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>
23 #include <linux/ratelimit.h>
25 #include "free-space-cache.h"
26 #include "transaction.h"
28 #include "extent_io.h"
29 #include "inode-map.h"
32 #define BITS_PER_BITMAP (PAGE_CACHE_SIZE * 8)
33 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
35 struct btrfs_trim_range {
38 struct list_head list;
41 static int link_free_space(struct btrfs_free_space_ctl *ctl,
42 struct btrfs_free_space *info);
43 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
44 struct btrfs_free_space *info);
46 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
47 struct btrfs_path *path,
51 struct btrfs_key location;
52 struct btrfs_disk_key disk_key;
53 struct btrfs_free_space_header *header;
54 struct extent_buffer *leaf;
55 struct inode *inode = NULL;
58 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
62 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
66 btrfs_release_path(path);
67 return ERR_PTR(-ENOENT);
70 leaf = path->nodes[0];
71 header = btrfs_item_ptr(leaf, path->slots[0],
72 struct btrfs_free_space_header);
73 btrfs_free_space_key(leaf, header, &disk_key);
74 btrfs_disk_key_to_cpu(&location, &disk_key);
75 btrfs_release_path(path);
77 inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
79 return ERR_PTR(-ENOENT);
82 if (is_bad_inode(inode)) {
84 return ERR_PTR(-ENOENT);
87 mapping_set_gfp_mask(inode->i_mapping,
88 mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS);
93 struct inode *lookup_free_space_inode(struct btrfs_root *root,
94 struct btrfs_block_group_cache
95 *block_group, struct btrfs_path *path)
97 struct inode *inode = NULL;
98 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
100 spin_lock(&block_group->lock);
101 if (block_group->inode)
102 inode = igrab(block_group->inode);
103 spin_unlock(&block_group->lock);
107 inode = __lookup_free_space_inode(root, path,
108 block_group->key.objectid);
112 spin_lock(&block_group->lock);
113 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
114 btrfs_info(root->fs_info,
115 "Old style space inode found, converting.");
116 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
117 BTRFS_INODE_NODATACOW;
118 block_group->disk_cache_state = BTRFS_DC_CLEAR;
121 if (!block_group->iref) {
122 block_group->inode = igrab(inode);
123 block_group->iref = 1;
125 spin_unlock(&block_group->lock);
130 static int __create_free_space_inode(struct btrfs_root *root,
131 struct btrfs_trans_handle *trans,
132 struct btrfs_path *path,
135 struct btrfs_key key;
136 struct btrfs_disk_key disk_key;
137 struct btrfs_free_space_header *header;
138 struct btrfs_inode_item *inode_item;
139 struct extent_buffer *leaf;
140 u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
143 ret = btrfs_insert_empty_inode(trans, root, path, ino);
147 /* We inline crc's for the free disk space cache */
148 if (ino != BTRFS_FREE_INO_OBJECTID)
149 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
151 leaf = path->nodes[0];
152 inode_item = btrfs_item_ptr(leaf, path->slots[0],
153 struct btrfs_inode_item);
154 btrfs_item_key(leaf, &disk_key, path->slots[0]);
155 memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
156 sizeof(*inode_item));
157 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
158 btrfs_set_inode_size(leaf, inode_item, 0);
159 btrfs_set_inode_nbytes(leaf, inode_item, 0);
160 btrfs_set_inode_uid(leaf, inode_item, 0);
161 btrfs_set_inode_gid(leaf, inode_item, 0);
162 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
163 btrfs_set_inode_flags(leaf, inode_item, flags);
164 btrfs_set_inode_nlink(leaf, inode_item, 1);
165 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
166 btrfs_set_inode_block_group(leaf, inode_item, offset);
167 btrfs_mark_buffer_dirty(leaf);
168 btrfs_release_path(path);
170 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
174 ret = btrfs_insert_empty_item(trans, root, path, &key,
175 sizeof(struct btrfs_free_space_header));
177 btrfs_release_path(path);
180 leaf = path->nodes[0];
181 header = btrfs_item_ptr(leaf, path->slots[0],
182 struct btrfs_free_space_header);
183 memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
184 btrfs_set_free_space_key(leaf, header, &disk_key);
185 btrfs_mark_buffer_dirty(leaf);
186 btrfs_release_path(path);
191 int create_free_space_inode(struct btrfs_root *root,
192 struct btrfs_trans_handle *trans,
193 struct btrfs_block_group_cache *block_group,
194 struct btrfs_path *path)
199 ret = btrfs_find_free_objectid(root, &ino);
203 return __create_free_space_inode(root, trans, path, ino,
204 block_group->key.objectid);
207 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
208 struct btrfs_block_rsv *rsv)
213 /* 1 for slack space, 1 for updating the inode */
214 needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
215 btrfs_calc_trans_metadata_size(root, 1);
217 spin_lock(&rsv->lock);
218 if (rsv->reserved < needed_bytes)
222 spin_unlock(&rsv->lock);
226 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
227 struct btrfs_trans_handle *trans,
232 btrfs_i_size_write(inode, 0);
233 truncate_pagecache(inode, 0);
236 * We don't need an orphan item because truncating the free space cache
237 * will never be split across transactions.
239 ret = btrfs_truncate_inode_items(trans, root, inode,
240 0, BTRFS_EXTENT_DATA_KEY);
242 btrfs_abort_transaction(trans, root, ret);
246 ret = btrfs_update_inode(trans, root, inode);
248 btrfs_abort_transaction(trans, root, ret);
253 static int readahead_cache(struct inode *inode)
255 struct file_ra_state *ra;
256 unsigned long last_index;
258 ra = kzalloc(sizeof(*ra), GFP_NOFS);
262 file_ra_state_init(ra, inode->i_mapping);
263 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
265 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
276 struct btrfs_root *root;
280 unsigned check_crcs:1;
283 static int io_ctl_init(struct io_ctl *io_ctl, struct inode *inode,
284 struct btrfs_root *root, int write)
289 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_CACHE_SIZE);
291 if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
294 /* Make sure we can fit our crcs into the first page */
295 if (write && check_crcs &&
296 (num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE)
299 memset(io_ctl, 0, sizeof(struct io_ctl));
301 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
305 io_ctl->num_pages = num_pages;
307 io_ctl->check_crcs = check_crcs;
312 static void io_ctl_free(struct io_ctl *io_ctl)
314 kfree(io_ctl->pages);
317 static void io_ctl_unmap_page(struct io_ctl *io_ctl)
320 kunmap(io_ctl->page);
326 static void io_ctl_map_page(struct io_ctl *io_ctl, int clear)
328 ASSERT(io_ctl->index < io_ctl->num_pages);
329 io_ctl->page = io_ctl->pages[io_ctl->index++];
330 io_ctl->cur = kmap(io_ctl->page);
331 io_ctl->orig = io_ctl->cur;
332 io_ctl->size = PAGE_CACHE_SIZE;
334 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
337 static void io_ctl_drop_pages(struct io_ctl *io_ctl)
341 io_ctl_unmap_page(io_ctl);
343 for (i = 0; i < io_ctl->num_pages; i++) {
344 if (io_ctl->pages[i]) {
345 ClearPageChecked(io_ctl->pages[i]);
346 unlock_page(io_ctl->pages[i]);
347 page_cache_release(io_ctl->pages[i]);
352 static int io_ctl_prepare_pages(struct io_ctl *io_ctl, struct inode *inode,
356 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
359 for (i = 0; i < io_ctl->num_pages; i++) {
360 page = find_or_create_page(inode->i_mapping, i, mask);
362 io_ctl_drop_pages(io_ctl);
365 io_ctl->pages[i] = page;
366 if (uptodate && !PageUptodate(page)) {
367 btrfs_readpage(NULL, page);
369 if (!PageUptodate(page)) {
370 btrfs_err(BTRFS_I(inode)->root->fs_info,
371 "error reading free space cache");
372 io_ctl_drop_pages(io_ctl);
378 for (i = 0; i < io_ctl->num_pages; i++) {
379 clear_page_dirty_for_io(io_ctl->pages[i]);
380 set_page_extent_mapped(io_ctl->pages[i]);
386 static void io_ctl_set_generation(struct io_ctl *io_ctl, u64 generation)
390 io_ctl_map_page(io_ctl, 1);
393 * Skip the csum areas. If we don't check crcs then we just have a
394 * 64bit chunk at the front of the first page.
396 if (io_ctl->check_crcs) {
397 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
398 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
400 io_ctl->cur += sizeof(u64);
401 io_ctl->size -= sizeof(u64) * 2;
405 *val = cpu_to_le64(generation);
406 io_ctl->cur += sizeof(u64);
409 static int io_ctl_check_generation(struct io_ctl *io_ctl, u64 generation)
414 * Skip the crc area. If we don't check crcs then we just have a 64bit
415 * chunk at the front of the first page.
417 if (io_ctl->check_crcs) {
418 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
419 io_ctl->size -= sizeof(u64) +
420 (sizeof(u32) * io_ctl->num_pages);
422 io_ctl->cur += sizeof(u64);
423 io_ctl->size -= sizeof(u64) * 2;
427 if (le64_to_cpu(*gen) != generation) {
428 printk_ratelimited(KERN_ERR "BTRFS: space cache generation "
429 "(%Lu) does not match inode (%Lu)\n", *gen,
431 io_ctl_unmap_page(io_ctl);
434 io_ctl->cur += sizeof(u64);
438 static void io_ctl_set_crc(struct io_ctl *io_ctl, int index)
444 if (!io_ctl->check_crcs) {
445 io_ctl_unmap_page(io_ctl);
450 offset = sizeof(u32) * io_ctl->num_pages;
452 crc = btrfs_csum_data(io_ctl->orig + offset, crc,
453 PAGE_CACHE_SIZE - offset);
454 btrfs_csum_final(crc, (char *)&crc);
455 io_ctl_unmap_page(io_ctl);
456 tmp = kmap(io_ctl->pages[0]);
459 kunmap(io_ctl->pages[0]);
462 static int io_ctl_check_crc(struct io_ctl *io_ctl, int index)
468 if (!io_ctl->check_crcs) {
469 io_ctl_map_page(io_ctl, 0);
474 offset = sizeof(u32) * io_ctl->num_pages;
476 tmp = kmap(io_ctl->pages[0]);
479 kunmap(io_ctl->pages[0]);
481 io_ctl_map_page(io_ctl, 0);
482 crc = btrfs_csum_data(io_ctl->orig + offset, crc,
483 PAGE_CACHE_SIZE - offset);
484 btrfs_csum_final(crc, (char *)&crc);
486 printk_ratelimited(KERN_ERR "BTRFS: csum mismatch on free "
488 io_ctl_unmap_page(io_ctl);
495 static int io_ctl_add_entry(struct io_ctl *io_ctl, u64 offset, u64 bytes,
498 struct btrfs_free_space_entry *entry;
504 entry->offset = cpu_to_le64(offset);
505 entry->bytes = cpu_to_le64(bytes);
506 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
507 BTRFS_FREE_SPACE_EXTENT;
508 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
509 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
511 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
514 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
516 /* No more pages to map */
517 if (io_ctl->index >= io_ctl->num_pages)
520 /* map the next page */
521 io_ctl_map_page(io_ctl, 1);
525 static int io_ctl_add_bitmap(struct io_ctl *io_ctl, void *bitmap)
531 * If we aren't at the start of the current page, unmap this one and
532 * map the next one if there is any left.
534 if (io_ctl->cur != io_ctl->orig) {
535 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
536 if (io_ctl->index >= io_ctl->num_pages)
538 io_ctl_map_page(io_ctl, 0);
541 memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
542 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
543 if (io_ctl->index < io_ctl->num_pages)
544 io_ctl_map_page(io_ctl, 0);
548 static void io_ctl_zero_remaining_pages(struct io_ctl *io_ctl)
551 * If we're not on the boundary we know we've modified the page and we
552 * need to crc the page.
554 if (io_ctl->cur != io_ctl->orig)
555 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
557 io_ctl_unmap_page(io_ctl);
559 while (io_ctl->index < io_ctl->num_pages) {
560 io_ctl_map_page(io_ctl, 1);
561 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
565 static int io_ctl_read_entry(struct io_ctl *io_ctl,
566 struct btrfs_free_space *entry, u8 *type)
568 struct btrfs_free_space_entry *e;
572 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
578 entry->offset = le64_to_cpu(e->offset);
579 entry->bytes = le64_to_cpu(e->bytes);
581 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
582 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
584 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
587 io_ctl_unmap_page(io_ctl);
592 static int io_ctl_read_bitmap(struct io_ctl *io_ctl,
593 struct btrfs_free_space *entry)
597 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
601 memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
602 io_ctl_unmap_page(io_ctl);
608 * Since we attach pinned extents after the fact we can have contiguous sections
609 * of free space that are split up in entries. This poses a problem with the
610 * tree logging stuff since it could have allocated across what appears to be 2
611 * entries since we would have merged the entries when adding the pinned extents
612 * back to the free space cache. So run through the space cache that we just
613 * loaded and merge contiguous entries. This will make the log replay stuff not
614 * blow up and it will make for nicer allocator behavior.
616 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
618 struct btrfs_free_space *e, *prev = NULL;
622 spin_lock(&ctl->tree_lock);
623 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
624 e = rb_entry(n, struct btrfs_free_space, offset_index);
627 if (e->bitmap || prev->bitmap)
629 if (prev->offset + prev->bytes == e->offset) {
630 unlink_free_space(ctl, prev);
631 unlink_free_space(ctl, e);
632 prev->bytes += e->bytes;
633 kmem_cache_free(btrfs_free_space_cachep, e);
634 link_free_space(ctl, prev);
636 spin_unlock(&ctl->tree_lock);
642 spin_unlock(&ctl->tree_lock);
645 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
646 struct btrfs_free_space_ctl *ctl,
647 struct btrfs_path *path, u64 offset)
649 struct btrfs_free_space_header *header;
650 struct extent_buffer *leaf;
651 struct io_ctl io_ctl;
652 struct btrfs_key key;
653 struct btrfs_free_space *e, *n;
661 /* Nothing in the space cache, goodbye */
662 if (!i_size_read(inode))
665 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
669 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
673 btrfs_release_path(path);
679 leaf = path->nodes[0];
680 header = btrfs_item_ptr(leaf, path->slots[0],
681 struct btrfs_free_space_header);
682 num_entries = btrfs_free_space_entries(leaf, header);
683 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
684 generation = btrfs_free_space_generation(leaf, header);
685 btrfs_release_path(path);
687 if (!BTRFS_I(inode)->generation) {
688 btrfs_info(root->fs_info,
689 "The free space cache file (%llu) is invalid. skip it\n",
694 if (BTRFS_I(inode)->generation != generation) {
695 btrfs_err(root->fs_info,
696 "free space inode generation (%llu) "
697 "did not match free space cache generation (%llu)",
698 BTRFS_I(inode)->generation, generation);
705 ret = io_ctl_init(&io_ctl, inode, root, 0);
709 ret = readahead_cache(inode);
713 ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
717 ret = io_ctl_check_crc(&io_ctl, 0);
721 ret = io_ctl_check_generation(&io_ctl, generation);
725 while (num_entries) {
726 e = kmem_cache_zalloc(btrfs_free_space_cachep,
731 ret = io_ctl_read_entry(&io_ctl, e, &type);
733 kmem_cache_free(btrfs_free_space_cachep, e);
738 kmem_cache_free(btrfs_free_space_cachep, e);
742 if (type == BTRFS_FREE_SPACE_EXTENT) {
743 spin_lock(&ctl->tree_lock);
744 ret = link_free_space(ctl, e);
745 spin_unlock(&ctl->tree_lock);
747 btrfs_err(root->fs_info,
748 "Duplicate entries in free space cache, dumping");
749 kmem_cache_free(btrfs_free_space_cachep, e);
755 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
758 btrfs_free_space_cachep, e);
761 spin_lock(&ctl->tree_lock);
762 ret = link_free_space(ctl, e);
763 ctl->total_bitmaps++;
764 ctl->op->recalc_thresholds(ctl);
765 spin_unlock(&ctl->tree_lock);
767 btrfs_err(root->fs_info,
768 "Duplicate entries in free space cache, dumping");
769 kmem_cache_free(btrfs_free_space_cachep, e);
772 list_add_tail(&e->list, &bitmaps);
778 io_ctl_unmap_page(&io_ctl);
781 * We add the bitmaps at the end of the entries in order that
782 * the bitmap entries are added to the cache.
784 list_for_each_entry_safe(e, n, &bitmaps, list) {
785 list_del_init(&e->list);
786 ret = io_ctl_read_bitmap(&io_ctl, e);
791 io_ctl_drop_pages(&io_ctl);
792 merge_space_tree(ctl);
795 io_ctl_free(&io_ctl);
798 io_ctl_drop_pages(&io_ctl);
799 __btrfs_remove_free_space_cache(ctl);
803 int load_free_space_cache(struct btrfs_fs_info *fs_info,
804 struct btrfs_block_group_cache *block_group)
806 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
807 struct btrfs_root *root = fs_info->tree_root;
809 struct btrfs_path *path;
812 u64 used = btrfs_block_group_used(&block_group->item);
815 * If this block group has been marked to be cleared for one reason or
816 * another then we can't trust the on disk cache, so just return.
818 spin_lock(&block_group->lock);
819 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
820 spin_unlock(&block_group->lock);
823 spin_unlock(&block_group->lock);
825 path = btrfs_alloc_path();
828 path->search_commit_root = 1;
829 path->skip_locking = 1;
831 inode = lookup_free_space_inode(root, block_group, path);
833 btrfs_free_path(path);
837 /* We may have converted the inode and made the cache invalid. */
838 spin_lock(&block_group->lock);
839 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
840 spin_unlock(&block_group->lock);
841 btrfs_free_path(path);
844 spin_unlock(&block_group->lock);
846 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
847 path, block_group->key.objectid);
848 btrfs_free_path(path);
852 spin_lock(&ctl->tree_lock);
853 matched = (ctl->free_space == (block_group->key.offset - used -
854 block_group->bytes_super));
855 spin_unlock(&ctl->tree_lock);
858 __btrfs_remove_free_space_cache(ctl);
859 btrfs_warn(fs_info, "block group %llu has wrong amount of free space",
860 block_group->key.objectid);
865 /* This cache is bogus, make sure it gets cleared */
866 spin_lock(&block_group->lock);
867 block_group->disk_cache_state = BTRFS_DC_CLEAR;
868 spin_unlock(&block_group->lock);
871 btrfs_warn(fs_info, "failed to load free space cache for block group %llu, rebuild it now",
872 block_group->key.objectid);
879 static noinline_for_stack
880 int write_cache_extent_entries(struct io_ctl *io_ctl,
881 struct btrfs_free_space_ctl *ctl,
882 struct btrfs_block_group_cache *block_group,
883 int *entries, int *bitmaps,
884 struct list_head *bitmap_list)
887 struct btrfs_free_cluster *cluster = NULL;
888 struct rb_node *node = rb_first(&ctl->free_space_offset);
889 struct btrfs_trim_range *trim_entry;
891 /* Get the cluster for this block_group if it exists */
892 if (block_group && !list_empty(&block_group->cluster_list)) {
893 cluster = list_entry(block_group->cluster_list.next,
894 struct btrfs_free_cluster,
898 if (!node && cluster) {
899 node = rb_first(&cluster->root);
903 /* Write out the extent entries */
905 struct btrfs_free_space *e;
907 e = rb_entry(node, struct btrfs_free_space, offset_index);
910 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
916 list_add_tail(&e->list, bitmap_list);
919 node = rb_next(node);
920 if (!node && cluster) {
921 node = rb_first(&cluster->root);
927 * Make sure we don't miss any range that was removed from our rbtree
928 * because trimming is running. Otherwise after a umount+mount (or crash
929 * after committing the transaction) we would leak free space and get
930 * an inconsistent free space cache report from fsck.
932 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
933 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
934 trim_entry->bytes, NULL);
945 static noinline_for_stack int
946 update_cache_item(struct btrfs_trans_handle *trans,
947 struct btrfs_root *root,
949 struct btrfs_path *path, u64 offset,
950 int entries, int bitmaps)
952 struct btrfs_key key;
953 struct btrfs_free_space_header *header;
954 struct extent_buffer *leaf;
957 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
961 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
963 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
964 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
968 leaf = path->nodes[0];
970 struct btrfs_key found_key;
971 ASSERT(path->slots[0]);
973 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
974 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
975 found_key.offset != offset) {
976 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
978 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
980 btrfs_release_path(path);
985 BTRFS_I(inode)->generation = trans->transid;
986 header = btrfs_item_ptr(leaf, path->slots[0],
987 struct btrfs_free_space_header);
988 btrfs_set_free_space_entries(leaf, header, entries);
989 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
990 btrfs_set_free_space_generation(leaf, header, trans->transid);
991 btrfs_mark_buffer_dirty(leaf);
992 btrfs_release_path(path);
1000 static noinline_for_stack int
1001 write_pinned_extent_entries(struct btrfs_root *root,
1002 struct btrfs_block_group_cache *block_group,
1003 struct io_ctl *io_ctl,
1006 u64 start, extent_start, extent_end, len;
1007 struct extent_io_tree *unpin = NULL;
1014 * We want to add any pinned extents to our free space cache
1015 * so we don't leak the space
1017 * We shouldn't have switched the pinned extents yet so this is the
1020 unpin = root->fs_info->pinned_extents;
1022 start = block_group->key.objectid;
1024 while (start < block_group->key.objectid + block_group->key.offset) {
1025 ret = find_first_extent_bit(unpin, start,
1026 &extent_start, &extent_end,
1027 EXTENT_DIRTY, NULL);
1031 /* This pinned extent is out of our range */
1032 if (extent_start >= block_group->key.objectid +
1033 block_group->key.offset)
1036 extent_start = max(extent_start, start);
1037 extent_end = min(block_group->key.objectid +
1038 block_group->key.offset, extent_end + 1);
1039 len = extent_end - extent_start;
1042 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1052 static noinline_for_stack int
1053 write_bitmap_entries(struct io_ctl *io_ctl, struct list_head *bitmap_list)
1055 struct list_head *pos, *n;
1058 /* Write out the bitmaps */
1059 list_for_each_safe(pos, n, bitmap_list) {
1060 struct btrfs_free_space *entry =
1061 list_entry(pos, struct btrfs_free_space, list);
1063 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1066 list_del_init(&entry->list);
1072 static int flush_dirty_cache(struct inode *inode)
1076 ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1078 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1079 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1085 static void noinline_for_stack
1086 cleanup_write_cache_enospc(struct inode *inode,
1087 struct io_ctl *io_ctl,
1088 struct extent_state **cached_state,
1089 struct list_head *bitmap_list)
1091 struct list_head *pos, *n;
1093 list_for_each_safe(pos, n, bitmap_list) {
1094 struct btrfs_free_space *entry =
1095 list_entry(pos, struct btrfs_free_space, list);
1096 list_del_init(&entry->list);
1098 io_ctl_drop_pages(io_ctl);
1099 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1100 i_size_read(inode) - 1, cached_state,
1105 * __btrfs_write_out_cache - write out cached info to an inode
1106 * @root - the root the inode belongs to
1107 * @ctl - the free space cache we are going to write out
1108 * @block_group - the block_group for this cache if it belongs to a block_group
1109 * @trans - the trans handle
1110 * @path - the path to use
1111 * @offset - the offset for the key we'll insert
1113 * This function writes out a free space cache struct to disk for quick recovery
1114 * on mount. This will return 0 if it was successfull in writing the cache out,
1115 * and -1 if it was not.
1117 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1118 struct btrfs_free_space_ctl *ctl,
1119 struct btrfs_block_group_cache *block_group,
1120 struct btrfs_trans_handle *trans,
1121 struct btrfs_path *path, u64 offset)
1123 struct extent_state *cached_state = NULL;
1124 struct io_ctl io_ctl;
1125 LIST_HEAD(bitmap_list);
1130 if (!i_size_read(inode))
1133 ret = io_ctl_init(&io_ctl, inode, root, 1);
1137 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1138 down_write(&block_group->data_rwsem);
1139 spin_lock(&block_group->lock);
1140 if (block_group->delalloc_bytes) {
1141 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1142 spin_unlock(&block_group->lock);
1143 up_write(&block_group->data_rwsem);
1144 BTRFS_I(inode)->generation = 0;
1148 spin_unlock(&block_group->lock);
1151 /* Lock all pages first so we can lock the extent safely. */
1152 io_ctl_prepare_pages(&io_ctl, inode, 0);
1154 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1157 io_ctl_set_generation(&io_ctl, trans->transid);
1159 mutex_lock(&ctl->cache_writeout_mutex);
1160 /* Write out the extent entries in the free space cache */
1161 ret = write_cache_extent_entries(&io_ctl, ctl,
1162 block_group, &entries, &bitmaps,
1165 mutex_unlock(&ctl->cache_writeout_mutex);
1170 * Some spaces that are freed in the current transaction are pinned,
1171 * they will be added into free space cache after the transaction is
1172 * committed, we shouldn't lose them.
1174 ret = write_pinned_extent_entries(root, block_group, &io_ctl, &entries);
1176 mutex_unlock(&ctl->cache_writeout_mutex);
1181 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1182 * locked while doing it because a concurrent trim can be manipulating
1183 * or freeing the bitmap.
1185 ret = write_bitmap_entries(&io_ctl, &bitmap_list);
1186 mutex_unlock(&ctl->cache_writeout_mutex);
1190 /* Zero out the rest of the pages just to make sure */
1191 io_ctl_zero_remaining_pages(&io_ctl);
1193 /* Everything is written out, now we dirty the pages in the file. */
1194 ret = btrfs_dirty_pages(root, inode, io_ctl.pages, io_ctl.num_pages,
1195 0, i_size_read(inode), &cached_state);
1199 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1200 up_write(&block_group->data_rwsem);
1202 * Release the pages and unlock the extent, we will flush
1205 io_ctl_drop_pages(&io_ctl);
1207 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1208 i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1210 /* Flush the dirty pages in the cache file. */
1211 ret = flush_dirty_cache(inode);
1215 /* Update the cache item to tell everyone this cache file is valid. */
1216 ret = update_cache_item(trans, root, inode, path, offset,
1219 io_ctl_free(&io_ctl);
1221 invalidate_inode_pages2(inode->i_mapping);
1222 BTRFS_I(inode)->generation = 0;
1224 btrfs_update_inode(trans, root, inode);
1228 cleanup_write_cache_enospc(inode, &io_ctl, &cached_state, &bitmap_list);
1230 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1231 up_write(&block_group->data_rwsem);
1236 int btrfs_write_out_cache(struct btrfs_root *root,
1237 struct btrfs_trans_handle *trans,
1238 struct btrfs_block_group_cache *block_group,
1239 struct btrfs_path *path)
1241 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1242 struct inode *inode;
1244 enum btrfs_disk_cache_state dcs = BTRFS_DC_WRITTEN;
1246 root = root->fs_info->tree_root;
1248 spin_lock(&block_group->lock);
1249 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1250 spin_unlock(&block_group->lock);
1254 if (block_group->delalloc_bytes) {
1255 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1256 spin_unlock(&block_group->lock);
1259 spin_unlock(&block_group->lock);
1261 inode = lookup_free_space_inode(root, block_group, path);
1265 ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
1266 path, block_group->key.objectid);
1268 dcs = BTRFS_DC_ERROR;
1271 btrfs_err(root->fs_info,
1272 "failed to write free space cache for block group %llu",
1273 block_group->key.objectid);
1277 spin_lock(&block_group->lock);
1278 block_group->disk_cache_state = dcs;
1279 spin_unlock(&block_group->lock);
1284 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1287 ASSERT(offset >= bitmap_start);
1288 offset -= bitmap_start;
1289 return (unsigned long)(div_u64(offset, unit));
1292 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1294 return (unsigned long)(div_u64(bytes, unit));
1297 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1301 u32 bytes_per_bitmap;
1303 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1304 bitmap_start = offset - ctl->start;
1305 bitmap_start = div_u64(bitmap_start, bytes_per_bitmap);
1306 bitmap_start *= bytes_per_bitmap;
1307 bitmap_start += ctl->start;
1309 return bitmap_start;
1312 static int tree_insert_offset(struct rb_root *root, u64 offset,
1313 struct rb_node *node, int bitmap)
1315 struct rb_node **p = &root->rb_node;
1316 struct rb_node *parent = NULL;
1317 struct btrfs_free_space *info;
1321 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1323 if (offset < info->offset) {
1325 } else if (offset > info->offset) {
1326 p = &(*p)->rb_right;
1329 * we could have a bitmap entry and an extent entry
1330 * share the same offset. If this is the case, we want
1331 * the extent entry to always be found first if we do a
1332 * linear search through the tree, since we want to have
1333 * the quickest allocation time, and allocating from an
1334 * extent is faster than allocating from a bitmap. So
1335 * if we're inserting a bitmap and we find an entry at
1336 * this offset, we want to go right, or after this entry
1337 * logically. If we are inserting an extent and we've
1338 * found a bitmap, we want to go left, or before
1346 p = &(*p)->rb_right;
1348 if (!info->bitmap) {
1357 rb_link_node(node, parent, p);
1358 rb_insert_color(node, root);
1364 * searches the tree for the given offset.
1366 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1367 * want a section that has at least bytes size and comes at or after the given
1370 static struct btrfs_free_space *
1371 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1372 u64 offset, int bitmap_only, int fuzzy)
1374 struct rb_node *n = ctl->free_space_offset.rb_node;
1375 struct btrfs_free_space *entry, *prev = NULL;
1377 /* find entry that is closest to the 'offset' */
1384 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1387 if (offset < entry->offset)
1389 else if (offset > entry->offset)
1402 * bitmap entry and extent entry may share same offset,
1403 * in that case, bitmap entry comes after extent entry.
1408 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1409 if (entry->offset != offset)
1412 WARN_ON(!entry->bitmap);
1415 if (entry->bitmap) {
1417 * if previous extent entry covers the offset,
1418 * we should return it instead of the bitmap entry
1420 n = rb_prev(&entry->offset_index);
1422 prev = rb_entry(n, struct btrfs_free_space,
1424 if (!prev->bitmap &&
1425 prev->offset + prev->bytes > offset)
1435 /* find last entry before the 'offset' */
1437 if (entry->offset > offset) {
1438 n = rb_prev(&entry->offset_index);
1440 entry = rb_entry(n, struct btrfs_free_space,
1442 ASSERT(entry->offset <= offset);
1451 if (entry->bitmap) {
1452 n = rb_prev(&entry->offset_index);
1454 prev = rb_entry(n, struct btrfs_free_space,
1456 if (!prev->bitmap &&
1457 prev->offset + prev->bytes > offset)
1460 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1462 } else if (entry->offset + entry->bytes > offset)
1469 if (entry->bitmap) {
1470 if (entry->offset + BITS_PER_BITMAP *
1474 if (entry->offset + entry->bytes > offset)
1478 n = rb_next(&entry->offset_index);
1481 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1487 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1488 struct btrfs_free_space *info)
1490 rb_erase(&info->offset_index, &ctl->free_space_offset);
1491 ctl->free_extents--;
1494 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1495 struct btrfs_free_space *info)
1497 __unlink_free_space(ctl, info);
1498 ctl->free_space -= info->bytes;
1501 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1502 struct btrfs_free_space *info)
1506 ASSERT(info->bytes || info->bitmap);
1507 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1508 &info->offset_index, (info->bitmap != NULL));
1512 ctl->free_space += info->bytes;
1513 ctl->free_extents++;
1517 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1519 struct btrfs_block_group_cache *block_group = ctl->private;
1523 u64 size = block_group->key.offset;
1524 u32 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1525 u32 max_bitmaps = div_u64(size + bytes_per_bg - 1, bytes_per_bg);
1527 max_bitmaps = max_t(u32, max_bitmaps, 1);
1529 ASSERT(ctl->total_bitmaps <= max_bitmaps);
1532 * The goal is to keep the total amount of memory used per 1gb of space
1533 * at or below 32k, so we need to adjust how much memory we allow to be
1534 * used by extent based free space tracking
1536 if (size < 1024 * 1024 * 1024)
1537 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1539 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1540 div_u64(size, 1024 * 1024 * 1024);
1543 * we want to account for 1 more bitmap than what we have so we can make
1544 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1545 * we add more bitmaps.
1547 bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1549 if (bitmap_bytes >= max_bytes) {
1550 ctl->extents_thresh = 0;
1555 * we want the extent entry threshold to always be at most 1/2 the max
1556 * bytes we can have, or whatever is less than that.
1558 extent_bytes = max_bytes - bitmap_bytes;
1559 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1561 ctl->extents_thresh =
1562 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1565 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1566 struct btrfs_free_space *info,
1567 u64 offset, u64 bytes)
1569 unsigned long start, count;
1571 start = offset_to_bit(info->offset, ctl->unit, offset);
1572 count = bytes_to_bits(bytes, ctl->unit);
1573 ASSERT(start + count <= BITS_PER_BITMAP);
1575 bitmap_clear(info->bitmap, start, count);
1577 info->bytes -= bytes;
1580 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1581 struct btrfs_free_space *info, u64 offset,
1584 __bitmap_clear_bits(ctl, info, offset, bytes);
1585 ctl->free_space -= bytes;
1588 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1589 struct btrfs_free_space *info, u64 offset,
1592 unsigned long start, count;
1594 start = offset_to_bit(info->offset, ctl->unit, offset);
1595 count = bytes_to_bits(bytes, ctl->unit);
1596 ASSERT(start + count <= BITS_PER_BITMAP);
1598 bitmap_set(info->bitmap, start, count);
1600 info->bytes += bytes;
1601 ctl->free_space += bytes;
1605 * If we can not find suitable extent, we will use bytes to record
1606 * the size of the max extent.
1608 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1609 struct btrfs_free_space *bitmap_info, u64 *offset,
1612 unsigned long found_bits = 0;
1613 unsigned long max_bits = 0;
1614 unsigned long bits, i;
1615 unsigned long next_zero;
1616 unsigned long extent_bits;
1618 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1619 max_t(u64, *offset, bitmap_info->offset));
1620 bits = bytes_to_bits(*bytes, ctl->unit);
1622 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1623 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1624 BITS_PER_BITMAP, i);
1625 extent_bits = next_zero - i;
1626 if (extent_bits >= bits) {
1627 found_bits = extent_bits;
1629 } else if (extent_bits > max_bits) {
1630 max_bits = extent_bits;
1636 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1637 *bytes = (u64)(found_bits) * ctl->unit;
1641 *bytes = (u64)(max_bits) * ctl->unit;
1645 /* Cache the size of the max extent in bytes */
1646 static struct btrfs_free_space *
1647 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1648 unsigned long align, u64 *max_extent_size)
1650 struct btrfs_free_space *entry;
1651 struct rb_node *node;
1656 if (!ctl->free_space_offset.rb_node)
1659 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1663 for (node = &entry->offset_index; node; node = rb_next(node)) {
1664 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1665 if (entry->bytes < *bytes) {
1666 if (entry->bytes > *max_extent_size)
1667 *max_extent_size = entry->bytes;
1671 /* make sure the space returned is big enough
1672 * to match our requested alignment
1674 if (*bytes >= align) {
1675 tmp = entry->offset - ctl->start + align - 1;
1677 tmp = tmp * align + ctl->start;
1678 align_off = tmp - entry->offset;
1681 tmp = entry->offset;
1684 if (entry->bytes < *bytes + align_off) {
1685 if (entry->bytes > *max_extent_size)
1686 *max_extent_size = entry->bytes;
1690 if (entry->bitmap) {
1693 ret = search_bitmap(ctl, entry, &tmp, &size);
1698 } else if (size > *max_extent_size) {
1699 *max_extent_size = size;
1705 *bytes = entry->bytes - align_off;
1712 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1713 struct btrfs_free_space *info, u64 offset)
1715 info->offset = offset_to_bitmap(ctl, offset);
1717 INIT_LIST_HEAD(&info->list);
1718 link_free_space(ctl, info);
1719 ctl->total_bitmaps++;
1721 ctl->op->recalc_thresholds(ctl);
1724 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1725 struct btrfs_free_space *bitmap_info)
1727 unlink_free_space(ctl, bitmap_info);
1728 kfree(bitmap_info->bitmap);
1729 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1730 ctl->total_bitmaps--;
1731 ctl->op->recalc_thresholds(ctl);
1734 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1735 struct btrfs_free_space *bitmap_info,
1736 u64 *offset, u64 *bytes)
1739 u64 search_start, search_bytes;
1743 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1746 * We need to search for bits in this bitmap. We could only cover some
1747 * of the extent in this bitmap thanks to how we add space, so we need
1748 * to search for as much as it as we can and clear that amount, and then
1749 * go searching for the next bit.
1751 search_start = *offset;
1752 search_bytes = ctl->unit;
1753 search_bytes = min(search_bytes, end - search_start + 1);
1754 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1755 if (ret < 0 || search_start != *offset)
1758 /* We may have found more bits than what we need */
1759 search_bytes = min(search_bytes, *bytes);
1761 /* Cannot clear past the end of the bitmap */
1762 search_bytes = min(search_bytes, end - search_start + 1);
1764 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1765 *offset += search_bytes;
1766 *bytes -= search_bytes;
1769 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1770 if (!bitmap_info->bytes)
1771 free_bitmap(ctl, bitmap_info);
1774 * no entry after this bitmap, but we still have bytes to
1775 * remove, so something has gone wrong.
1780 bitmap_info = rb_entry(next, struct btrfs_free_space,
1784 * if the next entry isn't a bitmap we need to return to let the
1785 * extent stuff do its work.
1787 if (!bitmap_info->bitmap)
1791 * Ok the next item is a bitmap, but it may not actually hold
1792 * the information for the rest of this free space stuff, so
1793 * look for it, and if we don't find it return so we can try
1794 * everything over again.
1796 search_start = *offset;
1797 search_bytes = ctl->unit;
1798 ret = search_bitmap(ctl, bitmap_info, &search_start,
1800 if (ret < 0 || search_start != *offset)
1804 } else if (!bitmap_info->bytes)
1805 free_bitmap(ctl, bitmap_info);
1810 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1811 struct btrfs_free_space *info, u64 offset,
1814 u64 bytes_to_set = 0;
1817 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1819 bytes_to_set = min(end - offset, bytes);
1821 bitmap_set_bits(ctl, info, offset, bytes_to_set);
1823 return bytes_to_set;
1827 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1828 struct btrfs_free_space *info)
1830 struct btrfs_block_group_cache *block_group = ctl->private;
1833 * If we are below the extents threshold then we can add this as an
1834 * extent, and don't have to deal with the bitmap
1836 if (ctl->free_extents < ctl->extents_thresh) {
1838 * If this block group has some small extents we don't want to
1839 * use up all of our free slots in the cache with them, we want
1840 * to reserve them to larger extents, however if we have plent
1841 * of cache left then go ahead an dadd them, no sense in adding
1842 * the overhead of a bitmap if we don't have to.
1844 if (info->bytes <= block_group->sectorsize * 4) {
1845 if (ctl->free_extents * 2 <= ctl->extents_thresh)
1853 * The original block groups from mkfs can be really small, like 8
1854 * megabytes, so don't bother with a bitmap for those entries. However
1855 * some block groups can be smaller than what a bitmap would cover but
1856 * are still large enough that they could overflow the 32k memory limit,
1857 * so allow those block groups to still be allowed to have a bitmap
1860 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
1866 static struct btrfs_free_space_op free_space_op = {
1867 .recalc_thresholds = recalculate_thresholds,
1868 .use_bitmap = use_bitmap,
1871 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1872 struct btrfs_free_space *info)
1874 struct btrfs_free_space *bitmap_info;
1875 struct btrfs_block_group_cache *block_group = NULL;
1877 u64 bytes, offset, bytes_added;
1880 bytes = info->bytes;
1881 offset = info->offset;
1883 if (!ctl->op->use_bitmap(ctl, info))
1886 if (ctl->op == &free_space_op)
1887 block_group = ctl->private;
1890 * Since we link bitmaps right into the cluster we need to see if we
1891 * have a cluster here, and if so and it has our bitmap we need to add
1892 * the free space to that bitmap.
1894 if (block_group && !list_empty(&block_group->cluster_list)) {
1895 struct btrfs_free_cluster *cluster;
1896 struct rb_node *node;
1897 struct btrfs_free_space *entry;
1899 cluster = list_entry(block_group->cluster_list.next,
1900 struct btrfs_free_cluster,
1902 spin_lock(&cluster->lock);
1903 node = rb_first(&cluster->root);
1905 spin_unlock(&cluster->lock);
1906 goto no_cluster_bitmap;
1909 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1910 if (!entry->bitmap) {
1911 spin_unlock(&cluster->lock);
1912 goto no_cluster_bitmap;
1915 if (entry->offset == offset_to_bitmap(ctl, offset)) {
1916 bytes_added = add_bytes_to_bitmap(ctl, entry,
1918 bytes -= bytes_added;
1919 offset += bytes_added;
1921 spin_unlock(&cluster->lock);
1929 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1936 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1937 bytes -= bytes_added;
1938 offset += bytes_added;
1948 if (info && info->bitmap) {
1949 add_new_bitmap(ctl, info, offset);
1954 spin_unlock(&ctl->tree_lock);
1956 /* no pre-allocated info, allocate a new one */
1958 info = kmem_cache_zalloc(btrfs_free_space_cachep,
1961 spin_lock(&ctl->tree_lock);
1967 /* allocate the bitmap */
1968 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1969 spin_lock(&ctl->tree_lock);
1970 if (!info->bitmap) {
1980 kfree(info->bitmap);
1981 kmem_cache_free(btrfs_free_space_cachep, info);
1987 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1988 struct btrfs_free_space *info, bool update_stat)
1990 struct btrfs_free_space *left_info;
1991 struct btrfs_free_space *right_info;
1992 bool merged = false;
1993 u64 offset = info->offset;
1994 u64 bytes = info->bytes;
1997 * first we want to see if there is free space adjacent to the range we
1998 * are adding, if there is remove that struct and add a new one to
1999 * cover the entire range
2001 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2002 if (right_info && rb_prev(&right_info->offset_index))
2003 left_info = rb_entry(rb_prev(&right_info->offset_index),
2004 struct btrfs_free_space, offset_index);
2006 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2008 if (right_info && !right_info->bitmap) {
2010 unlink_free_space(ctl, right_info);
2012 __unlink_free_space(ctl, right_info);
2013 info->bytes += right_info->bytes;
2014 kmem_cache_free(btrfs_free_space_cachep, right_info);
2018 if (left_info && !left_info->bitmap &&
2019 left_info->offset + left_info->bytes == offset) {
2021 unlink_free_space(ctl, left_info);
2023 __unlink_free_space(ctl, left_info);
2024 info->offset = left_info->offset;
2025 info->bytes += left_info->bytes;
2026 kmem_cache_free(btrfs_free_space_cachep, left_info);
2033 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2034 struct btrfs_free_space *info,
2037 struct btrfs_free_space *bitmap;
2040 const u64 end = info->offset + info->bytes;
2041 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2044 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2048 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2049 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2052 bytes = (j - i) * ctl->unit;
2053 info->bytes += bytes;
2056 bitmap_clear_bits(ctl, bitmap, end, bytes);
2058 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2061 free_bitmap(ctl, bitmap);
2066 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2067 struct btrfs_free_space *info,
2070 struct btrfs_free_space *bitmap;
2074 unsigned long prev_j;
2077 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2078 /* If we're on a boundary, try the previous logical bitmap. */
2079 if (bitmap_offset == info->offset) {
2080 if (info->offset == 0)
2082 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2085 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2089 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2091 prev_j = (unsigned long)-1;
2092 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2100 if (prev_j == (unsigned long)-1)
2101 bytes = (i + 1) * ctl->unit;
2103 bytes = (i - prev_j) * ctl->unit;
2105 info->offset -= bytes;
2106 info->bytes += bytes;
2109 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2111 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2114 free_bitmap(ctl, bitmap);
2120 * We prefer always to allocate from extent entries, both for clustered and
2121 * non-clustered allocation requests. So when attempting to add a new extent
2122 * entry, try to see if there's adjacent free space in bitmap entries, and if
2123 * there is, migrate that space from the bitmaps to the extent.
2124 * Like this we get better chances of satisfying space allocation requests
2125 * because we attempt to satisfy them based on a single cache entry, and never
2126 * on 2 or more entries - even if the entries represent a contiguous free space
2127 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2130 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2131 struct btrfs_free_space *info,
2135 * Only work with disconnected entries, as we can change their offset,
2136 * and must be extent entries.
2138 ASSERT(!info->bitmap);
2139 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2141 if (ctl->total_bitmaps > 0) {
2143 bool stole_front = false;
2145 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2146 if (ctl->total_bitmaps > 0)
2147 stole_front = steal_from_bitmap_to_front(ctl, info,
2150 if (stole_end || stole_front)
2151 try_merge_free_space(ctl, info, update_stat);
2155 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2156 u64 offset, u64 bytes)
2158 struct btrfs_free_space *info;
2161 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2165 info->offset = offset;
2166 info->bytes = bytes;
2167 RB_CLEAR_NODE(&info->offset_index);
2169 spin_lock(&ctl->tree_lock);
2171 if (try_merge_free_space(ctl, info, true))
2175 * There was no extent directly to the left or right of this new
2176 * extent then we know we're going to have to allocate a new extent, so
2177 * before we do that see if we need to drop this into a bitmap
2179 ret = insert_into_bitmap(ctl, info);
2188 * Only steal free space from adjacent bitmaps if we're sure we're not
2189 * going to add the new free space to existing bitmap entries - because
2190 * that would mean unnecessary work that would be reverted. Therefore
2191 * attempt to steal space from bitmaps if we're adding an extent entry.
2193 steal_from_bitmap(ctl, info, true);
2195 ret = link_free_space(ctl, info);
2197 kmem_cache_free(btrfs_free_space_cachep, info);
2199 spin_unlock(&ctl->tree_lock);
2202 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2203 ASSERT(ret != -EEXIST);
2209 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2210 u64 offset, u64 bytes)
2212 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2213 struct btrfs_free_space *info;
2215 bool re_search = false;
2217 spin_lock(&ctl->tree_lock);
2224 info = tree_search_offset(ctl, offset, 0, 0);
2227 * oops didn't find an extent that matched the space we wanted
2228 * to remove, look for a bitmap instead
2230 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2234 * If we found a partial bit of our free space in a
2235 * bitmap but then couldn't find the other part this may
2236 * be a problem, so WARN about it.
2244 if (!info->bitmap) {
2245 unlink_free_space(ctl, info);
2246 if (offset == info->offset) {
2247 u64 to_free = min(bytes, info->bytes);
2249 info->bytes -= to_free;
2250 info->offset += to_free;
2252 ret = link_free_space(ctl, info);
2255 kmem_cache_free(btrfs_free_space_cachep, info);
2262 u64 old_end = info->bytes + info->offset;
2264 info->bytes = offset - info->offset;
2265 ret = link_free_space(ctl, info);
2270 /* Not enough bytes in this entry to satisfy us */
2271 if (old_end < offset + bytes) {
2272 bytes -= old_end - offset;
2275 } else if (old_end == offset + bytes) {
2279 spin_unlock(&ctl->tree_lock);
2281 ret = btrfs_add_free_space(block_group, offset + bytes,
2282 old_end - (offset + bytes));
2288 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2289 if (ret == -EAGAIN) {
2294 spin_unlock(&ctl->tree_lock);
2299 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2302 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2303 struct btrfs_free_space *info;
2307 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2308 info = rb_entry(n, struct btrfs_free_space, offset_index);
2309 if (info->bytes >= bytes && !block_group->ro)
2311 btrfs_crit(block_group->fs_info,
2312 "entry offset %llu, bytes %llu, bitmap %s",
2313 info->offset, info->bytes,
2314 (info->bitmap) ? "yes" : "no");
2316 btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2317 list_empty(&block_group->cluster_list) ? "no" : "yes");
2318 btrfs_info(block_group->fs_info,
2319 "%d blocks of free space at or bigger than bytes is", count);
2322 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2324 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2326 spin_lock_init(&ctl->tree_lock);
2327 ctl->unit = block_group->sectorsize;
2328 ctl->start = block_group->key.objectid;
2329 ctl->private = block_group;
2330 ctl->op = &free_space_op;
2331 INIT_LIST_HEAD(&ctl->trimming_ranges);
2332 mutex_init(&ctl->cache_writeout_mutex);
2335 * we only want to have 32k of ram per block group for keeping
2336 * track of free space, and if we pass 1/2 of that we want to
2337 * start converting things over to using bitmaps
2339 ctl->extents_thresh = ((1024 * 32) / 2) /
2340 sizeof(struct btrfs_free_space);
2344 * for a given cluster, put all of its extents back into the free
2345 * space cache. If the block group passed doesn't match the block group
2346 * pointed to by the cluster, someone else raced in and freed the
2347 * cluster already. In that case, we just return without changing anything
2350 __btrfs_return_cluster_to_free_space(
2351 struct btrfs_block_group_cache *block_group,
2352 struct btrfs_free_cluster *cluster)
2354 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2355 struct btrfs_free_space *entry;
2356 struct rb_node *node;
2358 spin_lock(&cluster->lock);
2359 if (cluster->block_group != block_group)
2362 cluster->block_group = NULL;
2363 cluster->window_start = 0;
2364 list_del_init(&cluster->block_group_list);
2366 node = rb_first(&cluster->root);
2370 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2371 node = rb_next(&entry->offset_index);
2372 rb_erase(&entry->offset_index, &cluster->root);
2373 RB_CLEAR_NODE(&entry->offset_index);
2375 bitmap = (entry->bitmap != NULL);
2377 try_merge_free_space(ctl, entry, false);
2378 steal_from_bitmap(ctl, entry, false);
2380 tree_insert_offset(&ctl->free_space_offset,
2381 entry->offset, &entry->offset_index, bitmap);
2383 cluster->root = RB_ROOT;
2386 spin_unlock(&cluster->lock);
2387 btrfs_put_block_group(block_group);
2391 static void __btrfs_remove_free_space_cache_locked(
2392 struct btrfs_free_space_ctl *ctl)
2394 struct btrfs_free_space *info;
2395 struct rb_node *node;
2397 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2398 info = rb_entry(node, struct btrfs_free_space, offset_index);
2399 if (!info->bitmap) {
2400 unlink_free_space(ctl, info);
2401 kmem_cache_free(btrfs_free_space_cachep, info);
2403 free_bitmap(ctl, info);
2406 cond_resched_lock(&ctl->tree_lock);
2410 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2412 spin_lock(&ctl->tree_lock);
2413 __btrfs_remove_free_space_cache_locked(ctl);
2414 spin_unlock(&ctl->tree_lock);
2417 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2419 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2420 struct btrfs_free_cluster *cluster;
2421 struct list_head *head;
2423 spin_lock(&ctl->tree_lock);
2424 while ((head = block_group->cluster_list.next) !=
2425 &block_group->cluster_list) {
2426 cluster = list_entry(head, struct btrfs_free_cluster,
2429 WARN_ON(cluster->block_group != block_group);
2430 __btrfs_return_cluster_to_free_space(block_group, cluster);
2432 cond_resched_lock(&ctl->tree_lock);
2434 __btrfs_remove_free_space_cache_locked(ctl);
2435 spin_unlock(&ctl->tree_lock);
2439 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2440 u64 offset, u64 bytes, u64 empty_size,
2441 u64 *max_extent_size)
2443 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2444 struct btrfs_free_space *entry = NULL;
2445 u64 bytes_search = bytes + empty_size;
2448 u64 align_gap_len = 0;
2450 spin_lock(&ctl->tree_lock);
2451 entry = find_free_space(ctl, &offset, &bytes_search,
2452 block_group->full_stripe_len, max_extent_size);
2457 if (entry->bitmap) {
2458 bitmap_clear_bits(ctl, entry, offset, bytes);
2460 free_bitmap(ctl, entry);
2462 unlink_free_space(ctl, entry);
2463 align_gap_len = offset - entry->offset;
2464 align_gap = entry->offset;
2466 entry->offset = offset + bytes;
2467 WARN_ON(entry->bytes < bytes + align_gap_len);
2469 entry->bytes -= bytes + align_gap_len;
2471 kmem_cache_free(btrfs_free_space_cachep, entry);
2473 link_free_space(ctl, entry);
2476 spin_unlock(&ctl->tree_lock);
2479 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2484 * given a cluster, put all of its extents back into the free space
2485 * cache. If a block group is passed, this function will only free
2486 * a cluster that belongs to the passed block group.
2488 * Otherwise, it'll get a reference on the block group pointed to by the
2489 * cluster and remove the cluster from it.
2491 int btrfs_return_cluster_to_free_space(
2492 struct btrfs_block_group_cache *block_group,
2493 struct btrfs_free_cluster *cluster)
2495 struct btrfs_free_space_ctl *ctl;
2498 /* first, get a safe pointer to the block group */
2499 spin_lock(&cluster->lock);
2501 block_group = cluster->block_group;
2503 spin_unlock(&cluster->lock);
2506 } else if (cluster->block_group != block_group) {
2507 /* someone else has already freed it don't redo their work */
2508 spin_unlock(&cluster->lock);
2511 atomic_inc(&block_group->count);
2512 spin_unlock(&cluster->lock);
2514 ctl = block_group->free_space_ctl;
2516 /* now return any extents the cluster had on it */
2517 spin_lock(&ctl->tree_lock);
2518 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2519 spin_unlock(&ctl->tree_lock);
2521 /* finally drop our ref */
2522 btrfs_put_block_group(block_group);
2526 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2527 struct btrfs_free_cluster *cluster,
2528 struct btrfs_free_space *entry,
2529 u64 bytes, u64 min_start,
2530 u64 *max_extent_size)
2532 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2534 u64 search_start = cluster->window_start;
2535 u64 search_bytes = bytes;
2538 search_start = min_start;
2539 search_bytes = bytes;
2541 err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2543 if (search_bytes > *max_extent_size)
2544 *max_extent_size = search_bytes;
2549 __bitmap_clear_bits(ctl, entry, ret, bytes);
2555 * given a cluster, try to allocate 'bytes' from it, returns 0
2556 * if it couldn't find anything suitably large, or a logical disk offset
2557 * if things worked out
2559 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2560 struct btrfs_free_cluster *cluster, u64 bytes,
2561 u64 min_start, u64 *max_extent_size)
2563 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2564 struct btrfs_free_space *entry = NULL;
2565 struct rb_node *node;
2568 spin_lock(&cluster->lock);
2569 if (bytes > cluster->max_size)
2572 if (cluster->block_group != block_group)
2575 node = rb_first(&cluster->root);
2579 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2581 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2582 *max_extent_size = entry->bytes;
2584 if (entry->bytes < bytes ||
2585 (!entry->bitmap && entry->offset < min_start)) {
2586 node = rb_next(&entry->offset_index);
2589 entry = rb_entry(node, struct btrfs_free_space,
2594 if (entry->bitmap) {
2595 ret = btrfs_alloc_from_bitmap(block_group,
2596 cluster, entry, bytes,
2597 cluster->window_start,
2600 node = rb_next(&entry->offset_index);
2603 entry = rb_entry(node, struct btrfs_free_space,
2607 cluster->window_start += bytes;
2609 ret = entry->offset;
2611 entry->offset += bytes;
2612 entry->bytes -= bytes;
2615 if (entry->bytes == 0)
2616 rb_erase(&entry->offset_index, &cluster->root);
2620 spin_unlock(&cluster->lock);
2625 spin_lock(&ctl->tree_lock);
2627 ctl->free_space -= bytes;
2628 if (entry->bytes == 0) {
2629 ctl->free_extents--;
2630 if (entry->bitmap) {
2631 kfree(entry->bitmap);
2632 ctl->total_bitmaps--;
2633 ctl->op->recalc_thresholds(ctl);
2635 kmem_cache_free(btrfs_free_space_cachep, entry);
2638 spin_unlock(&ctl->tree_lock);
2643 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2644 struct btrfs_free_space *entry,
2645 struct btrfs_free_cluster *cluster,
2646 u64 offset, u64 bytes,
2647 u64 cont1_bytes, u64 min_bytes)
2649 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2650 unsigned long next_zero;
2652 unsigned long want_bits;
2653 unsigned long min_bits;
2654 unsigned long found_bits;
2655 unsigned long start = 0;
2656 unsigned long total_found = 0;
2659 i = offset_to_bit(entry->offset, ctl->unit,
2660 max_t(u64, offset, entry->offset));
2661 want_bits = bytes_to_bits(bytes, ctl->unit);
2662 min_bits = bytes_to_bits(min_bytes, ctl->unit);
2666 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2667 next_zero = find_next_zero_bit(entry->bitmap,
2668 BITS_PER_BITMAP, i);
2669 if (next_zero - i >= min_bits) {
2670 found_bits = next_zero - i;
2681 cluster->max_size = 0;
2684 total_found += found_bits;
2686 if (cluster->max_size < found_bits * ctl->unit)
2687 cluster->max_size = found_bits * ctl->unit;
2689 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2694 cluster->window_start = start * ctl->unit + entry->offset;
2695 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2696 ret = tree_insert_offset(&cluster->root, entry->offset,
2697 &entry->offset_index, 1);
2698 ASSERT(!ret); /* -EEXIST; Logic error */
2700 trace_btrfs_setup_cluster(block_group, cluster,
2701 total_found * ctl->unit, 1);
2706 * This searches the block group for just extents to fill the cluster with.
2707 * Try to find a cluster with at least bytes total bytes, at least one
2708 * extent of cont1_bytes, and other clusters of at least min_bytes.
2711 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2712 struct btrfs_free_cluster *cluster,
2713 struct list_head *bitmaps, u64 offset, u64 bytes,
2714 u64 cont1_bytes, u64 min_bytes)
2716 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2717 struct btrfs_free_space *first = NULL;
2718 struct btrfs_free_space *entry = NULL;
2719 struct btrfs_free_space *last;
2720 struct rb_node *node;
2725 entry = tree_search_offset(ctl, offset, 0, 1);
2730 * We don't want bitmaps, so just move along until we find a normal
2733 while (entry->bitmap || entry->bytes < min_bytes) {
2734 if (entry->bitmap && list_empty(&entry->list))
2735 list_add_tail(&entry->list, bitmaps);
2736 node = rb_next(&entry->offset_index);
2739 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2742 window_free = entry->bytes;
2743 max_extent = entry->bytes;
2747 for (node = rb_next(&entry->offset_index); node;
2748 node = rb_next(&entry->offset_index)) {
2749 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2751 if (entry->bitmap) {
2752 if (list_empty(&entry->list))
2753 list_add_tail(&entry->list, bitmaps);
2757 if (entry->bytes < min_bytes)
2761 window_free += entry->bytes;
2762 if (entry->bytes > max_extent)
2763 max_extent = entry->bytes;
2766 if (window_free < bytes || max_extent < cont1_bytes)
2769 cluster->window_start = first->offset;
2771 node = &first->offset_index;
2774 * now we've found our entries, pull them out of the free space
2775 * cache and put them into the cluster rbtree
2780 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2781 node = rb_next(&entry->offset_index);
2782 if (entry->bitmap || entry->bytes < min_bytes)
2785 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2786 ret = tree_insert_offset(&cluster->root, entry->offset,
2787 &entry->offset_index, 0);
2788 total_size += entry->bytes;
2789 ASSERT(!ret); /* -EEXIST; Logic error */
2790 } while (node && entry != last);
2792 cluster->max_size = max_extent;
2793 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2798 * This specifically looks for bitmaps that may work in the cluster, we assume
2799 * that we have already failed to find extents that will work.
2802 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2803 struct btrfs_free_cluster *cluster,
2804 struct list_head *bitmaps, u64 offset, u64 bytes,
2805 u64 cont1_bytes, u64 min_bytes)
2807 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2808 struct btrfs_free_space *entry;
2810 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2812 if (ctl->total_bitmaps == 0)
2816 * The bitmap that covers offset won't be in the list unless offset
2817 * is just its start offset.
2819 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2820 if (entry->offset != bitmap_offset) {
2821 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2822 if (entry && list_empty(&entry->list))
2823 list_add(&entry->list, bitmaps);
2826 list_for_each_entry(entry, bitmaps, list) {
2827 if (entry->bytes < bytes)
2829 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2830 bytes, cont1_bytes, min_bytes);
2836 * The bitmaps list has all the bitmaps that record free space
2837 * starting after offset, so no more search is required.
2843 * here we try to find a cluster of blocks in a block group. The goal
2844 * is to find at least bytes+empty_size.
2845 * We might not find them all in one contiguous area.
2847 * returns zero and sets up cluster if things worked out, otherwise
2848 * it returns -enospc
2850 int btrfs_find_space_cluster(struct btrfs_root *root,
2851 struct btrfs_block_group_cache *block_group,
2852 struct btrfs_free_cluster *cluster,
2853 u64 offset, u64 bytes, u64 empty_size)
2855 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2856 struct btrfs_free_space *entry, *tmp;
2863 * Choose the minimum extent size we'll require for this
2864 * cluster. For SSD_SPREAD, don't allow any fragmentation.
2865 * For metadata, allow allocates with smaller extents. For
2866 * data, keep it dense.
2868 if (btrfs_test_opt(root, SSD_SPREAD)) {
2869 cont1_bytes = min_bytes = bytes + empty_size;
2870 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2871 cont1_bytes = bytes;
2872 min_bytes = block_group->sectorsize;
2874 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
2875 min_bytes = block_group->sectorsize;
2878 spin_lock(&ctl->tree_lock);
2881 * If we know we don't have enough space to make a cluster don't even
2882 * bother doing all the work to try and find one.
2884 if (ctl->free_space < bytes) {
2885 spin_unlock(&ctl->tree_lock);
2889 spin_lock(&cluster->lock);
2891 /* someone already found a cluster, hooray */
2892 if (cluster->block_group) {
2897 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
2900 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2902 cont1_bytes, min_bytes);
2904 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2905 offset, bytes + empty_size,
2906 cont1_bytes, min_bytes);
2908 /* Clear our temporary list */
2909 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2910 list_del_init(&entry->list);
2913 atomic_inc(&block_group->count);
2914 list_add_tail(&cluster->block_group_list,
2915 &block_group->cluster_list);
2916 cluster->block_group = block_group;
2918 trace_btrfs_failed_cluster_setup(block_group);
2921 spin_unlock(&cluster->lock);
2922 spin_unlock(&ctl->tree_lock);
2928 * simple code to zero out a cluster
2930 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2932 spin_lock_init(&cluster->lock);
2933 spin_lock_init(&cluster->refill_lock);
2934 cluster->root = RB_ROOT;
2935 cluster->max_size = 0;
2936 INIT_LIST_HEAD(&cluster->block_group_list);
2937 cluster->block_group = NULL;
2940 static int do_trimming(struct btrfs_block_group_cache *block_group,
2941 u64 *total_trimmed, u64 start, u64 bytes,
2942 u64 reserved_start, u64 reserved_bytes,
2943 struct btrfs_trim_range *trim_entry)
2945 struct btrfs_space_info *space_info = block_group->space_info;
2946 struct btrfs_fs_info *fs_info = block_group->fs_info;
2947 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2952 spin_lock(&space_info->lock);
2953 spin_lock(&block_group->lock);
2954 if (!block_group->ro) {
2955 block_group->reserved += reserved_bytes;
2956 space_info->bytes_reserved += reserved_bytes;
2959 spin_unlock(&block_group->lock);
2960 spin_unlock(&space_info->lock);
2962 ret = btrfs_discard_extent(fs_info->extent_root,
2963 start, bytes, &trimmed);
2965 *total_trimmed += trimmed;
2967 mutex_lock(&ctl->cache_writeout_mutex);
2968 btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
2969 list_del(&trim_entry->list);
2970 mutex_unlock(&ctl->cache_writeout_mutex);
2973 spin_lock(&space_info->lock);
2974 spin_lock(&block_group->lock);
2975 if (block_group->ro)
2976 space_info->bytes_readonly += reserved_bytes;
2977 block_group->reserved -= reserved_bytes;
2978 space_info->bytes_reserved -= reserved_bytes;
2979 spin_unlock(&space_info->lock);
2980 spin_unlock(&block_group->lock);
2986 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
2987 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
2989 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2990 struct btrfs_free_space *entry;
2991 struct rb_node *node;
2997 while (start < end) {
2998 struct btrfs_trim_range trim_entry;
3000 mutex_lock(&ctl->cache_writeout_mutex);
3001 spin_lock(&ctl->tree_lock);
3003 if (ctl->free_space < minlen) {
3004 spin_unlock(&ctl->tree_lock);
3005 mutex_unlock(&ctl->cache_writeout_mutex);
3009 entry = tree_search_offset(ctl, start, 0, 1);
3011 spin_unlock(&ctl->tree_lock);
3012 mutex_unlock(&ctl->cache_writeout_mutex);
3017 while (entry->bitmap) {
3018 node = rb_next(&entry->offset_index);
3020 spin_unlock(&ctl->tree_lock);
3021 mutex_unlock(&ctl->cache_writeout_mutex);
3024 entry = rb_entry(node, struct btrfs_free_space,
3028 if (entry->offset >= end) {
3029 spin_unlock(&ctl->tree_lock);
3030 mutex_unlock(&ctl->cache_writeout_mutex);
3034 extent_start = entry->offset;
3035 extent_bytes = entry->bytes;
3036 start = max(start, extent_start);
3037 bytes = min(extent_start + extent_bytes, end) - start;
3038 if (bytes < minlen) {
3039 spin_unlock(&ctl->tree_lock);
3040 mutex_unlock(&ctl->cache_writeout_mutex);
3044 unlink_free_space(ctl, entry);
3045 kmem_cache_free(btrfs_free_space_cachep, entry);
3047 spin_unlock(&ctl->tree_lock);
3048 trim_entry.start = extent_start;
3049 trim_entry.bytes = extent_bytes;
3050 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3051 mutex_unlock(&ctl->cache_writeout_mutex);
3053 ret = do_trimming(block_group, total_trimmed, start, bytes,
3054 extent_start, extent_bytes, &trim_entry);
3060 if (fatal_signal_pending(current)) {
3071 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3072 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3074 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3075 struct btrfs_free_space *entry;
3079 u64 offset = offset_to_bitmap(ctl, start);
3081 while (offset < end) {
3082 bool next_bitmap = false;
3083 struct btrfs_trim_range trim_entry;
3085 mutex_lock(&ctl->cache_writeout_mutex);
3086 spin_lock(&ctl->tree_lock);
3088 if (ctl->free_space < minlen) {
3089 spin_unlock(&ctl->tree_lock);
3090 mutex_unlock(&ctl->cache_writeout_mutex);
3094 entry = tree_search_offset(ctl, offset, 1, 0);
3096 spin_unlock(&ctl->tree_lock);
3097 mutex_unlock(&ctl->cache_writeout_mutex);
3103 ret2 = search_bitmap(ctl, entry, &start, &bytes);
3104 if (ret2 || start >= end) {
3105 spin_unlock(&ctl->tree_lock);
3106 mutex_unlock(&ctl->cache_writeout_mutex);
3111 bytes = min(bytes, end - start);
3112 if (bytes < minlen) {
3113 spin_unlock(&ctl->tree_lock);
3114 mutex_unlock(&ctl->cache_writeout_mutex);
3118 bitmap_clear_bits(ctl, entry, start, bytes);
3119 if (entry->bytes == 0)
3120 free_bitmap(ctl, entry);
3122 spin_unlock(&ctl->tree_lock);
3123 trim_entry.start = start;
3124 trim_entry.bytes = bytes;
3125 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3126 mutex_unlock(&ctl->cache_writeout_mutex);
3128 ret = do_trimming(block_group, total_trimmed, start, bytes,
3129 start, bytes, &trim_entry);
3134 offset += BITS_PER_BITMAP * ctl->unit;
3137 if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3138 offset += BITS_PER_BITMAP * ctl->unit;
3141 if (fatal_signal_pending(current)) {
3152 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3153 u64 *trimmed, u64 start, u64 end, u64 minlen)
3159 spin_lock(&block_group->lock);
3160 if (block_group->removed) {
3161 spin_unlock(&block_group->lock);
3164 atomic_inc(&block_group->trimming);
3165 spin_unlock(&block_group->lock);
3167 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3171 ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3173 spin_lock(&block_group->lock);
3174 if (atomic_dec_and_test(&block_group->trimming) &&
3175 block_group->removed) {
3176 struct extent_map_tree *em_tree;
3177 struct extent_map *em;
3179 spin_unlock(&block_group->lock);
3181 lock_chunks(block_group->fs_info->chunk_root);
3182 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3183 write_lock(&em_tree->lock);
3184 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3186 BUG_ON(!em); /* logic error, can't happen */
3188 * remove_extent_mapping() will delete us from the pinned_chunks
3189 * list, which is protected by the chunk mutex.
3191 remove_extent_mapping(em_tree, em);
3192 write_unlock(&em_tree->lock);
3193 unlock_chunks(block_group->fs_info->chunk_root);
3195 /* once for us and once for the tree */
3196 free_extent_map(em);
3197 free_extent_map(em);
3200 * We've left one free space entry and other tasks trimming
3201 * this block group have left 1 entry each one. Free them.
3203 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3205 spin_unlock(&block_group->lock);
3212 * Find the left-most item in the cache tree, and then return the
3213 * smallest inode number in the item.
3215 * Note: the returned inode number may not be the smallest one in
3216 * the tree, if the left-most item is a bitmap.
3218 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3220 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3221 struct btrfs_free_space *entry = NULL;
3224 spin_lock(&ctl->tree_lock);
3226 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3229 entry = rb_entry(rb_first(&ctl->free_space_offset),
3230 struct btrfs_free_space, offset_index);
3232 if (!entry->bitmap) {
3233 ino = entry->offset;
3235 unlink_free_space(ctl, entry);
3239 kmem_cache_free(btrfs_free_space_cachep, entry);
3241 link_free_space(ctl, entry);
3247 ret = search_bitmap(ctl, entry, &offset, &count);
3248 /* Logic error; Should be empty if it can't find anything */
3252 bitmap_clear_bits(ctl, entry, offset, 1);
3253 if (entry->bytes == 0)
3254 free_bitmap(ctl, entry);
3257 spin_unlock(&ctl->tree_lock);
3262 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3263 struct btrfs_path *path)
3265 struct inode *inode = NULL;
3267 spin_lock(&root->ino_cache_lock);
3268 if (root->ino_cache_inode)
3269 inode = igrab(root->ino_cache_inode);
3270 spin_unlock(&root->ino_cache_lock);
3274 inode = __lookup_free_space_inode(root, path, 0);
3278 spin_lock(&root->ino_cache_lock);
3279 if (!btrfs_fs_closing(root->fs_info))
3280 root->ino_cache_inode = igrab(inode);
3281 spin_unlock(&root->ino_cache_lock);
3286 int create_free_ino_inode(struct btrfs_root *root,
3287 struct btrfs_trans_handle *trans,
3288 struct btrfs_path *path)
3290 return __create_free_space_inode(root, trans, path,
3291 BTRFS_FREE_INO_OBJECTID, 0);
3294 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3296 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3297 struct btrfs_path *path;
3298 struct inode *inode;
3300 u64 root_gen = btrfs_root_generation(&root->root_item);
3302 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3306 * If we're unmounting then just return, since this does a search on the
3307 * normal root and not the commit root and we could deadlock.
3309 if (btrfs_fs_closing(fs_info))
3312 path = btrfs_alloc_path();
3316 inode = lookup_free_ino_inode(root, path);
3320 if (root_gen != BTRFS_I(inode)->generation)
3323 ret = __load_free_space_cache(root, inode, ctl, path, 0);
3327 "failed to load free ino cache for root %llu",
3328 root->root_key.objectid);
3332 btrfs_free_path(path);
3336 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3337 struct btrfs_trans_handle *trans,
3338 struct btrfs_path *path,
3339 struct inode *inode)
3341 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3344 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3347 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
3349 btrfs_delalloc_release_metadata(inode, inode->i_size);
3351 btrfs_err(root->fs_info,
3352 "failed to write free ino cache for root %llu",
3353 root->root_key.objectid);
3360 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3362 * Use this if you need to make a bitmap or extent entry specifically, it
3363 * doesn't do any of the merging that add_free_space does, this acts a lot like
3364 * how the free space cache loading stuff works, so you can get really weird
3367 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3368 u64 offset, u64 bytes, bool bitmap)
3370 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3371 struct btrfs_free_space *info = NULL, *bitmap_info;
3378 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3384 spin_lock(&ctl->tree_lock);
3385 info->offset = offset;
3386 info->bytes = bytes;
3387 ret = link_free_space(ctl, info);
3388 spin_unlock(&ctl->tree_lock);
3390 kmem_cache_free(btrfs_free_space_cachep, info);
3395 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3397 kmem_cache_free(btrfs_free_space_cachep, info);
3402 spin_lock(&ctl->tree_lock);
3403 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3408 add_new_bitmap(ctl, info, offset);
3413 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3414 bytes -= bytes_added;
3415 offset += bytes_added;
3416 spin_unlock(&ctl->tree_lock);
3422 kmem_cache_free(btrfs_free_space_cachep, info);
3429 * Checks to see if the given range is in the free space cache. This is really
3430 * just used to check the absence of space, so if there is free space in the
3431 * range at all we will return 1.
3433 int test_check_exists(struct btrfs_block_group_cache *cache,
3434 u64 offset, u64 bytes)
3436 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3437 struct btrfs_free_space *info;
3440 spin_lock(&ctl->tree_lock);
3441 info = tree_search_offset(ctl, offset, 0, 0);
3443 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3451 u64 bit_off, bit_bytes;
3453 struct btrfs_free_space *tmp;
3456 bit_bytes = ctl->unit;
3457 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes);
3459 if (bit_off == offset) {
3462 } else if (bit_off > offset &&
3463 offset + bytes > bit_off) {
3469 n = rb_prev(&info->offset_index);
3471 tmp = rb_entry(n, struct btrfs_free_space,
3473 if (tmp->offset + tmp->bytes < offset)
3475 if (offset + bytes < tmp->offset) {
3476 n = rb_prev(&info->offset_index);
3483 n = rb_next(&info->offset_index);
3485 tmp = rb_entry(n, struct btrfs_free_space,
3487 if (offset + bytes < tmp->offset)
3489 if (tmp->offset + tmp->bytes < offset) {
3490 n = rb_next(&info->offset_index);
3501 if (info->offset == offset) {
3506 if (offset > info->offset && offset < info->offset + info->bytes)
3509 spin_unlock(&ctl->tree_lock);
3512 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */