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) &
89 ~(GFP_NOFS & ~__GFP_HIGHMEM));
94 struct inode *lookup_free_space_inode(struct btrfs_root *root,
95 struct btrfs_block_group_cache
96 *block_group, struct btrfs_path *path)
98 struct inode *inode = NULL;
99 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
101 spin_lock(&block_group->lock);
102 if (block_group->inode)
103 inode = igrab(block_group->inode);
104 spin_unlock(&block_group->lock);
108 inode = __lookup_free_space_inode(root, path,
109 block_group->key.objectid);
113 spin_lock(&block_group->lock);
114 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
115 btrfs_info(root->fs_info,
116 "Old style space inode found, converting.");
117 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
118 BTRFS_INODE_NODATACOW;
119 block_group->disk_cache_state = BTRFS_DC_CLEAR;
122 if (!block_group->iref) {
123 block_group->inode = igrab(inode);
124 block_group->iref = 1;
126 spin_unlock(&block_group->lock);
131 static int __create_free_space_inode(struct btrfs_root *root,
132 struct btrfs_trans_handle *trans,
133 struct btrfs_path *path,
136 struct btrfs_key key;
137 struct btrfs_disk_key disk_key;
138 struct btrfs_free_space_header *header;
139 struct btrfs_inode_item *inode_item;
140 struct extent_buffer *leaf;
141 u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
144 ret = btrfs_insert_empty_inode(trans, root, path, ino);
148 /* We inline crc's for the free disk space cache */
149 if (ino != BTRFS_FREE_INO_OBJECTID)
150 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
152 leaf = path->nodes[0];
153 inode_item = btrfs_item_ptr(leaf, path->slots[0],
154 struct btrfs_inode_item);
155 btrfs_item_key(leaf, &disk_key, path->slots[0]);
156 memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
157 sizeof(*inode_item));
158 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
159 btrfs_set_inode_size(leaf, inode_item, 0);
160 btrfs_set_inode_nbytes(leaf, inode_item, 0);
161 btrfs_set_inode_uid(leaf, inode_item, 0);
162 btrfs_set_inode_gid(leaf, inode_item, 0);
163 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
164 btrfs_set_inode_flags(leaf, inode_item, flags);
165 btrfs_set_inode_nlink(leaf, inode_item, 1);
166 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
167 btrfs_set_inode_block_group(leaf, inode_item, offset);
168 btrfs_mark_buffer_dirty(leaf);
169 btrfs_release_path(path);
171 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);
181 leaf = path->nodes[0];
182 header = btrfs_item_ptr(leaf, path->slots[0],
183 struct btrfs_free_space_header);
184 memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
185 btrfs_set_free_space_key(leaf, header, &disk_key);
186 btrfs_mark_buffer_dirty(leaf);
187 btrfs_release_path(path);
192 int create_free_space_inode(struct btrfs_root *root,
193 struct btrfs_trans_handle *trans,
194 struct btrfs_block_group_cache *block_group,
195 struct btrfs_path *path)
200 ret = btrfs_find_free_objectid(root, &ino);
204 return __create_free_space_inode(root, trans, path, ino,
205 block_group->key.objectid);
208 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
209 struct btrfs_block_rsv *rsv)
214 /* 1 for slack space, 1 for updating the inode */
215 needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
216 btrfs_calc_trans_metadata_size(root, 1);
218 spin_lock(&rsv->lock);
219 if (rsv->reserved < needed_bytes)
223 spin_unlock(&rsv->lock);
227 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
228 struct btrfs_trans_handle *trans,
229 struct btrfs_block_group_cache *block_group,
233 struct btrfs_path *path = btrfs_alloc_path();
241 mutex_lock(&trans->transaction->cache_write_mutex);
242 if (!list_empty(&block_group->io_list)) {
243 list_del_init(&block_group->io_list);
245 btrfs_wait_cache_io(root, trans, block_group,
246 &block_group->io_ctl, path,
247 block_group->key.objectid);
248 btrfs_put_block_group(block_group);
252 * now that we've truncated the cache away, its no longer
255 spin_lock(&block_group->lock);
256 block_group->disk_cache_state = BTRFS_DC_CLEAR;
257 spin_unlock(&block_group->lock);
259 btrfs_free_path(path);
261 btrfs_i_size_write(inode, 0);
262 truncate_pagecache(inode, 0);
265 * We don't need an orphan item because truncating the free space cache
266 * will never be split across transactions.
267 * We don't need to check for -EAGAIN because we're a free space
270 ret = btrfs_truncate_inode_items(trans, root, inode,
271 0, BTRFS_EXTENT_DATA_KEY);
273 mutex_unlock(&trans->transaction->cache_write_mutex);
274 btrfs_abort_transaction(trans, root, ret);
278 ret = btrfs_update_inode(trans, root, inode);
281 mutex_unlock(&trans->transaction->cache_write_mutex);
285 btrfs_abort_transaction(trans, root, ret);
290 static int readahead_cache(struct inode *inode)
292 struct file_ra_state *ra;
293 unsigned long last_index;
295 ra = kzalloc(sizeof(*ra), GFP_NOFS);
299 file_ra_state_init(ra, inode->i_mapping);
300 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
302 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
309 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
310 struct btrfs_root *root, int write)
315 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_CACHE_SIZE);
317 if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
320 /* Make sure we can fit our crcs into the first page */
321 if (write && check_crcs &&
322 (num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE)
325 memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
327 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
331 io_ctl->num_pages = num_pages;
333 io_ctl->check_crcs = check_crcs;
334 io_ctl->inode = inode;
339 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
341 kfree(io_ctl->pages);
342 io_ctl->pages = NULL;
345 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
353 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
355 ASSERT(io_ctl->index < io_ctl->num_pages);
356 io_ctl->page = io_ctl->pages[io_ctl->index++];
357 io_ctl->cur = page_address(io_ctl->page);
358 io_ctl->orig = io_ctl->cur;
359 io_ctl->size = PAGE_CACHE_SIZE;
361 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
364 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
368 io_ctl_unmap_page(io_ctl);
370 for (i = 0; i < io_ctl->num_pages; i++) {
371 if (io_ctl->pages[i]) {
372 ClearPageChecked(io_ctl->pages[i]);
373 unlock_page(io_ctl->pages[i]);
374 page_cache_release(io_ctl->pages[i]);
379 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
383 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
386 for (i = 0; i < io_ctl->num_pages; i++) {
387 page = find_or_create_page(inode->i_mapping, i, mask);
389 io_ctl_drop_pages(io_ctl);
392 io_ctl->pages[i] = page;
393 if (uptodate && !PageUptodate(page)) {
394 btrfs_readpage(NULL, page);
396 if (!PageUptodate(page)) {
397 btrfs_err(BTRFS_I(inode)->root->fs_info,
398 "error reading free space cache");
399 io_ctl_drop_pages(io_ctl);
405 for (i = 0; i < io_ctl->num_pages; i++) {
406 clear_page_dirty_for_io(io_ctl->pages[i]);
407 set_page_extent_mapped(io_ctl->pages[i]);
413 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
417 io_ctl_map_page(io_ctl, 1);
420 * Skip the csum areas. If we don't check crcs then we just have a
421 * 64bit chunk at the front of the first page.
423 if (io_ctl->check_crcs) {
424 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
425 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
427 io_ctl->cur += sizeof(u64);
428 io_ctl->size -= sizeof(u64) * 2;
432 *val = cpu_to_le64(generation);
433 io_ctl->cur += sizeof(u64);
436 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
441 * Skip the crc area. If we don't check crcs then we just have a 64bit
442 * chunk at the front of the first page.
444 if (io_ctl->check_crcs) {
445 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
446 io_ctl->size -= sizeof(u64) +
447 (sizeof(u32) * io_ctl->num_pages);
449 io_ctl->cur += sizeof(u64);
450 io_ctl->size -= sizeof(u64) * 2;
454 if (le64_to_cpu(*gen) != generation) {
455 printk_ratelimited(KERN_ERR "BTRFS: space cache generation "
456 "(%Lu) does not match inode (%Lu)\n", *gen,
458 io_ctl_unmap_page(io_ctl);
461 io_ctl->cur += sizeof(u64);
465 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
471 if (!io_ctl->check_crcs) {
472 io_ctl_unmap_page(io_ctl);
477 offset = sizeof(u32) * io_ctl->num_pages;
479 crc = btrfs_csum_data(io_ctl->orig + offset, crc,
480 PAGE_CACHE_SIZE - offset);
481 btrfs_csum_final(crc, (char *)&crc);
482 io_ctl_unmap_page(io_ctl);
483 tmp = page_address(io_ctl->pages[0]);
488 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
494 if (!io_ctl->check_crcs) {
495 io_ctl_map_page(io_ctl, 0);
500 offset = sizeof(u32) * io_ctl->num_pages;
502 tmp = page_address(io_ctl->pages[0]);
506 io_ctl_map_page(io_ctl, 0);
507 crc = btrfs_csum_data(io_ctl->orig + offset, crc,
508 PAGE_CACHE_SIZE - offset);
509 btrfs_csum_final(crc, (char *)&crc);
511 printk_ratelimited(KERN_ERR "BTRFS: csum mismatch on free "
513 io_ctl_unmap_page(io_ctl);
520 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
523 struct btrfs_free_space_entry *entry;
529 entry->offset = cpu_to_le64(offset);
530 entry->bytes = cpu_to_le64(bytes);
531 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
532 BTRFS_FREE_SPACE_EXTENT;
533 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
534 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
536 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
539 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
541 /* No more pages to map */
542 if (io_ctl->index >= io_ctl->num_pages)
545 /* map the next page */
546 io_ctl_map_page(io_ctl, 1);
550 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
556 * If we aren't at the start of the current page, unmap this one and
557 * map the next one if there is any left.
559 if (io_ctl->cur != io_ctl->orig) {
560 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
561 if (io_ctl->index >= io_ctl->num_pages)
563 io_ctl_map_page(io_ctl, 0);
566 memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
567 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
568 if (io_ctl->index < io_ctl->num_pages)
569 io_ctl_map_page(io_ctl, 0);
573 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
576 * If we're not on the boundary we know we've modified the page and we
577 * need to crc the page.
579 if (io_ctl->cur != io_ctl->orig)
580 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
582 io_ctl_unmap_page(io_ctl);
584 while (io_ctl->index < io_ctl->num_pages) {
585 io_ctl_map_page(io_ctl, 1);
586 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
590 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
591 struct btrfs_free_space *entry, u8 *type)
593 struct btrfs_free_space_entry *e;
597 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
603 entry->offset = le64_to_cpu(e->offset);
604 entry->bytes = le64_to_cpu(e->bytes);
606 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
607 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
609 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
612 io_ctl_unmap_page(io_ctl);
617 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
618 struct btrfs_free_space *entry)
622 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
626 memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
627 io_ctl_unmap_page(io_ctl);
633 * Since we attach pinned extents after the fact we can have contiguous sections
634 * of free space that are split up in entries. This poses a problem with the
635 * tree logging stuff since it could have allocated across what appears to be 2
636 * entries since we would have merged the entries when adding the pinned extents
637 * back to the free space cache. So run through the space cache that we just
638 * loaded and merge contiguous entries. This will make the log replay stuff not
639 * blow up and it will make for nicer allocator behavior.
641 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
643 struct btrfs_free_space *e, *prev = NULL;
647 spin_lock(&ctl->tree_lock);
648 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
649 e = rb_entry(n, struct btrfs_free_space, offset_index);
652 if (e->bitmap || prev->bitmap)
654 if (prev->offset + prev->bytes == e->offset) {
655 unlink_free_space(ctl, prev);
656 unlink_free_space(ctl, e);
657 prev->bytes += e->bytes;
658 kmem_cache_free(btrfs_free_space_cachep, e);
659 link_free_space(ctl, prev);
661 spin_unlock(&ctl->tree_lock);
667 spin_unlock(&ctl->tree_lock);
670 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
671 struct btrfs_free_space_ctl *ctl,
672 struct btrfs_path *path, u64 offset)
674 struct btrfs_free_space_header *header;
675 struct extent_buffer *leaf;
676 struct btrfs_io_ctl io_ctl;
677 struct btrfs_key key;
678 struct btrfs_free_space *e, *n;
686 /* Nothing in the space cache, goodbye */
687 if (!i_size_read(inode))
690 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
694 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
698 btrfs_release_path(path);
704 leaf = path->nodes[0];
705 header = btrfs_item_ptr(leaf, path->slots[0],
706 struct btrfs_free_space_header);
707 num_entries = btrfs_free_space_entries(leaf, header);
708 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
709 generation = btrfs_free_space_generation(leaf, header);
710 btrfs_release_path(path);
712 if (!BTRFS_I(inode)->generation) {
713 btrfs_info(root->fs_info,
714 "The free space cache file (%llu) is invalid. skip it\n",
719 if (BTRFS_I(inode)->generation != generation) {
720 btrfs_err(root->fs_info,
721 "free space inode generation (%llu) "
722 "did not match free space cache generation (%llu)",
723 BTRFS_I(inode)->generation, generation);
730 ret = io_ctl_init(&io_ctl, inode, root, 0);
734 ret = readahead_cache(inode);
738 ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
742 ret = io_ctl_check_crc(&io_ctl, 0);
746 ret = io_ctl_check_generation(&io_ctl, generation);
750 while (num_entries) {
751 e = kmem_cache_zalloc(btrfs_free_space_cachep,
756 ret = io_ctl_read_entry(&io_ctl, e, &type);
758 kmem_cache_free(btrfs_free_space_cachep, e);
763 kmem_cache_free(btrfs_free_space_cachep, e);
767 if (type == BTRFS_FREE_SPACE_EXTENT) {
768 spin_lock(&ctl->tree_lock);
769 ret = link_free_space(ctl, e);
770 spin_unlock(&ctl->tree_lock);
772 btrfs_err(root->fs_info,
773 "Duplicate entries in free space cache, dumping");
774 kmem_cache_free(btrfs_free_space_cachep, e);
780 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
783 btrfs_free_space_cachep, e);
786 spin_lock(&ctl->tree_lock);
787 ret = link_free_space(ctl, e);
788 ctl->total_bitmaps++;
789 ctl->op->recalc_thresholds(ctl);
790 spin_unlock(&ctl->tree_lock);
792 btrfs_err(root->fs_info,
793 "Duplicate entries in free space cache, dumping");
794 kmem_cache_free(btrfs_free_space_cachep, e);
797 list_add_tail(&e->list, &bitmaps);
803 io_ctl_unmap_page(&io_ctl);
806 * We add the bitmaps at the end of the entries in order that
807 * the bitmap entries are added to the cache.
809 list_for_each_entry_safe(e, n, &bitmaps, list) {
810 list_del_init(&e->list);
811 ret = io_ctl_read_bitmap(&io_ctl, e);
816 io_ctl_drop_pages(&io_ctl);
817 merge_space_tree(ctl);
820 io_ctl_free(&io_ctl);
823 io_ctl_drop_pages(&io_ctl);
824 __btrfs_remove_free_space_cache(ctl);
828 int load_free_space_cache(struct btrfs_fs_info *fs_info,
829 struct btrfs_block_group_cache *block_group)
831 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
832 struct btrfs_root *root = fs_info->tree_root;
834 struct btrfs_path *path;
837 u64 used = btrfs_block_group_used(&block_group->item);
840 * If this block group has been marked to be cleared for one reason or
841 * another then we can't trust the on disk cache, so just return.
843 spin_lock(&block_group->lock);
844 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
845 spin_unlock(&block_group->lock);
848 spin_unlock(&block_group->lock);
850 path = btrfs_alloc_path();
853 path->search_commit_root = 1;
854 path->skip_locking = 1;
856 inode = lookup_free_space_inode(root, block_group, path);
858 btrfs_free_path(path);
862 /* We may have converted the inode and made the cache invalid. */
863 spin_lock(&block_group->lock);
864 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
865 spin_unlock(&block_group->lock);
866 btrfs_free_path(path);
869 spin_unlock(&block_group->lock);
871 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
872 path, block_group->key.objectid);
873 btrfs_free_path(path);
877 spin_lock(&ctl->tree_lock);
878 matched = (ctl->free_space == (block_group->key.offset - used -
879 block_group->bytes_super));
880 spin_unlock(&ctl->tree_lock);
883 __btrfs_remove_free_space_cache(ctl);
884 btrfs_warn(fs_info, "block group %llu has wrong amount of free space",
885 block_group->key.objectid);
890 /* This cache is bogus, make sure it gets cleared */
891 spin_lock(&block_group->lock);
892 block_group->disk_cache_state = BTRFS_DC_CLEAR;
893 spin_unlock(&block_group->lock);
896 btrfs_warn(fs_info, "failed to load free space cache for block group %llu, rebuild it now",
897 block_group->key.objectid);
904 static noinline_for_stack
905 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
906 struct btrfs_free_space_ctl *ctl,
907 struct btrfs_block_group_cache *block_group,
908 int *entries, int *bitmaps,
909 struct list_head *bitmap_list)
912 struct btrfs_free_cluster *cluster = NULL;
913 struct btrfs_free_cluster *cluster_locked = NULL;
914 struct rb_node *node = rb_first(&ctl->free_space_offset);
915 struct btrfs_trim_range *trim_entry;
917 /* Get the cluster for this block_group if it exists */
918 if (block_group && !list_empty(&block_group->cluster_list)) {
919 cluster = list_entry(block_group->cluster_list.next,
920 struct btrfs_free_cluster,
924 if (!node && cluster) {
925 cluster_locked = cluster;
926 spin_lock(&cluster_locked->lock);
927 node = rb_first(&cluster->root);
931 /* Write out the extent entries */
933 struct btrfs_free_space *e;
935 e = rb_entry(node, struct btrfs_free_space, offset_index);
938 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
944 list_add_tail(&e->list, bitmap_list);
947 node = rb_next(node);
948 if (!node && cluster) {
949 node = rb_first(&cluster->root);
950 cluster_locked = cluster;
951 spin_lock(&cluster_locked->lock);
955 if (cluster_locked) {
956 spin_unlock(&cluster_locked->lock);
957 cluster_locked = NULL;
961 * Make sure we don't miss any range that was removed from our rbtree
962 * because trimming is running. Otherwise after a umount+mount (or crash
963 * after committing the transaction) we would leak free space and get
964 * an inconsistent free space cache report from fsck.
966 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
967 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
968 trim_entry->bytes, NULL);
977 spin_unlock(&cluster_locked->lock);
981 static noinline_for_stack int
982 update_cache_item(struct btrfs_trans_handle *trans,
983 struct btrfs_root *root,
985 struct btrfs_path *path, u64 offset,
986 int entries, int bitmaps)
988 struct btrfs_key key;
989 struct btrfs_free_space_header *header;
990 struct extent_buffer *leaf;
993 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
997 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
999 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1000 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1004 leaf = path->nodes[0];
1006 struct btrfs_key found_key;
1007 ASSERT(path->slots[0]);
1009 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1010 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1011 found_key.offset != offset) {
1012 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1014 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1016 btrfs_release_path(path);
1021 BTRFS_I(inode)->generation = trans->transid;
1022 header = btrfs_item_ptr(leaf, path->slots[0],
1023 struct btrfs_free_space_header);
1024 btrfs_set_free_space_entries(leaf, header, entries);
1025 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1026 btrfs_set_free_space_generation(leaf, header, trans->transid);
1027 btrfs_mark_buffer_dirty(leaf);
1028 btrfs_release_path(path);
1036 static noinline_for_stack int
1037 write_pinned_extent_entries(struct btrfs_root *root,
1038 struct btrfs_block_group_cache *block_group,
1039 struct btrfs_io_ctl *io_ctl,
1042 u64 start, extent_start, extent_end, len;
1043 struct extent_io_tree *unpin = NULL;
1050 * We want to add any pinned extents to our free space cache
1051 * so we don't leak the space
1053 * We shouldn't have switched the pinned extents yet so this is the
1056 unpin = root->fs_info->pinned_extents;
1058 start = block_group->key.objectid;
1060 while (start < block_group->key.objectid + block_group->key.offset) {
1061 ret = find_first_extent_bit(unpin, start,
1062 &extent_start, &extent_end,
1063 EXTENT_DIRTY, NULL);
1067 /* This pinned extent is out of our range */
1068 if (extent_start >= block_group->key.objectid +
1069 block_group->key.offset)
1072 extent_start = max(extent_start, start);
1073 extent_end = min(block_group->key.objectid +
1074 block_group->key.offset, extent_end + 1);
1075 len = extent_end - extent_start;
1078 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1088 static noinline_for_stack int
1089 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1091 struct list_head *pos, *n;
1094 /* Write out the bitmaps */
1095 list_for_each_safe(pos, n, bitmap_list) {
1096 struct btrfs_free_space *entry =
1097 list_entry(pos, struct btrfs_free_space, list);
1099 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1102 list_del_init(&entry->list);
1108 static int flush_dirty_cache(struct inode *inode)
1112 ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1114 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1115 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1121 static void noinline_for_stack
1122 cleanup_write_cache_enospc(struct inode *inode,
1123 struct btrfs_io_ctl *io_ctl,
1124 struct extent_state **cached_state,
1125 struct list_head *bitmap_list)
1127 struct list_head *pos, *n;
1129 list_for_each_safe(pos, n, bitmap_list) {
1130 struct btrfs_free_space *entry =
1131 list_entry(pos, struct btrfs_free_space, list);
1132 list_del_init(&entry->list);
1134 io_ctl_drop_pages(io_ctl);
1135 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1136 i_size_read(inode) - 1, cached_state,
1140 int btrfs_wait_cache_io(struct btrfs_root *root,
1141 struct btrfs_trans_handle *trans,
1142 struct btrfs_block_group_cache *block_group,
1143 struct btrfs_io_ctl *io_ctl,
1144 struct btrfs_path *path, u64 offset)
1147 struct inode *inode = io_ctl->inode;
1152 root = root->fs_info->tree_root;
1154 /* Flush the dirty pages in the cache file. */
1155 ret = flush_dirty_cache(inode);
1159 /* Update the cache item to tell everyone this cache file is valid. */
1160 ret = update_cache_item(trans, root, inode, path, offset,
1161 io_ctl->entries, io_ctl->bitmaps);
1163 io_ctl_free(io_ctl);
1165 invalidate_inode_pages2(inode->i_mapping);
1166 BTRFS_I(inode)->generation = 0;
1169 btrfs_err(root->fs_info,
1170 "failed to write free space cache for block group %llu",
1171 block_group->key.objectid);
1175 btrfs_update_inode(trans, root, inode);
1178 /* the dirty list is protected by the dirty_bgs_lock */
1179 spin_lock(&trans->transaction->dirty_bgs_lock);
1181 /* the disk_cache_state is protected by the block group lock */
1182 spin_lock(&block_group->lock);
1185 * only mark this as written if we didn't get put back on
1186 * the dirty list while waiting for IO. Otherwise our
1187 * cache state won't be right, and we won't get written again
1189 if (!ret && list_empty(&block_group->dirty_list))
1190 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1192 block_group->disk_cache_state = BTRFS_DC_ERROR;
1194 spin_unlock(&block_group->lock);
1195 spin_unlock(&trans->transaction->dirty_bgs_lock);
1196 io_ctl->inode = NULL;
1205 * __btrfs_write_out_cache - write out cached info to an inode
1206 * @root - the root the inode belongs to
1207 * @ctl - the free space cache we are going to write out
1208 * @block_group - the block_group for this cache if it belongs to a block_group
1209 * @trans - the trans handle
1210 * @path - the path to use
1211 * @offset - the offset for the key we'll insert
1213 * This function writes out a free space cache struct to disk for quick recovery
1214 * on mount. This will return 0 if it was successfull in writing the cache out,
1215 * and -1 if it was not.
1217 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1218 struct btrfs_free_space_ctl *ctl,
1219 struct btrfs_block_group_cache *block_group,
1220 struct btrfs_io_ctl *io_ctl,
1221 struct btrfs_trans_handle *trans,
1222 struct btrfs_path *path, u64 offset)
1224 struct extent_state *cached_state = NULL;
1225 LIST_HEAD(bitmap_list);
1231 if (!i_size_read(inode))
1234 WARN_ON(io_ctl->pages);
1235 ret = io_ctl_init(io_ctl, inode, root, 1);
1239 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1240 down_write(&block_group->data_rwsem);
1241 spin_lock(&block_group->lock);
1242 if (block_group->delalloc_bytes) {
1243 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1244 spin_unlock(&block_group->lock);
1245 up_write(&block_group->data_rwsem);
1246 BTRFS_I(inode)->generation = 0;
1251 spin_unlock(&block_group->lock);
1254 /* Lock all pages first so we can lock the extent safely. */
1255 io_ctl_prepare_pages(io_ctl, inode, 0);
1257 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1260 io_ctl_set_generation(io_ctl, trans->transid);
1262 mutex_lock(&ctl->cache_writeout_mutex);
1263 /* Write out the extent entries in the free space cache */
1264 spin_lock(&ctl->tree_lock);
1265 ret = write_cache_extent_entries(io_ctl, ctl,
1266 block_group, &entries, &bitmaps,
1268 spin_unlock(&ctl->tree_lock);
1270 mutex_unlock(&ctl->cache_writeout_mutex);
1275 * Some spaces that are freed in the current transaction are pinned,
1276 * they will be added into free space cache after the transaction is
1277 * committed, we shouldn't lose them.
1279 * If this changes while we are working we'll get added back to
1280 * the dirty list and redo it. No locking needed
1282 ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1284 mutex_unlock(&ctl->cache_writeout_mutex);
1289 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1290 * locked while doing it because a concurrent trim can be manipulating
1291 * or freeing the bitmap.
1293 spin_lock(&ctl->tree_lock);
1294 ret = write_bitmap_entries(io_ctl, &bitmap_list);
1295 spin_unlock(&ctl->tree_lock);
1296 mutex_unlock(&ctl->cache_writeout_mutex);
1300 /* Zero out the rest of the pages just to make sure */
1301 io_ctl_zero_remaining_pages(io_ctl);
1303 /* Everything is written out, now we dirty the pages in the file. */
1304 ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1305 0, i_size_read(inode), &cached_state);
1309 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1310 up_write(&block_group->data_rwsem);
1312 * Release the pages and unlock the extent, we will flush
1315 io_ctl_drop_pages(io_ctl);
1317 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1318 i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1321 * at this point the pages are under IO and we're happy,
1322 * The caller is responsible for waiting on them and updating the
1323 * the cache and the inode
1325 io_ctl->entries = entries;
1326 io_ctl->bitmaps = bitmaps;
1328 ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1335 io_ctl->inode = NULL;
1336 io_ctl_free(io_ctl);
1338 invalidate_inode_pages2(inode->i_mapping);
1339 BTRFS_I(inode)->generation = 0;
1341 btrfs_update_inode(trans, root, inode);
1347 cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1349 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1350 up_write(&block_group->data_rwsem);
1355 int btrfs_write_out_cache(struct btrfs_root *root,
1356 struct btrfs_trans_handle *trans,
1357 struct btrfs_block_group_cache *block_group,
1358 struct btrfs_path *path)
1360 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1361 struct inode *inode;
1364 root = root->fs_info->tree_root;
1366 spin_lock(&block_group->lock);
1367 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1368 spin_unlock(&block_group->lock);
1371 spin_unlock(&block_group->lock);
1373 inode = lookup_free_space_inode(root, block_group, path);
1377 ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1378 &block_group->io_ctl, trans,
1379 path, block_group->key.objectid);
1382 btrfs_err(root->fs_info,
1383 "failed to write free space cache for block group %llu",
1384 block_group->key.objectid);
1386 spin_lock(&block_group->lock);
1387 block_group->disk_cache_state = BTRFS_DC_ERROR;
1388 spin_unlock(&block_group->lock);
1390 block_group->io_ctl.inode = NULL;
1395 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1396 * to wait for IO and put the inode
1402 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1405 ASSERT(offset >= bitmap_start);
1406 offset -= bitmap_start;
1407 return (unsigned long)(div_u64(offset, unit));
1410 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1412 return (unsigned long)(div_u64(bytes, unit));
1415 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1419 u32 bytes_per_bitmap;
1421 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1422 bitmap_start = offset - ctl->start;
1423 bitmap_start = div_u64(bitmap_start, bytes_per_bitmap);
1424 bitmap_start *= bytes_per_bitmap;
1425 bitmap_start += ctl->start;
1427 return bitmap_start;
1430 static int tree_insert_offset(struct rb_root *root, u64 offset,
1431 struct rb_node *node, int bitmap)
1433 struct rb_node **p = &root->rb_node;
1434 struct rb_node *parent = NULL;
1435 struct btrfs_free_space *info;
1439 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1441 if (offset < info->offset) {
1443 } else if (offset > info->offset) {
1444 p = &(*p)->rb_right;
1447 * we could have a bitmap entry and an extent entry
1448 * share the same offset. If this is the case, we want
1449 * the extent entry to always be found first if we do a
1450 * linear search through the tree, since we want to have
1451 * the quickest allocation time, and allocating from an
1452 * extent is faster than allocating from a bitmap. So
1453 * if we're inserting a bitmap and we find an entry at
1454 * this offset, we want to go right, or after this entry
1455 * logically. If we are inserting an extent and we've
1456 * found a bitmap, we want to go left, or before
1464 p = &(*p)->rb_right;
1466 if (!info->bitmap) {
1475 rb_link_node(node, parent, p);
1476 rb_insert_color(node, root);
1482 * searches the tree for the given offset.
1484 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1485 * want a section that has at least bytes size and comes at or after the given
1488 static struct btrfs_free_space *
1489 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1490 u64 offset, int bitmap_only, int fuzzy)
1492 struct rb_node *n = ctl->free_space_offset.rb_node;
1493 struct btrfs_free_space *entry, *prev = NULL;
1495 /* find entry that is closest to the 'offset' */
1502 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1505 if (offset < entry->offset)
1507 else if (offset > entry->offset)
1520 * bitmap entry and extent entry may share same offset,
1521 * in that case, bitmap entry comes after extent entry.
1526 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1527 if (entry->offset != offset)
1530 WARN_ON(!entry->bitmap);
1533 if (entry->bitmap) {
1535 * if previous extent entry covers the offset,
1536 * we should return it instead of the bitmap entry
1538 n = rb_prev(&entry->offset_index);
1540 prev = rb_entry(n, struct btrfs_free_space,
1542 if (!prev->bitmap &&
1543 prev->offset + prev->bytes > offset)
1553 /* find last entry before the 'offset' */
1555 if (entry->offset > offset) {
1556 n = rb_prev(&entry->offset_index);
1558 entry = rb_entry(n, struct btrfs_free_space,
1560 ASSERT(entry->offset <= offset);
1569 if (entry->bitmap) {
1570 n = rb_prev(&entry->offset_index);
1572 prev = rb_entry(n, struct btrfs_free_space,
1574 if (!prev->bitmap &&
1575 prev->offset + prev->bytes > offset)
1578 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1580 } else if (entry->offset + entry->bytes > offset)
1587 if (entry->bitmap) {
1588 if (entry->offset + BITS_PER_BITMAP *
1592 if (entry->offset + entry->bytes > offset)
1596 n = rb_next(&entry->offset_index);
1599 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1605 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1606 struct btrfs_free_space *info)
1608 rb_erase(&info->offset_index, &ctl->free_space_offset);
1609 ctl->free_extents--;
1612 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1613 struct btrfs_free_space *info)
1615 __unlink_free_space(ctl, info);
1616 ctl->free_space -= info->bytes;
1619 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1620 struct btrfs_free_space *info)
1624 ASSERT(info->bytes || info->bitmap);
1625 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1626 &info->offset_index, (info->bitmap != NULL));
1630 ctl->free_space += info->bytes;
1631 ctl->free_extents++;
1635 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1637 struct btrfs_block_group_cache *block_group = ctl->private;
1641 u64 size = block_group->key.offset;
1642 u32 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1643 u32 max_bitmaps = div_u64(size + bytes_per_bg - 1, bytes_per_bg);
1645 max_bitmaps = max_t(u32, max_bitmaps, 1);
1647 ASSERT(ctl->total_bitmaps <= max_bitmaps);
1650 * The goal is to keep the total amount of memory used per 1gb of space
1651 * at or below 32k, so we need to adjust how much memory we allow to be
1652 * used by extent based free space tracking
1654 if (size < 1024 * 1024 * 1024)
1655 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1657 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1658 div_u64(size, 1024 * 1024 * 1024);
1661 * we want to account for 1 more bitmap than what we have so we can make
1662 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1663 * we add more bitmaps.
1665 bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1667 if (bitmap_bytes >= max_bytes) {
1668 ctl->extents_thresh = 0;
1673 * we want the extent entry threshold to always be at most 1/2 the max
1674 * bytes we can have, or whatever is less than that.
1676 extent_bytes = max_bytes - bitmap_bytes;
1677 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1679 ctl->extents_thresh =
1680 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1683 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1684 struct btrfs_free_space *info,
1685 u64 offset, u64 bytes)
1687 unsigned long start, count;
1689 start = offset_to_bit(info->offset, ctl->unit, offset);
1690 count = bytes_to_bits(bytes, ctl->unit);
1691 ASSERT(start + count <= BITS_PER_BITMAP);
1693 bitmap_clear(info->bitmap, start, count);
1695 info->bytes -= bytes;
1698 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1699 struct btrfs_free_space *info, u64 offset,
1702 __bitmap_clear_bits(ctl, info, offset, bytes);
1703 ctl->free_space -= bytes;
1706 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1707 struct btrfs_free_space *info, u64 offset,
1710 unsigned long start, count;
1712 start = offset_to_bit(info->offset, ctl->unit, offset);
1713 count = bytes_to_bits(bytes, ctl->unit);
1714 ASSERT(start + count <= BITS_PER_BITMAP);
1716 bitmap_set(info->bitmap, start, count);
1718 info->bytes += bytes;
1719 ctl->free_space += bytes;
1723 * If we can not find suitable extent, we will use bytes to record
1724 * the size of the max extent.
1726 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1727 struct btrfs_free_space *bitmap_info, u64 *offset,
1730 unsigned long found_bits = 0;
1731 unsigned long max_bits = 0;
1732 unsigned long bits, i;
1733 unsigned long next_zero;
1734 unsigned long extent_bits;
1736 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1737 max_t(u64, *offset, bitmap_info->offset));
1738 bits = bytes_to_bits(*bytes, ctl->unit);
1740 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1741 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1742 BITS_PER_BITMAP, i);
1743 extent_bits = next_zero - i;
1744 if (extent_bits >= bits) {
1745 found_bits = extent_bits;
1747 } else if (extent_bits > max_bits) {
1748 max_bits = extent_bits;
1754 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1755 *bytes = (u64)(found_bits) * ctl->unit;
1759 *bytes = (u64)(max_bits) * ctl->unit;
1763 /* Cache the size of the max extent in bytes */
1764 static struct btrfs_free_space *
1765 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1766 unsigned long align, u64 *max_extent_size)
1768 struct btrfs_free_space *entry;
1769 struct rb_node *node;
1774 if (!ctl->free_space_offset.rb_node)
1777 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1781 for (node = &entry->offset_index; node; node = rb_next(node)) {
1782 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1783 if (entry->bytes < *bytes) {
1784 if (entry->bytes > *max_extent_size)
1785 *max_extent_size = entry->bytes;
1789 /* make sure the space returned is big enough
1790 * to match our requested alignment
1792 if (*bytes >= align) {
1793 tmp = entry->offset - ctl->start + align - 1;
1794 tmp = div64_u64(tmp, align);
1795 tmp = tmp * align + ctl->start;
1796 align_off = tmp - entry->offset;
1799 tmp = entry->offset;
1802 if (entry->bytes < *bytes + align_off) {
1803 if (entry->bytes > *max_extent_size)
1804 *max_extent_size = entry->bytes;
1808 if (entry->bitmap) {
1811 ret = search_bitmap(ctl, entry, &tmp, &size);
1816 } else if (size > *max_extent_size) {
1817 *max_extent_size = size;
1823 *bytes = entry->bytes - align_off;
1830 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1831 struct btrfs_free_space *info, u64 offset)
1833 info->offset = offset_to_bitmap(ctl, offset);
1835 INIT_LIST_HEAD(&info->list);
1836 link_free_space(ctl, info);
1837 ctl->total_bitmaps++;
1839 ctl->op->recalc_thresholds(ctl);
1842 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1843 struct btrfs_free_space *bitmap_info)
1845 unlink_free_space(ctl, bitmap_info);
1846 kfree(bitmap_info->bitmap);
1847 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1848 ctl->total_bitmaps--;
1849 ctl->op->recalc_thresholds(ctl);
1852 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1853 struct btrfs_free_space *bitmap_info,
1854 u64 *offset, u64 *bytes)
1857 u64 search_start, search_bytes;
1861 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1864 * We need to search for bits in this bitmap. We could only cover some
1865 * of the extent in this bitmap thanks to how we add space, so we need
1866 * to search for as much as it as we can and clear that amount, and then
1867 * go searching for the next bit.
1869 search_start = *offset;
1870 search_bytes = ctl->unit;
1871 search_bytes = min(search_bytes, end - search_start + 1);
1872 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1873 if (ret < 0 || search_start != *offset)
1876 /* We may have found more bits than what we need */
1877 search_bytes = min(search_bytes, *bytes);
1879 /* Cannot clear past the end of the bitmap */
1880 search_bytes = min(search_bytes, end - search_start + 1);
1882 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1883 *offset += search_bytes;
1884 *bytes -= search_bytes;
1887 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1888 if (!bitmap_info->bytes)
1889 free_bitmap(ctl, bitmap_info);
1892 * no entry after this bitmap, but we still have bytes to
1893 * remove, so something has gone wrong.
1898 bitmap_info = rb_entry(next, struct btrfs_free_space,
1902 * if the next entry isn't a bitmap we need to return to let the
1903 * extent stuff do its work.
1905 if (!bitmap_info->bitmap)
1909 * Ok the next item is a bitmap, but it may not actually hold
1910 * the information for the rest of this free space stuff, so
1911 * look for it, and if we don't find it return so we can try
1912 * everything over again.
1914 search_start = *offset;
1915 search_bytes = ctl->unit;
1916 ret = search_bitmap(ctl, bitmap_info, &search_start,
1918 if (ret < 0 || search_start != *offset)
1922 } else if (!bitmap_info->bytes)
1923 free_bitmap(ctl, bitmap_info);
1928 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1929 struct btrfs_free_space *info, u64 offset,
1932 u64 bytes_to_set = 0;
1935 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1937 bytes_to_set = min(end - offset, bytes);
1939 bitmap_set_bits(ctl, info, offset, bytes_to_set);
1941 return bytes_to_set;
1945 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1946 struct btrfs_free_space *info)
1948 struct btrfs_block_group_cache *block_group = ctl->private;
1951 * If we are below the extents threshold then we can add this as an
1952 * extent, and don't have to deal with the bitmap
1954 if (ctl->free_extents < ctl->extents_thresh) {
1956 * If this block group has some small extents we don't want to
1957 * use up all of our free slots in the cache with them, we want
1958 * to reserve them to larger extents, however if we have plent
1959 * of cache left then go ahead an dadd them, no sense in adding
1960 * the overhead of a bitmap if we don't have to.
1962 if (info->bytes <= block_group->sectorsize * 4) {
1963 if (ctl->free_extents * 2 <= ctl->extents_thresh)
1971 * The original block groups from mkfs can be really small, like 8
1972 * megabytes, so don't bother with a bitmap for those entries. However
1973 * some block groups can be smaller than what a bitmap would cover but
1974 * are still large enough that they could overflow the 32k memory limit,
1975 * so allow those block groups to still be allowed to have a bitmap
1978 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
1984 static struct btrfs_free_space_op free_space_op = {
1985 .recalc_thresholds = recalculate_thresholds,
1986 .use_bitmap = use_bitmap,
1989 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1990 struct btrfs_free_space *info)
1992 struct btrfs_free_space *bitmap_info;
1993 struct btrfs_block_group_cache *block_group = NULL;
1995 u64 bytes, offset, bytes_added;
1998 bytes = info->bytes;
1999 offset = info->offset;
2001 if (!ctl->op->use_bitmap(ctl, info))
2004 if (ctl->op == &free_space_op)
2005 block_group = ctl->private;
2008 * Since we link bitmaps right into the cluster we need to see if we
2009 * have a cluster here, and if so and it has our bitmap we need to add
2010 * the free space to that bitmap.
2012 if (block_group && !list_empty(&block_group->cluster_list)) {
2013 struct btrfs_free_cluster *cluster;
2014 struct rb_node *node;
2015 struct btrfs_free_space *entry;
2017 cluster = list_entry(block_group->cluster_list.next,
2018 struct btrfs_free_cluster,
2020 spin_lock(&cluster->lock);
2021 node = rb_first(&cluster->root);
2023 spin_unlock(&cluster->lock);
2024 goto no_cluster_bitmap;
2027 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2028 if (!entry->bitmap) {
2029 spin_unlock(&cluster->lock);
2030 goto no_cluster_bitmap;
2033 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2034 bytes_added = add_bytes_to_bitmap(ctl, entry,
2036 bytes -= bytes_added;
2037 offset += bytes_added;
2039 spin_unlock(&cluster->lock);
2047 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2054 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2055 bytes -= bytes_added;
2056 offset += bytes_added;
2066 if (info && info->bitmap) {
2067 add_new_bitmap(ctl, info, offset);
2072 spin_unlock(&ctl->tree_lock);
2074 /* no pre-allocated info, allocate a new one */
2076 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2079 spin_lock(&ctl->tree_lock);
2085 /* allocate the bitmap */
2086 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
2087 spin_lock(&ctl->tree_lock);
2088 if (!info->bitmap) {
2098 kfree(info->bitmap);
2099 kmem_cache_free(btrfs_free_space_cachep, info);
2105 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2106 struct btrfs_free_space *info, bool update_stat)
2108 struct btrfs_free_space *left_info;
2109 struct btrfs_free_space *right_info;
2110 bool merged = false;
2111 u64 offset = info->offset;
2112 u64 bytes = info->bytes;
2115 * first we want to see if there is free space adjacent to the range we
2116 * are adding, if there is remove that struct and add a new one to
2117 * cover the entire range
2119 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2120 if (right_info && rb_prev(&right_info->offset_index))
2121 left_info = rb_entry(rb_prev(&right_info->offset_index),
2122 struct btrfs_free_space, offset_index);
2124 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2126 if (right_info && !right_info->bitmap) {
2128 unlink_free_space(ctl, right_info);
2130 __unlink_free_space(ctl, right_info);
2131 info->bytes += right_info->bytes;
2132 kmem_cache_free(btrfs_free_space_cachep, right_info);
2136 if (left_info && !left_info->bitmap &&
2137 left_info->offset + left_info->bytes == offset) {
2139 unlink_free_space(ctl, left_info);
2141 __unlink_free_space(ctl, left_info);
2142 info->offset = left_info->offset;
2143 info->bytes += left_info->bytes;
2144 kmem_cache_free(btrfs_free_space_cachep, left_info);
2151 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2152 struct btrfs_free_space *info,
2155 struct btrfs_free_space *bitmap;
2158 const u64 end = info->offset + info->bytes;
2159 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2162 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2166 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2167 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2170 bytes = (j - i) * ctl->unit;
2171 info->bytes += bytes;
2174 bitmap_clear_bits(ctl, bitmap, end, bytes);
2176 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2179 free_bitmap(ctl, bitmap);
2184 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2185 struct btrfs_free_space *info,
2188 struct btrfs_free_space *bitmap;
2192 unsigned long prev_j;
2195 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2196 /* If we're on a boundary, try the previous logical bitmap. */
2197 if (bitmap_offset == info->offset) {
2198 if (info->offset == 0)
2200 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2203 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2207 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2209 prev_j = (unsigned long)-1;
2210 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2218 if (prev_j == (unsigned long)-1)
2219 bytes = (i + 1) * ctl->unit;
2221 bytes = (i - prev_j) * ctl->unit;
2223 info->offset -= bytes;
2224 info->bytes += bytes;
2227 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2229 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2232 free_bitmap(ctl, bitmap);
2238 * We prefer always to allocate from extent entries, both for clustered and
2239 * non-clustered allocation requests. So when attempting to add a new extent
2240 * entry, try to see if there's adjacent free space in bitmap entries, and if
2241 * there is, migrate that space from the bitmaps to the extent.
2242 * Like this we get better chances of satisfying space allocation requests
2243 * because we attempt to satisfy them based on a single cache entry, and never
2244 * on 2 or more entries - even if the entries represent a contiguous free space
2245 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2248 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2249 struct btrfs_free_space *info,
2253 * Only work with disconnected entries, as we can change their offset,
2254 * and must be extent entries.
2256 ASSERT(!info->bitmap);
2257 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2259 if (ctl->total_bitmaps > 0) {
2261 bool stole_front = false;
2263 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2264 if (ctl->total_bitmaps > 0)
2265 stole_front = steal_from_bitmap_to_front(ctl, info,
2268 if (stole_end || stole_front)
2269 try_merge_free_space(ctl, info, update_stat);
2273 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2274 u64 offset, u64 bytes)
2276 struct btrfs_free_space *info;
2279 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2283 info->offset = offset;
2284 info->bytes = bytes;
2285 RB_CLEAR_NODE(&info->offset_index);
2287 spin_lock(&ctl->tree_lock);
2289 if (try_merge_free_space(ctl, info, true))
2293 * There was no extent directly to the left or right of this new
2294 * extent then we know we're going to have to allocate a new extent, so
2295 * before we do that see if we need to drop this into a bitmap
2297 ret = insert_into_bitmap(ctl, info);
2306 * Only steal free space from adjacent bitmaps if we're sure we're not
2307 * going to add the new free space to existing bitmap entries - because
2308 * that would mean unnecessary work that would be reverted. Therefore
2309 * attempt to steal space from bitmaps if we're adding an extent entry.
2311 steal_from_bitmap(ctl, info, true);
2313 ret = link_free_space(ctl, info);
2315 kmem_cache_free(btrfs_free_space_cachep, info);
2317 spin_unlock(&ctl->tree_lock);
2320 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2321 ASSERT(ret != -EEXIST);
2327 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2328 u64 offset, u64 bytes)
2330 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2331 struct btrfs_free_space *info;
2333 bool re_search = false;
2335 spin_lock(&ctl->tree_lock);
2342 info = tree_search_offset(ctl, offset, 0, 0);
2345 * oops didn't find an extent that matched the space we wanted
2346 * to remove, look for a bitmap instead
2348 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2352 * If we found a partial bit of our free space in a
2353 * bitmap but then couldn't find the other part this may
2354 * be a problem, so WARN about it.
2362 if (!info->bitmap) {
2363 unlink_free_space(ctl, info);
2364 if (offset == info->offset) {
2365 u64 to_free = min(bytes, info->bytes);
2367 info->bytes -= to_free;
2368 info->offset += to_free;
2370 ret = link_free_space(ctl, info);
2373 kmem_cache_free(btrfs_free_space_cachep, info);
2380 u64 old_end = info->bytes + info->offset;
2382 info->bytes = offset - info->offset;
2383 ret = link_free_space(ctl, info);
2388 /* Not enough bytes in this entry to satisfy us */
2389 if (old_end < offset + bytes) {
2390 bytes -= old_end - offset;
2393 } else if (old_end == offset + bytes) {
2397 spin_unlock(&ctl->tree_lock);
2399 ret = btrfs_add_free_space(block_group, offset + bytes,
2400 old_end - (offset + bytes));
2406 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2407 if (ret == -EAGAIN) {
2412 spin_unlock(&ctl->tree_lock);
2417 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2420 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2421 struct btrfs_free_space *info;
2425 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2426 info = rb_entry(n, struct btrfs_free_space, offset_index);
2427 if (info->bytes >= bytes && !block_group->ro)
2429 btrfs_crit(block_group->fs_info,
2430 "entry offset %llu, bytes %llu, bitmap %s",
2431 info->offset, info->bytes,
2432 (info->bitmap) ? "yes" : "no");
2434 btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2435 list_empty(&block_group->cluster_list) ? "no" : "yes");
2436 btrfs_info(block_group->fs_info,
2437 "%d blocks of free space at or bigger than bytes is", count);
2440 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2442 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2444 spin_lock_init(&ctl->tree_lock);
2445 ctl->unit = block_group->sectorsize;
2446 ctl->start = block_group->key.objectid;
2447 ctl->private = block_group;
2448 ctl->op = &free_space_op;
2449 INIT_LIST_HEAD(&ctl->trimming_ranges);
2450 mutex_init(&ctl->cache_writeout_mutex);
2453 * we only want to have 32k of ram per block group for keeping
2454 * track of free space, and if we pass 1/2 of that we want to
2455 * start converting things over to using bitmaps
2457 ctl->extents_thresh = ((1024 * 32) / 2) /
2458 sizeof(struct btrfs_free_space);
2462 * for a given cluster, put all of its extents back into the free
2463 * space cache. If the block group passed doesn't match the block group
2464 * pointed to by the cluster, someone else raced in and freed the
2465 * cluster already. In that case, we just return without changing anything
2468 __btrfs_return_cluster_to_free_space(
2469 struct btrfs_block_group_cache *block_group,
2470 struct btrfs_free_cluster *cluster)
2472 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2473 struct btrfs_free_space *entry;
2474 struct rb_node *node;
2476 spin_lock(&cluster->lock);
2477 if (cluster->block_group != block_group)
2480 cluster->block_group = NULL;
2481 cluster->window_start = 0;
2482 list_del_init(&cluster->block_group_list);
2484 node = rb_first(&cluster->root);
2488 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2489 node = rb_next(&entry->offset_index);
2490 rb_erase(&entry->offset_index, &cluster->root);
2491 RB_CLEAR_NODE(&entry->offset_index);
2493 bitmap = (entry->bitmap != NULL);
2495 try_merge_free_space(ctl, entry, false);
2496 steal_from_bitmap(ctl, entry, false);
2498 tree_insert_offset(&ctl->free_space_offset,
2499 entry->offset, &entry->offset_index, bitmap);
2501 cluster->root = RB_ROOT;
2504 spin_unlock(&cluster->lock);
2505 btrfs_put_block_group(block_group);
2509 static void __btrfs_remove_free_space_cache_locked(
2510 struct btrfs_free_space_ctl *ctl)
2512 struct btrfs_free_space *info;
2513 struct rb_node *node;
2515 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2516 info = rb_entry(node, struct btrfs_free_space, offset_index);
2517 if (!info->bitmap) {
2518 unlink_free_space(ctl, info);
2519 kmem_cache_free(btrfs_free_space_cachep, info);
2521 free_bitmap(ctl, info);
2524 cond_resched_lock(&ctl->tree_lock);
2528 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2530 spin_lock(&ctl->tree_lock);
2531 __btrfs_remove_free_space_cache_locked(ctl);
2532 spin_unlock(&ctl->tree_lock);
2535 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2537 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2538 struct btrfs_free_cluster *cluster;
2539 struct list_head *head;
2541 spin_lock(&ctl->tree_lock);
2542 while ((head = block_group->cluster_list.next) !=
2543 &block_group->cluster_list) {
2544 cluster = list_entry(head, struct btrfs_free_cluster,
2547 WARN_ON(cluster->block_group != block_group);
2548 __btrfs_return_cluster_to_free_space(block_group, cluster);
2550 cond_resched_lock(&ctl->tree_lock);
2552 __btrfs_remove_free_space_cache_locked(ctl);
2553 spin_unlock(&ctl->tree_lock);
2557 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2558 u64 offset, u64 bytes, u64 empty_size,
2559 u64 *max_extent_size)
2561 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2562 struct btrfs_free_space *entry = NULL;
2563 u64 bytes_search = bytes + empty_size;
2566 u64 align_gap_len = 0;
2568 spin_lock(&ctl->tree_lock);
2569 entry = find_free_space(ctl, &offset, &bytes_search,
2570 block_group->full_stripe_len, max_extent_size);
2575 if (entry->bitmap) {
2576 bitmap_clear_bits(ctl, entry, offset, bytes);
2578 free_bitmap(ctl, entry);
2580 unlink_free_space(ctl, entry);
2581 align_gap_len = offset - entry->offset;
2582 align_gap = entry->offset;
2584 entry->offset = offset + bytes;
2585 WARN_ON(entry->bytes < bytes + align_gap_len);
2587 entry->bytes -= bytes + align_gap_len;
2589 kmem_cache_free(btrfs_free_space_cachep, entry);
2591 link_free_space(ctl, entry);
2594 spin_unlock(&ctl->tree_lock);
2597 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2602 * given a cluster, put all of its extents back into the free space
2603 * cache. If a block group is passed, this function will only free
2604 * a cluster that belongs to the passed block group.
2606 * Otherwise, it'll get a reference on the block group pointed to by the
2607 * cluster and remove the cluster from it.
2609 int btrfs_return_cluster_to_free_space(
2610 struct btrfs_block_group_cache *block_group,
2611 struct btrfs_free_cluster *cluster)
2613 struct btrfs_free_space_ctl *ctl;
2616 /* first, get a safe pointer to the block group */
2617 spin_lock(&cluster->lock);
2619 block_group = cluster->block_group;
2621 spin_unlock(&cluster->lock);
2624 } else if (cluster->block_group != block_group) {
2625 /* someone else has already freed it don't redo their work */
2626 spin_unlock(&cluster->lock);
2629 atomic_inc(&block_group->count);
2630 spin_unlock(&cluster->lock);
2632 ctl = block_group->free_space_ctl;
2634 /* now return any extents the cluster had on it */
2635 spin_lock(&ctl->tree_lock);
2636 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2637 spin_unlock(&ctl->tree_lock);
2639 /* finally drop our ref */
2640 btrfs_put_block_group(block_group);
2644 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2645 struct btrfs_free_cluster *cluster,
2646 struct btrfs_free_space *entry,
2647 u64 bytes, u64 min_start,
2648 u64 *max_extent_size)
2650 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2652 u64 search_start = cluster->window_start;
2653 u64 search_bytes = bytes;
2656 search_start = min_start;
2657 search_bytes = bytes;
2659 err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2661 if (search_bytes > *max_extent_size)
2662 *max_extent_size = search_bytes;
2667 __bitmap_clear_bits(ctl, entry, ret, bytes);
2673 * given a cluster, try to allocate 'bytes' from it, returns 0
2674 * if it couldn't find anything suitably large, or a logical disk offset
2675 * if things worked out
2677 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2678 struct btrfs_free_cluster *cluster, u64 bytes,
2679 u64 min_start, u64 *max_extent_size)
2681 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2682 struct btrfs_free_space *entry = NULL;
2683 struct rb_node *node;
2686 spin_lock(&cluster->lock);
2687 if (bytes > cluster->max_size)
2690 if (cluster->block_group != block_group)
2693 node = rb_first(&cluster->root);
2697 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2699 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2700 *max_extent_size = entry->bytes;
2702 if (entry->bytes < bytes ||
2703 (!entry->bitmap && entry->offset < min_start)) {
2704 node = rb_next(&entry->offset_index);
2707 entry = rb_entry(node, struct btrfs_free_space,
2712 if (entry->bitmap) {
2713 ret = btrfs_alloc_from_bitmap(block_group,
2714 cluster, entry, bytes,
2715 cluster->window_start,
2718 node = rb_next(&entry->offset_index);
2721 entry = rb_entry(node, struct btrfs_free_space,
2725 cluster->window_start += bytes;
2727 ret = entry->offset;
2729 entry->offset += bytes;
2730 entry->bytes -= bytes;
2733 if (entry->bytes == 0)
2734 rb_erase(&entry->offset_index, &cluster->root);
2738 spin_unlock(&cluster->lock);
2743 spin_lock(&ctl->tree_lock);
2745 ctl->free_space -= bytes;
2746 if (entry->bytes == 0) {
2747 ctl->free_extents--;
2748 if (entry->bitmap) {
2749 kfree(entry->bitmap);
2750 ctl->total_bitmaps--;
2751 ctl->op->recalc_thresholds(ctl);
2753 kmem_cache_free(btrfs_free_space_cachep, entry);
2756 spin_unlock(&ctl->tree_lock);
2761 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2762 struct btrfs_free_space *entry,
2763 struct btrfs_free_cluster *cluster,
2764 u64 offset, u64 bytes,
2765 u64 cont1_bytes, u64 min_bytes)
2767 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2768 unsigned long next_zero;
2770 unsigned long want_bits;
2771 unsigned long min_bits;
2772 unsigned long found_bits;
2773 unsigned long start = 0;
2774 unsigned long total_found = 0;
2777 i = offset_to_bit(entry->offset, ctl->unit,
2778 max_t(u64, offset, entry->offset));
2779 want_bits = bytes_to_bits(bytes, ctl->unit);
2780 min_bits = bytes_to_bits(min_bytes, ctl->unit);
2784 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2785 next_zero = find_next_zero_bit(entry->bitmap,
2786 BITS_PER_BITMAP, i);
2787 if (next_zero - i >= min_bits) {
2788 found_bits = next_zero - i;
2799 cluster->max_size = 0;
2802 total_found += found_bits;
2804 if (cluster->max_size < found_bits * ctl->unit)
2805 cluster->max_size = found_bits * ctl->unit;
2807 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2812 cluster->window_start = start * ctl->unit + entry->offset;
2813 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2814 ret = tree_insert_offset(&cluster->root, entry->offset,
2815 &entry->offset_index, 1);
2816 ASSERT(!ret); /* -EEXIST; Logic error */
2818 trace_btrfs_setup_cluster(block_group, cluster,
2819 total_found * ctl->unit, 1);
2824 * This searches the block group for just extents to fill the cluster with.
2825 * Try to find a cluster with at least bytes total bytes, at least one
2826 * extent of cont1_bytes, and other clusters of at least min_bytes.
2829 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2830 struct btrfs_free_cluster *cluster,
2831 struct list_head *bitmaps, u64 offset, u64 bytes,
2832 u64 cont1_bytes, u64 min_bytes)
2834 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2835 struct btrfs_free_space *first = NULL;
2836 struct btrfs_free_space *entry = NULL;
2837 struct btrfs_free_space *last;
2838 struct rb_node *node;
2843 entry = tree_search_offset(ctl, offset, 0, 1);
2848 * We don't want bitmaps, so just move along until we find a normal
2851 while (entry->bitmap || entry->bytes < min_bytes) {
2852 if (entry->bitmap && list_empty(&entry->list))
2853 list_add_tail(&entry->list, bitmaps);
2854 node = rb_next(&entry->offset_index);
2857 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2860 window_free = entry->bytes;
2861 max_extent = entry->bytes;
2865 for (node = rb_next(&entry->offset_index); node;
2866 node = rb_next(&entry->offset_index)) {
2867 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2869 if (entry->bitmap) {
2870 if (list_empty(&entry->list))
2871 list_add_tail(&entry->list, bitmaps);
2875 if (entry->bytes < min_bytes)
2879 window_free += entry->bytes;
2880 if (entry->bytes > max_extent)
2881 max_extent = entry->bytes;
2884 if (window_free < bytes || max_extent < cont1_bytes)
2887 cluster->window_start = first->offset;
2889 node = &first->offset_index;
2892 * now we've found our entries, pull them out of the free space
2893 * cache and put them into the cluster rbtree
2898 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2899 node = rb_next(&entry->offset_index);
2900 if (entry->bitmap || entry->bytes < min_bytes)
2903 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2904 ret = tree_insert_offset(&cluster->root, entry->offset,
2905 &entry->offset_index, 0);
2906 total_size += entry->bytes;
2907 ASSERT(!ret); /* -EEXIST; Logic error */
2908 } while (node && entry != last);
2910 cluster->max_size = max_extent;
2911 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2916 * This specifically looks for bitmaps that may work in the cluster, we assume
2917 * that we have already failed to find extents that will work.
2920 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2921 struct btrfs_free_cluster *cluster,
2922 struct list_head *bitmaps, u64 offset, u64 bytes,
2923 u64 cont1_bytes, u64 min_bytes)
2925 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2926 struct btrfs_free_space *entry;
2928 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2930 if (ctl->total_bitmaps == 0)
2934 * The bitmap that covers offset won't be in the list unless offset
2935 * is just its start offset.
2937 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2938 if (entry->offset != bitmap_offset) {
2939 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2940 if (entry && list_empty(&entry->list))
2941 list_add(&entry->list, bitmaps);
2944 list_for_each_entry(entry, bitmaps, list) {
2945 if (entry->bytes < bytes)
2947 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2948 bytes, cont1_bytes, min_bytes);
2954 * The bitmaps list has all the bitmaps that record free space
2955 * starting after offset, so no more search is required.
2961 * here we try to find a cluster of blocks in a block group. The goal
2962 * is to find at least bytes+empty_size.
2963 * We might not find them all in one contiguous area.
2965 * returns zero and sets up cluster if things worked out, otherwise
2966 * it returns -enospc
2968 int btrfs_find_space_cluster(struct btrfs_root *root,
2969 struct btrfs_block_group_cache *block_group,
2970 struct btrfs_free_cluster *cluster,
2971 u64 offset, u64 bytes, u64 empty_size)
2973 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2974 struct btrfs_free_space *entry, *tmp;
2981 * Choose the minimum extent size we'll require for this
2982 * cluster. For SSD_SPREAD, don't allow any fragmentation.
2983 * For metadata, allow allocates with smaller extents. For
2984 * data, keep it dense.
2986 if (btrfs_test_opt(root, SSD_SPREAD)) {
2987 cont1_bytes = min_bytes = bytes + empty_size;
2988 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2989 cont1_bytes = bytes;
2990 min_bytes = block_group->sectorsize;
2992 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
2993 min_bytes = block_group->sectorsize;
2996 spin_lock(&ctl->tree_lock);
2999 * If we know we don't have enough space to make a cluster don't even
3000 * bother doing all the work to try and find one.
3002 if (ctl->free_space < bytes) {
3003 spin_unlock(&ctl->tree_lock);
3007 spin_lock(&cluster->lock);
3009 /* someone already found a cluster, hooray */
3010 if (cluster->block_group) {
3015 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3018 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3020 cont1_bytes, min_bytes);
3022 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3023 offset, bytes + empty_size,
3024 cont1_bytes, min_bytes);
3026 /* Clear our temporary list */
3027 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3028 list_del_init(&entry->list);
3031 atomic_inc(&block_group->count);
3032 list_add_tail(&cluster->block_group_list,
3033 &block_group->cluster_list);
3034 cluster->block_group = block_group;
3036 trace_btrfs_failed_cluster_setup(block_group);
3039 spin_unlock(&cluster->lock);
3040 spin_unlock(&ctl->tree_lock);
3046 * simple code to zero out a cluster
3048 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3050 spin_lock_init(&cluster->lock);
3051 spin_lock_init(&cluster->refill_lock);
3052 cluster->root = RB_ROOT;
3053 cluster->max_size = 0;
3054 INIT_LIST_HEAD(&cluster->block_group_list);
3055 cluster->block_group = NULL;
3058 static int do_trimming(struct btrfs_block_group_cache *block_group,
3059 u64 *total_trimmed, u64 start, u64 bytes,
3060 u64 reserved_start, u64 reserved_bytes,
3061 struct btrfs_trim_range *trim_entry)
3063 struct btrfs_space_info *space_info = block_group->space_info;
3064 struct btrfs_fs_info *fs_info = block_group->fs_info;
3065 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3070 spin_lock(&space_info->lock);
3071 spin_lock(&block_group->lock);
3072 if (!block_group->ro) {
3073 block_group->reserved += reserved_bytes;
3074 space_info->bytes_reserved += reserved_bytes;
3077 spin_unlock(&block_group->lock);
3078 spin_unlock(&space_info->lock);
3080 ret = btrfs_discard_extent(fs_info->extent_root,
3081 start, bytes, &trimmed);
3083 *total_trimmed += trimmed;
3085 mutex_lock(&ctl->cache_writeout_mutex);
3086 btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3087 list_del(&trim_entry->list);
3088 mutex_unlock(&ctl->cache_writeout_mutex);
3091 spin_lock(&space_info->lock);
3092 spin_lock(&block_group->lock);
3093 if (block_group->ro)
3094 space_info->bytes_readonly += reserved_bytes;
3095 block_group->reserved -= reserved_bytes;
3096 space_info->bytes_reserved -= reserved_bytes;
3097 spin_unlock(&space_info->lock);
3098 spin_unlock(&block_group->lock);
3104 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3105 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3107 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3108 struct btrfs_free_space *entry;
3109 struct rb_node *node;
3115 while (start < end) {
3116 struct btrfs_trim_range trim_entry;
3118 mutex_lock(&ctl->cache_writeout_mutex);
3119 spin_lock(&ctl->tree_lock);
3121 if (ctl->free_space < minlen) {
3122 spin_unlock(&ctl->tree_lock);
3123 mutex_unlock(&ctl->cache_writeout_mutex);
3127 entry = tree_search_offset(ctl, start, 0, 1);
3129 spin_unlock(&ctl->tree_lock);
3130 mutex_unlock(&ctl->cache_writeout_mutex);
3135 while (entry->bitmap) {
3136 node = rb_next(&entry->offset_index);
3138 spin_unlock(&ctl->tree_lock);
3139 mutex_unlock(&ctl->cache_writeout_mutex);
3142 entry = rb_entry(node, struct btrfs_free_space,
3146 if (entry->offset >= end) {
3147 spin_unlock(&ctl->tree_lock);
3148 mutex_unlock(&ctl->cache_writeout_mutex);
3152 extent_start = entry->offset;
3153 extent_bytes = entry->bytes;
3154 start = max(start, extent_start);
3155 bytes = min(extent_start + extent_bytes, end) - start;
3156 if (bytes < minlen) {
3157 spin_unlock(&ctl->tree_lock);
3158 mutex_unlock(&ctl->cache_writeout_mutex);
3162 unlink_free_space(ctl, entry);
3163 kmem_cache_free(btrfs_free_space_cachep, entry);
3165 spin_unlock(&ctl->tree_lock);
3166 trim_entry.start = extent_start;
3167 trim_entry.bytes = extent_bytes;
3168 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3169 mutex_unlock(&ctl->cache_writeout_mutex);
3171 ret = do_trimming(block_group, total_trimmed, start, bytes,
3172 extent_start, extent_bytes, &trim_entry);
3178 if (fatal_signal_pending(current)) {
3189 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3190 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3192 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3193 struct btrfs_free_space *entry;
3197 u64 offset = offset_to_bitmap(ctl, start);
3199 while (offset < end) {
3200 bool next_bitmap = false;
3201 struct btrfs_trim_range trim_entry;
3203 mutex_lock(&ctl->cache_writeout_mutex);
3204 spin_lock(&ctl->tree_lock);
3206 if (ctl->free_space < minlen) {
3207 spin_unlock(&ctl->tree_lock);
3208 mutex_unlock(&ctl->cache_writeout_mutex);
3212 entry = tree_search_offset(ctl, offset, 1, 0);
3214 spin_unlock(&ctl->tree_lock);
3215 mutex_unlock(&ctl->cache_writeout_mutex);
3221 ret2 = search_bitmap(ctl, entry, &start, &bytes);
3222 if (ret2 || start >= end) {
3223 spin_unlock(&ctl->tree_lock);
3224 mutex_unlock(&ctl->cache_writeout_mutex);
3229 bytes = min(bytes, end - start);
3230 if (bytes < minlen) {
3231 spin_unlock(&ctl->tree_lock);
3232 mutex_unlock(&ctl->cache_writeout_mutex);
3236 bitmap_clear_bits(ctl, entry, start, bytes);
3237 if (entry->bytes == 0)
3238 free_bitmap(ctl, entry);
3240 spin_unlock(&ctl->tree_lock);
3241 trim_entry.start = start;
3242 trim_entry.bytes = bytes;
3243 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3244 mutex_unlock(&ctl->cache_writeout_mutex);
3246 ret = do_trimming(block_group, total_trimmed, start, bytes,
3247 start, bytes, &trim_entry);
3252 offset += BITS_PER_BITMAP * ctl->unit;
3255 if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3256 offset += BITS_PER_BITMAP * ctl->unit;
3259 if (fatal_signal_pending(current)) {
3270 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3271 u64 *trimmed, u64 start, u64 end, u64 minlen)
3277 spin_lock(&block_group->lock);
3278 if (block_group->removed) {
3279 spin_unlock(&block_group->lock);
3282 atomic_inc(&block_group->trimming);
3283 spin_unlock(&block_group->lock);
3285 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3289 ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3291 spin_lock(&block_group->lock);
3292 if (atomic_dec_and_test(&block_group->trimming) &&
3293 block_group->removed) {
3294 struct extent_map_tree *em_tree;
3295 struct extent_map *em;
3297 spin_unlock(&block_group->lock);
3299 lock_chunks(block_group->fs_info->chunk_root);
3300 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3301 write_lock(&em_tree->lock);
3302 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3304 BUG_ON(!em); /* logic error, can't happen */
3306 * remove_extent_mapping() will delete us from the pinned_chunks
3307 * list, which is protected by the chunk mutex.
3309 remove_extent_mapping(em_tree, em);
3310 write_unlock(&em_tree->lock);
3311 unlock_chunks(block_group->fs_info->chunk_root);
3313 /* once for us and once for the tree */
3314 free_extent_map(em);
3315 free_extent_map(em);
3318 * We've left one free space entry and other tasks trimming
3319 * this block group have left 1 entry each one. Free them.
3321 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3323 spin_unlock(&block_group->lock);
3330 * Find the left-most item in the cache tree, and then return the
3331 * smallest inode number in the item.
3333 * Note: the returned inode number may not be the smallest one in
3334 * the tree, if the left-most item is a bitmap.
3336 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3338 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3339 struct btrfs_free_space *entry = NULL;
3342 spin_lock(&ctl->tree_lock);
3344 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3347 entry = rb_entry(rb_first(&ctl->free_space_offset),
3348 struct btrfs_free_space, offset_index);
3350 if (!entry->bitmap) {
3351 ino = entry->offset;
3353 unlink_free_space(ctl, entry);
3357 kmem_cache_free(btrfs_free_space_cachep, entry);
3359 link_free_space(ctl, entry);
3365 ret = search_bitmap(ctl, entry, &offset, &count);
3366 /* Logic error; Should be empty if it can't find anything */
3370 bitmap_clear_bits(ctl, entry, offset, 1);
3371 if (entry->bytes == 0)
3372 free_bitmap(ctl, entry);
3375 spin_unlock(&ctl->tree_lock);
3380 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3381 struct btrfs_path *path)
3383 struct inode *inode = NULL;
3385 spin_lock(&root->ino_cache_lock);
3386 if (root->ino_cache_inode)
3387 inode = igrab(root->ino_cache_inode);
3388 spin_unlock(&root->ino_cache_lock);
3392 inode = __lookup_free_space_inode(root, path, 0);
3396 spin_lock(&root->ino_cache_lock);
3397 if (!btrfs_fs_closing(root->fs_info))
3398 root->ino_cache_inode = igrab(inode);
3399 spin_unlock(&root->ino_cache_lock);
3404 int create_free_ino_inode(struct btrfs_root *root,
3405 struct btrfs_trans_handle *trans,
3406 struct btrfs_path *path)
3408 return __create_free_space_inode(root, trans, path,
3409 BTRFS_FREE_INO_OBJECTID, 0);
3412 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3414 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3415 struct btrfs_path *path;
3416 struct inode *inode;
3418 u64 root_gen = btrfs_root_generation(&root->root_item);
3420 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3424 * If we're unmounting then just return, since this does a search on the
3425 * normal root and not the commit root and we could deadlock.
3427 if (btrfs_fs_closing(fs_info))
3430 path = btrfs_alloc_path();
3434 inode = lookup_free_ino_inode(root, path);
3438 if (root_gen != BTRFS_I(inode)->generation)
3441 ret = __load_free_space_cache(root, inode, ctl, path, 0);
3445 "failed to load free ino cache for root %llu",
3446 root->root_key.objectid);
3450 btrfs_free_path(path);
3454 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3455 struct btrfs_trans_handle *trans,
3456 struct btrfs_path *path,
3457 struct inode *inode)
3459 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3461 struct btrfs_io_ctl io_ctl;
3463 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3466 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3468 btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3470 btrfs_delalloc_release_metadata(inode, inode->i_size);
3472 btrfs_err(root->fs_info,
3473 "failed to write free ino cache for root %llu",
3474 root->root_key.objectid);
3481 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3483 * Use this if you need to make a bitmap or extent entry specifically, it
3484 * doesn't do any of the merging that add_free_space does, this acts a lot like
3485 * how the free space cache loading stuff works, so you can get really weird
3488 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3489 u64 offset, u64 bytes, bool bitmap)
3491 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3492 struct btrfs_free_space *info = NULL, *bitmap_info;
3499 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3505 spin_lock(&ctl->tree_lock);
3506 info->offset = offset;
3507 info->bytes = bytes;
3508 ret = link_free_space(ctl, info);
3509 spin_unlock(&ctl->tree_lock);
3511 kmem_cache_free(btrfs_free_space_cachep, info);
3516 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3518 kmem_cache_free(btrfs_free_space_cachep, info);
3523 spin_lock(&ctl->tree_lock);
3524 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3529 add_new_bitmap(ctl, info, offset);
3534 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3535 bytes -= bytes_added;
3536 offset += bytes_added;
3537 spin_unlock(&ctl->tree_lock);
3543 kmem_cache_free(btrfs_free_space_cachep, info);
3550 * Checks to see if the given range is in the free space cache. This is really
3551 * just used to check the absence of space, so if there is free space in the
3552 * range at all we will return 1.
3554 int test_check_exists(struct btrfs_block_group_cache *cache,
3555 u64 offset, u64 bytes)
3557 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3558 struct btrfs_free_space *info;
3561 spin_lock(&ctl->tree_lock);
3562 info = tree_search_offset(ctl, offset, 0, 0);
3564 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3572 u64 bit_off, bit_bytes;
3574 struct btrfs_free_space *tmp;
3577 bit_bytes = ctl->unit;
3578 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes);
3580 if (bit_off == offset) {
3583 } else if (bit_off > offset &&
3584 offset + bytes > bit_off) {
3590 n = rb_prev(&info->offset_index);
3592 tmp = rb_entry(n, struct btrfs_free_space,
3594 if (tmp->offset + tmp->bytes < offset)
3596 if (offset + bytes < tmp->offset) {
3597 n = rb_prev(&info->offset_index);
3604 n = rb_next(&info->offset_index);
3606 tmp = rb_entry(n, struct btrfs_free_space,
3608 if (offset + bytes < tmp->offset)
3610 if (tmp->offset + tmp->bytes < offset) {
3611 n = rb_next(&info->offset_index);
3622 if (info->offset == offset) {
3627 if (offset > info->offset && offset < info->offset + info->bytes)
3630 spin_unlock(&ctl->tree_lock);
3633 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */