap(struct btrfs_free_space_ctl *ctl,
  51			 struct btrfs_free_space *bitmap_info, u64 *offset,
  52			 u64 *bytes, bool for_alloc);
  53static void free_bitmap(struct btrfs_free_space_ctl *ctl,
  54			struct btrfs_free_space *bitmap_info);
  55static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
  56			      struct btrfs_free_space *info, u64 offset,
  57			      u64 bytes, bool update_stats);
  58
  59static void btrfs_crc32c_final(u32 crc, u8 *result)
  60{
  61	put_unaligned_le32(~crc, result);
  62}
  63
  64static void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
  65{
  66	struct btrfs_free_space *info;
  67	struct rb_node *node;
  68
  69	while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
  70		info = rb_entry(node, struct btrfs_free_space, offset_index);
  71		if (!info->bitmap) {
  72			unlink_free_space(ctl, info, true);
  73			kmem_cache_free(btrfs_free_space_cachep, info);
  74		} else {
  75			free_bitmap(ctl, info);
  76		}
  77
  78		cond_resched_lock(&ctl->tree_lock);
  79	}
  80}
  81
  82static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  83					       struct btrfs_path *path,
  84					       u64 offset)
  85{
  86	struct btrfs_key key;
  87	struct btrfs_key location;
  88	struct btrfs_disk_key disk_key;
  89	struct btrfs_free_space_header *header;
  90	struct extent_buffer *leaf;
  91	struct btrfs_inode *inode;
  92	unsigned nofs_flag;
  93	int ret;
  94
  95	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  96	key.type = 0;
  97	key.offset = offset;
  98
  99	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 100	if (ret < 0)
 101		return ERR_PTR(ret);
 102	if (ret > 0) {
 103		btrfs_release_path(path);
 104		return ERR_PTR(-ENOENT);
 105	}
 106
 107	leaf = path->nodes[0];
 108	header = btrfs_item_ptr(leaf, path->slots[0],
 109				struct btrfs_free_space_header);
 110	btrfs_free_space_key(leaf, header, &disk_key);
 111	btrfs_disk_key_to_cpu(&location, &disk_key);
 112	btrfs_release_path(path);
 113
 114	/*
 115	 * We are often under a trans handle at this point, so we need to make
 116	 * sure NOFS is set to keep us from deadlocking.
 117	 */
 118	nofs_flag = memalloc_nofs_save();
 119	inode = btrfs_iget_path(location.objectid, root, path);
 120	btrfs_release_path(path);
 121	memalloc_nofs_restore(nofs_flag);
 122	if (IS_ERR(inode))
 123		return ERR_CAST(inode);
 124
 125	mapping_set_gfp_mask(inode->vfs_inode.i_mapping,
 126			mapping_gfp_constraint(inode->vfs_inode.i_mapping,
 127			~(__GFP_FS | __GFP_HIGHMEM)));
 128
 129	return &inode->vfs_inode;
 130}
 131
 132struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
 133		struct btrfs_path *path)
 134{
 135	struct btrfs_fs_info *fs_info = block_group->fs_info;
 136	struct inode *inode = NULL;
 137	u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 138
 139	spin_lock(&block_group->lock);
 140	if (block_group->inode)
 141		inode = igrab(&block_group->inode->vfs_inode);
 142	spin_unlock(&block_group->lock);
 143	if (inode)
 144		return inode;
 145
 146	inode = __lookup_free_space_inode(fs_info->tree_root, path,
 147					  block_group->start);
 148	if (IS_ERR(inode))
 149		return inode;
 150
 151	spin_lock(&block_group->lock);
 152	if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 153		btrfs_info(fs_info, "Old style space inode found, converting.");
 154		BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 155			BTRFS_INODE_NODATACOW;
 156		block_group->disk_cache_state = BTRFS_DC_CLEAR;
 157	}
 158
 159	if (!test_and_set_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags))
 160		block_group->inode = BTRFS_I(igrab(inode));
 161	spin_unlock(&block_group->lock);
 162
 163	return inode;
 164}
 165
 166static int __create_free_space_inode(struct btrfs_root *root,
 167				     struct btrfs_trans_handle *trans,
 168				     struct btrfs_path *path,
 169				     u64 ino, u64 offset)
 170{
 171	struct btrfs_key key;
 172	struct btrfs_disk_key disk_key;
 173	struct btrfs_free_space_header *header;
 174	struct btrfs_inode_item *inode_item;
 175	struct extent_buffer *leaf;
 176	/* We inline CRCs for the free disk space cache */
 177	const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
 178			  BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 179	int ret;
 180
 181	ret = btrfs_insert_empty_inode(trans, root, path, ino);
 182	if (ret)
 183		return ret;
 184
 185	leaf = path->nodes[0];
 186	inode_item = btrfs_item_ptr(leaf, path->slots[0],
 187				    struct btrfs_inode_item);
 188	btrfs_item_key(leaf, &disk_key, path->slots[0]);
 189	memzero_extent_buffer(leaf, (unsigned long)inode_item,
 190			     sizeof(*inode_item));
 191	btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 192	btrfs_set_inode_size(leaf, inode_item, 0);
 193	btrfs_set_inode_nbytes(leaf, inode_item, 0);
 194	btrfs_set_inode_uid(leaf, inode_item, 0);
 195	btrfs_set_inode_gid(leaf, inode_item, 0);
 196	btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 197	btrfs_set_inode_flags(leaf, inode_item, flags);
 198	btrfs_set_inode_nlink(leaf, inode_item, 1);
 199	btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 200	btrfs_set_inode_block_group(leaf, inode_item, offset);
 201	btrfs_release_path(path);
 202
 203	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 204	key.type = 0;
 205	key.offset = offset;
 206	ret = btrfs_insert_empty_item(trans, root, path, &key,
 207				      sizeof(struct btrfs_free_space_header));
 208	if (ret < 0) {
 209		btrfs_release_path(path);
 210		return ret;
 211	}
 212
 213	leaf = path->nodes[0];
 214	header = btrfs_item_ptr(leaf, path->slots[0],
 215				struct btrfs_free_space_header);
 216	memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 217	btrfs_set_free_space_key(leaf, header, &disk_key);
 218	btrfs_release_path(path);
 219
 220	return 0;
 221}
 222
 223int create_free_space_inode(struct btrfs_trans_handle *trans,
 224			    struct btrfs_block_group *block_group,
 225			    struct btrfs_path *path)
 226{
 227	int ret;
 228	u64 ino;
 229
 230	ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
 231	if (ret < 0)
 232		return ret;
 233
 234	return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 235					 ino, block_group->start);
 236}
 237
 238/*
 239 * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
 240 * handles lookup, otherwise it takes ownership and iputs the inode.
 241 * Don't reuse an inode pointer after passing it into this function.
 242 */
 243int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
 244				  struct inode *inode,
 245				  struct btrfs_block_group *block_group)
 246{
 247	BTRFS_PATH_AUTO_FREE(path);
 248	struct btrfs_key key;
 249	int ret = 0;
 250
 251	path = btrfs_alloc_path();
 252	if (!path)
 253		return -ENOMEM;
 254
 255	if (!inode)
 256		inode = lookup_free_space_inode(block_group, path);
 257	if (IS_ERR(inode)) {
 258		if (PTR_ERR(inode) != -ENOENT)
 259			ret = PTR_ERR(inode);
 260		return ret;
 261	}
 262	ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 263	if (ret) {
 264		btrfs_add_delayed_iput(BTRFS_I(inode));
 265		return ret;
 266	}
 267	clear_nlink(inode);
 268	/* One for the block groups ref */
 269	spin_lock(&block_group->lock);
 270	if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags)) {
 271		block_group->inode = NULL;
 272		spin_unlock(&block_group->lock);
 273		iput(inode);
 274	} else {
 275		spin_unlock(&block_group->lock);
 276	}
 277	/* One for the lookup ref */
 278	btrfs_add_delayed_iput(BTRFS_I(inode));
 279
 280	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 281	key.type = 0;
 282	key.offset = block_group->start;
 283	ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
 284				-1, 1);
 285	if (ret) {
 286		if (ret > 0)
 287			ret = 0;
 288		return ret;
 289	}
 290	return btrfs_del_item(trans, trans->fs_info->tree_root, path);
 291}
 292
 293int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 294				    struct btrfs_block_group *block_group,
 295				    struct inode *vfs_inode)
 296{
 297	struct btrfs_truncate_control control = {
 298		.inode = BTRFS_I(vfs_inode),
 299		.new_size = 0,
 300		.ino = btrfs_ino(BTRFS_I(vfs_inode)),
 301		.min_type = BTRFS_EXTENT_DATA_KEY,
 302		.clear_extent_range = true,
 303	};
 304	struct btrfs_inode *inode = BTRFS_I(vfs_inode);
 305	struct btrfs_root *root = inode->root;
 306	struct extent_state *cached_state = NULL;
 307	int ret = 0;
 308	bool locked = false;
 309
 310	if (block_group) {
 311		BTRFS_PATH_AUTO_FREE(path);
 312
 313		path = btrfs_alloc_path();
 314		if (!path) {
 315			ret = -ENOMEM;
 316			goto fail;
 317		}
 318		locked = true;
 319		mutex_lock(&trans->transaction->cache_write_mutex);
 320		if (!list_empty(&block_group->io_list)) {
 321			list_del_init(&block_group->io_list);
 322
 323			btrfs_wait_cache_io(trans, block_group, path);
 324			btrfs_put_block_group(block_group);
 325		}
 326
 327		/*
 328		 * now that we've truncated the cache away, its no longer
 329		 * setup or written
 330		 */
 331		spin_lock(&block_group->lock);
 332		block_group->disk_cache_state = BTRFS_DC_CLEAR;
 333		spin_unlock(&block_group->lock);
 334	}
 335
 336	btrfs_i_size_write(inode, 0);
 337	truncate_pagecache(vfs_inode, 0);
 338
 339	btrfs_lock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 340	btrfs_drop_extent_map_range(inode, 0, (u64)-1, false);
 341
 342	/*
 343	 * We skip the throttling logic for free space cache inodes, so we don't
 344	 * need to check for -EAGAIN.
 345	 */
 346	ret = btrfs_truncate_inode_items(trans, root, &control);
 347
 348	inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
 349	btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
 350
 351	btrfs_unlock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 352	if (ret)
 353		goto fail;
 354
 355	ret = btrfs_update_inode(trans, inode);
 356
 357fail:
 358	if (locked)
 359		mutex_unlock(&trans->transaction->cache_write_mutex);
 360	if (ret)
 361		btrfs_abort_transaction(trans, ret);
 362
 363	return ret;
 364}
 365
 366static void readahead_cache(struct inode *inode)
 367{
 368	struct file_ra_state ra;
 369	pgoff_t last_index;
 370
 371	file_ra_state_init(&ra, inode->i_mapping);
 372	last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 373
 374	page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
 375}
 376
 377static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 378		       int write)
 379{
 380	int num_pages;
 381
 382	num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 383
 384	/* Make sure we can fit our crcs and generation into the first page */
 385	if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 386		return -ENOSPC;
 387
 388	memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 389
 390	io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 391	if (!io_ctl->pages)
 392		return -ENOMEM;
 393
 394	io_ctl->num_pages = num_pages;
 395	io_ctl->fs_info = inode_to_fs_info(inode);
 396	io_ctl->inode = inode;
 397
 398	return 0;
 399}
 400ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 401
 402static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 403{
 404	kfree(io_ctl->pages);
 405	io_ctl->pages = NULL;
 406}
 407
 408static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 409{
 410	if (io_ctl->cur) {
 411		io_ctl->cur = NULL;
 412		io_ctl->orig = NULL;
 413	}
 414}
 415
 416static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 417{
 418	ASSERT(io_ctl->index < io_ctl->num_pages);
 419	io_ctl->page = io_ctl->pages[io_ctl->index++];
 420	io_ctl->cur = page_address(io_ctl->page);
 421	io_ctl->orig = io_ctl->cur;
 422	io_ctl->size = PAGE_SIZE;
 423	if (clear)
 424		clear_page(io_ctl->cur);
 425}
 426
 427static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 428{
 429	int i;
 430
 431	io_ctl_unmap_page(io_ctl);
 432
 433	for (i = 0; i < io_ctl->num_pages; i++) {
 434		if (io_ctl->pages[i]) {
 435			btrfs_folio_clear_checked(io_ctl->fs_info,
 436					page_folio(io_ctl->pages[i]),
 437					page_offset(io_ctl->pages[i]),
 438					PAGE_SIZE);
 439			unlock_page(io_ctl->pages[i]);
 440			put_page(io_ctl->pages[i]);
 441		}
 442	}
 443}
 444
 445static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 446{
 447	struct folio *folio;
 448	struct inode *inode = io_ctl->inode;
 449	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 450	int i;
 451
 452	for (i = 0; i < io_ctl->num_pages; i++) {
 453		int ret;
 454
 455		folio = __filemap_get_folio(inode->i_mapping, i,
 456					    FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
 457					    mask);
 458		if (IS_ERR(folio)) {
 459			io_ctl_drop_pages(io_ctl);
 460			return PTR_ERR(folio);
 461		}
 462
 463		ret = set_folio_extent_mapped(folio);
 464		if (ret < 0) {
 465			folio_unlock(folio);
 466			folio_put(folio);
 467			io_ctl_drop_pages(io_ctl);
 468			return ret;
 469		}
 470
 471		io_ctl->pages[i] = &folio->page;
 472		if (uptodate && !folio_test_uptodate(folio)) {
 473			btrfs_read_folio(NULL, folio);
 474			folio_lock(folio);
 475			if (folio->mapping != inode->i_mapping) {
 476				btrfs_err(BTRFS_I(inode)->root->fs_info,
 477					  "free space cache page truncated");
 478				io_ctl_drop_pages(io_ctl);
 479				return -EIO;
 480			}
 481			if (!folio_test_uptodate(folio)) {
 482				btrfs_err(BTRFS_I(inode)->root->fs_info,
 483					   "error reading free space cache");
 484				io_ctl_drop_pages(io_ctl);
 485				return -EIO;
 486			}
 487		}
 488	}
 489
 490	for (i = 0; i < io_ctl->num_pages; i++)
 491		clear_page_dirty_for_io(io_ctl->pages[i]);
 492
 493	return 0;
 494}
 495
 496static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 497{
 498	io_ctl_map_page(io_ctl, 1);
 499
 500	/*
 501	 * Skip the csum areas.  If we don't check crcs then we just have a
 502	 * 64bit chunk at the front of the first page.
 503	 */
 504	io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 505	io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 506
 507	put_unaligned_le64(generation, io_ctl->cur);
 508	io_ctl->cur += sizeof(u64);
 509}
 510
 511static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 512{
 513	u64 cache_gen;
 514
 515	/*
 516	 * Skip the crc area.  If we don't check crcs then we just have a 64bit
 517	 * chunk at the front of the first page.
 518	 */
 519	io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 520	io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 521
 522	cache_gen = get_unaligned_le64(io_ctl->cur);
 523	if (cache_gen != generation) {
 524		btrfs_err_rl(io_ctl->fs_info,
 525			"space cache generation (%llu) does not match inode (%llu)",
 526				cache_gen, generation);
 527		io_ctl_unmap_page(io_ctl);
 528		return -EIO;
 529	}
 530	io_ctl->cur += sizeof(u64);
 531	return 0;
 532}
 533
 534static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 535{
 536	u32 *tmp;
 537	u32 crc = ~(u32)0;
 538	unsigned offset = 0;
 539
 540	if (index == 0)
 541		offset = sizeof(u32) * io_ctl->num_pages;
 542
 543	crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 544	btrfs_crc32c_final(crc, (u8 *)&crc);
 545	io_ctl_unmap_page(io_ctl);
 546	tmp = page_address(io_ctl->pages[0]);
 547	tmp += index;
 548	*tmp = crc;
 549}
 550
 551static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 552{
 553	u32 *tmp, val;
 554	u32 crc = ~(u32)0;
 555	unsigned offset = 0;
 556
 557	if (index == 0)
 558		offset = sizeof(u32) * io_ctl->num_pages;
 559
 560	tmp = page_address(io_ctl->pages[0]);
 561	tmp += index;
 562	val = *tmp;
 563
 564	io_ctl_map_page(io_ctl, 0);
 565	crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 566	btrfs_crc32c_final(crc, (u8 *)&crc);
 567	if (val != crc) {
 568		btrfs_err_rl(io_ctl->fs_info,
 569			"csum mismatch on free space cache");
 570		io_ctl_unmap_page(io_ctl);
 571		return -EIO;
 572	}
 573
 574	return 0;
 575}
 576
 577static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 578			    void *bitmap)
 579{
 580	struct btrfs_free_space_entry *entry;
 581
 582	if (!io_ctl->cur)
 583		return -ENOSPC;
 584
 585	entry = io_ctl->cur;
 586	put_unaligned_le64(offset, &entry->offset);
 587	put_unaligned_le64(bytes, &entry->bytes);
 588	entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 589		BTRFS_FREE_SPACE_EXTENT;
 590	io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 591	io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 592
 593	if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 594		return 0;
 595
 596	io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 597
 598	/* No more pages to map */
 599	if (io_ctl->index >= io_ctl->num_pages)
 600		return 0;
 601
 602	/* map the next page */
 603	io_ctl_map_page(io_ctl, 1);
 604	return 0;
 605}
 606
 607static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 608{
 609	if (!io_ctl->cur)
 610		return -ENOSPC;
 611
 612	/*
 613	 * If we aren't at the start of the current page, unmap this one and
 614	 * map the next one if there is any left.
 615	 */
 616	if (io_ctl->cur != io_ctl->orig) {
 617		io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 618		if (io_ctl->index >= io_ctl->num_pages)
 619			return -ENOSPC;
 620		io_ctl_map_page(io_ctl, 0);
 621	}
 622
 623	copy_page(io_ctl->cur, bitmap);
 624	io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 625	if (io_ctl->index < io_ctl->num_pages)
 626		io_ctl_map_page(io_ctl, 0);
 627	return 0;
 628}
 629
 630static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 631{
 632	/*
 633	 * If we're not on the boundary we know we've modified the page and we
 634	 * need to crc the page.
 635	 */
 636	if (io_ctl->cur != io_ctl->orig)
 637		io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 638	else
 639		io_ctl_unmap_page(io_ctl);
 640
 641	while (io_ctl->index < io_ctl->num_pages) {
 642		io_ctl_map_page(io_ctl, 1);
 643		io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 644	}
 645}
 646
 647static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 648			    struct btrfs_free_space *entry, u8 *type)
 649{
 650	struct btrfs_free_space_entry *e;
 651	int ret;
 652
 653	if (!io_ctl->cur) {
 654		ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 655		if (ret)
 656			return ret;
 657	}
 658
 659	e = io_ctl->cur;
 660	entry->offset = get_unaligned_le64(&e->offset);
 661	entry->bytes = get_unaligned_le64(&e->bytes);
 662	*type = e->type;
 663	io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 664	io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 665
 666	if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 667		return 0;
 668
 669	io_ctl_unmap_page(io_ctl);
 670
 671	return 0;
 672}
 673
 674static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 675			      struct btrfs_free_space *entry)
 676{
 677	int ret;
 678
 679	ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 680	if (ret)
 681		return ret;
 682
 683	copy_page(entry->bitmap, io_ctl->cur);
 684	io_ctl_unmap_page(io_ctl);
 685
 686	return 0;
 687}
 688
 689static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 690{
 691	struct btrfs_block_group *block_group = ctl->block_group;
 692	u64 max_bytes;
 693	u64 bitmap_bytes;
 694	u64 extent_bytes;
 695	u64 size = block_group->length;
 696	u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
 697	u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
 698
 699	max_bitmaps = max_t(u64, max_bitmaps, 1);
 700
 701	if (ctl->total_bitmaps > max_bitmaps)
 702		btrfs_err(block_group->fs_info,
 703"invalid free space control: bg start=%llu len=%llu total_bitmaps=%u unit=%u max_bitmaps=%llu bytes_per_bg=%llu",
 704			  block_group->start, block_group->length,
 705			  ctl->total_bitmaps, ctl->unit, max_bitmaps,
 706			  bytes_per_bg);
 707	ASSERT(ctl->total_bitmaps <= max_bitmaps);
 708
 709	/*
 710	 * We are trying to keep the total amount of memory used per 1GiB of
 711	 * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
 712	 * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
 713	 * bitmaps, we may end up using more memory than this.
 714	 */
 715	if (size < SZ_1G)
 716		max_bytes = MAX_CACHE_BYTES_PER_GIG;
 717	else
 718		max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
 719
 720	bitmap_bytes = ctl->total_bitmaps * ctl->unit;
 721
 722	/*
 723	 * we want the extent entry threshold to always be at most 1/2 the max
 724	 * bytes we can have, or whatever is less than that.
 725	 */
 726	extent_bytes = max_bytes - bitmap_bytes;
 727	extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
 728
 729	ctl->extents_thresh =
 730		div_u64(extent_bytes, sizeof(struct btrfs_free_space));
 731}
 732
 733static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 734				   struct btrfs_free_space_ctl *ctl,
 735				   struct btrfs_path *path, u64 offset)
 736{
 737	struct btrfs_fs_info *fs_info = root->fs_info;
 738	struct btrfs_free_space_header *header;
 739	struct extent_buffer *leaf;
 740	struct btrfs_io_ctl io_ctl;
 741	struct btrfs_key key;
 742	struct btrfs_free_space *e, *n;
 743	LIST_HEAD(bitmaps);
 744	u64 num_entries;
 745	u64 num_bitmaps;
 746	u64 generation;
 747	u8 type;
 748	int ret = 0;
 749
 750	/* Nothing in the space cache, goodbye */
 751	if (!i_size_read(inode))
 752		return 0;
 753
 754	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 755	key.type = 0;
 756	key.offset = offset;
 757
 758	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 759	if (ret < 0)
 760		return 0;
 761	else if (ret > 0) {
 762		btrfs_release_path(path);
 763		return 0;
 764	}
 765
 766	ret = -1;
 767
 768	leaf = path->nodes[0];
 769	header = btrfs_item_ptr(leaf, path->slots[0],
 770				struct btrfs_free_space_header);
 771	num_entries = btrfs_free_space_entries(leaf, header);
 772	num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 773	generation = btrfs_free_space_generation(leaf, header);
 774	btrfs_release_path(path);
 775
 776	if (!BTRFS_I(inode)->generation) {
 777		btrfs_info(fs_info,
 778			   "the free space cache file (%llu) is invalid, skip it",
 779			   offset);
 780		return 0;
 781	}
 782
 783	if (BTRFS_I(inode)->generation != generation) {
 784		btrfs_err(fs_info,
 785			  "free space inode generation (%llu) did not match free space cache generation (%llu)",
 786			  BTRFS_I(inode)->generation, generation);
 787		return 0;
 788	}
 789
 790	if (!num_entries)
 791		return 0;
 792
 793	ret = io_ctl_init(&io_ctl, inode, 0);
 794	if (ret)
 795		return ret;
 796
 797	readahead_cache(inode);
 798
 799	ret = io_ctl_prepare_pages(&io_ctl, true);
 800	if (ret)
 801		goto out;
 802
 803	ret = io_ctl_check_crc(&io_ctl, 0);
 804	if (ret)
 805		goto free_cache;
 806
 807	ret = io_ctl_check_generation(&io_ctl, generation);
 808	if (ret)
 809		goto free_cache;
 810
 811	while (num_entries) {
 812		e = kmem_cache_zalloc(btrfs_free_space_cachep,
 813				      GFP_NOFS);
 814		if (!e) {
 815			ret = -ENOMEM;
 816			goto free_cache;
 817		}
 818
 819		ret = io_ctl_read_entry(&io_ctl, e, &type);
 820		if (ret) {
 821			kmem_cache_free(btrfs_free_space_cachep, e);
 822			goto free_cache;
 823		}
 824
 825		if (!e->bytes) {
 826			ret = -1;
 827			kmem_cache_free(btrfs_free_space_cachep, e);
 828			goto free_cache;
 829		}
 830
 831		if (type == BTRFS_FREE_SPACE_EXTENT) {
 832			spin_lock(&ctl->tree_lock);
 833			ret = link_free_space(ctl, e);
 834			spin_unlock(&ctl->tree_lock);
 835			if (ret) {
 836				btrfs_err(fs_info,
 837					"Duplicate entries in free space cache, dumping");
 838				kmem_cache_free(btrfs_free_space_cachep, e);
 839				goto free_cache;
 840			}
 841		} else {
 842			ASSERT(num_bitmaps);
 843			num_bitmaps--;
 844			e->bitmap = kmem_cache_zalloc(
 845					btrfs_free_space_bitmap_cachep, GFP_NOFS);
 846			if (!e->bitmap) {
 847				ret = -ENOMEM;
 848				kmem_cache_free(
 849					btrfs_free_space_cachep, e);
 850				goto free_cache;
 851			}
 852			spin_lock(&ctl->tree_lock);
 853			ret = link_free_space(ctl, e);
 854			if (ret) {
 855				spin_unlock(&ctl->tree_lock);
 856				btrfs_err(fs_info,
 857					"Duplicate entries in free space cache, dumping");
 858				kmem_cache_free(btrfs_free_space_bitmap_cachep, e->bitmap);
 859				kmem_cache_free(btrfs_free_space_cachep, e);
 860				goto free_cache;
 861			}
 862			ctl->total_bitmaps++;
 863			recalculate_thresholds(ctl);
 864			spin_unlock(&ctl->tree_lock);
 865			list_add_tail(&e->list, &bitmaps);
 866		}
 867
 868		num_entries--;
 869	}
 870
 871	io_ctl_unmap_page(&io_ctl);
 872
 873	/*
 874	 * We add the bitmaps at the end of the entries in order that
 875	 * the bitmap entries are added to the cache.
 876	 */
 877	list_for_each_entry_safe(e, n, &bitmaps, list) {
 878		list_del_init(&e->list);
 879		ret = io_ctl_read_bitmap(&io_ctl, e);
 880		if (ret)
 881			goto free_cache;
 882	}
 883
 884	io_ctl_drop_pages(&io_ctl);
 885	ret = 1;
 886out:
 887	io_ctl_free(&io_ctl);
 888	return ret;
 889free_cache:
 890	io_ctl_drop_pages(&io_ctl);
 891
 892	spin_lock(&ctl->tree_lock);
 893	__btrfs_remove_free_space_cache(ctl);
 894	spin_unlock(&ctl->tree_lock);
 895	goto out;
 896}
 897
 898static int copy_free_space_cache(struct btrfs_block_group *block_group,
 899				 struct btrfs_free_space_ctl *ctl)
 900{
 901	struct btrfs_free_space *info;
 902	struct rb_node *n;
 903	int ret = 0;
 904
 905	while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
 906		info = rb_entry(n, struct btrfs_free_space, offset_index);
 907		if (!info->bitmap) {
 908			const u64 offset = info->offset;
 909			const u64 bytes = info->bytes;
 910
 911			unlink_free_space(ctl, info, true);
 912			spin_unlock(&ctl->tree_lock);
 913			kmem_cache_free(btrfs_free_space_cachep, info);
 914			ret = btrfs_add_free_space(block_group, offset, bytes);
 915			spin_lock(&ctl->tree_lock);
 916		} else {
 917			u64 offset = info->offset;
 918			u64 bytes = ctl->unit;
 919
 920			ret = search_bitmap(ctl, info, &offset, &bytes, false);
 921			if (ret == 0) {
 922				bitmap_clear_bits(ctl, info, offset, bytes, true);
 923				spin_unlock(&ctl->tree_lock);
 924				ret = btrfs_add_free_space(block_group, offset,
 925							   bytes);
 926				spin_lock(&ctl->tree_lock);
 927			} else {
 928				free_bitmap(ctl, info);
 929				ret = 0;
 930			}
 931		}
 932		cond_resched_lock(&ctl->tree_lock);
 933	}
 934	return ret;
 935}
 936
 937static struct lock_class_key btrfs_free_space_inode_key;
 938
 939int load_free_space_cache(struct btrfs_block_group *block_group)
 940{
 941	struct btrfs_fs_info *fs_info = block_group->fs_info;
 942	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 943	struct btrfs_free_space_ctl tmp_ctl = {};
 944	struct inode *inode;
 945	struct btrfs_path *path;
 946	int ret = 0;
 947	bool matched;
 948	u64 used = block_group->used;
 949
 950	/*
 951	 * Because we could potentially discard our loaded free space, we want
 952	 * to load everything into a temporary structure first, and then if it's
 953	 * valid copy it all into the actual free space ctl.
 954	 */
 955	btrfs_init_free_space_ctl(block_group, &tmp_ctl);
 956
 957	/*
 958	 * If this block group has been marked to be cleared for one reason or
 959	 * another then we can't trust the on disk cache, so just return.
 960	 */
 961	spin_lock(&block_group->lock);
 962	if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 963		spin_unlock(&block_group->lock);
 964		return 0;
 965	}
 966	spin_unlock(&block_group->lock);
 967
 968	path = btrfs_alloc_path();
 969	if (!path)
 970		return 0;
 971	path->search_commit_root = true;
 972	path->skip_locking = true;
 973
 974	/*
 975	 * We must pass a path with search_commit_root set to btrfs_iget in
 976	 * order to avoid a deadlock when allocating extents for the tree root.
 977	 *
 978	 * When we are COWing an extent buffer from the tree root, when looking
 979	 * for a free extent, at extent-tree.c:find_free_extent(), we can find
 980	 * block group without its free space cache loaded. When we find one
 981	 * we must load its space cache which requires reading its free space
 982	 * cache's inode item from the root tree. If this inode item is located
 983	 * in the same leaf that we started COWing before, then we end up in
 984	 * deadlock on the extent buffer (trying to read lock it when we
 985	 * previously write locked it).
 986	 *
 987	 * It's safe to read the inode item using the commit root because
 988	 * block groups, once loaded, stay in memory forever (until they are
 989	 * removed) as well as their space caches once loaded. New block groups
 990	 * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 991	 * we will never try to read their inode item while the fs is mounted.
 992	 */
 993	inode = lookup_free_space_inode(block_group, path);
 994	if (IS_ERR(inode)) {
 995		btrfs_free_path(path);
 996		return 0;
 997	}
 998
 999	/* We may have converted the inode and made the cache invalid. */
1000	spin_lock(&block_group->lock);
1001	if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
1002		spin_unlock(&block_group->lock);
1003		btrfs_free_path(path);
1004		goto out;
1005	}
1006	spin_unlock(&block_group->lock);
1007
1008	/*
1009	 * Reinitialize the class of struct inode's mapping->invalidate_lock for
1010	 * free space inodes to prevent false positives related to locks for normal
1011	 * inodes.
1012	 */
1013	lockdep_set_class(&(&inode->i_data)->invalidate_lock,
1014			  &btrfs_free_space_inode_key);
1015
1016	ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
1017				      path, block_group->start);
1018	btrfs_free_path(path);
1019	if (ret <= 0)
1020		goto out;
1021
1022	matched = (tmp_ctl.free_space == (block_group->length - used -
1023					  block_group->bytes_super));
1024
1025	if (matched) {
1026		spin_lock(&tmp_ctl.tree_lock);
1027		ret = copy_free_space_cache(block_group, &tmp_ctl);
1028		spin_unlock(&tmp_ctl.tree_lock);
1029		/*
1030		 * ret == 1 means we successfully loaded the free space cache,
1031		 * so we need to re-set it here.
1032		 */
1033		if (ret == 0)
1034			ret = 1;
1035	} else {
1036		/*
1037		 * We need to call the _locked variant so we don't try to update
1038		 * the discard counters.
1039		 */
1040		spin_lock(&tmp_ctl.tree_lock);
1041		__btrfs_remove_free_space_cache(&tmp_ctl);
1042		spin_unlock(&tmp_ctl.tree_lock);
1043		btrfs_warn(fs_info,
1044			   "block group %llu has wrong amount of free space",
1045			   block_group->start);
1046		ret = -1;
1047	}
1048out:
1049	if (ret < 0) {
1050		/* This cache is bogus, make sure it gets cleared */
1051		spin_lock(&block_group->lock);
1052		block_group->disk_cache_state = BTRFS_DC_CLEAR;
1053		spin_unlock(&block_group->lock);
1054		ret = 0;
1055
1056		btrfs_warn(fs_info,
1057			   "failed to load free space cache for block group %llu, rebuilding it now",
1058			   block_group->start);
1059	}
1060
1061	spin_lock(&ctl->tree_lock);
1062	btrfs_discard_update_discardable(block_group);
1063	spin_unlock(&ctl->tree_lock);
1064	iput(inode);
1065	return ret;
1066}
1067
1068static noinline_for_stack
1069int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1070			      struct btrfs_free_space_ctl *ctl,
1071			      struct btrfs_block_group *block_group,
1072			      int *entries, int *bitmaps,
1073			      struct list_head *bitmap_list)
1074{
1075	int ret;
1076	struct btrfs_free_cluster *cluster = NULL;
1077	struct btrfs_free_cluster *cluster_locked = NULL;
1078	struct rb_node *node = rb_first(&ctl->free_space_offset);
1079	struct btrfs_trim_range *trim_entry;
1080
1081	/* Get the cluster for this block_group if it exists */
1082	if (block_group && !list_empty(&block_group->cluster_list)) {
1083		cluster = list_first_entry(&block_group->cluster_list,
1084					   struct btrfs_free_cluster, block_group_list);
1085	}
1086
1087	if (!node && cluster) {
1088		cluster_locked = cluster;
1089		spin_lock(&cluster_locked->lock);
1090		node = rb_first(&cluster->root);
1091		cluster = NULL;
1092	}
1093
1094	/* Write out the extent entries */
1095	while (node) {
1096		struct btrfs_free_space *e;
1097
1098		e = rb_entry(node, struct btrfs_free_space, offset_index);
1099		*entries += 1;
1100
1101		ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1102				       e->bitmap);
1103		if (ret)
1104			goto fail;
1105
1106		if (e->bitmap) {
1107			list_add_tail(&e->list, bitmap_list);
1108			*bitmaps += 1;
1109		}
1110		node = rb_next(node);
1111		if (!node && cluster) {
1112			node = rb_first(&cluster->root);
1113			cluster_locked = cluster;
1114			spin_lock(&cluster_locked->lock);
1115			cluster = NULL;
1116		}
1117	}
1118	if (cluster_locked) {
1119		spin_unlock(&cluster_locked->lock);
1120		cluster_locked = NULL;
1121	}
1122
1123	/*
1124	 * Make sure we don't miss any range that was removed from our rbtree
1125	 * because trimming is running. Otherwise after a umount+mount (or crash
1126	 * after committing the transaction) we would leak free space and get
1127	 * an inconsistent free space cache report from fsck.
1128	 */
1129	list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1130		ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1131				       trim_entry->bytes, NULL);
1132		if (ret)
1133			goto fail;
1134		*entries += 1;
1135	}
1136
1137	return 0;
1138fail:
1139	if (cluster_locked)
1140		spin_unlock(&cluster_locked->lock);
1141	return -ENOSPC;
1142}
1143
1144static noinline_for_stack int
1145update_cache_item(struct btrfs_trans_handle *trans,
1146		  struct btrfs_root *root,
1147		  struct inode *inode,
1148		  struct btrfs_path *path, u64 offset,
1149		  int entries, int bitmaps)
1150{
1151	struct btrfs_key key;
1152	struct btrfs_free_space_header *header;
1153	struct extent_buffer *leaf;
1154	int ret;
1155
1156	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1157	key.type = 0;
1158	key.offset = offset;
1159
1160	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1161	if (ret < 0) {
1162		btrfs_clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1163				       EXTENT_DELALLOC, NULL);
1164		goto fail;
1165	}
1166	leaf = path->nodes[0];
1167	if (ret > 0) {
1168		struct btrfs_key found_key;
1169		ASSERT(path->slots[0]);
1170		path->slots[0]--;
1171		btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1172		if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1173		    found_key.offset != offset) {
1174			btrfs_clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1175					       inode->i_size - 1, EXTENT_DELALLOC,
1176					       NULL);
1177			btrfs_release_path(path);
1178			goto fail;
1179		}
1180	}
1181
1182	BTRFS_I(inode)->generation = trans->transid;
1183	header = btrfs_item_ptr(leaf, path->slots[0],
1184				struct btrfs_free_space_header);
1185	btrfs_set_free_space_entries(leaf, header, entries);
1186	btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1187	btrfs_set_free_space_generation(leaf, header, trans->transid);
1188	btrfs_release_path(path);
1189
1190	return 0;
1191
1192fail:
1193	return -1;
1194}
1195
1196static noinline_for_stack int write_pinned_extent_entries(
1197			    struct btrfs_trans_handle *trans,
1198			    struct btrfs_block_group *block_group,
1199			    struct btrfs_io_ctl *io_ctl,
1200			    int *entries)
1201{
1202	u64 start, extent_start, extent_end, len;
1203	struct extent_io_tree *unpin = NULL;
1204	int ret;
1205
1206	if (!block_group)
1207		return 0;
1208
1209	/*
1210	 * We want to add any pinned extents to our free space cache
1211	 * so we don't leak the space
1212	 *
1213	 * We shouldn't have switched the pinned extents yet so this is the
1214	 * right one
1215	 */
1216	unpin = &trans->transaction->pinned_extents;
1217
1218	start = block_group->start;
1219
1220	while (start < block_group->start + block_group->length) {
1221		if (!btrfs_find_first_extent_bit(unpin, start,
1222						 &extent_start, &extent_end,
1223						 EXTENT_DIRTY, NULL))
1224			return 0;
1225
1226		/* This pinned extent is out of our range */
1227		if (extent_start >= block_group->start + block_group->length)
1228			return 0;
1229
1230		extent_start = max(extent_start, start);
1231		extent_end = min(block_group->start + block_group->length,
1232				 extent_end + 1);
1233		len = extent_end - extent_start;
1234
1235		*entries += 1;
1236		ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1237		if (ret)
1238			return -ENOSPC;
1239
1240		start = extent_end;
1241	}
1242
1243	return 0;
1244}
1245
1246static noinline_for_stack int
1247write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1248{
1249	struct btrfs_free_space *entry, *next;
1250	int ret;
1251
1252	/* Write out the bitmaps */
1253	list_for_each_entry_safe(entry, next, bitmap_list, list) {
1254		ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1255		if (ret)
1256			return -ENOSPC;
1257		list_del_init(&entry->list);
1258	}
1259
1260	return 0;
1261}
1262
1263static int flush_dirty_cache(struct inode *inode)
1264{
1265	int ret;
1266
1267	ret = btrfs_wait_ordered_range(BTRFS_I(inode), 0, (u64)-1);
1268	if (ret)
1269		btrfs_clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1270				       EXTENT_DELALLOC, NULL);
1271
1272	return ret;
1273}
1274
1275static void noinline_for_stack
1276cleanup_bitmap_list(struct list_head *bitmap_list)
1277{
1278	struct btrfs_free_space *entry, *next;
1279
1280	list_for_each_entry_safe(entry, next, bitmap_list, list)
1281		list_del_init(&entry->list);
1282}
1283
1284static void noinline_for_stack
1285cleanup_write_cache_enospc(struct inode *inode,
1286			   struct btrfs_io_ctl *io_ctl,
1287			   struct extent_state **cached_state)
1288{
1289	io_ctl_drop_pages(io_ctl);
1290	btrfs_unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1291			    cached_state);
1292}
1293
1294static int __btrfs_wait_cache_io(struct btrfs_root *root,
1295				 struct btrfs_trans_handle *trans,
1296				 struct btrfs_block_group *block_group,
1297				 struct btrfs_io_ctl *io_ctl,
1298				 struct btrfs_path *path, u64 offset)
1299{
1300	int ret;
1301	struct inode *inode = io_ctl->inode;
1302
1303	if (!inode)
1304		return 0;
1305
1306	/* Flush the dirty pages in the cache file. */
1307	ret = flush_dirty_cache(inode);
1308	if (ret)
1309		goto out;
1310
1311	/* Update the cache item to tell everyone this cache file is valid. */
1312	ret = update_cache_item(trans, root, inode, path, offset,
1313				io_ctl->entries, io_ctl->bitmaps);
1314out:
1315	if (ret) {
1316		invalidate_inode_pages2(inode->i_mapping);
1317		BTRFS_I(inode)->generation = 0;
1318		if (block_group)
1319			btrfs_debug(root->fs_info,
1320	  "failed to write free space cache for block group %llu error %d",
1321				  block_group->start, ret);
1322	}
1323	btrfs_update_inode(trans, BTRFS_I(inode));
1324
1325	if (block_group) {
1326		/* the dirty list is protected by the dirty_bgs_lock */
1327		spin_lock(&trans->transaction->dirty_bgs_lock);
1328
1329		/* the disk_cache_state is protected by the block group lock */
1330		spin_lock(&block_group->lock);
1331
1332		/*
1333		 * only mark this as written if we didn't get put back on
1334		 * the dirty list while waiting for IO.   Otherwise our
1335		 * cache state won't be right, and we won't get written again
1336		 */
1337		if (!ret && list_empty(&block_group->dirty_list))
1338			block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1339		else if (ret)
1340			block_group->disk_cache_state = BTRFS_DC_ERROR;
1341
1342		spin_unlock(&block_group->lock);
1343		spin_unlock(&trans->transaction->dirty_bgs_lock);
1344		io_ctl->inode = NULL;
1345		iput(inode);
1346	}
1347
1348	return ret;
1349
1350}
1351
1352int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1353			struct btrfs_block_group *block_group,
1354			struct btrfs_path *path)
1355{
1356	return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1357				     block_group, &block_group->io_ctl,
1358				     path, block_group->start);
1359}
1360
1361/*
1362 * Write out cached info to an inode.
1363 *
1364 * @inode:       freespace inode we are writing out
1365 * @ctl:         free space cache we are going to write out
1366 * @block_group: block_group for this cache if it belongs to a block_group
1367 * @io_ctl:      holds context for the io
1368 * @trans:       the trans handle
1369 *
1370 * This function writes out a free space cache struct to disk for quick recovery
1371 * on mount.  This will return 0 if it was successful in writing the cache out,
1372 * or an errno if it was not.
1373 */
1374static int __btrfs_write_out_cache(struct inode *inode,
1375				   struct btrfs_free_space_ctl *ctl,
1376				   struct btrfs_block_group *block_group,
1377				   struct btrfs_io_ctl *io_ctl,
1378				   struct btrfs_trans_handle *trans)
1379{
1380	struct extent_state *cached_state = NULL;
1381	LIST_HEAD(bitmap_list);
1382	int entries = 0;
1383	int bitmaps = 0;
1384	int ret;
1385	int must_iput = 0;
1386	int i_size;
1387
1388	if (!i_size_read(inode))
1389		return -EIO;
1390
1391	WARN_ON(io_ctl->pages);
1392	ret = io_ctl_init(io_ctl, inode, 1);
1393	if (ret)
1394		return ret;
1395
1396	if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1397		down_write(&block_group->data_rwsem);
1398		spin_lock(&block_group->lock);
1399		if (block_group->delalloc_bytes) {
1400			block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1401			spin_unlock(&block_group->lock);
1402			up_write(&block_group->data_rwsem);
1403			BTRFS_I(inode)->generation = 0;
1404			ret = 0;
1405			must_iput = 1;
1406			goto out;
1407		}
1408		spin_unlock(&block_group->lock);
1409	}
1410
1411	/* Lock all pages first so we can lock the extent safely. */
1412	ret = io_ctl_prepare_pages(io_ctl, false);
1413	if (ret)
1414		goto out_unlock;
1415
1416	btrfs_lock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1417			  &cached_state);
1418
1419	io_ctl_set_generation(io_ctl, trans->transid);
1420
1421	mutex_lock(&ctl->cache_writeout_mutex);
1422	/* Write out the extent entries in the free space cache */
1423	spin_lock(&ctl->tree_lock);
1424	ret = write_cache_extent_entries(io_ctl, ctl,
1425					 block_group, &entries, &bitmaps,
1426					 &bitmap_list);
1427	if (ret)
1428		goto out_nospc_locked;
1429
1430	/*
1431	 * Some spaces that are freed in the current transaction are pinned,
1432	 * they will be added into free space cache after the transaction is
1433	 * committed, we shouldn't lose them.
1434	 *
1435	 * If this changes while we are working we'll get added back to
1436	 * the dirty list and redo it.  No locking needed
1437	 */
1438	ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1439	if (ret)
1440		goto out_nospc_locked;
1441
1442	/*
1443	 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1444	 * locked while doing it because a concurrent trim can be manipulating
1445	 * or freeing the bitmap.
1446	 */
1447	ret = write_bitmap_entries(io_ctl, &bitmap_list);
1448	spin_unlock(&ctl->tree_lock);
1449	mutex_unlock(&ctl->cache_writeout_mutex);
1450	if (ret)
1451		goto out_nospc;
1452
1453	/* Zero out the rest of the pages just to make sure */
1454	io_ctl_zero_remaining_pages(io_ctl);
1455
1456	/* Everything is written out, now we dirty the pages in the file. */
1457	i_size = i_size_read(inode);
1458	for (int i = 0; i < round_up(i_size, PAGE_SIZE) / PAGE_SIZE; i++) {
1459		u64 dirty_start = i * PAGE_SIZE;
1460		u64 dirty_len = min_t(u64, dirty_start + PAGE_SIZE, i_size) - dirty_start;
1461
1462		ret = btrfs_dirty_folio(BTRFS_I(inode), page_folio(io_ctl->pages[i]),
1463					dirty_start, dirty_len, &cached_state, false);
1464		if (ret < 0)
1465			goto out_nospc;
1466	}
1467
1468	if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1469		up_write(&block_group->data_rwsem);
1470	/*
1471	 * Release the pages and unlock the extent, we will flush
1472	 * them out later
1473	 */
1474	io_ctl_drop_pages(io_ctl);
1475	io_ctl_free(io_ctl);
1476
1477	btrfs_unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1478			    &cached_state);
1479
1480	/*
1481	 * at this point the pages are under IO and we're happy,
1482	 * The caller is responsible for waiting on them and updating
1483	 * the cache and the inode
1484	 */
1485	io_ctl->entries = entries;
1486	io_ctl->bitmaps = bitmaps;
1487
1488	ret = btrfs_fdatawrite_range(BTRFS_I(inode), 0, (u64)-1);
1489	if (ret)
1490		goto out;
1491
1492	return 0;
1493
1494out_nospc_locked:
1495	cleanup_bitmap_list(&bitmap_list);
1496	spin_unlock(&ctl->tree_lock);
1497	mutex_unlock(&ctl->cache_writeout_mutex);
1498
1499out_nospc:
1500	cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1501
1502out_unlock:
1503	if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1504		up_write(&block_group->data_rwsem);
1505
1506out:
1507	io_ctl->inode = NULL;
1508	io_ctl_free(io_ctl);
1509	if (ret) {
1510		invalidate_inode_pages2(inode->i_mapping);
1511		BTRFS_I(inode)->generation = 0;
1512	}
1513	btrfs_update_inode(trans, BTRFS_I(inode));
1514	if (must_iput)
1515		iput(inode);
1516	return ret;
1517}
1518
1519int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1520			  struct btrfs_block_group *block_group,
1521			  struct btrfs_path *path)
1522{
1523	struct btrfs_fs_info *fs_info = trans->fs_info;
1524	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1525	struct inode *inode;
1526	int ret = 0;
1527
1528	spin_lock(&block_group->lock);
1529	if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1530		spin_unlock(&block_group->lock);
1531		return 0;
1532	}
1533	spin_unlock(&block_group->lock);
1534
1535	inode = lookup_free_space_inode(block_group, path);
1536	if (IS_ERR(inode))
1537		return 0;
1538
1539	ret = __btrfs_write_out_cache(inode, ctl, block_group,
1540				      &block_group->io_ctl, trans);
1541	if (ret) {
1542		btrfs_debug(fs_info,
1543	  "failed to write free space cache for block group %llu error %d",
1544			  block_group->start, ret);
1545		spin_lock(&block_group->lock);
1546		block_group->disk_cache_state = BTRFS_DC_ERROR;
1547		spin_unlock(&block_group->lock);
1548
1549		block_group->io_ctl.inode = NULL;
1550		iput(inode);
1551	}
1552
1553	/*
1554	 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1555	 * to wait for IO and put the inode
1556	 */
1557
1558	return ret;
1559}
1560
1561static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1562					  u64 offset)
1563{
1564	ASSERT(offset >= bitmap_start);
1565	offset -= bitmap_start;
1566	return (unsigned long)(div_u64(offset, unit));
1567}
1568
1569static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1570{
1571	return (unsigned long)(div_u64(bytes, unit));
1572}
1573
1574static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1575				   u64 offset)
1576{
1577	u64 bitmap_start;
1578	u64 bytes_per_bitmap;
1579
1580	bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1581	bitmap_start = offset - ctl->start;
1582	bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1583	bitmap_start *= bytes_per_bitmap;
1584	bitmap_start += ctl->start;
1585
1586	return bitmap_start;
1587}
1588
1589static int tree_insert_offset(struct btrfs_free_space_ctl *ctl,
1590			      struct btrfs_free_cluster *cluster,
1591			      struct btrfs_free_space *new_entry)
1592{
1593	struct rb_root *root;
1594	struct rb_node **p;
1595	struct rb_node *parent = NULL;
1596
1597	lockdep_assert_held(&ctl->tree_lock);
1598
1599	if (cluster) {
1600		lockdep_assert_held(&cluster->lock);
1601		root = &cluster->root;
1602	} else {
1603		root = &ctl->free_space_offset;
1604	}
1605
1606	p = &root->rb_node;
1607
1608	while (*p) {
1609		struct btrfs_free_space *info;
1610
1611		parent = *p;
1612		info = rb_entry(parent, struct btrfs_free_space, offset_index);
1613
1614		if (new_entry->offset < info->offset) {
1615			p = &(*p)->rb_left;
1616		} else if (new_entry->offset > info->offset) {
1617			p = &(*p)->rb_right;
1618		} else {
1619			/*
1620			 * we could have a bitmap entry and an extent entry
1621			 * share the same offset.  If this is the case, we want
1622			 * the extent entry to always be found first if we do a
1623			 * linear search through the tree, since we want to have
1624			 * the quickest allocation time, and allocating from an
1625			 * extent is faster than allocating from a bitmap.  So
1626			 * if we're inserting a bitmap and we find an entry at
1627			 * this offset, we want to go right, or after this entry
1628			 * logically.  If we are inserting an extent and we've
1629			 * found a bitmap, we want to go left, or before
1630			 * logically.
1631			 */
1632			if (new_entry->bitmap) {
1633				if (info->bitmap) {
1634					WARN_ON_ONCE(1);
1635					return -EEXIST;
1636				}
1637				p = &(*p)->rb_right;
1638			} else {
1639				if (!info->bitmap) {
1640					WARN_ON_ONCE(1);
1641					return -EEXIST;
1642				}
1643				p = &(*p)->rb_left;
1644			}
1645		}
1646	}
1647
1648	rb_link_node(&new_entry->offset_index, parent, p);
1649	rb_insert_color(&new_entry->offset_index, root);
1650
1651	return 0;
1652}
1653
1654/*
1655 * This is a little subtle.  We *only* have ->max_extent_size set if we actually
1656 * searched through the bitmap and figured out the largest ->max_extent_size,
1657 * otherwise it's 0.  In the case that it's 0 we don't want to tell the
1658 * allocator the wrong thing, we want to use the actual real max_extent_size
1659 * we've found already if it's larger, or we want to use ->bytes.
1660 *
1661 * This matters because find_free_space() will skip entries who's ->bytes is
1662 * less than the required bytes.  So if we didn't search down this bitmap, we
1663 * may pick some previous entry that has a smaller ->max_extent_size than we
1664 * have.  For example, assume we have two entries, one that has
1665 * ->max_extent_size set to 4K and ->bytes set to 1M.  A second entry hasn't set
1666 * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous.  We will
1667 *  call into find_free_space(), and return with max_extent_size == 4K, because
1668 *  that first bitmap entry had ->max_extent_size set, but the second one did
1669 *  not.  If instead we returned 8K we'd come in searching for 8K, and find the
1670 *  8K contiguous range.
1671 *
1672 *  Consider the other case, we have 2 8K chunks in that second entry and still
1673 *  don't have ->max_extent_size set.  We'll return 16K, and the next time the
1674 *  allocator comes in it'll fully search our second bitmap, and this time it'll
1675 *  get an uptodate value of 8K as the maximum chunk size.  Then we'll get the
1676 *  right allocation the next loop through.
1677 */
1678static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1679{
1680	if (entry->bitmap && entry->max_extent_size)
1681		return entry->max_extent_size;
1682	return entry->bytes;
1683}
1684
1685/*
1686 * We want the largest entry to be leftmost, so this is inverted from what you'd
1687 * normally expect.
1688 */
1689static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1690{
1691	const struct btrfs_free_space *entry, *exist;
1692
1693	entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1694	exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1695	return get_max_extent_size(exist) < get_max_extent_size(entry);
1696}
1697
1698/*
1699 * searches the tree for the given offset.
1700 *
1701 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1702 * want a section that has at least bytes size and comes at or after the given
1703 * offset.
1704 */
1705static struct btrfs_free_space *
1706tree_search_offset(struct btrfs_free_space_ctl *ctl,
1707		   u64 offset, int bitmap_only, int fuzzy)
1708{
1709	struct rb_node *n = ctl->free_space_offset.rb_node;
1710	struct btrfs_free_space *entry = NULL, *prev = NULL;
1711
1712	lockdep_assert_held(&ctl->tree_lock);
1713
1714	/* find entry that is closest to the 'offset' */
1715	while (n) {
1716		entry = rb_entry(n, struct btrfs_free_space, offset_index);
1717		prev = entry;
1718
1719		if (offset < entry->offset)
1720			n = n->rb_left;
1721		else if (offset > entry->offset)
1722			n = n->rb_right;
1723		else
1724			break;
1725
1726		entry = NULL;
1727	}
1728
1729	if (bitmap_only) {
1730		if (!entry)
1731			return NULL;
1732		if (entry->bitmap)
1733			return entry;
1734
1735		/*
1736		 * bitmap entry and extent entry may share same offset,
1737		 * in that case, bitmap entry comes after extent entry.
1738		 */
1739		n = rb_next(n);
1740		if (!n)
1741			return NULL;
1742		entry = rb_entry(n, struct btrfs_free_space, offset_index);
1743		if (entry->offset != offset)
1744			return NULL;
1745
1746		WARN_ON(!entry->bitmap);
1747		return entry;
1748	} else if (entry) {
1749		if (entry->bitmap) {
1750			/*
1751			 * if previous extent entry covers the offset,
1752			 * we should return it instead of the bitmap entry
1753			 */
1754			n = rb_prev(&entry->offset_index);
1755			if (n) {
1756				prev = rb_entry(n, struct btrfs_free_space,
1757						offset_index);
1758				if (!prev->bitmap &&
1759				    prev->offset + prev->bytes > offset)
1760					entry = prev;
1761			}
1762		}
1763		return entry;
1764	}
1765
1766	if (!prev)
1767		return NULL;
1768
1769	/* find last entry before the 'offset' */
1770	entry = prev;
1771	if (entry->offset > offset) {
1772		n = rb_prev(&entry->offset_index);
1773		if (n) {
1774			entry = rb_entry(n, struct btrfs_free_space,
1775					offset_index);
1776			ASSERT(entry->offset <= offset);
1777		} else {
1778			if (fuzzy)
1779				return entry;
1780			else
1781				return NULL;
1782		}
1783	}
1784
1785	if (entry->bitmap) {
1786		n = rb_prev(&entry->offset_index);
1787		if (n) {
1788			prev = rb_entry(n, struct btrfs_free_space,
1789					offset_index);
1790			if (!prev->bitmap &&
1791			    prev->offset + prev->bytes > offset)
1792				return prev;
1793		}
1794		if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1795			return entry;
1796	} else if (entry->offset + entry->bytes > offset)
1797		return entry;
1798
1799	if (!fuzzy)
1800		return NULL;
1801
1802	while (1) {
1803		n = rb_next(&entry->offset_index);
1804		if (!n)
1805			return NULL;
1806		entry = rb_entry(n, struct btrfs_free_space, offset_index);
1807		if (entry->bitmap) {
1808			if (entry->offset + BITS_PER_BITMAP *
1809			    ctl->unit > offset)
1810				break;
1811		} else {
1812			if (entry->offset + entry->bytes > offset)
1813				break;
1814		}
1815	}
1816	return entry;
1817}
1818
1819static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1820				     struct btrfs_free_space *info,
1821				     bool update_stat)
1822{
1823	lockdep_assert_held(&ctl->tree_lock);
1824
1825	rb_erase(&info->offset_index, &ctl->free_space_offset);
1826	rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1827	ctl->free_extents--;
1828
1829	if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1830		ctl->discardable_extents[BTRFS_STAT_CURR]--;
1831		ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1832	}
1833
1834	if (update_stat)
1835		ctl->free_space -= info->bytes;
1836}
1837
1838static int link_free_space(struct btrfs_free_space_ctl *ctl,
1839			   struct btrfs_free_space *info)
1840{
1841	int ret = 0;
1842
1843	lockdep_assert_held(&ctl->tree_lock);
1844
1845	ASSERT(info->bytes || info->bitmap);
1846	ret = tree_insert_offset(ctl, NULL, info);
1847	if (ret)
1848		return ret;
1849
1850	rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1851
1852	if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1853		ctl->discardable_extents[BTRFS_STAT_CURR]++;
1854		ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1855	}
1856
1857	ctl->free_space += info->bytes;
1858	ctl->free_extents++;
1859	return ret;
1860}
1861
1862static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1863				struct btrfs_free_space *info)
1864{
1865	ASSERT(info->bitmap);
1866
1867	/*
1868	 * If our entry is empty it's because we're on a cluster and we don't
1869	 * want to re-link it into our ctl bytes index.
1870	 */
1871	if (RB_EMPTY_NODE(&info->bytes_index))
1872		return;
1873
1874	lockdep_assert_held(&ctl->tree_lock);
1875
1876	rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1877	rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1878}
1879
1880static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1881				     struct btrfs_free_space *info,
1882				     u64 offset, u64 bytes, bool update_stat)
1883{
1884	unsigned long start, count, end;
1885	int extent_delta = -1;
1886
1887	start = offset_to_bit(info->offset, ctl->unit, offset);
1888	count = bytes_to_bits(bytes, ctl->unit);
1889	end = start + count;
1890	ASSERT(end <= BITS_PER_BITMAP);
1891
1892	bitmap_clear(info->bitmap, start, count);
1893
1894	info->bytes -= bytes;
1895	if (info->max_extent_size > ctl->unit)
1896		info->max_extent_size = 0;
1897
1898	relink_bitmap_entry(ctl, info);
1899
1900	if (start && test_bit(start - 1, info->bitmap))
1901		extent_delta++;
1902
1903	if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1904		extent_delta++;
1905
1906	info->bitmap_extents += extent_delta;
1907	if (!btrfs_free_space_trimmed(info)) {
1908		ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1909		ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1910	}
1911
1912	if (update_stat)
1913		ctl->free_space -= bytes;
1914}
1915
1916static void btrfs_bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1917				  struct btrfs_free_space *info, u64 offset,
1918				  u64 bytes)
1919{
1920	unsigned long start, count, end;
1921	int extent_delta = 1;
1922
1923	start = offset_to_bit(info->offset, ctl->unit, offset);
1924	count = bytes_to_bits(bytes, ctl->unit);
1925	end = start + count;
1926	ASSERT(end <= BITS_PER_BITMAP);
1927
1928	bitmap_set(info->bitmap, start, count);
1929
1930	/*
1931	 * We set some bytes, we have no idea what the max extent size is
1932	 * anymore.
1933	 */
1934	info->max_extent_size = 0;
1935	info->bytes += bytes;
1936	ctl->free_space += bytes;
1937
1938	relink_bitmap_entry(ctl, info);
1939
1940	if (start && test_bit(start - 1, info->bitmap))
1941		extent_delta--;
1942
1943	if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1944		extent_delta--;
1945
1946	info->bitmap_extents += extent_delta;
1947	if (!btrfs_free_space_trimmed(info)) {
1948		ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1949		ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1950	}
1951}
1952
1953/*
1954 * If we can not find suitable extent, we will use bytes to record
1955 * the size of the max extent.
1956 */
1957static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1958			 struct btrfs_free_space *bitmap_info, u64 *offset,
1959			 u64 *bytes, bool for_alloc)
1960{
1961	unsigned long found_bits = 0;
1962	unsigned long max_bits = 0;
1963	unsigned long bits, i;
1964	unsigned long next_zero;
1965	unsigned long extent_bits;
1966
1967	/*
1968	 * Skip searching the bitmap if we don't have a contiguous section that
1969	 * is large enough for this allocation.
1970	 */
1971	if (for_alloc &&
1972	    bitmap_info->max_extent_size &&
1973	    bitmap_info->max_extent_size < *bytes) {
1974		*bytes = bitmap_info->max_extent_size;
1975		return -1;
1976	}
1977
1978	i = offset_to_bit(bitmap_info->offset, ctl->unit,
1979			  max_t(u64, *offset, bitmap_info->offset));
1980	bits = bytes_to_bits(*bytes, ctl->unit);
1981
1982	for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1983		if (for_alloc && bits == 1) {
1984			found_bits = 1;
1985			break;
1986		}
1987		next_zero = find_next_zero_bit(bitmap_info->bitmap,
1988					       BITS_PER_BITMAP, i);
1989		extent_bits = next_zero - i;
1990		if (extent_bits >= bits) {
1991			found_bits = extent_bits;
1992			break;
1993		} else if (extent_bits > max_bits) {
1994			max_bits = extent_bits;
1995		}
1996		i = next_zero;
1997	}
1998
1999	if (found_bits) {
2000		*offset = (u64)(i * ctl->unit) + bitmap_info->offset;
2001		*bytes = (u64)(found_bits) * ctl->unit;
2002		return 0;
2003	}
2004
2005	*bytes = (u64)(max_bits) * ctl->unit;
2006	bitmap_info->max_extent_size = *bytes;
2007	relink_bitmap_entry(ctl, bitmap_info);
2008	return -1;
2009}
2010
2011/* Cache the size of the max extent in bytes */
2012static struct btrfs_free_space *
2013find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
2014		unsigned long align, u64 *max_extent_size, bool use_bytes_index)
2015{
2016	struct btrfs_free_space *entry;
2017	struct rb_node *node;
2018	u64 tmp;
2019	u64 align_off;
2020	int ret;
2021
2022	if (!ctl->free_space_offset.rb_node)
2023		goto out;
2024again:
2025	if (use_bytes_index) {
2026		node = rb_first_cached(&ctl->free_space_bytes);
2027	} else {
2028		entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
2029					   0, 1);
2030		if (!entry)
2031			goto out;
2032		node = &entry->offset_index;
2033	}
2034
2035	for (; node; node = rb_next(node)) {
2036		if (use_bytes_index)
2037			entry = rb_entry(node, struct btrfs_free_space,
2038					 bytes_index);
2039		else
2040			entry = rb_entry(node, struct btrfs_free_space,
2041					 offset_index);
2042
2043		/*
2044		 * If we are using the bytes index then all subsequent entries
2045		 * in this tree are going to be < bytes, so simply set the max
2046		 * extent size and exit the loop.
2047		 *
2048		 * If we're using the offset index then we need to keep going
2049		 * through the rest of the tree.
2050		 */
2051		if (entry->bytes < *bytes) {
2052			*max_extent_size = max(get_max_extent_size(entry),
2053					       *max_extent_size);
2054			if (use_bytes_index)
2055				break;
2056			continue;
2057		}
2058
2059		/* make sure the space returned is big enough
2060		 * to match our requested alignment
2061		 */
2062		if (*bytes >= align) {
2063			tmp = entry->offset - ctl->start + align - 1;
2064			tmp = div64_u64(tmp, align);
2065			tmp = tmp * align + ctl->start;
2066			align_off = tmp - entry->offset;
2067		} else {
2068			align_off = 0;
2069			tmp = entry->offset;
2070		}
2071
2072		/*
2073		 * We don't break here if we're using the bytes index because we
2074		 * may have another entry that has the correct alignment that is
2075		 * the right size, so we don't want to miss that possibility.
2076		 * At worst this adds another loop through the logic, but if we
2077		 * broke here we could prematurely ENOSPC.
2078		 */
2079		if (entry->bytes < *bytes + align_off) {
2080			*max_extent_size = max(get_max_extent_size(entry),
2081					       *max_extent_size);
2082			continue;
2083		}
2084
2085		if (entry->bitmap) {
2086			struct rb_node *old_next = rb_next(node);
2087			u64 size = *bytes;
2088
2089			ret = search_bitmap(ctl, entry, &tmp, &size, true);
2090			if (!ret) {
2091				*offset = tmp;
2092				*bytes = size;
2093				return entry;
2094			} else {
2095				*max_extent_size =
2096					max(get_max_extent_size(entry),
2097					    *max_extent_size);
2098			}
2099
2100			/*
2101			 * The bitmap may have gotten re-arranged in the space
2102			 * index here because the max_extent_size may have been
2103			 * updated.  Start from the beginning again if this
2104			 * happened.
2105			 */
2106			if (use_bytes_index && old_next != rb_next(node))
2107				goto again;
2108			continue;
2109		}
2110
2111		*offset = tmp;
2112		*bytes = entry->bytes - align_off;
2113		return entry;
2114	}
2115out:
2116	return NULL;
2117}
2118
2119static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2120			   struct btrfs_free_space *info, u64 offset)
2121{
2122	info->offset = offset_to_bitmap(ctl, offset);
2123	info->bytes = 0;
2124	info->bitmap_extents = 0;
2125	INIT_LIST_HEAD(&info->list);
2126	link_free_space(ctl, info);
2127	ctl->total_bitmaps++;
2128	recalculate_thresholds(ctl);
2129}
2130
2131static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2132			struct btrfs_free_space *bitmap_info)
2133{
2134	/*
2135	 * Normally when this is called, the bitmap is completely empty. However,
2136	 * if we are blowing up the free space cache for one reason or another
2137	 * via __btrfs_remove_free_space_cache(), then it may not be freed and
2138	 * we may leave stats on the table.
2139	 */
2140	if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2141		ctl->discardable_extents[BTRFS_STAT_CURR] -=
2142			bitmap_info->bitmap_extents;
2143		ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2144
2145	}
2146	unlink_free_space(ctl, bitmap_info, true);
2147	kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2148	kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2149	ctl->total_bitmaps--;
2150	recalculate_thresholds(ctl);
2151}
2152
2153static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2154			      struct btrfs_free_space *bitmap_info,
2155			      u64 *offset, u64 *bytes)
2156{
2157	u64 end;
2158	u64 search_start, search_bytes;
2159	int ret;
2160
2161again:
2162	end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2163
2164	/*
2165	 * We need to search for bits in this bitmap.  We could only cover some
2166	 * of the extent in this bitmap thanks to how we add space, so we need
2167	 * to search for as much as it as we can and clear that amount, and then
2168	 * go searching for the next bit.
2169	 */
2170	search_start = *offset;
2171	search_bytes = ctl->unit;
2172	search_bytes = min(search_bytes, end - search_start + 1);
2173	ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2174			    false);
2175	if (ret < 0 || search_start != *offset)
2176		return -EINVAL;
2177
2178	/* We may have found more bits than what we need */
2179	search_bytes = min(search_bytes, *bytes);
2180
2181	/* Cannot clear past the end of the bitmap */
2182	search_bytes = min(search_bytes, end - search_start + 1);
2183
2184	bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2185	*offset += search_bytes;
2186	*bytes -= search_bytes;
2187
2188	if (*bytes) {
2189		struct rb_node *next = rb_next(&bitmap_info->offset_index);
2190		if (!bitmap_info->bytes)
2191			free_bitmap(ctl, bitmap_info);
2192
2193		/*
2194		 * no entry after this bitmap, but we still have bytes to
2195		 * remove, so something has gone wrong.
2196		 */
2197		if (!next)
2198			return -EINVAL;
2199
2200		bitmap_info = rb_entry(next, struct btrfs_free_space,
2201				       offset_index);
2202
2203		/*
2204		 * if the next entry isn't a bitmap we need to return to let the
2205		 * extent stuff do its work.
2206		 */
2207		if (!bitmap_info->bitmap)
2208			return -EAGAIN;
2209
2210		/*
2211		 * Ok the next item is a bitmap, but it may not actually hold
2212		 * the information for the rest of this free space stuff, so
2213		 * look for it, and if we don't find it return so we can try
2214		 * everything over again.
2215		 */
2216		search_start = *offset;
2217		search_bytes = ctl->unit;
2218		ret = search_bitmap(ctl, bitmap_info, &search_start,
2219				    &search_bytes, false);
2220		if (ret < 0 || search_start != *offset)
2221			return -EAGAIN;
2222
2223		goto again;
2224	} else if (!bitmap_info->bytes)
2225		free_bitmap(ctl, bitmap_info);
2226
2227	return 0;
2228}
2229
2230static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2231			       struct btrfs_free_space *info, u64 offset,
2232			       u64 bytes, enum btrfs_trim_state trim_state)
2233{
2234	u64 bytes_to_set = 0;
2235	u64 end;
2236
2237	/*
2238	 * This is a tradeoff to make bitmap trim state minimal.  We mark the
2239	 * whole bitmap untrimmed if at any point we add untrimmed regions.
2240	 */
2241	if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2242		if (btrfs_free_space_trimmed(info)) {
2243			ctl->discardable_extents[BTRFS_STAT_CURR] +=
2244				info->bitmap_extents;
2245			ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2246		}
2247		info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2248	}
2249
2250	end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2251
2252	bytes_to_set = min(end - offset, bytes);
2253
2254	btrfs_bitmap_set_bits(ctl, info, offset, bytes_to_set);
2255
2256	return bytes_to_set;
2257
2258}
2259
2260static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2261		      struct btrfs_free_space *info)
2262{
2263	struct btrfs_block_group *block_group = ctl->block_group;
2264	struct btrfs_fs_info *fs_info = block_group->fs_info;
2265	bool forced = false;
2266
2267#ifdef CONFIG_BTRFS_DEBUG
2268	if (btrfs_should_fragment_free_space(block_group))
2269		forced = true;
2270#endif
2271
2272	/* This is a way to reclaim large regions from the bitmaps. */
2273	if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2274		return false;
2275
2276	/*
2277	 * If we are below the extents threshold then we can add this as an
2278	 * extent, and don't have to deal with the bitmap
2279	 */
2280	if (!forced && ctl->free_extents < ctl->extents_thresh) {
2281		/*
2282		 * If this block group has some small extents we don't want to
2283		 * use up all of our free slots in the cache with them, we want
2284		 * to reserve them to larger extents, however if we have plenty
2285		 * of cache left then go ahead and add them, no sense in adding
2286		 * the overhead of a bitmap if we don't have to.
2287		 */
2288		if (info->bytes <= fs_info->sectorsize * 8) {
2289			if (ctl->free_extents * 3 <= ctl->extents_thresh)
2290				return false;
2291		} else {
2292			return false;
2293		}
2294	}
2295
2296	/*
2297	 * The original block groups from mkfs can be really small, like 8
2298	 * megabytes, so don't bother with a bitmap for those entries.  However
2299	 * some block groups can be smaller than what a bitmap would cover but
2300	 * are still large enough that they could overflow the 32k memory limit,
2301	 * so allow those block groups to still be allowed to have a bitmap
2302	 * entry.
2303	 */
2304	if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2305		return false;
2306
2307	return true;
2308}
2309
2310static const struct btrfs_free_space_op free_space_op = {
2311	.use_bitmap		= use_bitmap,
2312};
2313
2314static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2315			      struct btrfs_free_space *info)
2316{
2317	struct btrfs_free_space *bitmap_info;
2318	struct btrfs_block_group *block_group = NULL;
2319	int added = 0;
2320	u64 bytes, offset, bytes_added;
2321	enum btrfs_trim_state trim_state;
2322	int ret;
2323
2324	bytes = info->bytes;
2325	offset = info->offset;
2326	trim_state = info->trim_state;
2327
2328	if (!ctl->op->use_bitmap(ctl, info))
2329		return 0;
2330
2331	if (ctl->op == &free_space_op)
2332		block_group = ctl->block_group;
2333again:
2334	/*
2335	 * Since we link bitmaps right into the cluster we need to see if we
2336	 * have a cluster here, and if so and it has our bitmap we need to add
2337	 * the free space to that bitmap.
2338	 */
2339	if (block_group && !list_empty(&block_group->cluster_list)) {
2340		struct btrfs_free_cluster *cluster;
2341		struct rb_node *node;
2342		struct btrfs_free_space *entry;
2343
2344		cluster = list_first_entry(&block_group->cluster_list,
2345					   struct btrfs_free_cluster, block_group_list);
2346		spin_lock(&cluster->lock);
2347		node = rb_first(&cluster->root);
2348		if (!node) {
2349			spin_unlock(&cluster->lock);
2350			goto no_cluster_bitmap;
2351		}
2352
2353		entry = rb_entry(node, struct btrfs_free_space, offset_index);
2354		if (!entry->bitmap) {
2355			spin_unlock(&cluster->lock);
2356			goto no_cluster_bitmap;
2357		}
2358
2359		if (entry->offset == offset_to_bitmap(ctl, offset)) {
2360			bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2361							  bytes, trim_state);
2362			bytes -= bytes_added;
2363			offset += bytes_added;
2364		}
2365		spin_unlock(&cluster->lock);
2366		if (!bytes) {
2367			ret = 1;
2368			goto out;
2369		}
2370	}
2371
2372no_cluster_bitmap:
2373	bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2374					 1, 0);
2375	if (!bitmap_info) {
2376		ASSERT(added == 0);
2377		goto new_bitmap;
2378	}
2379
2380	bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2381					  trim_state);
2382	bytes -= bytes_added;
2383	offset += bytes_added;
2384	added = 0;
2385
2386	if (!bytes) {
2387		ret = 1;
2388		goto out;
2389	} else
2390		goto again;
2391
2392new_bitmap:
2393	if (info && info->bitmap) {
2394		add_new_bitmap(ctl, info, offset);
2395		added = 1;
2396		info = NULL;
2397		goto again;
2398	} else {
2399		spin_unlock(&ctl->tree_lock);
2400
2401		/* no pre-allocated info, allocate a new one */
2402		if (!info) {
2403			info = kmem_cache_zalloc(btrfs_free_space_cachep,
2404						 GFP_NOFS);
2405			if (!info) {
2406				spin_lock(&ctl->tree_lock);
2407				ret = -ENOMEM;
2408				goto out;
2409			}
2410		}
2411
2412		/* allocate the bitmap */
2413		info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2414						 GFP_NOFS);
2415		info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2416		spin_lock(&ctl->tree_lock);
2417		if (!info->bitmap) {
2418			ret = -ENOMEM;
2419			goto out;
2420		}
2421		goto again;
2422	}
2423
2424out:
2425	if (info) {
2426		if (info->bitmap)
2427			kmem_cache_free(btrfs_free_space_bitmap_cachep,
2428					info->bitmap);
2429		kmem_cache_free(btrfs_free_space_cachep, info);
2430	}
2431
2432	return ret;
2433}
2434
2435/*
2436 * Free space merging rules:
2437 *  1) Merge trimmed areas together
2438 *  2) Let untrimmed areas coalesce with trimmed areas
2439 *  3) Always pull neighboring regions from bitmaps
2440 *
2441 * The above rules are for when we merge free space based on btrfs_trim_state.
2442 * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2443 * same reason: to promote larger extent regions which makes life easier for
2444 * find_free_extent().  Rule 2 enables coalescing based on the common path
2445 * being returning free space from btrfs_finish_extent_commit().  So when free
2446 * space is trimmed, it will prevent aggregating trimmed new region and
2447 * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2448 * and provide find_free_extent() with the largest extents possible hoping for
2449 * the reuse path.
2450 */
2451static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2452			  struct btrfs_free_space *info, bool update_stat)
2453{
2454	struct btrfs_free_space *left_info = NULL;
2455	struct btrfs_free_space *right_info;
2456	bool merged = false;
2457	u64 offset = info->offset;
2458	u64 bytes = info->bytes;
2459	const bool is_trimmed = btrfs_free_space_trimmed(info);
2460	struct rb_node *right_prev = NULL;
2461
2462	/*
2463	 * first we want to see if there is free space adjacent to the range we
2464	 * are adding, if there is remove that struct and add a new one to
2465	 * cover the entire range
2466	 */
2467	right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2468	if (right_info)
2469		right_prev = rb_prev(&right_info->offset_index);
2470
2471	if (right_prev)
2472		left_info = rb_entry(right_prev, struct btrfs_free_space, offset_index);
2473	else if (!right_info)
2474		left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2475
2476	/* See try_merge_free_space() comment. */
2477	if (right_info && !right_info->bitmap &&
2478	    (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2479		unlink_free_space(ctl, right_info, update_stat);
2480		info->bytes += right_info->bytes;
2481		kmem_cache_free(btrfs_free_space_cachep, right_info);
2482		merged = true;
2483	}
2484
2485	/* See try_merge_free_space() comment. */
2486	if (left_info && !left_info->bitmap &&
2487	    left_info->offset + left_info->bytes == offset &&
2488	    (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2489		unlink_free_space(ctl, left_info, update_stat);
2490		info->offset = left_info->offset;
2491		info->bytes += left_info->bytes;
2492		kmem_cache_free(btrfs_free_space_cachep, left_info);
2493		merged = true;
2494	}
2495
2496	return merged;
2497}
2498
2499static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2500				     struct btrfs_free_space *info,
2501				     bool update_stat)
2502{
2503	struct btrfs_free_space *bitmap;
2504	unsigned long i;
2505	unsigned long j;
2506	const u64 end = info->offset + info->bytes;
2507	const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2508	u64 bytes;
2509
2510	bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2511	if (!bitmap)
2512		return false;
2513
2514	i = offset_to_bit(bitmap->offset, ctl->unit, end);
2515	j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2516	if (j == i)
2517		return false;
2518	bytes = (j - i) * ctl->unit;
2519	info->bytes += bytes;
2520
2521	/* See try_merge_free_space() comment. */
2522	if (!btrfs_free_space_trimmed(bitmap))
2523		info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2524
2525	bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2526
2527	if (!bitmap->bytes)
2528		free_bitmap(ctl, bitmap);
2529
2530	return true;
2531}
2532
2533static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2534				       struct btrfs_free_space *info,
2535				       bool update_stat)
2536{
2537	struct btrfs_free_space *bitmap;
2538	u64 bitmap_offset;
2539	unsigned long i;
2540	unsigned long j;
2541	unsigned long prev_j;
2542	u64 bytes;
2543
2544	bitmap_offset = offset_to_bitmap(ctl, info->offset);
2545	/* If we're on a boundary, try the previous logical bitmap. */
2546	if (bitmap_offset == info->offset) {
2547		if (info->offset == 0)
2548			return false;
2549		bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2550	}
2551
2552	bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2553	if (!bitmap)
2554		return false;
2555
2556	i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2557	j = 0;
2558	prev_j = (unsigned long)-1;
2559	for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2560		if (j > i)
2561			break;
2562		prev_j = j;
2563	}
2564	if (prev_j == i)
2565		return false;
2566
2567	if (prev_j == (unsigned long)-1)
2568		bytes = (i + 1) * ctl->unit;
2569	else
2570		bytes = (i - prev_j) * ctl->unit;
2571
2572	info->offset -= bytes;
2573	info->bytes += bytes;
2574
2575	/* See try_merge_free_space() comment. */
2576	if (!btrfs_free_space_trimmed(bitmap))
2577		info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2578
2579	bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2580
2581	if (!bitmap->bytes)
2582		free_bitmap(ctl, bitmap);
2583
2584	return true;
2585}
2586
2587/*
2588 * We prefer always to allocate from extent entries, both for clustered and
2589 * non-clustered allocation requests. So when attempting to add a new extent
2590 * entry, try to see if there's adjacent free space in bitmap entries, and if
2591 * there is, migrate that space from the bitmaps to the extent.
2592 * Like this we get better chances of satisfying space allocation requests
2593 * because we attempt to satisfy them based on a single cache entry, and never
2594 * on 2 or more entries - even if the entries represent a contiguous free space
2595 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2596 * ends).
2597 */
2598static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2599			      struct btrfs_free_space *info,
2600			      bool update_stat)
2601{
2602	/*
2603	 * Only work with disconnected entries, as we can change their offset,
2604	 * and must be extent entries.
2605	 */
2606	ASSERT(!info->bitmap);
2607	ASSERT(RB_EMPTY_NODE(&info->offset_index));
2608
2609	if (ctl->total_bitmaps > 0) {
2610		bool stole_end;
2611		bool stole_front = false;
2612
2613		stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2614		if (ctl->total_bitmaps > 0)
2615			stole_front = steal_from_bitmap_to_front(ctl, info,
2616								 update_stat);
2617
2618		if (stole_end || stole_front)
2619			try_merge_free_space(ctl, info, update_stat);
2620	}
2621}
2622
2623static int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2624			   u64 offset, u64 bytes,
2625			   enum btrfs_trim_state trim_state)
2626{
2627	struct btrfs_fs_info *fs_info = block_group->fs_info;
2628	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2629	struct btrfs_free_space *info;
2630	int ret = 0;
2631	u64 filter_bytes = bytes;
2632
2633	ASSERT(!btrfs_is_zoned(fs_info));
2634
2635	info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2636	if (!info)
2637		return -ENOMEM;
2638
2639	info->offset = offset;
2640	info->bytes = bytes;
2641	info->trim_state = trim_state;
2642	RB_CLEAR_NODE(&info->offset_index);
2643	RB_CLEAR_NODE(&info->bytes_index);
2644
2645	spin_lock(&ctl->tree_lock);
2646
2647	if (try_merge_free_space(ctl, info, true))
2648		goto link;
2649
2650	/*
2651	 * There was no extent directly to the left or right of this new
2652	 * extent then we know we're going to have to allocate a new extent, so
2653	 * before we do that see if we need to drop this into a bitmap
2654	 */
2655	ret = insert_into_bitmap(ctl, info);
2656	if (ret < 0) {
2657		goto out;
2658	} else if (ret) {
2659		ret = 0;
2660		goto out;
2661	}
2662link:
2663	/*
2664	 * Only steal free space from adjacent bitmaps if we're sure we're not
2665	 * going to add the new free space to existing bitmap entries - because
2666	 * that would mean unnecessary work that would be reverted. Therefore
2667	 * attempt to steal space from bitmaps if we're adding an extent entry.
2668	 */
2669	steal_from_bitmap(ctl, info, true);
2670
2671	filter_bytes = max(filter_bytes, info->bytes);
2672
2673	ret = link_free_space(ctl, info);
2674	if (ret)
2675		kmem_cache_free(btrfs_free_space_cachep, info);
2676out:
2677	btrfs_discard_update_discardable(block_group);
2678	spin_unlock(&ctl->tree_lock);
2679
2680	if (ret) {
2681		btrfs_crit(fs_info, "unable to add free space :%d", ret);
2682		ASSERT(ret != -EEXIST);
2683	}
2684
2685	if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2686		btrfs_discard_check_filter(block_group, filter_bytes);
2687		btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2688	}
2689
2690	return ret;
2691}
2692
2693static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2694					u64 bytenr, u64 size, bool used)
2695{
2696	struct btrfs_space_info *sinfo = block_group->space_info;
2697	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2698	u64 offset = bytenr - block_group->start;
2699	u64 to_free, to_unusable;
2700	int bg_reclaim_threshold = 0;
2701	bool initial;
2702	u64 reclaimable_unusable;
2703
2704	spin_lock(&block_group->lock);
2705
2706	initial = ((size == block_group->length) && (block_group->alloc_offset == 0));
2707	WARN_ON(!initial && offset + size > block_group->zone_capacity);
2708	if (!initial)
2709		bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2710
2711	if (!used)
2712		to_free = size;
2713	else if (initial)
2714		to_free = block_group->zone_capacity;
2715	else if (offset >= block_group->alloc_offset)
2716		to_free = size;
2717	else if (offset + size <= block_group->alloc_offset)
2718		to_free = 0;
2719	else
2720		to_free = offset + size - block_group->alloc_offset;
2721	to_unusable = size - to_free;
2722
2723	spin_lock(&ctl->tree_lock);
2724	ctl->free_space += to_free;
2725	spin_unlock(&ctl->tree_lock);
2726	/*
2727	 * If the block group is read-only, we should account freed space into
2728	 * bytes_readonly.
2729	 */
2730	if (!block_group->ro) {
2731		block_group->zone_unusable += to_unusable;
2732		WARN_ON(block_group->zone_unusable > block_group->length);
2733	}
2734	if (!used) {
2735		block_group->alloc_offset -= size;
2736	}
2737
2738	reclaimable_unusable = block_group->zone_unusable -
2739			       (block_group->length - block_group->zone_capacity);
2740	/* All the region is now unusable. Mark it as unused and reclaim */
2741	if (block_group->zone_unusable == block_group->length) {
2742		btrfs_mark_bg_unused(block_group);
2743	} else if (bg_reclaim_threshold &&
2744		   reclaimable_unusable >=
2745		   mult_perc(block_group->zone_capacity, bg_reclaim_threshold)) {
2746		btrfs_mark_bg_to_reclaim(block_group);
2747	}
2748
2749	spin_unlock(&block_group->lock);
2750
2751	return 0;
2752}
2753
2754int btrfs_add_free_space(struct btrfs_block_group *block_group,
2755			 u64 bytenr, u64 size)
2756{
2757	enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2758
2759	if (btrfs_is_zoned(block_group->fs_info))
2760		return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2761						    true);
2762
2763	if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2764		trim_state = BTRFS_TRIM_STATE_TRIMMED;
2765
2766	return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2767}
2768
2769int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2770				u64 bytenr, u64 size)
2771{
2772	if (btrfs_is_zoned(block_group->fs_info))
2773		return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2774						    false);
2775
2776	return btrfs_add_free_space(block_group, bytenr, size);
2777}
2778
2779/*
2780 * This is a subtle distinction because when adding free space back in general,
2781 * we want it to be added as untrimmed for async. But in the case where we add
2782 * it on loading of a block group, we want to consider it trimmed.
2783 */
2784int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2785				       u64 bytenr, u64 size)
2786{
2787	enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2788
2789	if (btrfs_is_zoned(block_group->fs_info))
2790		return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2791						    true);
2792
2793	if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2794	    btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2795		trim_state = BTRFS_TRIM_STATE_TRIMMED;
2796
2797	return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2798}
2799
2800int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2801			    u64 offset, u64 bytes)
2802{
2803	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2804	struct btrfs_free_space *info;
2805	int ret;
2806	bool re_search = false;
2807
2808	if (btrfs_is_zoned(block_group->fs_info)) {
2809		/*
2810		 * This can happen with conventional zones when replaying log.
2811		 * Since the allocation info of tree-log nodes are not recorded
2812		 * to the extent-tree, calculate_alloc_pointer() failed to
2813		 * advance the allocation pointer after last allocated tree log
2814		 * node blocks.
2815		 *
2816		 * This function is called from
2817		 * btrfs_pin_extent_for_log_replay() when replaying the log.
2818		 * Advance the pointer not to overwrite the tree-log nodes.
2819		 */
2820		if (block_group->start + block_group->alloc_offset <
2821		    offset + bytes) {
2822			block_group->alloc_offset =
2823				offset + bytes - block_group->start;
2824		}
2825		return 0;
2826	}
2827
2828	spin_lock(&ctl->tree_lock);
2829
2830again:
2831	ret = 0;
2832	if (!bytes)
2833		goto out_lock;
2834
2835	info = tree_search_offset(ctl, offset, 0, 0);
2836	if (!info) {
2837		/*
2838		 * oops didn't find an extent that matched the space we wanted
2839		 * to remove, look for a bitmap instead
2840		 */
2841		info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2842					  1, 0);
2843		if (!info) {
2844			/*
2845			 * If we found a partial bit of our free space in a
2846			 * bitmap but then couldn't find the other part this may
2847			 * be a problem, so WARN about it.
2848			 */
2849			WARN_ON(re_search);
2850			goto out_lock;
2851		}
2852	}
2853
2854	re_search = false;
2855	if (!info->bitmap) {
2856		unlink_free_space(ctl, info, true);
2857		if (offset == info->offset) {
2858			u64 to_free = min(bytes, info->bytes);
2859
2860			info->bytes -= to_free;
2861			info->offset += to_free;
2862			if (info->bytes) {
2863				ret = link_free_space(ctl, info);
2864				WARN_ON(ret);
2865			} else {
2866				kmem_cache_free(btrfs_free_space_cachep, info);
2867			}
2868
2869			offset += to_free;
2870			bytes -= to_free;
2871			goto again;
2872		} else {
2873			u64 old_end = info->bytes + info->offset;
2874
2875			info->bytes = offset - info->offset;
2876			ret = link_free_space(ctl, info);
2877			WARN_ON(ret);
2878			if (ret)
2879				goto out_lock;
2880
2881			/* Not enough bytes in this entry to satisfy us */
2882			if (old_end < offset + bytes) {
2883				bytes -= old_end - offset;
2884				offset = old_end;
2885				goto again;
2886			} else if (old_end == offset + bytes) {
2887				/* all done */
2888				goto out_lock;
2889			}
2890			spin_unlock(&ctl->tree_lock);
2891
2892			ret = __btrfs_add_free_space(block_group,
2893						     offset + bytes,
2894						     old_end - (offset + bytes),
2895						     info->trim_state);
2896			WARN_ON(ret);
2897			goto out;
2898		}
2899	}
2900
2901	ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2902	if (ret == -EAGAIN) {
2903		re_search = true;
2904		goto again;
2905	}
2906out_lock:
2907	btrfs_discard_update_discardable(block_group);
2908	spin_unlock(&ctl->tree_lock);
2909out:
2910	return ret;
2911}
2912
2913void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2914			   u64 bytes)
2915{
2916	struct btrfs_fs_info *fs_info = block_group->fs_info;
2917	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2918	struct btrfs_free_space *info;
2919	struct rb_node *n;
2920	int count = 0;
2921
2922	/*
2923	 * Zoned btrfs does not use free space tree and cluster. Just print
2924	 * out the free space after the allocation offset.
2925	 */
2926	if (btrfs_is_zoned(fs_info)) {
2927		btrfs_info(fs_info, "free space %llu active %d",
2928			   block_group->zone_capacity - block_group->alloc_offset,
2929			   test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2930				    &block_group->runtime_flags));
2931		return;
2932	}
2933
2934	spin_lock(&ctl->tree_lock);
2935	for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2936		info = rb_entry(n, struct btrfs_free_space, offset_index);
2937		if (info->bytes >= bytes && !block_group->ro)
2938			count++;
2939		btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2940			   info->offset, info->bytes, str_yes_no(info->bitmap));
2941	}
2942	spin_unlock(&ctl->tree_lock);
2943	btrfs_info(fs_info, "block group has cluster?: %s",
2944	       str_no_yes(list_empty(&block_group->cluster_list)));
2945	btrfs_info(fs_info,
2946		   "%d free space entries at or bigger than %llu bytes",
2947		   count, bytes);
2948}
2949
2950void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2951			       struct btrfs_free_space_ctl *ctl)
2952{
2953	struct btrfs_fs_info *fs_info = block_group->fs_info;
2954
2955	spin_lock_init(&ctl->tree_lock);
2956	ctl->unit = fs_info->sectorsize;
2957	ctl->start = block_group->start;
2958	ctl->block_group = block_group;
2959	ctl->op = &free_space_op;
2960	ctl->free_space_bytes = RB_ROOT_CACHED;
2961	INIT_LIST_HEAD(&ctl->trimming_ranges);
2962	mutex_init(&ctl->cache_writeout_mutex);
2963
2964	/*
2965	 * we only want to have 32k of ram per block group for keeping
2966	 * track of free space, and if we pass 1/2 of that we want to
2967	 * start converting things over to using bitmaps
2968	 */
2969	ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2970}
2971
2972/*
2973 * for a given cluster, put all of its extents back into the free
2974 * space cache.  If the block group passed doesn't match the block group
2975 * pointed to by the cluster, someone else raced in and freed the
2976 * cluster already.  In that case, we just return without changing anything
2977 */
2978static void __btrfs_return_cluster_to_free_space(
2979			     struct btrfs_block_group *block_group,
2980			     struct btrfs_free_cluster *cluster)
2981{
2982	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2983	struct rb_node *node;
2984
2985	lockdep_assert_held(&ctl->tree_lock);
2986
2987	spin_lock(&cluster->lock);
2988	if (cluster->block_group != block_group) {
2989		spin_unlock(&cluster->lock);
2990		return;
2991	}
2992
2993	cluster->block_group = NULL;
2994	cluster->window_start = 0;
2995	list_del_init(&cluster->block_group_list);
2996
2997	node = rb_first(&cluster->root);
2998	while (node) {
2999		struct btrfs_free_space *entry;
3000
3001		entry = rb_entry(node, struct btrfs_free_space, offset_index);
3002		node = rb_next(&entry->offset_index);
3003		rb_erase(&entry->offset_index, &cluster->root);
3004		RB_CLEAR_NODE(&entry->offset_index);
3005
3006		if (!entry->bitmap) {
3007			/* Merging treats extents as if they were new */
3008			if (!btrfs_free_space_trimmed(entry)) {
3009				ctl->discardable_extents[BTRFS_STAT_CURR]--;
3010				ctl->discardable_bytes[BTRFS_STAT_CURR] -=
3011					entry->bytes;
3012			}
3013
3014			try_merge_free_space(ctl, entry, false);
3015			steal_from_bitmap(ctl, entry, false);
3016
3017			/* As we insert directly, update these statistics */
3018			if (!btrfs_free_space_trimmed(entry)) {
3019				ctl->discardable_extents[BTRFS_STAT_CURR]++;
3020				ctl->discardable_bytes[BTRFS_STAT_CURR] +=
3021					entry->bytes;
3022			}
3023		}
3024		tree_insert_offset(ctl, NULL, entry);
3025		rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
3026			      entry_less);
3027	}
3028	cluster->root = RB_ROOT;
3029	spin_unlock(&cluster->lock);
3030	btrfs_put_block_group(block_group);
3031}
3032
3033void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
3034{
3035	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3036	struct btrfs_free_cluster *cluster;
3037	struct list_head *head;
3038
3039	spin_lock(&ctl->tree_lock);
3040	while ((head = block_group->cluster_list.next) !=
3041	       &block_group->cluster_list) {
3042		cluster = list_entry(head, struct btrfs_free_cluster,
3043				     block_group_list);
3044
3045		WARN_ON(cluster->block_group != block_group);
3046		__btrfs_return_cluster_to_free_space(block_group, cluster);
3047
3048		cond_resched_lock(&ctl->tree_lock);
3049	}
3050	__btrfs_remove_free_space_cache(ctl);
3051	btrfs_discard_update_discardable(block_group);
3052	spin_unlock(&ctl->tree_lock);
3053
3054}
3055
3056/*
3057 * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3058 */
3059bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3060{
3061	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3062	struct btrfs_free_space *info;
3063	struct rb_node *node;
3064	bool ret = true;
3065
3066	spin_lock(&ctl->tree_lock);
3067	node = rb_first(&ctl->free_space_offset);
3068
3069	while (node) {
3070		info = rb_entry(node, struct btrfs_free_space, offset_index);
3071
3072		if (!btrfs_free_space_trimmed(info)) {
3073			ret = false;
3074			break;
3075		}
3076
3077		node = rb_next(node);
3078	}
3079
3080	spin_unlock(&ctl->tree_lock);
3081	return ret;
3082}
3083
3084u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3085			       u64 offset, u64 bytes, u64 empty_size,
3086			       u64 *max_extent_size)
3087{
3088	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3089	struct btrfs_discard_ctl *discard_ctl =
3090					&block_group->fs_info->discard_ctl;
3091	struct btrfs_free_space *entry = NULL;
3092	u64 bytes_search = bytes + empty_size;
3093	u64 ret = 0;
3094	u64 align_gap = 0;
3095	u64 align_gap_len = 0;
3096	enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3097	bool use_bytes_index = (offset == block_group->start);
3098
3099	ASSERT(!btrfs_is_zoned(block_group->fs_info));
3100
3101	spin_lock(&ctl->tree_lock);
3102	entry = find_free_space(ctl, &offset, &bytes_search,
3103				block_group->full_stripe_len, max_extent_size,
3104				use_bytes_index);
3105	if (!entry)
3106		goto out;
3107
3108	ret = offset;
3109	if (entry->bitmap) {
3110		bitmap_clear_bits(ctl, entry, offset, bytes, true);
3111
3112		if (!btrfs_free_space_trimmed(entry))
3113			atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3114
3115		if (!entry->bytes)
3116			free_bitmap(ctl, entry);
3117	} else {
3118		unlink_free_space(ctl, entry, true);
3119		align_gap_len = offset - entry->offset;
3120		align_gap = entry->offset;
3121		align_gap_trim_state = entry->trim_state;
3122
3123		if (!btrfs_free_space_trimmed(entry))
3124			atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3125
3126		entry->offset = offset + bytes;
3127		WARN_ON(entry->bytes < bytes + align_gap_len);
3128
3129		entry->bytes -= bytes + align_gap_len;
3130		if (!entry->bytes)
3131			kmem_cache_free(btrfs_free_space_cachep, entry);
3132		else
3133			link_free_space(ctl, entry);
3134	}
3135out:
3136	btrfs_discard_update_discardable(block_group);
3137	spin_unlock(&ctl->tree_lock);
3138
3139	if (align_gap_len)
3140		__btrfs_add_free_space(block_group, align_gap, align_gap_len,
3141				       align_gap_trim_state);
3142	return ret;
3143}
3144
3145/*
3146 * given a cluster, put all of its extents back into the free space
3147 * cache.  If a block group is passed, this function will only free
3148 * a cluster that belongs to the passed block group.
3149 *
3150 * Otherwise, it'll get a reference on the block group pointed to by the
3151 * cluster and remove the cluster from it.
3152 */
3153void btrfs_return_cluster_to_free_space(
3154			       struct btrfs_block_group *block_group,
3155			       struct btrfs_free_cluster *cluster)
3156{
3157	struct btrfs_free_space_ctl *ctl;
3158
3159	/* first, get a safe pointer to the block group */
3160	spin_lock(&cluster->lock);
3161	if (!block_group) {
3162		block_group = cluster->block_group;
3163		if (!block_group) {
3164			spin_unlock(&cluster->lock);
3165			return;
3166		}
3167	} else if (cluster->block_group != block_group) {
3168		/* someone else has already freed it don't redo their work */
3169		spin_unlock(&cluster->lock);
3170		return;
3171	}
3172	btrfs_get_block_group(block_group);
3173	spin_unlock(&cluster->lock);
3174
3175	ctl = block_group->free_space_ctl;
3176
3177	/* now return any extents the cluster had on it */
3178	spin_lock(&ctl->tree_lock);
3179	__btrfs_return_cluster_to_free_space(block_group, cluster);
3180	spin_unlock(&ctl->tree_lock);
3181
3182	btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3183
3184	/* finally drop our ref */
3185	btrfs_put_block_group(block_group);
3186}
3187
3188static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3189				   struct btrfs_free_cluster *cluster,
3190				   struct btrfs_free_space *entry,
3191				   u64 bytes, u64 min_start,
3192				   u64 *max_extent_size)
3193{
3194	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3195	int ret2;
3196	u64 search_start = cluster->window_start;
3197	u64 search_bytes = bytes;
3198	u64 ret = 0;
3199
3200	search_start = min_start;
3201	search_bytes = bytes;
3202
3203	ret2 = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3204	if (ret2) {
3205		*max_extent_size = max(get_max_extent_size(entry),
3206				       *max_extent_size);
3207		return 0;
3208	}
3209
3210	ret = search_start;
3211	bitmap_clear_bits(ctl, entry, ret, bytes, false);
3212
3213	return ret;
3214}
3215
3216/*
3217 * given a cluster, try to allocate 'bytes' from it, returns 0
3218 * if it couldn't find anything suitably large, or a logical disk offset
3219 * if things worked out
3220 */
3221u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3222			     struct btrfs_free_cluster *cluster, u64 bytes,
3223			     u64 min_start, u64 *max_extent_size)
3224{
3225	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3226	struct btrfs_discard_ctl *discard_ctl =
3227					&block_group->fs_info->discard_ctl;
3228	struct btrfs_free_space *entry = NULL;
3229	struct rb_node *node;
3230	u64 ret = 0;
3231
3232	ASSERT(!btrfs_is_zoned(block_group->fs_info));
3233
3234	spin_lock(&cluster->lock);
3235	if (bytes > cluster->max_size)
3236		goto out;
3237
3238	if (cluster->block_group != block_group)
3239		goto out;
3240
3241	node = rb_first(&cluster->root);
3242	if (!node)
3243		goto out;
3244
3245	entry = rb_entry(node, struct btrfs_free_space, offset_index);
3246	while (1) {
3247		if (entry->bytes < bytes)
3248			*max_extent_size = max(get_max_extent_size(entry),
3249					       *max_extent_size);
3250
3251		if (entry->bytes < bytes ||
3252		    (!entry->bitmap && entry->offset < min_start)) {
3253			node = rb_next(&entry->offset_index);
3254			if (!node)
3255				break;
3256			entry = rb_entry(node, struct btrfs_free_space,
3257					 offset_index);
3258			continue;
3259		}
3260
3261		if (entry->bitmap) {
3262			ret = btrfs_alloc_from_bitmap(block_group,
3263						      cluster, entry, bytes,
3264						      cluster->window_start,
3265						      max_extent_size);
3266			if (ret == 0) {
3267				node = rb_next(&entry->offset_index);
3268				if (!node)
3269					break;
3270				entry = rb_entry(node, struct btrfs_free_space,
3271						 offset_index);
3272				continue;
3273			}
3274			cluster->window_start += bytes;
3275		} else {
3276			ret = entry->offset;
3277
3278			entry->offset += bytes;
3279			entry->bytes -= bytes;
3280		}
3281
3282		break;
3283	}
3284out:
3285	spin_unlock(&cluster->lock);
3286
3287	if (!ret)
3288		return 0;
3289
3290	spin_lock(&ctl->tree_lock);
3291
3292	if (!btrfs_free_space_trimmed(entry))
3293		atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3294
3295	ctl->free_space -= bytes;
3296	if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3297		ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3298
3299	spin_lock(&cluster->lock);
3300	if (entry->bytes == 0) {
3301		rb_erase(&entry->offset_index, &cluster->root);
3302		ctl->free_extents--;
3303		if (entry->bitmap) {
3304			kmem_cache_free(btrfs_free_space_bitmap_cachep,
3305					entry->bitmap);
3306			ctl->total_bitmaps--;
3307			recalculate_thresholds(ctl);
3308		} else if (!btrfs_free_space_trimmed(entry)) {
3309			ctl->discardable_extents[BTRFS_STAT_CURR]--;
3310		}
3311		kmem_cache_free(btrfs_free_space_cachep, entry);
3312	}
3313
3314	spin_unlock(&cluster->lock);
3315	spin_unlock(&ctl->tree_lock);
3316
3317	return ret;
3318}
3319
3320static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3321				struct btrfs_free_space *entry,
3322				struct btrfs_free_cluster *cluster,
3323				u64 offset, u64 bytes,
3324				u64 cont1_bytes, u64 min_bytes)
3325{
3326	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3327	unsigned long next_zero;
3328	unsigned long i;
3329	unsigned long want_bits;
3330	unsigned long min_bits;
3331	unsigned long found_bits;
3332	unsigned long max_bits = 0;
3333	unsigned long start = 0;
3334	unsigned long total_found = 0;
3335	int ret;
3336
3337	lockdep_assert_held(&ctl->tree_lock);
3338
3339	i = offset_to_bit(entry->offset, ctl->unit,
3340			  max_t(u64, offset, entry->offset));
3341	want_bits = bytes_to_bits(bytes, ctl->unit);
3342	min_bits = bytes_to_bits(min_bytes, ctl->unit);
3343
3344	/*
3345	 * Don't bother looking for a cluster in this bitmap if it's heavily
3346	 * fragmented.
3347	 */
3348	if (entry->max_extent_size &&
3349	    entry->max_extent_size < cont1_bytes)
3350		return -ENOSPC;
3351again:
3352	found_bits = 0;
3353	for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3354		next_zero = find_next_zero_bit(entry->bitmap,
3355					       BITS_PER_BITMAP, i);
3356		if (next_zero - i >= min_bits) {
3357			found_bits = next_zero - i;
3358			if (found_bits > max_bits)
3359				max_bits = found_bits;
3360			break;
3361		}
3362		if (next_zero - i > max_bits)
3363			max_bits = next_zero - i;
3364		i = next_zero;
3365	}
3366
3367	if (!found_bits) {
3368		entry->max_extent_size = (u64)max_bits * ctl->unit;
3369		return -ENOSPC;
3370	}
3371
3372	if (!total_found) {
3373		start = i;
3374		cluster->max_size = 0;
3375	}
3376
3377	total_found += found_bits;
3378
3379	if (cluster->max_size < found_bits * ctl->unit)
3380		cluster->max_size = found_bits * ctl->unit;
3381
3382	if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3383		i = next_zero + 1;
3384		goto again;
3385	}
3386
3387	cluster->window_start = start * ctl->unit + entry->offset;
3388	rb_erase(&entry->offset_index, &ctl->free_space_offset);
3389	rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3390
3391	/*
3392	 * We need to know if we're currently on the normal space index when we
3393	 * manipulate the bitmap so that we know we need to remove and re-insert
3394	 * it into the space_index tree.  Clear the bytes_index node here so the
3395	 * bitmap manipulation helpers know not to mess with the space_index
3396	 * until this bitmap entry is added back into the normal cache.
3397	 */
3398	RB_CLEAR_NODE(&entry->bytes_index);
3399
3400	ret = tree_insert_offset(ctl, cluster, entry);
3401	ASSERT(!ret); /* -EEXIST; Logic error */
3402
3403	trace_btrfs_setup_cluster(block_group, cluster,
3404				  total_found * ctl->unit, 1);
3405	return 0;
3406}
3407
3408/*
3409 * This searches the block group for just extents to fill the cluster with.
3410 * Try to find a cluster with at least bytes total bytes, at least one
3411 * extent of cont1_bytes, and other clusters of at least min_bytes.
3412 */
3413static noinline int
3414setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3415			struct btrfs_free_cluster *cluster,
3416			struct list_head *bitmaps, u64 offset, u64 bytes,
3417			u64 cont1_bytes, u64 min_bytes)
3418{
3419	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3420	struct btrfs_free_space *first = NULL;
3421	struct btrfs_free_space *entry = NULL;
3422	struct btrfs_free_space *last;
3423	struct rb_node *node;
3424	u64 window_free;
3425	u64 max_extent;
3426	u64 total_size = 0;
3427
3428	lockdep_assert_held(&ctl->tree_lock);
3429
3430	entry = tree_search_offset(ctl, offset, 0, 1);
3431	if (!entry)
3432		return -ENOSPC;
3433
3434	/*
3435	 * We don't want bitmaps, so just move along until we find a normal
3436	 * extent entry.
3437	 */
3438	while (entry->bitmap || entry->bytes < min_bytes) {
3439		if (entry->bitmap && list_empty(&entry->list))
3440			list_add_tail(&entry->list, bitmaps);
3441		node = rb_next(&entry->offset_index);
3442		if (!node)
3443			return -ENOSPC;
3444		entry = rb_entry(node, struct btrfs_free_space, offset_index);
3445	}
3446
3447	window_free = entry->bytes;
3448	max_extent = entry->bytes;
3449	first = entry;
3450	last = entry;
3451
3452	for (node = rb_next(&entry->offset_index); node;
3453	     node = rb_next(&entry->offset_index)) {
3454		entry = rb_entry(node, struct btrfs_free_space, offset_index);
3455
3456		if (entry->bitmap) {
3457			if (list_empty(&entry->list))
3458				list_add_tail(&entry->list, bitmaps);
3459			continue;
3460		}
3461
3462		if (entry->bytes < min_bytes)
3463			continue;
3464
3465		last = entry;
3466		window_free += entry->bytes;
3467		if (entry->bytes > max_extent)
3468			max_extent = entry->bytes;
3469	}
3470
3471	if (window_free < bytes || max_extent < cont1_bytes)
3472		return -ENOSPC;
3473
3474	cluster->window_start = first->offset;
3475
3476	node = &first->offset_index;
3477
3478	/*
3479	 * now we've found our entries, pull them out of the free space
3480	 * cache and put them into the cluster rbtree
3481	 */
3482	do {
3483		int ret;
3484
3485		entry = rb_entry(node, struct btrfs_free_space, offset_index);
3486		node = rb_next(&entry->offset_index);
3487		if (entry->bitmap || entry->bytes < min_bytes)
3488			continue;
3489
3490		rb_erase(&entry->offset_index, &ctl->free_space_offset);
3491		rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3492		ret = tree_insert_offset(ctl, cluster, entry);
3493		total_size += entry->bytes;
3494		ASSERT(!ret); /* -EEXIST; Logic error */
3495	} while (node && entry != last);
3496
3497	cluster->max_size = max_extent;
3498	trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3499	return 0;
3500}
3501
3502/*
3503 * This specifically looks for bitmaps that may work in the cluster, we assume
3504 * that we have already failed to find extents that will work.
3505 */
3506static noinline int
3507setup_cluster_bitmap(struct btrfs_block_group *block_group,
3508		     struct btrfs_free_cluster *cluster,
3509		     struct list_head *bitmaps, u64 offset, u64 bytes,
3510		     u64 cont1_bytes, u64 min_bytes)
3511{
3512	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3513	struct btrfs_free_space *entry = NULL;
3514	int ret = -ENOSPC;
3515	u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3516
3517	if (ctl->total_bitmaps == 0)
3518		return -ENOSPC;
3519
3520	/*
3521	 * The bitmap that covers offset won't be in the list unless offset
3522	 * is just its start offset.
3523	 */
3524	if (!list_empty(bitmaps))
3525		entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3526
3527	if (!entry || entry->offset != bitmap_offset) {
3528		entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3529		if (entry && list_empty(&entry->list))
3530			list_add(&entry->list, bitmaps);
3531	}
3532
3533	list_for_each_entry(entry, bitmaps, list) {
3534		if (entry->bytes < bytes)
3535			continue;
3536		ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3537					   bytes, cont1_bytes, min_bytes);
3538		if (!ret)
3539			return 0;
3540	}
3541
3542	/*
3543	 * The bitmaps list has all the bitmaps that record free space
3544	 * starting after offset, so no more search is required.
3545	 */
3546	return -ENOSPC;
3547}
3548
3549/*
3550 * here we try to find a cluster of blocks in a block group.  The goal
3551 * is to find at least bytes+empty_size.
3552 * We might not find them all in one contiguous area.
3553 *
3554 * returns zero and sets up cluster if things worked out, otherwise
3555 * it returns -enospc
3556 */
3557int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3558			     struct btrfs_free_cluster *cluster,
3559			     u64 offset, u64 bytes, u64 empty_size)
3560{
3561	struct btrfs_fs_info *fs_info = block_group->fs_info;
3562	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3563	struct btrfs_free_space *entry, *tmp;
3564	LIST_HEAD(bitmaps);
3565	u64 min_bytes;
3566	u64 cont1_bytes;
3567	int ret;
3568
3569	/*
3570	 * Choose the minimum extent size we'll require for this
3571	 * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3572	 * For metadata, allow allocates with smaller extents.  For
3573	 * data, keep it dense.
3574	 */
3575	if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3576		cont1_bytes = bytes + empty_size;
3577		min_bytes = cont1_bytes;
3578	} else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3579		cont1_bytes = bytes;
3580		min_bytes = fs_info->sectorsize;
3581	} else {
3582		cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3583		min_bytes = fs_info->sectorsize;
3584	}
3585
3586	spin_lock(&ctl->tree_lock);
3587
3588	/*
3589	 * If we know we don't have enough space to make a cluster don't even
3590	 * bother doing all the work to try and find one.
3591	 */
3592	if (ctl->free_space < bytes) {
3593		spin_unlock(&ctl->tree_lock);
3594		return -ENOSPC;
3595	}
3596
3597	spin_lock(&cluster->lock);
3598
3599	/* someone already found a cluster, hooray */
3600	if (cluster->block_group) {
3601		ret = 0;
3602		goto out;
3603	}
3604
3605	trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3606				 min_bytes);
3607
3608	ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3609				      bytes + empty_size,
3610				      cont1_bytes, min_bytes);
3611	if (ret)
3612		ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3613					   offset, bytes + empty_size,
3614					   cont1_bytes, min_bytes);
3615
3616	/* Clear our temporary list */
3617	list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3618		list_del_init(&entry->list);
3619
3620	if (!ret) {
3621		btrfs_get_block_group(block_group);
3622		list_add_tail(&cluster->block_group_list,
3623			      &block_group->cluster_list);
3624		cluster->block_group = block_group;
3625	} else {
3626		trace_btrfs_failed_cluster_setup(block_group);
3627	}
3628out:
3629	spin_unlock(&cluster->lock);
3630	spin_unlock(&ctl->tree_lock);
3631
3632	return ret;
3633}
3634
3635/*
3636 * simple code to zero out a cluster
3637 */
3638void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3639{
3640	spin_lock_init(&cluster->lock);
3641	spin_lock_init(&cluster->refill_lock);
3642	cluster->root = RB_ROOT;
3643	cluster->max_size = 0;
3644	cluster->fragmented = false;
3645	INIT_LIST_HEAD(&cluster->block_group_list);
3646	cluster->block_group = NULL;
3647}
3648
3649static int do_trimming(struct btrfs_block_group *block_group,
3650		       u64 *total_trimmed, u64 start, u64 bytes,
3651		       u64 reserved_start, u64 reserved_bytes,
3652		       enum btrfs_trim_state reserved_trim_state,
3653		       struct btrfs_trim_range *trim_entry)
3654{
3655	struct btrfs_space_info *space_info = block_group->space_info;
3656	struct btrfs_fs_info *fs_info = block_group->fs_info;
3657	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3658	int ret;
3659	bool bg_ro;
3660	const u64 end = start + bytes;
3661	const u64 reserved_end = reserved_start + reserved_bytes;
3662	enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3663	u64 trimmed = 0;
3664
3665	spin_lock(&space_info->lock);
3666	spin_lock(&block_group->lock);
3667	bg_ro = block_group->ro;
3668	if (!bg_ro) {
3669		block_group->reserved += reserved_bytes;
3670		spin_unlock(&block_group->lock);
3671		space_info->bytes_reserved += reserved_bytes;
3672	} else {
3673		spin_unlock(&block_group->lock);
3674	}
3675	spin_unlock(&space_info->lock);
3676
3677	ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3678	if (!ret) {
3679		*total_trimmed += trimmed;
3680		trim_state = BTRFS_TRIM_STATE_TRIMMED;
3681	}
3682
3683	mutex_lock(&ctl->cache_writeout_mutex);
3684	if (reserved_start < start)
3685		__btrfs_add_free_space(block_group, reserved_start,
3686				       start - reserved_start,
3687				       reserved_trim_state);
3688	if (end < reserved_end)
3689		__btrfs_add_free_space(block_group, end, reserved_end - end,
3690				       reserved_trim_state);
3691	__btrfs_add_free_space(block_group, start, bytes, trim_state);
3692	list_del(&trim_entry->list);
3693	mutex_unlock(&ctl->cache_writeout_mutex);
3694
3695	if (!bg_ro) {
3696		spin_lock(&space_info->lock);
3697		spin_lock(&block_group->lock);
3698		bg_ro = block_group->ro;
3699		block_group->reserved -= reserved_bytes;
3700		spin_unlock(&block_group->lock);
3701
3702		space_info->bytes_reserved -= reserved_bytes;
3703		if (bg_ro)
3704			space_info->bytes_readonly += reserved_bytes;
3705		spin_unlock(&space_info->lock);
3706	}
3707
3708	return ret;
3709}
3710
3711/*
3712 * If @async is set, then we will trim 1 region and return.
3713 */
3714static int trim_no_bitmap(struct btrfs_block_group *block_group,
3715			  u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3716			  bool async)
3717{
3718	struct btrfs_discard_ctl *discard_ctl =
3719					&block_group->fs_info->discard_ctl;
3720	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3721	struct btrfs_free_space *entry;
3722	struct rb_node *node;
3723	int ret = 0;
3724	u64 extent_start;
3725	u64 extent_bytes;
3726	enum btrfs_trim_state extent_trim_state;
3727	u64 bytes;
3728	const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3729
3730	while (start < end) {
3731		struct btrfs_trim_range trim_entry;
3732
3733		mutex_lock(&ctl->cache_writeout_mutex);
3734		spin_lock(&ctl->tree_lock);
3735
3736		if (ctl->free_space < minlen)
3737			goto out_unlock;
3738
3739		entry = tree_search_offset(ctl, start, 0, 1);
3740		if (!entry)
3741			goto out_unlock;
3742
3743		/* Skip bitmaps and if async, already trimmed entries */
3744		while (entry->bitmap ||
3745		       (async && btrfs_free_space_trimmed(entry))) {
3746			node = rb_next(&entry->offset_index);
3747			if (!node)
3748				goto out_unlock;
3749			entry = rb_entry(node, struct btrfs_free_space,
3750					 offset_index);
3751		}
3752
3753		if (entry->offset >= end)
3754			goto out_unlock;
3755
3756		extent_start = entry->offset;
3757		extent_bytes = entry->bytes;
3758		extent_trim_state = entry->trim_state;
3759		if (async) {
3760			start = entry->offset;
3761			bytes = entry->bytes;
3762			if (bytes < minlen) {
3763				spin_unlock(&ctl->tree_lock);
3764				mutex_unlock(&ctl->cache_writeout_mutex);
3765				goto next;
3766			}
3767			unlink_free_space(ctl, entry, true);
3768			/*
3769			 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3770			 * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3771			 * X when we come back around.  So trim it now.
3772			 */
3773			if (max_discard_size &&
3774			    bytes >= (max_discard_size +
3775				      BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3776				bytes = max_discard_size;
3777				extent_bytes = max_discard_size;
3778				entry->offset += max_discard_size;
3779				entry->bytes -= max_discard_size;
3780				link_free_space(ctl, entry);
3781			} else {
3782				kmem_cache_free(btrfs_free_space_cachep, entry);
3783			}
3784		} else {
3785			start = max(start, extent_start);
3786			bytes = min(extent_start + extent_bytes, end) - start;
3787			if (bytes < minlen) {
3788				spin_unlock(&ctl->tree_lock);
3789				mutex_unlock(&ctl->cache_writeout_mutex);
3790				goto next;
3791			}
3792
3793			unlink_free_space(ctl, entry, true);
3794			kmem_cache_free(btrfs_free_space_cachep, entry);
3795		}
3796
3797		spin_unlock(&ctl->tree_lock);
3798		trim_entry.start = extent_start;
3799		trim_entry.bytes = extent_bytes;
3800		list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3801		mutex_unlock(&ctl->cache_writeout_mutex);
3802
3803		ret = do_trimming(block_group, total_trimmed, start, bytes,
3804				  extent_start, extent_bytes, extent_trim_state,
3805				  &trim_entry);
3806		if (ret) {
3807			block_group->discard_cursor = start + bytes;
3808			break;
3809		}
3810next:
3811		start += bytes;
3812		block_group->discard_cursor = start;
3813		if (async && *total_trimmed)
3814			break;
3815
3816		if (btrfs_trim_interrupted()) {
3817			ret = -ERESTARTSYS;
3818			break;
3819		}
3820
3821		cond_resched();
3822	}
3823
3824	return ret;
3825
3826out_unlock:
3827	block_group->discard_cursor = btrfs_block_group_end(block_group);
3828	spin_unlock(&ctl->tree_lock);
3829	mutex_unlock(&ctl->cache_writeout_mutex);
3830
3831	return ret;
3832}
3833
3834/*
3835 * If we break out of trimming a bitmap prematurely, we should reset the
3836 * trimming bit.  In a rather contrived case, it's possible to race here so
3837 * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3838 *
3839 * start = start of bitmap
3840 * end = near end of bitmap
3841 *
3842 * Thread 1:			Thread 2:
3843 * trim_bitmaps(start)
3844 *				trim_bitmaps(end)
3845 *				end_trimming_bitmap()
3846 * reset_trimming_bitmap()
3847 */
3848static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3849{
3850	struct btrfs_free_space *entry;
3851
3852	spin_lock(&ctl->tree_lock);
3853	entry = tree_search_offset(ctl, offset, 1, 0);
3854	if (entry) {
3855		if (btrfs_free_space_trimmed(entry)) {
3856			ctl->discardable_extents[BTRFS_STAT_CURR] +=
3857				entry->bitmap_extents;
3858			ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3859		}
3860		entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3861	}
3862
3863	spin_unlock(&ctl->tree_lock);
3864}
3865
3866static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3867				struct btrfs_free_space *entry)
3868{
3869	if (btrfs_free_space_trimming_bitmap(entry)) {
3870		entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3871		ctl->discardable_extents[BTRFS_STAT_CURR] -=
3872			entry->bitmap_extents;
3873		ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3874	}
3875}
3876
3877/*
3878 * If @async is set, then we will trim 1 region and return.
3879 */
3880static int trim_bitmaps(struct btrfs_block_group *block_group,
3881			u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3882			u64 maxlen, bool async)
3883{
3884	struct btrfs_discard_ctl *discard_ctl =
3885					&block_group->fs_info->discard_ctl;
3886	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3887	struct btrfs_free_space *entry;
3888	int ret = 0;
3889	int ret2;
3890	u64 bytes;
3891	u64 offset = offset_to_bitmap(ctl, start);
3892	const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3893
3894	while (offset < end) {
3895		bool next_bitmap = false;
3896		struct btrfs_trim_range trim_entry;
3897
3898		mutex_lock(&ctl->cache_writeout_mutex);
3899		spin_lock(&ctl->tree_lock);
3900
3901		if (ctl->free_space < minlen) {
3902			block_group->discard_cursor =
3903				btrfs_block_group_end(block_group);
3904			spin_unlock(&ctl->tree_lock);
3905			mutex_unlock(&ctl->cache_writeout_mutex);
3906			break;
3907		}
3908
3909		entry = tree_search_offset(ctl, offset, 1, 0);
3910		/*
3911		 * Bitmaps are marked trimmed lossily now to prevent constant
3912		 * discarding of the same bitmap (the reason why we are bound
3913		 * by the filters).  So, retrim the block group bitmaps when we
3914		 * are preparing to punt to the unused_bgs list.  This uses
3915		 * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3916		 * which is the only discard index which sets minlen to 0.
3917		 */
3918		if (!entry || (async && minlen && start == offset &&
3919			       btrfs_free_space_trimmed(entry))) {
3920			spin_unlock(&ctl->tree_lock);
3921			mutex_unlock(&ctl->cache_writeout_mutex);
3922			next_bitmap = true;
3923			goto next;
3924		}
3925
3926		/*
3927		 * Async discard bitmap trimming begins at by setting the start
3928		 * to be key.objectid and the offset_to_bitmap() aligns to the
3929		 * start of the bitmap.  This lets us know we are fully
3930		 * scanning the bitmap rather than only some portion of it.
3931		 */
3932		if (start == offset)
3933			entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3934
3935		bytes = minlen;
3936		ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3937		if (ret2 || start >= end) {
3938			/*
3939			 * We lossily consider a bitmap trimmed if we only skip
3940			 * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3941			 */
3942			if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3943				end_trimming_bitmap(ctl, entry);
3944			else
3945				entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3946			spin_unlock(&ctl->tree_lock);
3947			mutex_unlock(&ctl->cache_writeout_mutex);
3948			next_bitmap = true;
3949			goto next;
3950		}
3951
3952		/*
3953		 * We already trimmed a region, but are using the locking above
3954		 * to reset the trim_state.
3955		 */
3956		if (async && *total_trimmed) {
3957			spin_unlock(&ctl->tree_lock);
3958			mutex_unlock(&ctl->cache_writeout_mutex);
3959			goto out;
3960		}
3961
3962		bytes = min(bytes, end - start);
3963		if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3964			spin_unlock(&ctl->tree_lock);
3965			mutex_unlock(&ctl->cache_writeout_mutex);
3966			goto next;
3967		}
3968
3969		/*
3970		 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3971		 * If X < @minlen, we won't trim X when we come back around.
3972		 * So trim it now.  We differ here from trimming extents as we
3973		 * don't keep individual state per bit.
3974		 */
3975		if (async &&
3976		    max_discard_size &&
3977		    bytes > (max_discard_size + minlen))
3978			bytes = max_discard_size;
3979
3980		bitmap_clear_bits(ctl, entry, start, bytes, true);
3981		if (entry->bytes == 0)
3982			free_bitmap(ctl, entry);
3983
3984		spin_unlock(&ctl->tree_lock);
3985		trim_entry.start = start;
3986		trim_entry.bytes = bytes;
3987		list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3988		mutex_unlock(&ctl->cache_writeout_mutex);
3989
3990		ret = do_trimming(block_group, total_trimmed, start, bytes,
3991				  start, bytes, 0, &trim_entry);
3992		if (ret) {
3993			reset_trimming_bitmap(ctl, offset);
3994			block_group->discard_cursor =
3995				btrfs_block_group_end(block_group);
3996			break;
3997		}
3998next:
3999		if (next_bitmap) {
4000			offset += BITS_PER_BITMAP * ctl->unit;
4001			start = offset;
4002		} else {
4003			start += bytes;
4004		}
4005		block_group->discard_cursor = start;
4006
4007		if (btrfs_trim_interrupted()) {
4008			if (start != offset)
4009				reset_trimming_bitmap(ctl, offset);
4010			ret = -ERESTARTSYS;
4011			break;
4012		}
4013
4014		cond_resched();
4015	}
4016
4017	if (offset >= end)
4018		block_group->discard_cursor = end;
4019
4020out:
4021	return ret;
4022}
4023
4024int btrfs_trim_block_group(struct btrfs_block_group *block_group,
4025			   u64 *trimmed, u64 start, u64 end, u64 minlen)
4026{
4027	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
4028	int ret;
4029	u64 rem = 0;
4030
4031	ASSERT(!btrfs_is_zoned(block_group->fs_info));
4032
4033	*trimmed = 0;
4034
4035	spin_lock(&block_group->lock);
4036	if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4037		spin_unlock(&block_group->lock);
4038		return 0;
4039	}
4040	btrfs_freeze_block_group(block_group);
4041	spin_unlock(&block_group->lock);
4042
4043	ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4044	if (ret)
4045		goto out;
4046
4047	ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4048	div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4049	/* If we ended in the middle of a bitmap, reset the trimming flag */
4050	if (rem)
4051		reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4052out:
4053	btrfs_unfreeze_block_group(block_group);
4054	return ret;
4055}
4056
4057int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4058				   u64 *trimmed, u64 start, u64 end, u64 minlen,
4059				   bool async)
4060{
4061	int ret;
4062
4063	*trimmed = 0;
4064
4065	spin_lock(&block_group->lock);
4066	if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4067		spin_unlock(&block_group->lock);
4068		return 0;
4069	}
4070	btrfs_freeze_block_group(block_group);
4071	spin_unlock(&block_group->lock);
4072
4073	ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4074	btrfs_unfreeze_block_group(block_group);
4075
4076	return ret;
4077}
4078
4079int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4080				   u64 *trimmed, u64 start, u64 end, u64 minlen,
4081				   u64 maxlen, bool async)
4082{
4083	int ret;
4084
4085	*trimmed = 0;
4086
4087	spin_lock(&block_group->lock);
4088	if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4089		spin_unlock(&block_group->lock);
4090		return 0;
4091	}
4092	btrfs_freeze_block_group(block_group);
4093	spin_unlock(&block_group->lock);
4094
4095	ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4096			   async);
4097
4098	btrfs_unfreeze_block_group(block_group);
4099
4100	return ret;
4101}
4102
4103bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4104{
4105	return btrfs_super_cache_generation(fs_info->super_copy);
4106}
4107
4108static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4109				       struct btrfs_trans_handle *trans)
4110{
4111	struct btrfs_block_group *block_group;
4112	struct rb_node *node;
4113	int ret = 0;
4114
4115	btrfs_info(fs_info, "cleaning free space cache v1");
4116
4117	node = rb_first_cached(&fs_info->block_group_cache_tree);
4118	while (node) {
4119		block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4120		ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4121		if (ret)
4122			goto out;
4123		node = rb_next(node);
4124	}
4125out:
4126	return ret;
4127}
4128
4129int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4130{
4131	struct btrfs_trans_handle *trans;
4132	int ret;
4133
4134	/*
4135	 * update_super_roots will appropriately set or unset
4136	 * super_copy->cache_generation based on SPACE_CACHE and
4137	 * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4138	 * transaction commit whether we are enabling space cache v1 and don't
4139	 * have any other work to do, or are disabling it and removing free
4140	 * space inodes.
4141	 */
4142	trans = btrfs_start_transaction(fs_info->tree_root, 0);
4143	if (IS_ERR(trans))
4144		return PTR_ERR(trans);
4145
4146	if (!active) {
4147		set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4148		ret = cleanup_free_space_cache_v1(fs_info, trans);
4149		if (unlikely(ret)) {
4150			btrfs_abort_transaction(trans, ret);
4151			btrfs_end_transaction(trans);
4152			goto out;
4153		}
4154	}
4155
4156	ret = btrfs_commit_transaction(trans);
4157out:
4158	clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4159
4160	return ret;
4161}
4162
4163int __init btrfs_free_space_init(void)
4164{
4165	btrfs_free_space_cachep = KMEM_CACHE(btrfs_free_space, 0);
4166	if (!btrfs_free_space_cachep)
4167		return -ENOMEM;
4168
4169	btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
4170							PAGE_SIZE, PAGE_SIZE,
4171							0, NULL);
4172	if (!btrfs_free_space_bitmap_cachep) {
4173		kmem_cache_destroy(btrfs_free_space_cachep);
4174		return -ENOMEM;
4175	}
4176
4177	return 0;
4178}
4179
4180void __cold btrfs_free_space_exit(void)
4181{
4182	kmem_cache_destroy(btrfs_free_space_cachep);
4183	kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
4184}
4185
4186#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4187/*
4188 * Use this if you need to make a bitmap or extent entry specifically, it
4189 * doesn't do any of the merging that add_free_space does, this acts a lot like
4190 * how the free space cache loading stuff works, so you can get really weird
4191 * configurations.
4192 */
4193int test_add_free_space_entry(struct btrfs_block_group *cache,
4194			      u64 offset, u64 bytes, bool bitmap)
4195{
4196	struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4197	struct btrfs_free_space *info = NULL, *bitmap_info;
4198	void *map = NULL;
4199	enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4200	u64 bytes_added;
4201	int ret;
4202
4203again:
4204	if (!info) {
4205		info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4206		if (!info)
4207			return -ENOMEM;
4208	}
4209
4210	if (!bitmap) {
4211		spin_lock(&ctl->tree_lock);
4212		info->offset = offset;
4213		info->bytes = bytes;
4214		info->max_extent_size = 0;
4215		ret = link_free_space(ctl, info);
4216		spin_unlock(&ctl->tree_lock);
4217		if (ret)
4218			kmem_cache_free(btrfs_free_space_cachep, info);
4219		return ret;
4220	}
4221
4222	if (!map) {
4223		map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4224		if (!map) {
4225			kmem_cache_free(btrfs_free_space_cachep, info);
4226			return -ENOMEM;
4227		}
4228	}
4229
4230	spin_lock(&ctl->tree_lock);
4231	bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4232					 1, 0);
4233	if (!bitmap_info) {
4234		info->bitmap = map;
4235		map = NULL;
4236		add_new_bitmap(ctl, info, offset);
4237		bitmap_info = info;
4238		info = NULL;
4239	}
4240
4241	bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4242					  trim_state);
4243
4244	bytes -= bytes_added;
4245	offset += bytes_added;
4246	spin_unlock(&ctl->tree_lock);
4247
4248	if (bytes)
4249		goto again;
4250
4251	if (info)
4252		kmem_cache_free(btrfs_free_space_cachep, info);
4253	if (map)
4254		kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4255	return 0;
4256}
4257
4258/*
4259 * Checks to see if the given range is in the free space cache.  This is really
4260 * just used to check the absence of space, so if there is free space in the
4261 * range at all we will return 1.
4262 */
4263int test_check_exists(struct btrfs_block_group *cache,
4264		      u64 offset, u64 bytes)
4265{
4266	struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4267	struct btrfs_free_space *info;
4268	int ret = 0;
4269
4270	spin_lock(&ctl->tree_lock);
4271	info = tree_search_offset(ctl, offset, 0, 0);
4272	if (!info) {
4273		info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4274					  1, 0);
4275		if (!info)
4276			goto out;
4277	}
4278
4279have_info:
4280	if (info->bitmap) {
4281		u64 bit_off, bit_bytes;
4282		struct rb_node *n;
4283		struct btrfs_free_space *tmp;
4284
4285		bit_off = offset;
4286		bit_bytes = ctl->unit;
4287		ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4288		if (!ret) {
4289			if (bit_off == offset) {
4290				ret = 1;
4291				goto out;
4292			} else if (bit_off > offset &&
4293				   offset + bytes > bit_off) {
4294				ret = 1;
4295				goto out;
4296			}
4297		}
4298
4299		n = rb_prev(&info->offset_index);
4300		while (n) {
4301			tmp = rb_entry(n, struct btrfs_free_space,
4302				       offset_index);
4303			if (tmp->offset + tmp->bytes < offset)
4304				break;
4305			if (offset + bytes < tmp->offset) {
4306				n = rb_prev(&tmp->offset_index);
4307				continue;
4308			}
4309			info = tmp;
4310			goto have_info;
4311		}
4312
4313		n = rb_next(&info->offset_index);
4314		while (n) {
4315			tmp = rb_entry(n, struct btrfs_free_space,
4316				       offset_index);
4317			if (offset + bytes < tmp->offset)
4318				break;
4319			if (tmp->offset + tmp->bytes < offset) {
4320				n = rb_next(&tmp->offset_index);
4321				continue;
4322			}
4323			info = tmp;
4324			goto have_info;
4325		}
4326
4327		ret = 0;
4328		goto out;
4329	}
4330
4331	if (info->offset == offset) {
4332		ret = 1;
4333		goto out;
4334	}
4335
4336	if (offset > info->offset && offset < info->offset + info->bytes)
4337		ret = 1;
4338out:
4339	spin_unlock(&ctl->tree_lock);
4340	return ret;
4341}
4342#endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */