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_bitmap_list(struct list_head *bitmap_list)
1124 struct list_head *pos, *n;
1126 list_for_each_safe(pos, n, bitmap_list) {
1127 struct btrfs_free_space *entry =
1128 list_entry(pos, struct btrfs_free_space, list);
1129 list_del_init(&entry->list);
1133 static void noinline_for_stack
1134 cleanup_write_cache_enospc(struct inode *inode,
1135 struct btrfs_io_ctl *io_ctl,
1136 struct extent_state **cached_state,
1137 struct list_head *bitmap_list)
1139 io_ctl_drop_pages(io_ctl);
1140 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1141 i_size_read(inode) - 1, cached_state,
1145 int btrfs_wait_cache_io(struct btrfs_root *root,
1146 struct btrfs_trans_handle *trans,
1147 struct btrfs_block_group_cache *block_group,
1148 struct btrfs_io_ctl *io_ctl,
1149 struct btrfs_path *path, u64 offset)
1152 struct inode *inode = io_ctl->inode;
1158 root = root->fs_info->tree_root;
1160 /* Flush the dirty pages in the cache file. */
1161 ret = flush_dirty_cache(inode);
1165 /* Update the cache item to tell everyone this cache file is valid. */
1166 ret = update_cache_item(trans, root, inode, path, offset,
1167 io_ctl->entries, io_ctl->bitmaps);
1169 io_ctl_free(io_ctl);
1171 invalidate_inode_pages2(inode->i_mapping);
1172 BTRFS_I(inode)->generation = 0;
1175 btrfs_err(root->fs_info,
1176 "failed to write free space cache for block group %llu",
1177 block_group->key.objectid);
1181 btrfs_update_inode(trans, root, inode);
1184 /* the dirty list is protected by the dirty_bgs_lock */
1185 spin_lock(&trans->transaction->dirty_bgs_lock);
1187 /* the disk_cache_state is protected by the block group lock */
1188 spin_lock(&block_group->lock);
1191 * only mark this as written if we didn't get put back on
1192 * the dirty list while waiting for IO. Otherwise our
1193 * cache state won't be right, and we won't get written again
1195 if (!ret && list_empty(&block_group->dirty_list))
1196 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1198 block_group->disk_cache_state = BTRFS_DC_ERROR;
1200 spin_unlock(&block_group->lock);
1201 spin_unlock(&trans->transaction->dirty_bgs_lock);
1202 io_ctl->inode = NULL;
1211 * __btrfs_write_out_cache - write out cached info to an inode
1212 * @root - the root the inode belongs to
1213 * @ctl - the free space cache we are going to write out
1214 * @block_group - the block_group for this cache if it belongs to a block_group
1215 * @trans - the trans handle
1216 * @path - the path to use
1217 * @offset - the offset for the key we'll insert
1219 * This function writes out a free space cache struct to disk for quick recovery
1220 * on mount. This will return 0 if it was successfull in writing the cache out,
1221 * and -1 if it was not.
1223 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1224 struct btrfs_free_space_ctl *ctl,
1225 struct btrfs_block_group_cache *block_group,
1226 struct btrfs_io_ctl *io_ctl,
1227 struct btrfs_trans_handle *trans,
1228 struct btrfs_path *path, u64 offset)
1230 struct extent_state *cached_state = NULL;
1231 LIST_HEAD(bitmap_list);
1237 if (!i_size_read(inode))
1240 WARN_ON(io_ctl->pages);
1241 ret = io_ctl_init(io_ctl, inode, root, 1);
1245 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1246 down_write(&block_group->data_rwsem);
1247 spin_lock(&block_group->lock);
1248 if (block_group->delalloc_bytes) {
1249 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1250 spin_unlock(&block_group->lock);
1251 up_write(&block_group->data_rwsem);
1252 BTRFS_I(inode)->generation = 0;
1257 spin_unlock(&block_group->lock);
1260 /* Lock all pages first so we can lock the extent safely. */
1261 io_ctl_prepare_pages(io_ctl, inode, 0);
1263 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1266 io_ctl_set_generation(io_ctl, trans->transid);
1268 mutex_lock(&ctl->cache_writeout_mutex);
1269 /* Write out the extent entries in the free space cache */
1270 spin_lock(&ctl->tree_lock);
1271 ret = write_cache_extent_entries(io_ctl, ctl,
1272 block_group, &entries, &bitmaps,
1275 goto out_nospc_locked;
1278 * Some spaces that are freed in the current transaction are pinned,
1279 * they will be added into free space cache after the transaction is
1280 * committed, we shouldn't lose them.
1282 * If this changes while we are working we'll get added back to
1283 * the dirty list and redo it. No locking needed
1285 ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1287 goto out_nospc_locked;
1290 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1291 * locked while doing it because a concurrent trim can be manipulating
1292 * or freeing the bitmap.
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_bitmap_list(&bitmap_list);
1348 spin_unlock(&ctl->tree_lock);
1349 mutex_unlock(&ctl->cache_writeout_mutex);
1352 cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1354 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1355 up_write(&block_group->data_rwsem);
1360 int btrfs_write_out_cache(struct btrfs_root *root,
1361 struct btrfs_trans_handle *trans,
1362 struct btrfs_block_group_cache *block_group,
1363 struct btrfs_path *path)
1365 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1366 struct inode *inode;
1369 root = root->fs_info->tree_root;
1371 spin_lock(&block_group->lock);
1372 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1373 spin_unlock(&block_group->lock);
1376 spin_unlock(&block_group->lock);
1378 inode = lookup_free_space_inode(root, block_group, path);
1382 ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1383 &block_group->io_ctl, trans,
1384 path, block_group->key.objectid);
1387 btrfs_err(root->fs_info,
1388 "failed to write free space cache for block group %llu",
1389 block_group->key.objectid);
1391 spin_lock(&block_group->lock);
1392 block_group->disk_cache_state = BTRFS_DC_ERROR;
1393 spin_unlock(&block_group->lock);
1395 block_group->io_ctl.inode = NULL;
1400 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1401 * to wait for IO and put the inode
1407 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1410 ASSERT(offset >= bitmap_start);
1411 offset -= bitmap_start;
1412 return (unsigned long)(div_u64(offset, unit));
1415 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1417 return (unsigned long)(div_u64(bytes, unit));
1420 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1424 u32 bytes_per_bitmap;
1426 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1427 bitmap_start = offset - ctl->start;
1428 bitmap_start = div_u64(bitmap_start, bytes_per_bitmap);
1429 bitmap_start *= bytes_per_bitmap;
1430 bitmap_start += ctl->start;
1432 return bitmap_start;
1435 static int tree_insert_offset(struct rb_root *root, u64 offset,
1436 struct rb_node *node, int bitmap)
1438 struct rb_node **p = &root->rb_node;
1439 struct rb_node *parent = NULL;
1440 struct btrfs_free_space *info;
1444 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1446 if (offset < info->offset) {
1448 } else if (offset > info->offset) {
1449 p = &(*p)->rb_right;
1452 * we could have a bitmap entry and an extent entry
1453 * share the same offset. If this is the case, we want
1454 * the extent entry to always be found first if we do a
1455 * linear search through the tree, since we want to have
1456 * the quickest allocation time, and allocating from an
1457 * extent is faster than allocating from a bitmap. So
1458 * if we're inserting a bitmap and we find an entry at
1459 * this offset, we want to go right, or after this entry
1460 * logically. If we are inserting an extent and we've
1461 * found a bitmap, we want to go left, or before
1469 p = &(*p)->rb_right;
1471 if (!info->bitmap) {
1480 rb_link_node(node, parent, p);
1481 rb_insert_color(node, root);
1487 * searches the tree for the given offset.
1489 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1490 * want a section that has at least bytes size and comes at or after the given
1493 static struct btrfs_free_space *
1494 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1495 u64 offset, int bitmap_only, int fuzzy)
1497 struct rb_node *n = ctl->free_space_offset.rb_node;
1498 struct btrfs_free_space *entry, *prev = NULL;
1500 /* find entry that is closest to the 'offset' */
1507 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1510 if (offset < entry->offset)
1512 else if (offset > entry->offset)
1525 * bitmap entry and extent entry may share same offset,
1526 * in that case, bitmap entry comes after extent entry.
1531 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1532 if (entry->offset != offset)
1535 WARN_ON(!entry->bitmap);
1538 if (entry->bitmap) {
1540 * if previous extent entry covers the offset,
1541 * we should return it instead of the bitmap entry
1543 n = rb_prev(&entry->offset_index);
1545 prev = rb_entry(n, struct btrfs_free_space,
1547 if (!prev->bitmap &&
1548 prev->offset + prev->bytes > offset)
1558 /* find last entry before the 'offset' */
1560 if (entry->offset > offset) {
1561 n = rb_prev(&entry->offset_index);
1563 entry = rb_entry(n, struct btrfs_free_space,
1565 ASSERT(entry->offset <= offset);
1574 if (entry->bitmap) {
1575 n = rb_prev(&entry->offset_index);
1577 prev = rb_entry(n, struct btrfs_free_space,
1579 if (!prev->bitmap &&
1580 prev->offset + prev->bytes > offset)
1583 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1585 } else if (entry->offset + entry->bytes > offset)
1592 if (entry->bitmap) {
1593 if (entry->offset + BITS_PER_BITMAP *
1597 if (entry->offset + entry->bytes > offset)
1601 n = rb_next(&entry->offset_index);
1604 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1610 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1611 struct btrfs_free_space *info)
1613 rb_erase(&info->offset_index, &ctl->free_space_offset);
1614 ctl->free_extents--;
1617 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1618 struct btrfs_free_space *info)
1620 __unlink_free_space(ctl, info);
1621 ctl->free_space -= info->bytes;
1624 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1625 struct btrfs_free_space *info)
1629 ASSERT(info->bytes || info->bitmap);
1630 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1631 &info->offset_index, (info->bitmap != NULL));
1635 ctl->free_space += info->bytes;
1636 ctl->free_extents++;
1640 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1642 struct btrfs_block_group_cache *block_group = ctl->private;
1646 u64 size = block_group->key.offset;
1647 u32 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1648 u32 max_bitmaps = div_u64(size + bytes_per_bg - 1, bytes_per_bg);
1650 max_bitmaps = max_t(u32, max_bitmaps, 1);
1652 ASSERT(ctl->total_bitmaps <= max_bitmaps);
1655 * The goal is to keep the total amount of memory used per 1gb of space
1656 * at or below 32k, so we need to adjust how much memory we allow to be
1657 * used by extent based free space tracking
1659 if (size < 1024 * 1024 * 1024)
1660 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1662 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1663 div_u64(size, 1024 * 1024 * 1024);
1666 * we want to account for 1 more bitmap than what we have so we can make
1667 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1668 * we add more bitmaps.
1670 bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1672 if (bitmap_bytes >= max_bytes) {
1673 ctl->extents_thresh = 0;
1678 * we want the extent entry threshold to always be at most 1/2 the max
1679 * bytes we can have, or whatever is less than that.
1681 extent_bytes = max_bytes - bitmap_bytes;
1682 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1684 ctl->extents_thresh =
1685 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1688 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1689 struct btrfs_free_space *info,
1690 u64 offset, u64 bytes)
1692 unsigned long start, count;
1694 start = offset_to_bit(info->offset, ctl->unit, offset);
1695 count = bytes_to_bits(bytes, ctl->unit);
1696 ASSERT(start + count <= BITS_PER_BITMAP);
1698 bitmap_clear(info->bitmap, start, count);
1700 info->bytes -= bytes;
1703 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1704 struct btrfs_free_space *info, u64 offset,
1707 __bitmap_clear_bits(ctl, info, offset, bytes);
1708 ctl->free_space -= bytes;
1711 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1712 struct btrfs_free_space *info, u64 offset,
1715 unsigned long start, count;
1717 start = offset_to_bit(info->offset, ctl->unit, offset);
1718 count = bytes_to_bits(bytes, ctl->unit);
1719 ASSERT(start + count <= BITS_PER_BITMAP);
1721 bitmap_set(info->bitmap, start, count);
1723 info->bytes += bytes;
1724 ctl->free_space += bytes;
1728 * If we can not find suitable extent, we will use bytes to record
1729 * the size of the max extent.
1731 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1732 struct btrfs_free_space *bitmap_info, u64 *offset,
1735 unsigned long found_bits = 0;
1736 unsigned long max_bits = 0;
1737 unsigned long bits, i;
1738 unsigned long next_zero;
1739 unsigned long extent_bits;
1741 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1742 max_t(u64, *offset, bitmap_info->offset));
1743 bits = bytes_to_bits(*bytes, ctl->unit);
1745 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1746 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1747 BITS_PER_BITMAP, i);
1748 extent_bits = next_zero - i;
1749 if (extent_bits >= bits) {
1750 found_bits = extent_bits;
1752 } else if (extent_bits > max_bits) {
1753 max_bits = extent_bits;
1759 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1760 *bytes = (u64)(found_bits) * ctl->unit;
1764 *bytes = (u64)(max_bits) * ctl->unit;
1768 /* Cache the size of the max extent in bytes */
1769 static struct btrfs_free_space *
1770 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1771 unsigned long align, u64 *max_extent_size)
1773 struct btrfs_free_space *entry;
1774 struct rb_node *node;
1779 if (!ctl->free_space_offset.rb_node)
1782 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1786 for (node = &entry->offset_index; node; node = rb_next(node)) {
1787 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1788 if (entry->bytes < *bytes) {
1789 if (entry->bytes > *max_extent_size)
1790 *max_extent_size = entry->bytes;
1794 /* make sure the space returned is big enough
1795 * to match our requested alignment
1797 if (*bytes >= align) {
1798 tmp = entry->offset - ctl->start + align - 1;
1799 tmp = div64_u64(tmp, align);
1800 tmp = tmp * align + ctl->start;
1801 align_off = tmp - entry->offset;
1804 tmp = entry->offset;
1807 if (entry->bytes < *bytes + align_off) {
1808 if (entry->bytes > *max_extent_size)
1809 *max_extent_size = entry->bytes;
1813 if (entry->bitmap) {
1816 ret = search_bitmap(ctl, entry, &tmp, &size);
1821 } else if (size > *max_extent_size) {
1822 *max_extent_size = size;
1828 *bytes = entry->bytes - align_off;
1835 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1836 struct btrfs_free_space *info, u64 offset)
1838 info->offset = offset_to_bitmap(ctl, offset);
1840 INIT_LIST_HEAD(&info->list);
1841 link_free_space(ctl, info);
1842 ctl->total_bitmaps++;
1844 ctl->op->recalc_thresholds(ctl);
1847 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1848 struct btrfs_free_space *bitmap_info)
1850 unlink_free_space(ctl, bitmap_info);
1851 kfree(bitmap_info->bitmap);
1852 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1853 ctl->total_bitmaps--;
1854 ctl->op->recalc_thresholds(ctl);
1857 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1858 struct btrfs_free_space *bitmap_info,
1859 u64 *offset, u64 *bytes)
1862 u64 search_start, search_bytes;
1866 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1869 * We need to search for bits in this bitmap. We could only cover some
1870 * of the extent in this bitmap thanks to how we add space, so we need
1871 * to search for as much as it as we can and clear that amount, and then
1872 * go searching for the next bit.
1874 search_start = *offset;
1875 search_bytes = ctl->unit;
1876 search_bytes = min(search_bytes, end - search_start + 1);
1877 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1878 if (ret < 0 || search_start != *offset)
1881 /* We may have found more bits than what we need */
1882 search_bytes = min(search_bytes, *bytes);
1884 /* Cannot clear past the end of the bitmap */
1885 search_bytes = min(search_bytes, end - search_start + 1);
1887 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1888 *offset += search_bytes;
1889 *bytes -= search_bytes;
1892 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1893 if (!bitmap_info->bytes)
1894 free_bitmap(ctl, bitmap_info);
1897 * no entry after this bitmap, but we still have bytes to
1898 * remove, so something has gone wrong.
1903 bitmap_info = rb_entry(next, struct btrfs_free_space,
1907 * if the next entry isn't a bitmap we need to return to let the
1908 * extent stuff do its work.
1910 if (!bitmap_info->bitmap)
1914 * Ok the next item is a bitmap, but it may not actually hold
1915 * the information for the rest of this free space stuff, so
1916 * look for it, and if we don't find it return so we can try
1917 * everything over again.
1919 search_start = *offset;
1920 search_bytes = ctl->unit;
1921 ret = search_bitmap(ctl, bitmap_info, &search_start,
1923 if (ret < 0 || search_start != *offset)
1927 } else if (!bitmap_info->bytes)
1928 free_bitmap(ctl, bitmap_info);
1933 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1934 struct btrfs_free_space *info, u64 offset,
1937 u64 bytes_to_set = 0;
1940 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1942 bytes_to_set = min(end - offset, bytes);
1944 bitmap_set_bits(ctl, info, offset, bytes_to_set);
1946 return bytes_to_set;
1950 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1951 struct btrfs_free_space *info)
1953 struct btrfs_block_group_cache *block_group = ctl->private;
1956 * If we are below the extents threshold then we can add this as an
1957 * extent, and don't have to deal with the bitmap
1959 if (ctl->free_extents < ctl->extents_thresh) {
1961 * If this block group has some small extents we don't want to
1962 * use up all of our free slots in the cache with them, we want
1963 * to reserve them to larger extents, however if we have plent
1964 * of cache left then go ahead an dadd them, no sense in adding
1965 * the overhead of a bitmap if we don't have to.
1967 if (info->bytes <= block_group->sectorsize * 4) {
1968 if (ctl->free_extents * 2 <= ctl->extents_thresh)
1976 * The original block groups from mkfs can be really small, like 8
1977 * megabytes, so don't bother with a bitmap for those entries. However
1978 * some block groups can be smaller than what a bitmap would cover but
1979 * are still large enough that they could overflow the 32k memory limit,
1980 * so allow those block groups to still be allowed to have a bitmap
1983 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
1989 static struct btrfs_free_space_op free_space_op = {
1990 .recalc_thresholds = recalculate_thresholds,
1991 .use_bitmap = use_bitmap,
1994 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1995 struct btrfs_free_space *info)
1997 struct btrfs_free_space *bitmap_info;
1998 struct btrfs_block_group_cache *block_group = NULL;
2000 u64 bytes, offset, bytes_added;
2003 bytes = info->bytes;
2004 offset = info->offset;
2006 if (!ctl->op->use_bitmap(ctl, info))
2009 if (ctl->op == &free_space_op)
2010 block_group = ctl->private;
2013 * Since we link bitmaps right into the cluster we need to see if we
2014 * have a cluster here, and if so and it has our bitmap we need to add
2015 * the free space to that bitmap.
2017 if (block_group && !list_empty(&block_group->cluster_list)) {
2018 struct btrfs_free_cluster *cluster;
2019 struct rb_node *node;
2020 struct btrfs_free_space *entry;
2022 cluster = list_entry(block_group->cluster_list.next,
2023 struct btrfs_free_cluster,
2025 spin_lock(&cluster->lock);
2026 node = rb_first(&cluster->root);
2028 spin_unlock(&cluster->lock);
2029 goto no_cluster_bitmap;
2032 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2033 if (!entry->bitmap) {
2034 spin_unlock(&cluster->lock);
2035 goto no_cluster_bitmap;
2038 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2039 bytes_added = add_bytes_to_bitmap(ctl, entry,
2041 bytes -= bytes_added;
2042 offset += bytes_added;
2044 spin_unlock(&cluster->lock);
2052 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2059 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2060 bytes -= bytes_added;
2061 offset += bytes_added;
2071 if (info && info->bitmap) {
2072 add_new_bitmap(ctl, info, offset);
2077 spin_unlock(&ctl->tree_lock);
2079 /* no pre-allocated info, allocate a new one */
2081 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2084 spin_lock(&ctl->tree_lock);
2090 /* allocate the bitmap */
2091 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
2092 spin_lock(&ctl->tree_lock);
2093 if (!info->bitmap) {
2103 kfree(info->bitmap);
2104 kmem_cache_free(btrfs_free_space_cachep, info);
2110 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2111 struct btrfs_free_space *info, bool update_stat)
2113 struct btrfs_free_space *left_info;
2114 struct btrfs_free_space *right_info;
2115 bool merged = false;
2116 u64 offset = info->offset;
2117 u64 bytes = info->bytes;
2120 * first we want to see if there is free space adjacent to the range we
2121 * are adding, if there is remove that struct and add a new one to
2122 * cover the entire range
2124 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2125 if (right_info && rb_prev(&right_info->offset_index))
2126 left_info = rb_entry(rb_prev(&right_info->offset_index),
2127 struct btrfs_free_space, offset_index);
2129 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2131 if (right_info && !right_info->bitmap) {
2133 unlink_free_space(ctl, right_info);
2135 __unlink_free_space(ctl, right_info);
2136 info->bytes += right_info->bytes;
2137 kmem_cache_free(btrfs_free_space_cachep, right_info);
2141 if (left_info && !left_info->bitmap &&
2142 left_info->offset + left_info->bytes == offset) {
2144 unlink_free_space(ctl, left_info);
2146 __unlink_free_space(ctl, left_info);
2147 info->offset = left_info->offset;
2148 info->bytes += left_info->bytes;
2149 kmem_cache_free(btrfs_free_space_cachep, left_info);
2156 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2157 struct btrfs_free_space *info,
2160 struct btrfs_free_space *bitmap;
2163 const u64 end = info->offset + info->bytes;
2164 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2167 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2171 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2172 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2175 bytes = (j - i) * ctl->unit;
2176 info->bytes += bytes;
2179 bitmap_clear_bits(ctl, bitmap, end, bytes);
2181 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2184 free_bitmap(ctl, bitmap);
2189 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2190 struct btrfs_free_space *info,
2193 struct btrfs_free_space *bitmap;
2197 unsigned long prev_j;
2200 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2201 /* If we're on a boundary, try the previous logical bitmap. */
2202 if (bitmap_offset == info->offset) {
2203 if (info->offset == 0)
2205 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2208 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2212 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2214 prev_j = (unsigned long)-1;
2215 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2223 if (prev_j == (unsigned long)-1)
2224 bytes = (i + 1) * ctl->unit;
2226 bytes = (i - prev_j) * ctl->unit;
2228 info->offset -= bytes;
2229 info->bytes += bytes;
2232 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2234 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2237 free_bitmap(ctl, bitmap);
2243 * We prefer always to allocate from extent entries, both for clustered and
2244 * non-clustered allocation requests. So when attempting to add a new extent
2245 * entry, try to see if there's adjacent free space in bitmap entries, and if
2246 * there is, migrate that space from the bitmaps to the extent.
2247 * Like this we get better chances of satisfying space allocation requests
2248 * because we attempt to satisfy them based on a single cache entry, and never
2249 * on 2 or more entries - even if the entries represent a contiguous free space
2250 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2253 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2254 struct btrfs_free_space *info,
2258 * Only work with disconnected entries, as we can change their offset,
2259 * and must be extent entries.
2261 ASSERT(!info->bitmap);
2262 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2264 if (ctl->total_bitmaps > 0) {
2266 bool stole_front = false;
2268 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2269 if (ctl->total_bitmaps > 0)
2270 stole_front = steal_from_bitmap_to_front(ctl, info,
2273 if (stole_end || stole_front)
2274 try_merge_free_space(ctl, info, update_stat);
2278 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2279 u64 offset, u64 bytes)
2281 struct btrfs_free_space *info;
2284 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2288 info->offset = offset;
2289 info->bytes = bytes;
2290 RB_CLEAR_NODE(&info->offset_index);
2292 spin_lock(&ctl->tree_lock);
2294 if (try_merge_free_space(ctl, info, true))
2298 * There was no extent directly to the left or right of this new
2299 * extent then we know we're going to have to allocate a new extent, so
2300 * before we do that see if we need to drop this into a bitmap
2302 ret = insert_into_bitmap(ctl, info);
2311 * Only steal free space from adjacent bitmaps if we're sure we're not
2312 * going to add the new free space to existing bitmap entries - because
2313 * that would mean unnecessary work that would be reverted. Therefore
2314 * attempt to steal space from bitmaps if we're adding an extent entry.
2316 steal_from_bitmap(ctl, info, true);
2318 ret = link_free_space(ctl, info);
2320 kmem_cache_free(btrfs_free_space_cachep, info);
2322 spin_unlock(&ctl->tree_lock);
2325 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2326 ASSERT(ret != -EEXIST);
2332 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2333 u64 offset, u64 bytes)
2335 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2336 struct btrfs_free_space *info;
2338 bool re_search = false;
2340 spin_lock(&ctl->tree_lock);
2347 info = tree_search_offset(ctl, offset, 0, 0);
2350 * oops didn't find an extent that matched the space we wanted
2351 * to remove, look for a bitmap instead
2353 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2357 * If we found a partial bit of our free space in a
2358 * bitmap but then couldn't find the other part this may
2359 * be a problem, so WARN about it.
2367 if (!info->bitmap) {
2368 unlink_free_space(ctl, info);
2369 if (offset == info->offset) {
2370 u64 to_free = min(bytes, info->bytes);
2372 info->bytes -= to_free;
2373 info->offset += to_free;
2375 ret = link_free_space(ctl, info);
2378 kmem_cache_free(btrfs_free_space_cachep, info);
2385 u64 old_end = info->bytes + info->offset;
2387 info->bytes = offset - info->offset;
2388 ret = link_free_space(ctl, info);
2393 /* Not enough bytes in this entry to satisfy us */
2394 if (old_end < offset + bytes) {
2395 bytes -= old_end - offset;
2398 } else if (old_end == offset + bytes) {
2402 spin_unlock(&ctl->tree_lock);
2404 ret = btrfs_add_free_space(block_group, offset + bytes,
2405 old_end - (offset + bytes));
2411 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2412 if (ret == -EAGAIN) {
2417 spin_unlock(&ctl->tree_lock);
2422 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2425 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2426 struct btrfs_free_space *info;
2430 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2431 info = rb_entry(n, struct btrfs_free_space, offset_index);
2432 if (info->bytes >= bytes && !block_group->ro)
2434 btrfs_crit(block_group->fs_info,
2435 "entry offset %llu, bytes %llu, bitmap %s",
2436 info->offset, info->bytes,
2437 (info->bitmap) ? "yes" : "no");
2439 btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2440 list_empty(&block_group->cluster_list) ? "no" : "yes");
2441 btrfs_info(block_group->fs_info,
2442 "%d blocks of free space at or bigger than bytes is", count);
2445 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2447 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2449 spin_lock_init(&ctl->tree_lock);
2450 ctl->unit = block_group->sectorsize;
2451 ctl->start = block_group->key.objectid;
2452 ctl->private = block_group;
2453 ctl->op = &free_space_op;
2454 INIT_LIST_HEAD(&ctl->trimming_ranges);
2455 mutex_init(&ctl->cache_writeout_mutex);
2458 * we only want to have 32k of ram per block group for keeping
2459 * track of free space, and if we pass 1/2 of that we want to
2460 * start converting things over to using bitmaps
2462 ctl->extents_thresh = ((1024 * 32) / 2) /
2463 sizeof(struct btrfs_free_space);
2467 * for a given cluster, put all of its extents back into the free
2468 * space cache. If the block group passed doesn't match the block group
2469 * pointed to by the cluster, someone else raced in and freed the
2470 * cluster already. In that case, we just return without changing anything
2473 __btrfs_return_cluster_to_free_space(
2474 struct btrfs_block_group_cache *block_group,
2475 struct btrfs_free_cluster *cluster)
2477 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2478 struct btrfs_free_space *entry;
2479 struct rb_node *node;
2481 spin_lock(&cluster->lock);
2482 if (cluster->block_group != block_group)
2485 cluster->block_group = NULL;
2486 cluster->window_start = 0;
2487 list_del_init(&cluster->block_group_list);
2489 node = rb_first(&cluster->root);
2493 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2494 node = rb_next(&entry->offset_index);
2495 rb_erase(&entry->offset_index, &cluster->root);
2496 RB_CLEAR_NODE(&entry->offset_index);
2498 bitmap = (entry->bitmap != NULL);
2500 try_merge_free_space(ctl, entry, false);
2501 steal_from_bitmap(ctl, entry, false);
2503 tree_insert_offset(&ctl->free_space_offset,
2504 entry->offset, &entry->offset_index, bitmap);
2506 cluster->root = RB_ROOT;
2509 spin_unlock(&cluster->lock);
2510 btrfs_put_block_group(block_group);
2514 static void __btrfs_remove_free_space_cache_locked(
2515 struct btrfs_free_space_ctl *ctl)
2517 struct btrfs_free_space *info;
2518 struct rb_node *node;
2520 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2521 info = rb_entry(node, struct btrfs_free_space, offset_index);
2522 if (!info->bitmap) {
2523 unlink_free_space(ctl, info);
2524 kmem_cache_free(btrfs_free_space_cachep, info);
2526 free_bitmap(ctl, info);
2529 cond_resched_lock(&ctl->tree_lock);
2533 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2535 spin_lock(&ctl->tree_lock);
2536 __btrfs_remove_free_space_cache_locked(ctl);
2537 spin_unlock(&ctl->tree_lock);
2540 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2542 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2543 struct btrfs_free_cluster *cluster;
2544 struct list_head *head;
2546 spin_lock(&ctl->tree_lock);
2547 while ((head = block_group->cluster_list.next) !=
2548 &block_group->cluster_list) {
2549 cluster = list_entry(head, struct btrfs_free_cluster,
2552 WARN_ON(cluster->block_group != block_group);
2553 __btrfs_return_cluster_to_free_space(block_group, cluster);
2555 cond_resched_lock(&ctl->tree_lock);
2557 __btrfs_remove_free_space_cache_locked(ctl);
2558 spin_unlock(&ctl->tree_lock);
2562 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2563 u64 offset, u64 bytes, u64 empty_size,
2564 u64 *max_extent_size)
2566 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2567 struct btrfs_free_space *entry = NULL;
2568 u64 bytes_search = bytes + empty_size;
2571 u64 align_gap_len = 0;
2573 spin_lock(&ctl->tree_lock);
2574 entry = find_free_space(ctl, &offset, &bytes_search,
2575 block_group->full_stripe_len, max_extent_size);
2580 if (entry->bitmap) {
2581 bitmap_clear_bits(ctl, entry, offset, bytes);
2583 free_bitmap(ctl, entry);
2585 unlink_free_space(ctl, entry);
2586 align_gap_len = offset - entry->offset;
2587 align_gap = entry->offset;
2589 entry->offset = offset + bytes;
2590 WARN_ON(entry->bytes < bytes + align_gap_len);
2592 entry->bytes -= bytes + align_gap_len;
2594 kmem_cache_free(btrfs_free_space_cachep, entry);
2596 link_free_space(ctl, entry);
2599 spin_unlock(&ctl->tree_lock);
2602 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2607 * given a cluster, put all of its extents back into the free space
2608 * cache. If a block group is passed, this function will only free
2609 * a cluster that belongs to the passed block group.
2611 * Otherwise, it'll get a reference on the block group pointed to by the
2612 * cluster and remove the cluster from it.
2614 int btrfs_return_cluster_to_free_space(
2615 struct btrfs_block_group_cache *block_group,
2616 struct btrfs_free_cluster *cluster)
2618 struct btrfs_free_space_ctl *ctl;
2621 /* first, get a safe pointer to the block group */
2622 spin_lock(&cluster->lock);
2624 block_group = cluster->block_group;
2626 spin_unlock(&cluster->lock);
2629 } else if (cluster->block_group != block_group) {
2630 /* someone else has already freed it don't redo their work */
2631 spin_unlock(&cluster->lock);
2634 atomic_inc(&block_group->count);
2635 spin_unlock(&cluster->lock);
2637 ctl = block_group->free_space_ctl;
2639 /* now return any extents the cluster had on it */
2640 spin_lock(&ctl->tree_lock);
2641 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2642 spin_unlock(&ctl->tree_lock);
2644 /* finally drop our ref */
2645 btrfs_put_block_group(block_group);
2649 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2650 struct btrfs_free_cluster *cluster,
2651 struct btrfs_free_space *entry,
2652 u64 bytes, u64 min_start,
2653 u64 *max_extent_size)
2655 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2657 u64 search_start = cluster->window_start;
2658 u64 search_bytes = bytes;
2661 search_start = min_start;
2662 search_bytes = bytes;
2664 err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2666 if (search_bytes > *max_extent_size)
2667 *max_extent_size = search_bytes;
2672 __bitmap_clear_bits(ctl, entry, ret, bytes);
2678 * given a cluster, try to allocate 'bytes' from it, returns 0
2679 * if it couldn't find anything suitably large, or a logical disk offset
2680 * if things worked out
2682 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2683 struct btrfs_free_cluster *cluster, u64 bytes,
2684 u64 min_start, u64 *max_extent_size)
2686 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2687 struct btrfs_free_space *entry = NULL;
2688 struct rb_node *node;
2691 spin_lock(&cluster->lock);
2692 if (bytes > cluster->max_size)
2695 if (cluster->block_group != block_group)
2698 node = rb_first(&cluster->root);
2702 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2704 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2705 *max_extent_size = entry->bytes;
2707 if (entry->bytes < bytes ||
2708 (!entry->bitmap && entry->offset < min_start)) {
2709 node = rb_next(&entry->offset_index);
2712 entry = rb_entry(node, struct btrfs_free_space,
2717 if (entry->bitmap) {
2718 ret = btrfs_alloc_from_bitmap(block_group,
2719 cluster, entry, bytes,
2720 cluster->window_start,
2723 node = rb_next(&entry->offset_index);
2726 entry = rb_entry(node, struct btrfs_free_space,
2730 cluster->window_start += bytes;
2732 ret = entry->offset;
2734 entry->offset += bytes;
2735 entry->bytes -= bytes;
2738 if (entry->bytes == 0)
2739 rb_erase(&entry->offset_index, &cluster->root);
2743 spin_unlock(&cluster->lock);
2748 spin_lock(&ctl->tree_lock);
2750 ctl->free_space -= bytes;
2751 if (entry->bytes == 0) {
2752 ctl->free_extents--;
2753 if (entry->bitmap) {
2754 kfree(entry->bitmap);
2755 ctl->total_bitmaps--;
2756 ctl->op->recalc_thresholds(ctl);
2758 kmem_cache_free(btrfs_free_space_cachep, entry);
2761 spin_unlock(&ctl->tree_lock);
2766 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2767 struct btrfs_free_space *entry,
2768 struct btrfs_free_cluster *cluster,
2769 u64 offset, u64 bytes,
2770 u64 cont1_bytes, u64 min_bytes)
2772 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2773 unsigned long next_zero;
2775 unsigned long want_bits;
2776 unsigned long min_bits;
2777 unsigned long found_bits;
2778 unsigned long start = 0;
2779 unsigned long total_found = 0;
2782 i = offset_to_bit(entry->offset, ctl->unit,
2783 max_t(u64, offset, entry->offset));
2784 want_bits = bytes_to_bits(bytes, ctl->unit);
2785 min_bits = bytes_to_bits(min_bytes, ctl->unit);
2789 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2790 next_zero = find_next_zero_bit(entry->bitmap,
2791 BITS_PER_BITMAP, i);
2792 if (next_zero - i >= min_bits) {
2793 found_bits = next_zero - i;
2804 cluster->max_size = 0;
2807 total_found += found_bits;
2809 if (cluster->max_size < found_bits * ctl->unit)
2810 cluster->max_size = found_bits * ctl->unit;
2812 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2817 cluster->window_start = start * ctl->unit + entry->offset;
2818 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2819 ret = tree_insert_offset(&cluster->root, entry->offset,
2820 &entry->offset_index, 1);
2821 ASSERT(!ret); /* -EEXIST; Logic error */
2823 trace_btrfs_setup_cluster(block_group, cluster,
2824 total_found * ctl->unit, 1);
2829 * This searches the block group for just extents to fill the cluster with.
2830 * Try to find a cluster with at least bytes total bytes, at least one
2831 * extent of cont1_bytes, and other clusters of at least min_bytes.
2834 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2835 struct btrfs_free_cluster *cluster,
2836 struct list_head *bitmaps, u64 offset, u64 bytes,
2837 u64 cont1_bytes, u64 min_bytes)
2839 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2840 struct btrfs_free_space *first = NULL;
2841 struct btrfs_free_space *entry = NULL;
2842 struct btrfs_free_space *last;
2843 struct rb_node *node;
2848 entry = tree_search_offset(ctl, offset, 0, 1);
2853 * We don't want bitmaps, so just move along until we find a normal
2856 while (entry->bitmap || entry->bytes < min_bytes) {
2857 if (entry->bitmap && list_empty(&entry->list))
2858 list_add_tail(&entry->list, bitmaps);
2859 node = rb_next(&entry->offset_index);
2862 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2865 window_free = entry->bytes;
2866 max_extent = entry->bytes;
2870 for (node = rb_next(&entry->offset_index); node;
2871 node = rb_next(&entry->offset_index)) {
2872 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2874 if (entry->bitmap) {
2875 if (list_empty(&entry->list))
2876 list_add_tail(&entry->list, bitmaps);
2880 if (entry->bytes < min_bytes)
2884 window_free += entry->bytes;
2885 if (entry->bytes > max_extent)
2886 max_extent = entry->bytes;
2889 if (window_free < bytes || max_extent < cont1_bytes)
2892 cluster->window_start = first->offset;
2894 node = &first->offset_index;
2897 * now we've found our entries, pull them out of the free space
2898 * cache and put them into the cluster rbtree
2903 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2904 node = rb_next(&entry->offset_index);
2905 if (entry->bitmap || entry->bytes < min_bytes)
2908 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2909 ret = tree_insert_offset(&cluster->root, entry->offset,
2910 &entry->offset_index, 0);
2911 total_size += entry->bytes;
2912 ASSERT(!ret); /* -EEXIST; Logic error */
2913 } while (node && entry != last);
2915 cluster->max_size = max_extent;
2916 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2921 * This specifically looks for bitmaps that may work in the cluster, we assume
2922 * that we have already failed to find extents that will work.
2925 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2926 struct btrfs_free_cluster *cluster,
2927 struct list_head *bitmaps, u64 offset, u64 bytes,
2928 u64 cont1_bytes, u64 min_bytes)
2930 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2931 struct btrfs_free_space *entry;
2933 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2935 if (ctl->total_bitmaps == 0)
2939 * The bitmap that covers offset won't be in the list unless offset
2940 * is just its start offset.
2942 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2943 if (entry->offset != bitmap_offset) {
2944 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2945 if (entry && list_empty(&entry->list))
2946 list_add(&entry->list, bitmaps);
2949 list_for_each_entry(entry, bitmaps, list) {
2950 if (entry->bytes < bytes)
2952 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2953 bytes, cont1_bytes, min_bytes);
2959 * The bitmaps list has all the bitmaps that record free space
2960 * starting after offset, so no more search is required.
2966 * here we try to find a cluster of blocks in a block group. The goal
2967 * is to find at least bytes+empty_size.
2968 * We might not find them all in one contiguous area.
2970 * returns zero and sets up cluster if things worked out, otherwise
2971 * it returns -enospc
2973 int btrfs_find_space_cluster(struct btrfs_root *root,
2974 struct btrfs_block_group_cache *block_group,
2975 struct btrfs_free_cluster *cluster,
2976 u64 offset, u64 bytes, u64 empty_size)
2978 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2979 struct btrfs_free_space *entry, *tmp;
2986 * Choose the minimum extent size we'll require for this
2987 * cluster. For SSD_SPREAD, don't allow any fragmentation.
2988 * For metadata, allow allocates with smaller extents. For
2989 * data, keep it dense.
2991 if (btrfs_test_opt(root, SSD_SPREAD)) {
2992 cont1_bytes = min_bytes = bytes + empty_size;
2993 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2994 cont1_bytes = bytes;
2995 min_bytes = block_group->sectorsize;
2997 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
2998 min_bytes = block_group->sectorsize;
3001 spin_lock(&ctl->tree_lock);
3004 * If we know we don't have enough space to make a cluster don't even
3005 * bother doing all the work to try and find one.
3007 if (ctl->free_space < bytes) {
3008 spin_unlock(&ctl->tree_lock);
3012 spin_lock(&cluster->lock);
3014 /* someone already found a cluster, hooray */
3015 if (cluster->block_group) {
3020 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3023 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3025 cont1_bytes, min_bytes);
3027 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3028 offset, bytes + empty_size,
3029 cont1_bytes, min_bytes);
3031 /* Clear our temporary list */
3032 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3033 list_del_init(&entry->list);
3036 atomic_inc(&block_group->count);
3037 list_add_tail(&cluster->block_group_list,
3038 &block_group->cluster_list);
3039 cluster->block_group = block_group;
3041 trace_btrfs_failed_cluster_setup(block_group);
3044 spin_unlock(&cluster->lock);
3045 spin_unlock(&ctl->tree_lock);
3051 * simple code to zero out a cluster
3053 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3055 spin_lock_init(&cluster->lock);
3056 spin_lock_init(&cluster->refill_lock);
3057 cluster->root = RB_ROOT;
3058 cluster->max_size = 0;
3059 INIT_LIST_HEAD(&cluster->block_group_list);
3060 cluster->block_group = NULL;
3063 static int do_trimming(struct btrfs_block_group_cache *block_group,
3064 u64 *total_trimmed, u64 start, u64 bytes,
3065 u64 reserved_start, u64 reserved_bytes,
3066 struct btrfs_trim_range *trim_entry)
3068 struct btrfs_space_info *space_info = block_group->space_info;
3069 struct btrfs_fs_info *fs_info = block_group->fs_info;
3070 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3075 spin_lock(&space_info->lock);
3076 spin_lock(&block_group->lock);
3077 if (!block_group->ro) {
3078 block_group->reserved += reserved_bytes;
3079 space_info->bytes_reserved += reserved_bytes;
3082 spin_unlock(&block_group->lock);
3083 spin_unlock(&space_info->lock);
3085 ret = btrfs_discard_extent(fs_info->extent_root,
3086 start, bytes, &trimmed);
3088 *total_trimmed += trimmed;
3090 mutex_lock(&ctl->cache_writeout_mutex);
3091 btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3092 list_del(&trim_entry->list);
3093 mutex_unlock(&ctl->cache_writeout_mutex);
3096 spin_lock(&space_info->lock);
3097 spin_lock(&block_group->lock);
3098 if (block_group->ro)
3099 space_info->bytes_readonly += reserved_bytes;
3100 block_group->reserved -= reserved_bytes;
3101 space_info->bytes_reserved -= reserved_bytes;
3102 spin_unlock(&space_info->lock);
3103 spin_unlock(&block_group->lock);
3109 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3110 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3112 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3113 struct btrfs_free_space *entry;
3114 struct rb_node *node;
3120 while (start < end) {
3121 struct btrfs_trim_range trim_entry;
3123 mutex_lock(&ctl->cache_writeout_mutex);
3124 spin_lock(&ctl->tree_lock);
3126 if (ctl->free_space < minlen) {
3127 spin_unlock(&ctl->tree_lock);
3128 mutex_unlock(&ctl->cache_writeout_mutex);
3132 entry = tree_search_offset(ctl, start, 0, 1);
3134 spin_unlock(&ctl->tree_lock);
3135 mutex_unlock(&ctl->cache_writeout_mutex);
3140 while (entry->bitmap) {
3141 node = rb_next(&entry->offset_index);
3143 spin_unlock(&ctl->tree_lock);
3144 mutex_unlock(&ctl->cache_writeout_mutex);
3147 entry = rb_entry(node, struct btrfs_free_space,
3151 if (entry->offset >= end) {
3152 spin_unlock(&ctl->tree_lock);
3153 mutex_unlock(&ctl->cache_writeout_mutex);
3157 extent_start = entry->offset;
3158 extent_bytes = entry->bytes;
3159 start = max(start, extent_start);
3160 bytes = min(extent_start + extent_bytes, end) - start;
3161 if (bytes < minlen) {
3162 spin_unlock(&ctl->tree_lock);
3163 mutex_unlock(&ctl->cache_writeout_mutex);
3167 unlink_free_space(ctl, entry);
3168 kmem_cache_free(btrfs_free_space_cachep, entry);
3170 spin_unlock(&ctl->tree_lock);
3171 trim_entry.start = extent_start;
3172 trim_entry.bytes = extent_bytes;
3173 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3174 mutex_unlock(&ctl->cache_writeout_mutex);
3176 ret = do_trimming(block_group, total_trimmed, start, bytes,
3177 extent_start, extent_bytes, &trim_entry);
3183 if (fatal_signal_pending(current)) {
3194 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3195 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3197 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3198 struct btrfs_free_space *entry;
3202 u64 offset = offset_to_bitmap(ctl, start);
3204 while (offset < end) {
3205 bool next_bitmap = false;
3206 struct btrfs_trim_range trim_entry;
3208 mutex_lock(&ctl->cache_writeout_mutex);
3209 spin_lock(&ctl->tree_lock);
3211 if (ctl->free_space < minlen) {
3212 spin_unlock(&ctl->tree_lock);
3213 mutex_unlock(&ctl->cache_writeout_mutex);
3217 entry = tree_search_offset(ctl, offset, 1, 0);
3219 spin_unlock(&ctl->tree_lock);
3220 mutex_unlock(&ctl->cache_writeout_mutex);
3226 ret2 = search_bitmap(ctl, entry, &start, &bytes);
3227 if (ret2 || start >= end) {
3228 spin_unlock(&ctl->tree_lock);
3229 mutex_unlock(&ctl->cache_writeout_mutex);
3234 bytes = min(bytes, end - start);
3235 if (bytes < minlen) {
3236 spin_unlock(&ctl->tree_lock);
3237 mutex_unlock(&ctl->cache_writeout_mutex);
3241 bitmap_clear_bits(ctl, entry, start, bytes);
3242 if (entry->bytes == 0)
3243 free_bitmap(ctl, entry);
3245 spin_unlock(&ctl->tree_lock);
3246 trim_entry.start = start;
3247 trim_entry.bytes = bytes;
3248 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3249 mutex_unlock(&ctl->cache_writeout_mutex);
3251 ret = do_trimming(block_group, total_trimmed, start, bytes,
3252 start, bytes, &trim_entry);
3257 offset += BITS_PER_BITMAP * ctl->unit;
3260 if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3261 offset += BITS_PER_BITMAP * ctl->unit;
3264 if (fatal_signal_pending(current)) {
3275 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3276 u64 *trimmed, u64 start, u64 end, u64 minlen)
3282 spin_lock(&block_group->lock);
3283 if (block_group->removed) {
3284 spin_unlock(&block_group->lock);
3287 atomic_inc(&block_group->trimming);
3288 spin_unlock(&block_group->lock);
3290 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3294 ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3296 spin_lock(&block_group->lock);
3297 if (atomic_dec_and_test(&block_group->trimming) &&
3298 block_group->removed) {
3299 struct extent_map_tree *em_tree;
3300 struct extent_map *em;
3302 spin_unlock(&block_group->lock);
3304 lock_chunks(block_group->fs_info->chunk_root);
3305 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3306 write_lock(&em_tree->lock);
3307 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3309 BUG_ON(!em); /* logic error, can't happen */
3311 * remove_extent_mapping() will delete us from the pinned_chunks
3312 * list, which is protected by the chunk mutex.
3314 remove_extent_mapping(em_tree, em);
3315 write_unlock(&em_tree->lock);
3316 unlock_chunks(block_group->fs_info->chunk_root);
3318 /* once for us and once for the tree */
3319 free_extent_map(em);
3320 free_extent_map(em);
3323 * We've left one free space entry and other tasks trimming
3324 * this block group have left 1 entry each one. Free them.
3326 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3328 spin_unlock(&block_group->lock);
3335 * Find the left-most item in the cache tree, and then return the
3336 * smallest inode number in the item.
3338 * Note: the returned inode number may not be the smallest one in
3339 * the tree, if the left-most item is a bitmap.
3341 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3343 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3344 struct btrfs_free_space *entry = NULL;
3347 spin_lock(&ctl->tree_lock);
3349 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3352 entry = rb_entry(rb_first(&ctl->free_space_offset),
3353 struct btrfs_free_space, offset_index);
3355 if (!entry->bitmap) {
3356 ino = entry->offset;
3358 unlink_free_space(ctl, entry);
3362 kmem_cache_free(btrfs_free_space_cachep, entry);
3364 link_free_space(ctl, entry);
3370 ret = search_bitmap(ctl, entry, &offset, &count);
3371 /* Logic error; Should be empty if it can't find anything */
3375 bitmap_clear_bits(ctl, entry, offset, 1);
3376 if (entry->bytes == 0)
3377 free_bitmap(ctl, entry);
3380 spin_unlock(&ctl->tree_lock);
3385 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3386 struct btrfs_path *path)
3388 struct inode *inode = NULL;
3390 spin_lock(&root->ino_cache_lock);
3391 if (root->ino_cache_inode)
3392 inode = igrab(root->ino_cache_inode);
3393 spin_unlock(&root->ino_cache_lock);
3397 inode = __lookup_free_space_inode(root, path, 0);
3401 spin_lock(&root->ino_cache_lock);
3402 if (!btrfs_fs_closing(root->fs_info))
3403 root->ino_cache_inode = igrab(inode);
3404 spin_unlock(&root->ino_cache_lock);
3409 int create_free_ino_inode(struct btrfs_root *root,
3410 struct btrfs_trans_handle *trans,
3411 struct btrfs_path *path)
3413 return __create_free_space_inode(root, trans, path,
3414 BTRFS_FREE_INO_OBJECTID, 0);
3417 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3419 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3420 struct btrfs_path *path;
3421 struct inode *inode;
3423 u64 root_gen = btrfs_root_generation(&root->root_item);
3425 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3429 * If we're unmounting then just return, since this does a search on the
3430 * normal root and not the commit root and we could deadlock.
3432 if (btrfs_fs_closing(fs_info))
3435 path = btrfs_alloc_path();
3439 inode = lookup_free_ino_inode(root, path);
3443 if (root_gen != BTRFS_I(inode)->generation)
3446 ret = __load_free_space_cache(root, inode, ctl, path, 0);
3450 "failed to load free ino cache for root %llu",
3451 root->root_key.objectid);
3455 btrfs_free_path(path);
3459 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3460 struct btrfs_trans_handle *trans,
3461 struct btrfs_path *path,
3462 struct inode *inode)
3464 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3466 struct btrfs_io_ctl io_ctl;
3468 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3471 memset(&io_ctl, 0, sizeof(io_ctl));
3472 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3475 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3478 btrfs_delalloc_release_metadata(inode, inode->i_size);
3480 btrfs_err(root->fs_info,
3481 "failed to write free ino cache for root %llu",
3482 root->root_key.objectid);
3489 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3491 * Use this if you need to make a bitmap or extent entry specifically, it
3492 * doesn't do any of the merging that add_free_space does, this acts a lot like
3493 * how the free space cache loading stuff works, so you can get really weird
3496 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3497 u64 offset, u64 bytes, bool bitmap)
3499 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3500 struct btrfs_free_space *info = NULL, *bitmap_info;
3507 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3513 spin_lock(&ctl->tree_lock);
3514 info->offset = offset;
3515 info->bytes = bytes;
3516 ret = link_free_space(ctl, info);
3517 spin_unlock(&ctl->tree_lock);
3519 kmem_cache_free(btrfs_free_space_cachep, info);
3524 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3526 kmem_cache_free(btrfs_free_space_cachep, info);
3531 spin_lock(&ctl->tree_lock);
3532 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3537 add_new_bitmap(ctl, info, offset);
3542 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3543 bytes -= bytes_added;
3544 offset += bytes_added;
3545 spin_unlock(&ctl->tree_lock);
3551 kmem_cache_free(btrfs_free_space_cachep, info);
3558 * Checks to see if the given range is in the free space cache. This is really
3559 * just used to check the absence of space, so if there is free space in the
3560 * range at all we will return 1.
3562 int test_check_exists(struct btrfs_block_group_cache *cache,
3563 u64 offset, u64 bytes)
3565 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3566 struct btrfs_free_space *info;
3569 spin_lock(&ctl->tree_lock);
3570 info = tree_search_offset(ctl, offset, 0, 0);
3572 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3580 u64 bit_off, bit_bytes;
3582 struct btrfs_free_space *tmp;
3585 bit_bytes = ctl->unit;
3586 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes);
3588 if (bit_off == offset) {
3591 } else if (bit_off > offset &&
3592 offset + bytes > bit_off) {
3598 n = rb_prev(&info->offset_index);
3600 tmp = rb_entry(n, struct btrfs_free_space,
3602 if (tmp->offset + tmp->bytes < offset)
3604 if (offset + bytes < tmp->offset) {
3605 n = rb_prev(&info->offset_index);
3612 n = rb_next(&info->offset_index);
3614 tmp = rb_entry(n, struct btrfs_free_space,
3616 if (offset + bytes < tmp->offset)
3618 if (tmp->offset + tmp->bytes < offset) {
3619 n = rb_next(&info->offset_index);
3630 if (info->offset == offset) {
3635 if (offset > info->offset && offset < info->offset + info->bytes)
3638 spin_unlock(&ctl->tree_lock);
3641 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */