2 * Copyright (C) 2008 Red Hat. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
19 #include <linux/pagemap.h>
20 #include <linux/sched.h>
21 #include <linux/slab.h>
22 #include <linux/math64.h>
24 #include "free-space-cache.h"
25 #include "transaction.h"
27 #include "extent_io.h"
28 #include "inode-map.h"
30 #define BITS_PER_BITMAP (PAGE_CACHE_SIZE * 8)
31 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
33 static int link_free_space(struct btrfs_free_space_ctl *ctl,
34 struct btrfs_free_space *info);
36 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
37 struct btrfs_path *path,
41 struct btrfs_key location;
42 struct btrfs_disk_key disk_key;
43 struct btrfs_free_space_header *header;
44 struct extent_buffer *leaf;
45 struct inode *inode = NULL;
48 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
52 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
56 btrfs_release_path(path);
57 return ERR_PTR(-ENOENT);
60 leaf = path->nodes[0];
61 header = btrfs_item_ptr(leaf, path->slots[0],
62 struct btrfs_free_space_header);
63 btrfs_free_space_key(leaf, header, &disk_key);
64 btrfs_disk_key_to_cpu(&location, &disk_key);
65 btrfs_release_path(path);
67 inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
69 return ERR_PTR(-ENOENT);
72 if (is_bad_inode(inode)) {
74 return ERR_PTR(-ENOENT);
77 inode->i_mapping->flags &= ~__GFP_FS;
82 struct inode *lookup_free_space_inode(struct btrfs_root *root,
83 struct btrfs_block_group_cache
84 *block_group, struct btrfs_path *path)
86 struct inode *inode = NULL;
88 spin_lock(&block_group->lock);
89 if (block_group->inode)
90 inode = igrab(block_group->inode);
91 spin_unlock(&block_group->lock);
95 inode = __lookup_free_space_inode(root, path,
96 block_group->key.objectid);
100 spin_lock(&block_group->lock);
101 if (!root->fs_info->closing) {
102 block_group->inode = igrab(inode);
103 block_group->iref = 1;
105 spin_unlock(&block_group->lock);
110 int __create_free_space_inode(struct btrfs_root *root,
111 struct btrfs_trans_handle *trans,
112 struct btrfs_path *path, u64 ino, u64 offset)
114 struct btrfs_key key;
115 struct btrfs_disk_key disk_key;
116 struct btrfs_free_space_header *header;
117 struct btrfs_inode_item *inode_item;
118 struct extent_buffer *leaf;
121 ret = btrfs_insert_empty_inode(trans, root, path, ino);
125 leaf = path->nodes[0];
126 inode_item = btrfs_item_ptr(leaf, path->slots[0],
127 struct btrfs_inode_item);
128 btrfs_item_key(leaf, &disk_key, path->slots[0]);
129 memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
130 sizeof(*inode_item));
131 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
132 btrfs_set_inode_size(leaf, inode_item, 0);
133 btrfs_set_inode_nbytes(leaf, inode_item, 0);
134 btrfs_set_inode_uid(leaf, inode_item, 0);
135 btrfs_set_inode_gid(leaf, inode_item, 0);
136 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
137 btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
138 BTRFS_INODE_PREALLOC | BTRFS_INODE_NODATASUM);
139 btrfs_set_inode_nlink(leaf, inode_item, 1);
140 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
141 btrfs_set_inode_block_group(leaf, inode_item, offset);
142 btrfs_mark_buffer_dirty(leaf);
143 btrfs_release_path(path);
145 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
149 ret = btrfs_insert_empty_item(trans, root, path, &key,
150 sizeof(struct btrfs_free_space_header));
152 btrfs_release_path(path);
155 leaf = path->nodes[0];
156 header = btrfs_item_ptr(leaf, path->slots[0],
157 struct btrfs_free_space_header);
158 memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
159 btrfs_set_free_space_key(leaf, header, &disk_key);
160 btrfs_mark_buffer_dirty(leaf);
161 btrfs_release_path(path);
166 int create_free_space_inode(struct btrfs_root *root,
167 struct btrfs_trans_handle *trans,
168 struct btrfs_block_group_cache *block_group,
169 struct btrfs_path *path)
174 ret = btrfs_find_free_objectid(root, &ino);
178 return __create_free_space_inode(root, trans, path, ino,
179 block_group->key.objectid);
182 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
183 struct btrfs_trans_handle *trans,
184 struct btrfs_path *path,
190 trans->block_rsv = root->orphan_block_rsv;
191 ret = btrfs_block_rsv_check(trans, root,
192 root->orphan_block_rsv,
197 oldsize = i_size_read(inode);
198 btrfs_i_size_write(inode, 0);
199 truncate_pagecache(inode, oldsize, 0);
202 * We don't need an orphan item because truncating the free space cache
203 * will never be split across transactions.
205 ret = btrfs_truncate_inode_items(trans, root, inode,
206 0, BTRFS_EXTENT_DATA_KEY);
212 ret = btrfs_update_inode(trans, root, inode);
216 static int readahead_cache(struct inode *inode)
218 struct file_ra_state *ra;
219 unsigned long last_index;
221 ra = kzalloc(sizeof(*ra), GFP_NOFS);
225 file_ra_state_init(ra, inode->i_mapping);
226 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
228 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
235 int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
236 struct btrfs_free_space_ctl *ctl,
237 struct btrfs_path *path, u64 offset)
239 struct btrfs_free_space_header *header;
240 struct extent_buffer *leaf;
242 u32 *checksums = NULL, *crc;
243 char *disk_crcs = NULL;
244 struct btrfs_key key;
245 struct list_head bitmaps;
249 u32 cur_crc = ~(u32)0;
251 unsigned long first_page_offset;
255 INIT_LIST_HEAD(&bitmaps);
257 /* Nothing in the space cache, goodbye */
258 if (!i_size_read(inode))
261 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
265 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
269 btrfs_release_path(path);
276 leaf = path->nodes[0];
277 header = btrfs_item_ptr(leaf, path->slots[0],
278 struct btrfs_free_space_header);
279 num_entries = btrfs_free_space_entries(leaf, header);
280 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
281 generation = btrfs_free_space_generation(leaf, header);
282 btrfs_release_path(path);
284 if (BTRFS_I(inode)->generation != generation) {
285 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
286 " not match free space cache generation (%llu)\n",
287 (unsigned long long)BTRFS_I(inode)->generation,
288 (unsigned long long)generation);
295 /* Setup everything for doing checksumming */
296 num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
297 checksums = crc = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
300 first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
301 disk_crcs = kzalloc(first_page_offset, GFP_NOFS);
305 ret = readahead_cache(inode);
310 struct btrfs_free_space_entry *entry;
311 struct btrfs_free_space *e;
313 unsigned long offset = 0;
314 unsigned long start_offset = 0;
317 if (!num_entries && !num_bitmaps)
321 start_offset = first_page_offset;
322 offset = start_offset;
325 page = grab_cache_page(inode->i_mapping, index);
329 if (!PageUptodate(page)) {
330 btrfs_readpage(NULL, page);
332 if (!PageUptodate(page)) {
334 page_cache_release(page);
335 printk(KERN_ERR "btrfs: error reading free "
345 memcpy(disk_crcs, addr, first_page_offset);
346 gen = addr + (sizeof(u32) * num_checksums);
347 if (*gen != BTRFS_I(inode)->generation) {
348 printk(KERN_ERR "btrfs: space cache generation"
349 " (%llu) does not match inode (%llu)\n",
350 (unsigned long long)*gen,
352 BTRFS_I(inode)->generation);
355 page_cache_release(page);
358 crc = (u32 *)disk_crcs;
360 entry = addr + start_offset;
362 /* First lets check our crc before we do anything fun */
364 cur_crc = btrfs_csum_data(root, addr + start_offset, cur_crc,
365 PAGE_CACHE_SIZE - start_offset);
366 btrfs_csum_final(cur_crc, (char *)&cur_crc);
367 if (cur_crc != *crc) {
368 printk(KERN_ERR "btrfs: crc mismatch for page %lu\n",
372 page_cache_release(page);
382 e = kmem_cache_zalloc(btrfs_free_space_cachep,
387 page_cache_release(page);
391 e->offset = le64_to_cpu(entry->offset);
392 e->bytes = le64_to_cpu(entry->bytes);
395 kmem_cache_free(btrfs_free_space_cachep, e);
397 page_cache_release(page);
401 if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
402 spin_lock(&ctl->tree_lock);
403 ret = link_free_space(ctl, e);
404 spin_unlock(&ctl->tree_lock);
406 printk(KERN_ERR "Duplicate entries in "
407 "free space cache, dumping\n");
410 page_cache_release(page);
414 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
418 btrfs_free_space_cachep, e);
420 page_cache_release(page);
423 spin_lock(&ctl->tree_lock);
424 ret2 = link_free_space(ctl, e);
425 ctl->total_bitmaps++;
426 ctl->op->recalc_thresholds(ctl);
427 spin_unlock(&ctl->tree_lock);
428 list_add_tail(&e->list, &bitmaps);
430 printk(KERN_ERR "Duplicate entries in "
431 "free space cache, dumping\n");
434 page_cache_release(page);
440 offset += sizeof(struct btrfs_free_space_entry);
441 if (offset + sizeof(struct btrfs_free_space_entry) >=
448 * We read an entry out of this page, we need to move on to the
457 * We add the bitmaps at the end of the entries in order that
458 * the bitmap entries are added to the cache.
460 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
461 list_del_init(&e->list);
462 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
467 page_cache_release(page);
477 __btrfs_remove_free_space_cache(ctl);
481 int load_free_space_cache(struct btrfs_fs_info *fs_info,
482 struct btrfs_block_group_cache *block_group)
484 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
485 struct btrfs_root *root = fs_info->tree_root;
487 struct btrfs_path *path;
490 u64 used = btrfs_block_group_used(&block_group->item);
493 * If we're unmounting then just return, since this does a search on the
494 * normal root and not the commit root and we could deadlock.
497 if (fs_info->closing)
501 * If this block group has been marked to be cleared for one reason or
502 * another then we can't trust the on disk cache, so just return.
504 spin_lock(&block_group->lock);
505 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
506 spin_unlock(&block_group->lock);
509 spin_unlock(&block_group->lock);
511 path = btrfs_alloc_path();
515 inode = lookup_free_space_inode(root, block_group, path);
517 btrfs_free_path(path);
521 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
522 path, block_group->key.objectid);
523 btrfs_free_path(path);
527 spin_lock(&ctl->tree_lock);
528 matched = (ctl->free_space == (block_group->key.offset - used -
529 block_group->bytes_super));
530 spin_unlock(&ctl->tree_lock);
533 __btrfs_remove_free_space_cache(ctl);
534 printk(KERN_ERR "block group %llu has an wrong amount of free "
535 "space\n", block_group->key.objectid);
540 /* This cache is bogus, make sure it gets cleared */
541 spin_lock(&block_group->lock);
542 block_group->disk_cache_state = BTRFS_DC_CLEAR;
543 spin_unlock(&block_group->lock);
546 printk(KERN_ERR "btrfs: failed to load free space cache "
547 "for block group %llu\n", block_group->key.objectid);
554 int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
555 struct btrfs_free_space_ctl *ctl,
556 struct btrfs_block_group_cache *block_group,
557 struct btrfs_trans_handle *trans,
558 struct btrfs_path *path, u64 offset)
560 struct btrfs_free_space_header *header;
561 struct extent_buffer *leaf;
562 struct rb_node *node;
563 struct list_head *pos, *n;
566 struct extent_state *cached_state = NULL;
567 struct btrfs_free_cluster *cluster = NULL;
568 struct extent_io_tree *unpin = NULL;
569 struct list_head bitmap_list;
570 struct btrfs_key key;
573 u32 *crc, *checksums;
574 unsigned long first_page_offset;
575 int index = 0, num_pages = 0;
579 bool next_page = false;
580 bool out_of_space = false;
582 INIT_LIST_HEAD(&bitmap_list);
584 node = rb_first(&ctl->free_space_offset);
588 if (!i_size_read(inode))
591 num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
593 filemap_write_and_wait(inode->i_mapping);
594 btrfs_wait_ordered_range(inode, inode->i_size &
595 ~(root->sectorsize - 1), (u64)-1);
597 /* We need a checksum per page. */
598 crc = checksums = kzalloc(sizeof(u32) * num_pages, GFP_NOFS);
602 pages = kzalloc(sizeof(struct page *) * num_pages, GFP_NOFS);
608 /* Since the first page has all of our checksums and our generation we
609 * need to calculate the offset into the page that we can start writing
612 first_page_offset = (sizeof(u32) * num_pages) + sizeof(u64);
614 /* Get the cluster for this block_group if it exists */
615 if (block_group && !list_empty(&block_group->cluster_list))
616 cluster = list_entry(block_group->cluster_list.next,
617 struct btrfs_free_cluster,
621 * We shouldn't have switched the pinned extents yet so this is the
624 unpin = root->fs_info->pinned_extents;
627 * Lock all pages first so we can lock the extent safely.
629 * NOTE: Because we hold the ref the entire time we're going to write to
630 * the page find_get_page should never fail, so we don't do a check
631 * after find_get_page at this point. Just putting this here so people
632 * know and don't freak out.
634 while (index < num_pages) {
635 page = grab_cache_page(inode->i_mapping, index);
639 for (i = 0; i < num_pages; i++) {
640 unlock_page(pages[i]);
641 page_cache_release(pages[i]);
650 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
651 0, &cached_state, GFP_NOFS);
654 * When searching for pinned extents, we need to start at our start
658 start = block_group->key.objectid;
660 /* Write out the extent entries */
662 struct btrfs_free_space_entry *entry;
664 unsigned long offset = 0;
665 unsigned long start_offset = 0;
670 start_offset = first_page_offset;
671 offset = start_offset;
674 if (index >= num_pages) {
682 entry = addr + start_offset;
684 memset(addr, 0, PAGE_CACHE_SIZE);
685 while (node && !next_page) {
686 struct btrfs_free_space *e;
688 e = rb_entry(node, struct btrfs_free_space, offset_index);
691 entry->offset = cpu_to_le64(e->offset);
692 entry->bytes = cpu_to_le64(e->bytes);
694 entry->type = BTRFS_FREE_SPACE_BITMAP;
695 list_add_tail(&e->list, &bitmap_list);
698 entry->type = BTRFS_FREE_SPACE_EXTENT;
700 node = rb_next(node);
701 if (!node && cluster) {
702 node = rb_first(&cluster->root);
705 offset += sizeof(struct btrfs_free_space_entry);
706 if (offset + sizeof(struct btrfs_free_space_entry) >=
713 * We want to add any pinned extents to our free space cache
714 * so we don't leak the space
716 while (block_group && !next_page &&
717 (start < block_group->key.objectid +
718 block_group->key.offset)) {
719 ret = find_first_extent_bit(unpin, start, &start, &end,
726 /* This pinned extent is out of our range */
727 if (start >= block_group->key.objectid +
728 block_group->key.offset)
731 len = block_group->key.objectid +
732 block_group->key.offset - start;
733 len = min(len, end + 1 - start);
736 entry->offset = cpu_to_le64(start);
737 entry->bytes = cpu_to_le64(len);
738 entry->type = BTRFS_FREE_SPACE_EXTENT;
741 offset += sizeof(struct btrfs_free_space_entry);
742 if (offset + sizeof(struct btrfs_free_space_entry) >=
748 *crc = btrfs_csum_data(root, addr + start_offset, *crc,
749 PAGE_CACHE_SIZE - start_offset);
752 btrfs_csum_final(*crc, (char *)crc);
755 bytes += PAGE_CACHE_SIZE;
758 } while (node || next_page);
760 /* Write out the bitmaps */
761 list_for_each_safe(pos, n, &bitmap_list) {
763 struct btrfs_free_space *entry =
764 list_entry(pos, struct btrfs_free_space, list);
766 if (index >= num_pages) {
773 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
775 *crc = btrfs_csum_data(root, addr, *crc, PAGE_CACHE_SIZE);
777 btrfs_csum_final(*crc, (char *)crc);
779 bytes += PAGE_CACHE_SIZE;
781 list_del_init(&entry->list);
786 btrfs_drop_pages(pages, num_pages);
787 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
788 i_size_read(inode) - 1, &cached_state,
794 /* Zero out the rest of the pages just to make sure */
795 while (index < num_pages) {
800 memset(addr, 0, PAGE_CACHE_SIZE);
802 bytes += PAGE_CACHE_SIZE;
806 /* Write the checksums and trans id to the first page */
814 memcpy(addr, checksums, sizeof(u32) * num_pages);
815 gen = addr + (sizeof(u32) * num_pages);
816 *gen = trans->transid;
820 ret = btrfs_dirty_pages(root, inode, pages, num_pages, 0,
821 bytes, &cached_state);
822 btrfs_drop_pages(pages, num_pages);
823 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
824 i_size_read(inode) - 1, &cached_state, GFP_NOFS);
831 BTRFS_I(inode)->generation = trans->transid;
833 filemap_write_and_wait(inode->i_mapping);
835 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
839 ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
842 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
843 EXTENT_DIRTY | EXTENT_DELALLOC |
844 EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
847 leaf = path->nodes[0];
849 struct btrfs_key found_key;
850 BUG_ON(!path->slots[0]);
852 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
853 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
854 found_key.offset != offset) {
856 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
857 EXTENT_DIRTY | EXTENT_DELALLOC |
858 EXTENT_DO_ACCOUNTING, 0, 0, NULL,
860 btrfs_release_path(path);
864 header = btrfs_item_ptr(leaf, path->slots[0],
865 struct btrfs_free_space_header);
866 btrfs_set_free_space_entries(leaf, header, entries);
867 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
868 btrfs_set_free_space_generation(leaf, header, trans->transid);
869 btrfs_mark_buffer_dirty(leaf);
870 btrfs_release_path(path);
876 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
877 BTRFS_I(inode)->generation = 0;
881 btrfs_update_inode(trans, root, inode);
885 int btrfs_write_out_cache(struct btrfs_root *root,
886 struct btrfs_trans_handle *trans,
887 struct btrfs_block_group_cache *block_group,
888 struct btrfs_path *path)
890 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
894 root = root->fs_info->tree_root;
896 spin_lock(&block_group->lock);
897 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
898 spin_unlock(&block_group->lock);
901 spin_unlock(&block_group->lock);
903 inode = lookup_free_space_inode(root, block_group, path);
907 ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
908 path, block_group->key.objectid);
910 spin_lock(&block_group->lock);
911 block_group->disk_cache_state = BTRFS_DC_ERROR;
912 spin_unlock(&block_group->lock);
915 printk(KERN_ERR "btrfs: failed to write free space cace "
916 "for block group %llu\n", block_group->key.objectid);
923 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
926 BUG_ON(offset < bitmap_start);
927 offset -= bitmap_start;
928 return (unsigned long)(div_u64(offset, unit));
931 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
933 return (unsigned long)(div_u64(bytes, unit));
936 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
940 u64 bytes_per_bitmap;
942 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
943 bitmap_start = offset - ctl->start;
944 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
945 bitmap_start *= bytes_per_bitmap;
946 bitmap_start += ctl->start;
951 static int tree_insert_offset(struct rb_root *root, u64 offset,
952 struct rb_node *node, int bitmap)
954 struct rb_node **p = &root->rb_node;
955 struct rb_node *parent = NULL;
956 struct btrfs_free_space *info;
960 info = rb_entry(parent, struct btrfs_free_space, offset_index);
962 if (offset < info->offset) {
964 } else if (offset > info->offset) {
968 * we could have a bitmap entry and an extent entry
969 * share the same offset. If this is the case, we want
970 * the extent entry to always be found first if we do a
971 * linear search through the tree, since we want to have
972 * the quickest allocation time, and allocating from an
973 * extent is faster than allocating from a bitmap. So
974 * if we're inserting a bitmap and we find an entry at
975 * this offset, we want to go right, or after this entry
976 * logically. If we are inserting an extent and we've
977 * found a bitmap, we want to go left, or before
996 rb_link_node(node, parent, p);
997 rb_insert_color(node, root);
1003 * searches the tree for the given offset.
1005 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1006 * want a section that has at least bytes size and comes at or after the given
1009 static struct btrfs_free_space *
1010 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1011 u64 offset, int bitmap_only, int fuzzy)
1013 struct rb_node *n = ctl->free_space_offset.rb_node;
1014 struct btrfs_free_space *entry, *prev = NULL;
1016 /* find entry that is closest to the 'offset' */
1023 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1026 if (offset < entry->offset)
1028 else if (offset > entry->offset)
1041 * bitmap entry and extent entry may share same offset,
1042 * in that case, bitmap entry comes after extent entry.
1047 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1048 if (entry->offset != offset)
1051 WARN_ON(!entry->bitmap);
1054 if (entry->bitmap) {
1056 * if previous extent entry covers the offset,
1057 * we should return it instead of the bitmap entry
1059 n = &entry->offset_index;
1064 prev = rb_entry(n, struct btrfs_free_space,
1066 if (!prev->bitmap) {
1067 if (prev->offset + prev->bytes > offset)
1079 /* find last entry before the 'offset' */
1081 if (entry->offset > offset) {
1082 n = rb_prev(&entry->offset_index);
1084 entry = rb_entry(n, struct btrfs_free_space,
1086 BUG_ON(entry->offset > offset);
1095 if (entry->bitmap) {
1096 n = &entry->offset_index;
1101 prev = rb_entry(n, struct btrfs_free_space,
1103 if (!prev->bitmap) {
1104 if (prev->offset + prev->bytes > offset)
1109 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1111 } else if (entry->offset + entry->bytes > offset)
1118 if (entry->bitmap) {
1119 if (entry->offset + BITS_PER_BITMAP *
1123 if (entry->offset + entry->bytes > offset)
1127 n = rb_next(&entry->offset_index);
1130 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1136 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1137 struct btrfs_free_space *info)
1139 rb_erase(&info->offset_index, &ctl->free_space_offset);
1140 ctl->free_extents--;
1143 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1144 struct btrfs_free_space *info)
1146 __unlink_free_space(ctl, info);
1147 ctl->free_space -= info->bytes;
1150 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1151 struct btrfs_free_space *info)
1155 BUG_ON(!info->bitmap && !info->bytes);
1156 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1157 &info->offset_index, (info->bitmap != NULL));
1161 ctl->free_space += info->bytes;
1162 ctl->free_extents++;
1166 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1168 struct btrfs_block_group_cache *block_group = ctl->private;
1172 u64 size = block_group->key.offset;
1173 u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1174 int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1176 BUG_ON(ctl->total_bitmaps > max_bitmaps);
1179 * The goal is to keep the total amount of memory used per 1gb of space
1180 * at or below 32k, so we need to adjust how much memory we allow to be
1181 * used by extent based free space tracking
1183 if (size < 1024 * 1024 * 1024)
1184 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1186 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1187 div64_u64(size, 1024 * 1024 * 1024);
1190 * we want to account for 1 more bitmap than what we have so we can make
1191 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1192 * we add more bitmaps.
1194 bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1196 if (bitmap_bytes >= max_bytes) {
1197 ctl->extents_thresh = 0;
1202 * we want the extent entry threshold to always be at most 1/2 the maxw
1203 * bytes we can have, or whatever is less than that.
1205 extent_bytes = max_bytes - bitmap_bytes;
1206 extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1208 ctl->extents_thresh =
1209 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1212 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1213 struct btrfs_free_space *info, u64 offset,
1216 unsigned long start, count;
1218 start = offset_to_bit(info->offset, ctl->unit, offset);
1219 count = bytes_to_bits(bytes, ctl->unit);
1220 BUG_ON(start + count > BITS_PER_BITMAP);
1222 bitmap_clear(info->bitmap, start, count);
1224 info->bytes -= bytes;
1225 ctl->free_space -= bytes;
1228 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1229 struct btrfs_free_space *info, u64 offset,
1232 unsigned long start, count;
1234 start = offset_to_bit(info->offset, ctl->unit, offset);
1235 count = bytes_to_bits(bytes, ctl->unit);
1236 BUG_ON(start + count > BITS_PER_BITMAP);
1238 bitmap_set(info->bitmap, start, count);
1240 info->bytes += bytes;
1241 ctl->free_space += bytes;
1244 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1245 struct btrfs_free_space *bitmap_info, u64 *offset,
1248 unsigned long found_bits = 0;
1249 unsigned long bits, i;
1250 unsigned long next_zero;
1252 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1253 max_t(u64, *offset, bitmap_info->offset));
1254 bits = bytes_to_bits(*bytes, ctl->unit);
1256 for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1257 i < BITS_PER_BITMAP;
1258 i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1259 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1260 BITS_PER_BITMAP, i);
1261 if ((next_zero - i) >= bits) {
1262 found_bits = next_zero - i;
1269 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1270 *bytes = (u64)(found_bits) * ctl->unit;
1277 static struct btrfs_free_space *
1278 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
1280 struct btrfs_free_space *entry;
1281 struct rb_node *node;
1284 if (!ctl->free_space_offset.rb_node)
1287 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1291 for (node = &entry->offset_index; node; node = rb_next(node)) {
1292 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1293 if (entry->bytes < *bytes)
1296 if (entry->bitmap) {
1297 ret = search_bitmap(ctl, entry, offset, bytes);
1303 *offset = entry->offset;
1304 *bytes = entry->bytes;
1311 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1312 struct btrfs_free_space *info, u64 offset)
1314 info->offset = offset_to_bitmap(ctl, offset);
1316 link_free_space(ctl, info);
1317 ctl->total_bitmaps++;
1319 ctl->op->recalc_thresholds(ctl);
1322 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1323 struct btrfs_free_space *bitmap_info)
1325 unlink_free_space(ctl, bitmap_info);
1326 kfree(bitmap_info->bitmap);
1327 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1328 ctl->total_bitmaps--;
1329 ctl->op->recalc_thresholds(ctl);
1332 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1333 struct btrfs_free_space *bitmap_info,
1334 u64 *offset, u64 *bytes)
1337 u64 search_start, search_bytes;
1341 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1344 * XXX - this can go away after a few releases.
1346 * since the only user of btrfs_remove_free_space is the tree logging
1347 * stuff, and the only way to test that is under crash conditions, we
1348 * want to have this debug stuff here just in case somethings not
1349 * working. Search the bitmap for the space we are trying to use to
1350 * make sure its actually there. If its not there then we need to stop
1351 * because something has gone wrong.
1353 search_start = *offset;
1354 search_bytes = *bytes;
1355 search_bytes = min(search_bytes, end - search_start + 1);
1356 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1357 BUG_ON(ret < 0 || search_start != *offset);
1359 if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1360 bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
1361 *bytes -= end - *offset + 1;
1363 } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1364 bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
1369 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1370 if (!bitmap_info->bytes)
1371 free_bitmap(ctl, bitmap_info);
1374 * no entry after this bitmap, but we still have bytes to
1375 * remove, so something has gone wrong.
1380 bitmap_info = rb_entry(next, struct btrfs_free_space,
1384 * if the next entry isn't a bitmap we need to return to let the
1385 * extent stuff do its work.
1387 if (!bitmap_info->bitmap)
1391 * Ok the next item is a bitmap, but it may not actually hold
1392 * the information for the rest of this free space stuff, so
1393 * look for it, and if we don't find it return so we can try
1394 * everything over again.
1396 search_start = *offset;
1397 search_bytes = *bytes;
1398 ret = search_bitmap(ctl, bitmap_info, &search_start,
1400 if (ret < 0 || search_start != *offset)
1404 } else if (!bitmap_info->bytes)
1405 free_bitmap(ctl, bitmap_info);
1410 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1411 struct btrfs_free_space *info)
1413 struct btrfs_block_group_cache *block_group = ctl->private;
1416 * If we are below the extents threshold then we can add this as an
1417 * extent, and don't have to deal with the bitmap
1419 if (ctl->free_extents < ctl->extents_thresh) {
1421 * If this block group has some small extents we don't want to
1422 * use up all of our free slots in the cache with them, we want
1423 * to reserve them to larger extents, however if we have plent
1424 * of cache left then go ahead an dadd them, no sense in adding
1425 * the overhead of a bitmap if we don't have to.
1427 if (info->bytes <= block_group->sectorsize * 4) {
1428 if (ctl->free_extents * 2 <= ctl->extents_thresh)
1436 * some block groups are so tiny they can't be enveloped by a bitmap, so
1437 * don't even bother to create a bitmap for this
1439 if (BITS_PER_BITMAP * block_group->sectorsize >
1440 block_group->key.offset)
1446 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1447 struct btrfs_free_space *info)
1449 struct btrfs_free_space *bitmap_info;
1451 u64 bytes, offset, end;
1454 bytes = info->bytes;
1455 offset = info->offset;
1457 if (!ctl->op->use_bitmap(ctl, info))
1461 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1468 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1470 if (offset >= bitmap_info->offset && offset + bytes > end) {
1471 bitmap_set_bits(ctl, bitmap_info, offset, end - offset);
1472 bytes -= end - offset;
1475 } else if (offset >= bitmap_info->offset && offset + bytes <= end) {
1476 bitmap_set_bits(ctl, bitmap_info, offset, bytes);
1489 if (info && info->bitmap) {
1490 add_new_bitmap(ctl, info, offset);
1495 spin_unlock(&ctl->tree_lock);
1497 /* no pre-allocated info, allocate a new one */
1499 info = kmem_cache_zalloc(btrfs_free_space_cachep,
1502 spin_lock(&ctl->tree_lock);
1508 /* allocate the bitmap */
1509 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1510 spin_lock(&ctl->tree_lock);
1511 if (!info->bitmap) {
1521 kfree(info->bitmap);
1522 kmem_cache_free(btrfs_free_space_cachep, info);
1528 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1529 struct btrfs_free_space *info, bool update_stat)
1531 struct btrfs_free_space *left_info;
1532 struct btrfs_free_space *right_info;
1533 bool merged = false;
1534 u64 offset = info->offset;
1535 u64 bytes = info->bytes;
1538 * first we want to see if there is free space adjacent to the range we
1539 * are adding, if there is remove that struct and add a new one to
1540 * cover the entire range
1542 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1543 if (right_info && rb_prev(&right_info->offset_index))
1544 left_info = rb_entry(rb_prev(&right_info->offset_index),
1545 struct btrfs_free_space, offset_index);
1547 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1549 if (right_info && !right_info->bitmap) {
1551 unlink_free_space(ctl, right_info);
1553 __unlink_free_space(ctl, right_info);
1554 info->bytes += right_info->bytes;
1555 kmem_cache_free(btrfs_free_space_cachep, right_info);
1559 if (left_info && !left_info->bitmap &&
1560 left_info->offset + left_info->bytes == offset) {
1562 unlink_free_space(ctl, left_info);
1564 __unlink_free_space(ctl, left_info);
1565 info->offset = left_info->offset;
1566 info->bytes += left_info->bytes;
1567 kmem_cache_free(btrfs_free_space_cachep, left_info);
1574 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1575 u64 offset, u64 bytes)
1577 struct btrfs_free_space *info;
1580 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1584 info->offset = offset;
1585 info->bytes = bytes;
1587 spin_lock(&ctl->tree_lock);
1589 if (try_merge_free_space(ctl, info, true))
1593 * There was no extent directly to the left or right of this new
1594 * extent then we know we're going to have to allocate a new extent, so
1595 * before we do that see if we need to drop this into a bitmap
1597 ret = insert_into_bitmap(ctl, info);
1605 ret = link_free_space(ctl, info);
1607 kmem_cache_free(btrfs_free_space_cachep, info);
1609 spin_unlock(&ctl->tree_lock);
1612 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1613 BUG_ON(ret == -EEXIST);
1619 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1620 u64 offset, u64 bytes)
1622 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1623 struct btrfs_free_space *info;
1624 struct btrfs_free_space *next_info = NULL;
1627 spin_lock(&ctl->tree_lock);
1630 info = tree_search_offset(ctl, offset, 0, 0);
1633 * oops didn't find an extent that matched the space we wanted
1634 * to remove, look for a bitmap instead
1636 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1644 if (info->bytes < bytes && rb_next(&info->offset_index)) {
1646 next_info = rb_entry(rb_next(&info->offset_index),
1647 struct btrfs_free_space,
1650 if (next_info->bitmap)
1651 end = next_info->offset +
1652 BITS_PER_BITMAP * ctl->unit - 1;
1654 end = next_info->offset + next_info->bytes;
1656 if (next_info->bytes < bytes ||
1657 next_info->offset > offset || offset > end) {
1658 printk(KERN_CRIT "Found free space at %llu, size %llu,"
1659 " trying to use %llu\n",
1660 (unsigned long long)info->offset,
1661 (unsigned long long)info->bytes,
1662 (unsigned long long)bytes);
1671 if (info->bytes == bytes) {
1672 unlink_free_space(ctl, info);
1674 kfree(info->bitmap);
1675 ctl->total_bitmaps--;
1677 kmem_cache_free(btrfs_free_space_cachep, info);
1681 if (!info->bitmap && info->offset == offset) {
1682 unlink_free_space(ctl, info);
1683 info->offset += bytes;
1684 info->bytes -= bytes;
1685 link_free_space(ctl, info);
1689 if (!info->bitmap && info->offset <= offset &&
1690 info->offset + info->bytes >= offset + bytes) {
1691 u64 old_start = info->offset;
1693 * we're freeing space in the middle of the info,
1694 * this can happen during tree log replay
1696 * first unlink the old info and then
1697 * insert it again after the hole we're creating
1699 unlink_free_space(ctl, info);
1700 if (offset + bytes < info->offset + info->bytes) {
1701 u64 old_end = info->offset + info->bytes;
1703 info->offset = offset + bytes;
1704 info->bytes = old_end - info->offset;
1705 ret = link_free_space(ctl, info);
1710 /* the hole we're creating ends at the end
1711 * of the info struct, just free the info
1713 kmem_cache_free(btrfs_free_space_cachep, info);
1715 spin_unlock(&ctl->tree_lock);
1717 /* step two, insert a new info struct to cover
1718 * anything before the hole
1720 ret = btrfs_add_free_space(block_group, old_start,
1721 offset - old_start);
1726 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1731 spin_unlock(&ctl->tree_lock);
1736 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1739 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1740 struct btrfs_free_space *info;
1744 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1745 info = rb_entry(n, struct btrfs_free_space, offset_index);
1746 if (info->bytes >= bytes)
1748 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1749 (unsigned long long)info->offset,
1750 (unsigned long long)info->bytes,
1751 (info->bitmap) ? "yes" : "no");
1753 printk(KERN_INFO "block group has cluster?: %s\n",
1754 list_empty(&block_group->cluster_list) ? "no" : "yes");
1755 printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1759 static struct btrfs_free_space_op free_space_op = {
1760 .recalc_thresholds = recalculate_thresholds,
1761 .use_bitmap = use_bitmap,
1764 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
1766 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1768 spin_lock_init(&ctl->tree_lock);
1769 ctl->unit = block_group->sectorsize;
1770 ctl->start = block_group->key.objectid;
1771 ctl->private = block_group;
1772 ctl->op = &free_space_op;
1775 * we only want to have 32k of ram per block group for keeping
1776 * track of free space, and if we pass 1/2 of that we want to
1777 * start converting things over to using bitmaps
1779 ctl->extents_thresh = ((1024 * 32) / 2) /
1780 sizeof(struct btrfs_free_space);
1784 * for a given cluster, put all of its extents back into the free
1785 * space cache. If the block group passed doesn't match the block group
1786 * pointed to by the cluster, someone else raced in and freed the
1787 * cluster already. In that case, we just return without changing anything
1790 __btrfs_return_cluster_to_free_space(
1791 struct btrfs_block_group_cache *block_group,
1792 struct btrfs_free_cluster *cluster)
1794 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1795 struct btrfs_free_space *entry;
1796 struct rb_node *node;
1798 spin_lock(&cluster->lock);
1799 if (cluster->block_group != block_group)
1802 cluster->block_group = NULL;
1803 cluster->window_start = 0;
1804 list_del_init(&cluster->block_group_list);
1806 node = rb_first(&cluster->root);
1810 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1811 node = rb_next(&entry->offset_index);
1812 rb_erase(&entry->offset_index, &cluster->root);
1814 bitmap = (entry->bitmap != NULL);
1816 try_merge_free_space(ctl, entry, false);
1817 tree_insert_offset(&ctl->free_space_offset,
1818 entry->offset, &entry->offset_index, bitmap);
1820 cluster->root = RB_ROOT;
1823 spin_unlock(&cluster->lock);
1824 btrfs_put_block_group(block_group);
1828 void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
1830 struct btrfs_free_space *info;
1831 struct rb_node *node;
1833 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
1834 info = rb_entry(node, struct btrfs_free_space, offset_index);
1835 unlink_free_space(ctl, info);
1836 kfree(info->bitmap);
1837 kmem_cache_free(btrfs_free_space_cachep, info);
1838 if (need_resched()) {
1839 spin_unlock(&ctl->tree_lock);
1841 spin_lock(&ctl->tree_lock);
1846 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
1848 spin_lock(&ctl->tree_lock);
1849 __btrfs_remove_free_space_cache_locked(ctl);
1850 spin_unlock(&ctl->tree_lock);
1853 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1855 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1856 struct btrfs_free_cluster *cluster;
1857 struct list_head *head;
1859 spin_lock(&ctl->tree_lock);
1860 while ((head = block_group->cluster_list.next) !=
1861 &block_group->cluster_list) {
1862 cluster = list_entry(head, struct btrfs_free_cluster,
1865 WARN_ON(cluster->block_group != block_group);
1866 __btrfs_return_cluster_to_free_space(block_group, cluster);
1867 if (need_resched()) {
1868 spin_unlock(&ctl->tree_lock);
1870 spin_lock(&ctl->tree_lock);
1873 __btrfs_remove_free_space_cache_locked(ctl);
1874 spin_unlock(&ctl->tree_lock);
1878 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1879 u64 offset, u64 bytes, u64 empty_size)
1881 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1882 struct btrfs_free_space *entry = NULL;
1883 u64 bytes_search = bytes + empty_size;
1886 spin_lock(&ctl->tree_lock);
1887 entry = find_free_space(ctl, &offset, &bytes_search);
1892 if (entry->bitmap) {
1893 bitmap_clear_bits(ctl, entry, offset, bytes);
1895 free_bitmap(ctl, entry);
1897 unlink_free_space(ctl, entry);
1898 entry->offset += bytes;
1899 entry->bytes -= bytes;
1901 kmem_cache_free(btrfs_free_space_cachep, entry);
1903 link_free_space(ctl, entry);
1907 spin_unlock(&ctl->tree_lock);
1913 * given a cluster, put all of its extents back into the free space
1914 * cache. If a block group is passed, this function will only free
1915 * a cluster that belongs to the passed block group.
1917 * Otherwise, it'll get a reference on the block group pointed to by the
1918 * cluster and remove the cluster from it.
1920 int btrfs_return_cluster_to_free_space(
1921 struct btrfs_block_group_cache *block_group,
1922 struct btrfs_free_cluster *cluster)
1924 struct btrfs_free_space_ctl *ctl;
1927 /* first, get a safe pointer to the block group */
1928 spin_lock(&cluster->lock);
1930 block_group = cluster->block_group;
1932 spin_unlock(&cluster->lock);
1935 } else if (cluster->block_group != block_group) {
1936 /* someone else has already freed it don't redo their work */
1937 spin_unlock(&cluster->lock);
1940 atomic_inc(&block_group->count);
1941 spin_unlock(&cluster->lock);
1943 ctl = block_group->free_space_ctl;
1945 /* now return any extents the cluster had on it */
1946 spin_lock(&ctl->tree_lock);
1947 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1948 spin_unlock(&ctl->tree_lock);
1950 /* finally drop our ref */
1951 btrfs_put_block_group(block_group);
1955 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1956 struct btrfs_free_cluster *cluster,
1957 struct btrfs_free_space *entry,
1958 u64 bytes, u64 min_start)
1960 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1962 u64 search_start = cluster->window_start;
1963 u64 search_bytes = bytes;
1966 search_start = min_start;
1967 search_bytes = bytes;
1969 err = search_bitmap(ctl, entry, &search_start, &search_bytes);
1974 bitmap_clear_bits(ctl, entry, ret, bytes);
1980 * given a cluster, try to allocate 'bytes' from it, returns 0
1981 * if it couldn't find anything suitably large, or a logical disk offset
1982 * if things worked out
1984 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
1985 struct btrfs_free_cluster *cluster, u64 bytes,
1988 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1989 struct btrfs_free_space *entry = NULL;
1990 struct rb_node *node;
1993 spin_lock(&cluster->lock);
1994 if (bytes > cluster->max_size)
1997 if (cluster->block_group != block_group)
2000 node = rb_first(&cluster->root);
2004 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2006 if (entry->bytes < bytes ||
2007 (!entry->bitmap && entry->offset < min_start)) {
2008 node = rb_next(&entry->offset_index);
2011 entry = rb_entry(node, struct btrfs_free_space,
2016 if (entry->bitmap) {
2017 ret = btrfs_alloc_from_bitmap(block_group,
2018 cluster, entry, bytes,
2021 node = rb_next(&entry->offset_index);
2024 entry = rb_entry(node, struct btrfs_free_space,
2030 ret = entry->offset;
2032 entry->offset += bytes;
2033 entry->bytes -= bytes;
2036 if (entry->bytes == 0)
2037 rb_erase(&entry->offset_index, &cluster->root);
2041 spin_unlock(&cluster->lock);
2046 spin_lock(&ctl->tree_lock);
2048 ctl->free_space -= bytes;
2049 if (entry->bytes == 0) {
2050 ctl->free_extents--;
2051 if (entry->bitmap) {
2052 kfree(entry->bitmap);
2053 ctl->total_bitmaps--;
2054 ctl->op->recalc_thresholds(ctl);
2056 kmem_cache_free(btrfs_free_space_cachep, entry);
2059 spin_unlock(&ctl->tree_lock);
2064 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2065 struct btrfs_free_space *entry,
2066 struct btrfs_free_cluster *cluster,
2067 u64 offset, u64 bytes, u64 min_bytes)
2069 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2070 unsigned long next_zero;
2072 unsigned long search_bits;
2073 unsigned long total_bits;
2074 unsigned long found_bits;
2075 unsigned long start = 0;
2076 unsigned long total_found = 0;
2080 i = offset_to_bit(entry->offset, block_group->sectorsize,
2081 max_t(u64, offset, entry->offset));
2082 search_bits = bytes_to_bits(bytes, block_group->sectorsize);
2083 total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
2087 for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
2088 i < BITS_PER_BITMAP;
2089 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
2090 next_zero = find_next_zero_bit(entry->bitmap,
2091 BITS_PER_BITMAP, i);
2092 if (next_zero - i >= search_bits) {
2093 found_bits = next_zero - i;
2107 total_found += found_bits;
2109 if (cluster->max_size < found_bits * block_group->sectorsize)
2110 cluster->max_size = found_bits * block_group->sectorsize;
2112 if (total_found < total_bits) {
2113 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2114 if (i - start > total_bits * 2) {
2116 cluster->max_size = 0;
2122 cluster->window_start = start * block_group->sectorsize +
2124 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2125 ret = tree_insert_offset(&cluster->root, entry->offset,
2126 &entry->offset_index, 1);
2133 * This searches the block group for just extents to fill the cluster with.
2135 static int setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2136 struct btrfs_free_cluster *cluster,
2137 u64 offset, u64 bytes, u64 min_bytes)
2139 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2140 struct btrfs_free_space *first = NULL;
2141 struct btrfs_free_space *entry = NULL;
2142 struct btrfs_free_space *prev = NULL;
2143 struct btrfs_free_space *last;
2144 struct rb_node *node;
2148 u64 max_gap = 128 * 1024;
2150 entry = tree_search_offset(ctl, offset, 0, 1);
2155 * We don't want bitmaps, so just move along until we find a normal
2158 while (entry->bitmap) {
2159 node = rb_next(&entry->offset_index);
2162 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2165 window_start = entry->offset;
2166 window_free = entry->bytes;
2167 max_extent = entry->bytes;
2172 while (window_free <= min_bytes) {
2173 node = rb_next(&entry->offset_index);
2176 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2181 * we haven't filled the empty size and the window is
2182 * very large. reset and try again
2184 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2185 entry->offset - window_start > (min_bytes * 2)) {
2187 window_start = entry->offset;
2188 window_free = entry->bytes;
2190 max_extent = entry->bytes;
2193 window_free += entry->bytes;
2194 if (entry->bytes > max_extent)
2195 max_extent = entry->bytes;
2200 cluster->window_start = first->offset;
2202 node = &first->offset_index;
2205 * now we've found our entries, pull them out of the free space
2206 * cache and put them into the cluster rbtree
2211 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2212 node = rb_next(&entry->offset_index);
2216 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2217 ret = tree_insert_offset(&cluster->root, entry->offset,
2218 &entry->offset_index, 0);
2220 } while (node && entry != last);
2222 cluster->max_size = max_extent;
2228 * This specifically looks for bitmaps that may work in the cluster, we assume
2229 * that we have already failed to find extents that will work.
2231 static int setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2232 struct btrfs_free_cluster *cluster,
2233 u64 offset, u64 bytes, u64 min_bytes)
2235 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2236 struct btrfs_free_space *entry;
2237 struct rb_node *node;
2240 if (ctl->total_bitmaps == 0)
2243 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, offset), 0, 1);
2247 node = &entry->offset_index;
2249 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2250 node = rb_next(&entry->offset_index);
2253 if (entry->bytes < min_bytes)
2255 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2257 } while (ret && node);
2263 * here we try to find a cluster of blocks in a block group. The goal
2264 * is to find at least bytes free and up to empty_size + bytes free.
2265 * We might not find them all in one contiguous area.
2267 * returns zero and sets up cluster if things worked out, otherwise
2268 * it returns -enospc
2270 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2271 struct btrfs_root *root,
2272 struct btrfs_block_group_cache *block_group,
2273 struct btrfs_free_cluster *cluster,
2274 u64 offset, u64 bytes, u64 empty_size)
2276 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2280 /* for metadata, allow allocates with more holes */
2281 if (btrfs_test_opt(root, SSD_SPREAD)) {
2282 min_bytes = bytes + empty_size;
2283 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2285 * we want to do larger allocations when we are
2286 * flushing out the delayed refs, it helps prevent
2287 * making more work as we go along.
2289 if (trans->transaction->delayed_refs.flushing)
2290 min_bytes = max(bytes, (bytes + empty_size) >> 1);
2292 min_bytes = max(bytes, (bytes + empty_size) >> 4);
2294 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2296 spin_lock(&ctl->tree_lock);
2299 * If we know we don't have enough space to make a cluster don't even
2300 * bother doing all the work to try and find one.
2302 if (ctl->free_space < min_bytes) {
2303 spin_unlock(&ctl->tree_lock);
2307 spin_lock(&cluster->lock);
2309 /* someone already found a cluster, hooray */
2310 if (cluster->block_group) {
2315 ret = setup_cluster_no_bitmap(block_group, cluster, offset, bytes,
2318 ret = setup_cluster_bitmap(block_group, cluster, offset,
2322 atomic_inc(&block_group->count);
2323 list_add_tail(&cluster->block_group_list,
2324 &block_group->cluster_list);
2325 cluster->block_group = block_group;
2328 spin_unlock(&cluster->lock);
2329 spin_unlock(&ctl->tree_lock);
2335 * simple code to zero out a cluster
2337 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2339 spin_lock_init(&cluster->lock);
2340 spin_lock_init(&cluster->refill_lock);
2341 cluster->root = RB_ROOT;
2342 cluster->max_size = 0;
2343 INIT_LIST_HEAD(&cluster->block_group_list);
2344 cluster->block_group = NULL;
2347 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2348 u64 *trimmed, u64 start, u64 end, u64 minlen)
2350 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2351 struct btrfs_free_space *entry = NULL;
2352 struct btrfs_fs_info *fs_info = block_group->fs_info;
2354 u64 actually_trimmed;
2359 while (start < end) {
2360 spin_lock(&ctl->tree_lock);
2362 if (ctl->free_space < minlen) {
2363 spin_unlock(&ctl->tree_lock);
2367 entry = tree_search_offset(ctl, start, 0, 1);
2369 entry = tree_search_offset(ctl,
2370 offset_to_bitmap(ctl, start),
2373 if (!entry || entry->offset >= end) {
2374 spin_unlock(&ctl->tree_lock);
2378 if (entry->bitmap) {
2379 ret = search_bitmap(ctl, entry, &start, &bytes);
2382 spin_unlock(&ctl->tree_lock);
2385 bytes = min(bytes, end - start);
2386 bitmap_clear_bits(ctl, entry, start, bytes);
2387 if (entry->bytes == 0)
2388 free_bitmap(ctl, entry);
2390 start = entry->offset + BITS_PER_BITMAP *
2391 block_group->sectorsize;
2392 spin_unlock(&ctl->tree_lock);
2397 start = entry->offset;
2398 bytes = min(entry->bytes, end - start);
2399 unlink_free_space(ctl, entry);
2400 kmem_cache_free(btrfs_free_space_cachep, entry);
2403 spin_unlock(&ctl->tree_lock);
2405 if (bytes >= minlen) {
2407 update_ret = btrfs_update_reserved_bytes(block_group,
2410 ret = btrfs_error_discard_extent(fs_info->extent_root,
2415 btrfs_add_free_space(block_group, start, bytes);
2417 btrfs_update_reserved_bytes(block_group,
2422 *trimmed += actually_trimmed;
2427 if (fatal_signal_pending(current)) {
2439 * Find the left-most item in the cache tree, and then return the
2440 * smallest inode number in the item.
2442 * Note: the returned inode number may not be the smallest one in
2443 * the tree, if the left-most item is a bitmap.
2445 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2447 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2448 struct btrfs_free_space *entry = NULL;
2451 spin_lock(&ctl->tree_lock);
2453 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2456 entry = rb_entry(rb_first(&ctl->free_space_offset),
2457 struct btrfs_free_space, offset_index);
2459 if (!entry->bitmap) {
2460 ino = entry->offset;
2462 unlink_free_space(ctl, entry);
2466 kmem_cache_free(btrfs_free_space_cachep, entry);
2468 link_free_space(ctl, entry);
2474 ret = search_bitmap(ctl, entry, &offset, &count);
2478 bitmap_clear_bits(ctl, entry, offset, 1);
2479 if (entry->bytes == 0)
2480 free_bitmap(ctl, entry);
2483 spin_unlock(&ctl->tree_lock);
2488 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2489 struct btrfs_path *path)
2491 struct inode *inode = NULL;
2493 spin_lock(&root->cache_lock);
2494 if (root->cache_inode)
2495 inode = igrab(root->cache_inode);
2496 spin_unlock(&root->cache_lock);
2500 inode = __lookup_free_space_inode(root, path, 0);
2504 spin_lock(&root->cache_lock);
2505 if (!root->fs_info->closing)
2506 root->cache_inode = igrab(inode);
2507 spin_unlock(&root->cache_lock);
2512 int create_free_ino_inode(struct btrfs_root *root,
2513 struct btrfs_trans_handle *trans,
2514 struct btrfs_path *path)
2516 return __create_free_space_inode(root, trans, path,
2517 BTRFS_FREE_INO_OBJECTID, 0);
2520 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2522 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2523 struct btrfs_path *path;
2524 struct inode *inode;
2526 u64 root_gen = btrfs_root_generation(&root->root_item);
2529 * If we're unmounting then just return, since this does a search on the
2530 * normal root and not the commit root and we could deadlock.
2533 if (fs_info->closing)
2536 path = btrfs_alloc_path();
2540 inode = lookup_free_ino_inode(root, path);
2544 if (root_gen != BTRFS_I(inode)->generation)
2547 ret = __load_free_space_cache(root, inode, ctl, path, 0);
2550 printk(KERN_ERR "btrfs: failed to load free ino cache for "
2551 "root %llu\n", root->root_key.objectid);
2555 btrfs_free_path(path);
2559 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2560 struct btrfs_trans_handle *trans,
2561 struct btrfs_path *path)
2563 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2564 struct inode *inode;
2567 inode = lookup_free_ino_inode(root, path);
2571 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2573 printk(KERN_ERR "btrfs: failed to write free ino cache "
2574 "for root %llu\n", root->root_key.objectid);