Btrfs: check free space in block group before searching for a cluster
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / btrfs / free-space-cache.c
blobf631df870f64bd42fc158cb44df3bb56c57da5a6
1 /*
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 "ctree.h"
24 #include "free-space-cache.h"
25 #include "transaction.h"
26 #include "disk-io.h"
28 #define BITS_PER_BITMAP (PAGE_CACHE_SIZE * 8)
29 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
31 static void recalculate_thresholds(struct btrfs_block_group_cache
32 *block_group);
33 static int link_free_space(struct btrfs_block_group_cache *block_group,
34 struct btrfs_free_space *info);
36 struct inode *lookup_free_space_inode(struct btrfs_root *root,
37 struct btrfs_block_group_cache
38 *block_group, struct btrfs_path *path)
40 struct btrfs_key key;
41 struct btrfs_key location;
42 struct btrfs_disk_key disk_key;
43 struct btrfs_free_space_header *header;
44 struct extent_buffer *leaf;
45 struct inode *inode = NULL;
46 int ret;
48 spin_lock(&block_group->lock);
49 if (block_group->inode)
50 inode = igrab(block_group->inode);
51 spin_unlock(&block_group->lock);
52 if (inode)
53 return inode;
55 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
56 key.offset = block_group->key.objectid;
57 key.type = 0;
59 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
60 if (ret < 0)
61 return ERR_PTR(ret);
62 if (ret > 0) {
63 btrfs_release_path(root, path);
64 return ERR_PTR(-ENOENT);
67 leaf = path->nodes[0];
68 header = btrfs_item_ptr(leaf, path->slots[0],
69 struct btrfs_free_space_header);
70 btrfs_free_space_key(leaf, header, &disk_key);
71 btrfs_disk_key_to_cpu(&location, &disk_key);
72 btrfs_release_path(root, path);
74 inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
75 if (!inode)
76 return ERR_PTR(-ENOENT);
77 if (IS_ERR(inode))
78 return inode;
79 if (is_bad_inode(inode)) {
80 iput(inode);
81 return ERR_PTR(-ENOENT);
84 spin_lock(&block_group->lock);
85 if (!root->fs_info->closing) {
86 block_group->inode = igrab(inode);
87 block_group->iref = 1;
89 spin_unlock(&block_group->lock);
91 return inode;
94 int create_free_space_inode(struct btrfs_root *root,
95 struct btrfs_trans_handle *trans,
96 struct btrfs_block_group_cache *block_group,
97 struct btrfs_path *path)
99 struct btrfs_key key;
100 struct btrfs_disk_key disk_key;
101 struct btrfs_free_space_header *header;
102 struct btrfs_inode_item *inode_item;
103 struct extent_buffer *leaf;
104 u64 objectid;
105 int ret;
107 ret = btrfs_find_free_objectid(trans, root, 0, &objectid);
108 if (ret < 0)
109 return ret;
111 ret = btrfs_insert_empty_inode(trans, root, path, objectid);
112 if (ret)
113 return ret;
115 leaf = path->nodes[0];
116 inode_item = btrfs_item_ptr(leaf, path->slots[0],
117 struct btrfs_inode_item);
118 btrfs_item_key(leaf, &disk_key, path->slots[0]);
119 memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
120 sizeof(*inode_item));
121 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
122 btrfs_set_inode_size(leaf, inode_item, 0);
123 btrfs_set_inode_nbytes(leaf, inode_item, 0);
124 btrfs_set_inode_uid(leaf, inode_item, 0);
125 btrfs_set_inode_gid(leaf, inode_item, 0);
126 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
127 btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
128 BTRFS_INODE_PREALLOC | BTRFS_INODE_NODATASUM);
129 btrfs_set_inode_nlink(leaf, inode_item, 1);
130 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
131 btrfs_set_inode_block_group(leaf, inode_item,
132 block_group->key.objectid);
133 btrfs_mark_buffer_dirty(leaf);
134 btrfs_release_path(root, path);
136 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
137 key.offset = block_group->key.objectid;
138 key.type = 0;
140 ret = btrfs_insert_empty_item(trans, root, path, &key,
141 sizeof(struct btrfs_free_space_header));
142 if (ret < 0) {
143 btrfs_release_path(root, path);
144 return ret;
146 leaf = path->nodes[0];
147 header = btrfs_item_ptr(leaf, path->slots[0],
148 struct btrfs_free_space_header);
149 memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
150 btrfs_set_free_space_key(leaf, header, &disk_key);
151 btrfs_mark_buffer_dirty(leaf);
152 btrfs_release_path(root, path);
154 return 0;
157 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
158 struct btrfs_trans_handle *trans,
159 struct btrfs_path *path,
160 struct inode *inode)
162 loff_t oldsize;
163 int ret = 0;
165 trans->block_rsv = root->orphan_block_rsv;
166 ret = btrfs_block_rsv_check(trans, root,
167 root->orphan_block_rsv,
168 0, 5);
169 if (ret)
170 return ret;
172 oldsize = i_size_read(inode);
173 btrfs_i_size_write(inode, 0);
174 truncate_pagecache(inode, oldsize, 0);
177 * We don't need an orphan item because truncating the free space cache
178 * will never be split across transactions.
180 ret = btrfs_truncate_inode_items(trans, root, inode,
181 0, BTRFS_EXTENT_DATA_KEY);
182 if (ret) {
183 WARN_ON(1);
184 return ret;
187 return btrfs_update_inode(trans, root, inode);
190 static int readahead_cache(struct inode *inode)
192 struct file_ra_state *ra;
193 unsigned long last_index;
195 ra = kzalloc(sizeof(*ra), GFP_NOFS);
196 if (!ra)
197 return -ENOMEM;
199 file_ra_state_init(ra, inode->i_mapping);
200 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
202 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
204 kfree(ra);
206 return 0;
209 int load_free_space_cache(struct btrfs_fs_info *fs_info,
210 struct btrfs_block_group_cache *block_group)
212 struct btrfs_root *root = fs_info->tree_root;
213 struct inode *inode;
214 struct btrfs_free_space_header *header;
215 struct extent_buffer *leaf;
216 struct page *page;
217 struct btrfs_path *path;
218 u32 *checksums = NULL, *crc;
219 char *disk_crcs = NULL;
220 struct btrfs_key key;
221 struct list_head bitmaps;
222 u64 num_entries;
223 u64 num_bitmaps;
224 u64 generation;
225 u32 cur_crc = ~(u32)0;
226 pgoff_t index = 0;
227 unsigned long first_page_offset;
228 int num_checksums;
229 int ret = 0;
232 * If we're unmounting then just return, since this does a search on the
233 * normal root and not the commit root and we could deadlock.
235 smp_mb();
236 if (fs_info->closing)
237 return 0;
240 * If this block group has been marked to be cleared for one reason or
241 * another then we can't trust the on disk cache, so just return.
243 spin_lock(&block_group->lock);
244 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
245 spin_unlock(&block_group->lock);
246 return 0;
248 spin_unlock(&block_group->lock);
250 INIT_LIST_HEAD(&bitmaps);
252 path = btrfs_alloc_path();
253 if (!path)
254 return 0;
256 inode = lookup_free_space_inode(root, block_group, path);
257 if (IS_ERR(inode)) {
258 btrfs_free_path(path);
259 return 0;
262 /* Nothing in the space cache, goodbye */
263 if (!i_size_read(inode)) {
264 btrfs_free_path(path);
265 goto out;
268 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
269 key.offset = block_group->key.objectid;
270 key.type = 0;
272 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
273 if (ret) {
274 btrfs_free_path(path);
275 goto out;
278 leaf = path->nodes[0];
279 header = btrfs_item_ptr(leaf, path->slots[0],
280 struct btrfs_free_space_header);
281 num_entries = btrfs_free_space_entries(leaf, header);
282 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
283 generation = btrfs_free_space_generation(leaf, header);
284 btrfs_free_path(path);
286 if (BTRFS_I(inode)->generation != generation) {
287 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
288 " not match free space cache generation (%llu) for "
289 "block group %llu\n",
290 (unsigned long long)BTRFS_I(inode)->generation,
291 (unsigned long long)generation,
292 (unsigned long long)block_group->key.objectid);
293 goto free_cache;
296 if (!num_entries)
297 goto out;
299 /* Setup everything for doing checksumming */
300 num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
301 checksums = crc = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
302 if (!checksums)
303 goto out;
304 first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
305 disk_crcs = kzalloc(first_page_offset, GFP_NOFS);
306 if (!disk_crcs)
307 goto out;
309 ret = readahead_cache(inode);
310 if (ret) {
311 ret = 0;
312 goto out;
315 while (1) {
316 struct btrfs_free_space_entry *entry;
317 struct btrfs_free_space *e;
318 void *addr;
319 unsigned long offset = 0;
320 unsigned long start_offset = 0;
321 int need_loop = 0;
323 if (!num_entries && !num_bitmaps)
324 break;
326 if (index == 0) {
327 start_offset = first_page_offset;
328 offset = start_offset;
331 page = grab_cache_page(inode->i_mapping, index);
332 if (!page) {
333 ret = 0;
334 goto free_cache;
337 if (!PageUptodate(page)) {
338 btrfs_readpage(NULL, page);
339 lock_page(page);
340 if (!PageUptodate(page)) {
341 unlock_page(page);
342 page_cache_release(page);
343 printk(KERN_ERR "btrfs: error reading free "
344 "space cache: %llu\n",
345 (unsigned long long)
346 block_group->key.objectid);
347 goto free_cache;
350 addr = kmap(page);
352 if (index == 0) {
353 u64 *gen;
355 memcpy(disk_crcs, addr, first_page_offset);
356 gen = addr + (sizeof(u32) * num_checksums);
357 if (*gen != BTRFS_I(inode)->generation) {
358 printk(KERN_ERR "btrfs: space cache generation"
359 " (%llu) does not match inode (%llu) "
360 "for block group %llu\n",
361 (unsigned long long)*gen,
362 (unsigned long long)
363 BTRFS_I(inode)->generation,
364 (unsigned long long)
365 block_group->key.objectid);
366 kunmap(page);
367 unlock_page(page);
368 page_cache_release(page);
369 goto free_cache;
371 crc = (u32 *)disk_crcs;
373 entry = addr + start_offset;
375 /* First lets check our crc before we do anything fun */
376 cur_crc = ~(u32)0;
377 cur_crc = btrfs_csum_data(root, addr + start_offset, cur_crc,
378 PAGE_CACHE_SIZE - start_offset);
379 btrfs_csum_final(cur_crc, (char *)&cur_crc);
380 if (cur_crc != *crc) {
381 printk(KERN_ERR "btrfs: crc mismatch for page %lu in "
382 "block group %llu\n", index,
383 (unsigned long long)block_group->key.objectid);
384 kunmap(page);
385 unlock_page(page);
386 page_cache_release(page);
387 goto free_cache;
389 crc++;
391 while (1) {
392 if (!num_entries)
393 break;
395 need_loop = 1;
396 e = kmem_cache_zalloc(btrfs_free_space_cachep,
397 GFP_NOFS);
398 if (!e) {
399 kunmap(page);
400 unlock_page(page);
401 page_cache_release(page);
402 goto free_cache;
405 e->offset = le64_to_cpu(entry->offset);
406 e->bytes = le64_to_cpu(entry->bytes);
407 if (!e->bytes) {
408 kunmap(page);
409 kmem_cache_free(btrfs_free_space_cachep, e);
410 unlock_page(page);
411 page_cache_release(page);
412 goto free_cache;
415 if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
416 spin_lock(&block_group->tree_lock);
417 ret = link_free_space(block_group, e);
418 spin_unlock(&block_group->tree_lock);
419 BUG_ON(ret);
420 } else {
421 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
422 if (!e->bitmap) {
423 kunmap(page);
424 kmem_cache_free(
425 btrfs_free_space_cachep, e);
426 unlock_page(page);
427 page_cache_release(page);
428 goto free_cache;
430 spin_lock(&block_group->tree_lock);
431 ret = link_free_space(block_group, e);
432 block_group->total_bitmaps++;
433 recalculate_thresholds(block_group);
434 spin_unlock(&block_group->tree_lock);
435 list_add_tail(&e->list, &bitmaps);
438 num_entries--;
439 offset += sizeof(struct btrfs_free_space_entry);
440 if (offset + sizeof(struct btrfs_free_space_entry) >=
441 PAGE_CACHE_SIZE)
442 break;
443 entry++;
447 * We read an entry out of this page, we need to move on to the
448 * next page.
450 if (need_loop) {
451 kunmap(page);
452 goto next;
456 * We add the bitmaps at the end of the entries in order that
457 * the bitmap entries are added to the cache.
459 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
460 list_del_init(&e->list);
461 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
462 kunmap(page);
463 num_bitmaps--;
464 next:
465 unlock_page(page);
466 page_cache_release(page);
467 index++;
470 ret = 1;
471 out:
472 kfree(checksums);
473 kfree(disk_crcs);
474 iput(inode);
475 return ret;
477 free_cache:
478 /* This cache is bogus, make sure it gets cleared */
479 spin_lock(&block_group->lock);
480 block_group->disk_cache_state = BTRFS_DC_CLEAR;
481 spin_unlock(&block_group->lock);
482 btrfs_remove_free_space_cache(block_group);
483 goto out;
486 int btrfs_write_out_cache(struct btrfs_root *root,
487 struct btrfs_trans_handle *trans,
488 struct btrfs_block_group_cache *block_group,
489 struct btrfs_path *path)
491 struct btrfs_free_space_header *header;
492 struct extent_buffer *leaf;
493 struct inode *inode;
494 struct rb_node *node;
495 struct list_head *pos, *n;
496 struct page *page;
497 struct extent_state *cached_state = NULL;
498 struct list_head bitmap_list;
499 struct btrfs_key key;
500 u64 bytes = 0;
501 u32 *crc, *checksums;
502 pgoff_t index = 0, last_index = 0;
503 unsigned long first_page_offset;
504 int num_checksums;
505 int entries = 0;
506 int bitmaps = 0;
507 int ret = 0;
509 root = root->fs_info->tree_root;
511 INIT_LIST_HEAD(&bitmap_list);
513 spin_lock(&block_group->lock);
514 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
515 spin_unlock(&block_group->lock);
516 return 0;
518 spin_unlock(&block_group->lock);
520 inode = lookup_free_space_inode(root, block_group, path);
521 if (IS_ERR(inode))
522 return 0;
524 if (!i_size_read(inode)) {
525 iput(inode);
526 return 0;
529 node = rb_first(&block_group->free_space_offset);
530 if (!node) {
531 iput(inode);
532 return 0;
535 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
536 filemap_write_and_wait(inode->i_mapping);
537 btrfs_wait_ordered_range(inode, inode->i_size &
538 ~(root->sectorsize - 1), (u64)-1);
540 /* We need a checksum per page. */
541 num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
542 crc = checksums = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
543 if (!crc) {
544 iput(inode);
545 return 0;
548 /* Since the first page has all of our checksums and our generation we
549 * need to calculate the offset into the page that we can start writing
550 * our entries.
552 first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
555 * Lock all pages first so we can lock the extent safely.
557 * NOTE: Because we hold the ref the entire time we're going to write to
558 * the page find_get_page should never fail, so we don't do a check
559 * after find_get_page at this point. Just putting this here so people
560 * know and don't freak out.
562 while (index <= last_index) {
563 page = grab_cache_page(inode->i_mapping, index);
564 if (!page) {
565 pgoff_t i = 0;
567 while (i < index) {
568 page = find_get_page(inode->i_mapping, i);
569 unlock_page(page);
570 page_cache_release(page);
571 page_cache_release(page);
572 i++;
574 goto out_free;
576 index++;
579 index = 0;
580 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
581 0, &cached_state, GFP_NOFS);
583 /* Write out the extent entries */
584 do {
585 struct btrfs_free_space_entry *entry;
586 void *addr;
587 unsigned long offset = 0;
588 unsigned long start_offset = 0;
590 if (index == 0) {
591 start_offset = first_page_offset;
592 offset = start_offset;
595 page = find_get_page(inode->i_mapping, index);
597 addr = kmap(page);
598 entry = addr + start_offset;
600 memset(addr, 0, PAGE_CACHE_SIZE);
601 while (1) {
602 struct btrfs_free_space *e;
604 e = rb_entry(node, struct btrfs_free_space, offset_index);
605 entries++;
607 entry->offset = cpu_to_le64(e->offset);
608 entry->bytes = cpu_to_le64(e->bytes);
609 if (e->bitmap) {
610 entry->type = BTRFS_FREE_SPACE_BITMAP;
611 list_add_tail(&e->list, &bitmap_list);
612 bitmaps++;
613 } else {
614 entry->type = BTRFS_FREE_SPACE_EXTENT;
616 node = rb_next(node);
617 if (!node)
618 break;
619 offset += sizeof(struct btrfs_free_space_entry);
620 if (offset + sizeof(struct btrfs_free_space_entry) >=
621 PAGE_CACHE_SIZE)
622 break;
623 entry++;
625 *crc = ~(u32)0;
626 *crc = btrfs_csum_data(root, addr + start_offset, *crc,
627 PAGE_CACHE_SIZE - start_offset);
628 kunmap(page);
630 btrfs_csum_final(*crc, (char *)crc);
631 crc++;
633 bytes += PAGE_CACHE_SIZE;
635 ClearPageChecked(page);
636 set_page_extent_mapped(page);
637 SetPageUptodate(page);
638 set_page_dirty(page);
641 * We need to release our reference we got for grab_cache_page,
642 * except for the first page which will hold our checksums, we
643 * do that below.
645 if (index != 0) {
646 unlock_page(page);
647 page_cache_release(page);
650 page_cache_release(page);
652 index++;
653 } while (node);
655 /* Write out the bitmaps */
656 list_for_each_safe(pos, n, &bitmap_list) {
657 void *addr;
658 struct btrfs_free_space *entry =
659 list_entry(pos, struct btrfs_free_space, list);
661 page = find_get_page(inode->i_mapping, index);
663 addr = kmap(page);
664 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
665 *crc = ~(u32)0;
666 *crc = btrfs_csum_data(root, addr, *crc, PAGE_CACHE_SIZE);
667 kunmap(page);
668 btrfs_csum_final(*crc, (char *)crc);
669 crc++;
670 bytes += PAGE_CACHE_SIZE;
672 ClearPageChecked(page);
673 set_page_extent_mapped(page);
674 SetPageUptodate(page);
675 set_page_dirty(page);
676 unlock_page(page);
677 page_cache_release(page);
678 page_cache_release(page);
679 list_del_init(&entry->list);
680 index++;
683 /* Zero out the rest of the pages just to make sure */
684 while (index <= last_index) {
685 void *addr;
687 page = find_get_page(inode->i_mapping, index);
689 addr = kmap(page);
690 memset(addr, 0, PAGE_CACHE_SIZE);
691 kunmap(page);
692 ClearPageChecked(page);
693 set_page_extent_mapped(page);
694 SetPageUptodate(page);
695 set_page_dirty(page);
696 unlock_page(page);
697 page_cache_release(page);
698 page_cache_release(page);
699 bytes += PAGE_CACHE_SIZE;
700 index++;
703 btrfs_set_extent_delalloc(inode, 0, bytes - 1, &cached_state);
705 /* Write the checksums and trans id to the first page */
707 void *addr;
708 u64 *gen;
710 page = find_get_page(inode->i_mapping, 0);
712 addr = kmap(page);
713 memcpy(addr, checksums, sizeof(u32) * num_checksums);
714 gen = addr + (sizeof(u32) * num_checksums);
715 *gen = trans->transid;
716 kunmap(page);
717 ClearPageChecked(page);
718 set_page_extent_mapped(page);
719 SetPageUptodate(page);
720 set_page_dirty(page);
721 unlock_page(page);
722 page_cache_release(page);
723 page_cache_release(page);
725 BTRFS_I(inode)->generation = trans->transid;
727 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
728 i_size_read(inode) - 1, &cached_state, GFP_NOFS);
730 filemap_write_and_wait(inode->i_mapping);
732 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
733 key.offset = block_group->key.objectid;
734 key.type = 0;
736 ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
737 if (ret < 0) {
738 ret = 0;
739 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
740 EXTENT_DIRTY | EXTENT_DELALLOC |
741 EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
742 goto out_free;
744 leaf = path->nodes[0];
745 if (ret > 0) {
746 struct btrfs_key found_key;
747 BUG_ON(!path->slots[0]);
748 path->slots[0]--;
749 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
750 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
751 found_key.offset != block_group->key.objectid) {
752 ret = 0;
753 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
754 EXTENT_DIRTY | EXTENT_DELALLOC |
755 EXTENT_DO_ACCOUNTING, 0, 0, NULL,
756 GFP_NOFS);
757 btrfs_release_path(root, path);
758 goto out_free;
761 header = btrfs_item_ptr(leaf, path->slots[0],
762 struct btrfs_free_space_header);
763 btrfs_set_free_space_entries(leaf, header, entries);
764 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
765 btrfs_set_free_space_generation(leaf, header, trans->transid);
766 btrfs_mark_buffer_dirty(leaf);
767 btrfs_release_path(root, path);
769 ret = 1;
771 out_free:
772 if (ret == 0) {
773 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
774 spin_lock(&block_group->lock);
775 block_group->disk_cache_state = BTRFS_DC_ERROR;
776 spin_unlock(&block_group->lock);
777 BTRFS_I(inode)->generation = 0;
779 kfree(checksums);
780 btrfs_update_inode(trans, root, inode);
781 iput(inode);
782 return ret;
785 static inline unsigned long offset_to_bit(u64 bitmap_start, u64 sectorsize,
786 u64 offset)
788 BUG_ON(offset < bitmap_start);
789 offset -= bitmap_start;
790 return (unsigned long)(div64_u64(offset, sectorsize));
793 static inline unsigned long bytes_to_bits(u64 bytes, u64 sectorsize)
795 return (unsigned long)(div64_u64(bytes, sectorsize));
798 static inline u64 offset_to_bitmap(struct btrfs_block_group_cache *block_group,
799 u64 offset)
801 u64 bitmap_start;
802 u64 bytes_per_bitmap;
804 bytes_per_bitmap = BITS_PER_BITMAP * block_group->sectorsize;
805 bitmap_start = offset - block_group->key.objectid;
806 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
807 bitmap_start *= bytes_per_bitmap;
808 bitmap_start += block_group->key.objectid;
810 return bitmap_start;
813 static int tree_insert_offset(struct rb_root *root, u64 offset,
814 struct rb_node *node, int bitmap)
816 struct rb_node **p = &root->rb_node;
817 struct rb_node *parent = NULL;
818 struct btrfs_free_space *info;
820 while (*p) {
821 parent = *p;
822 info = rb_entry(parent, struct btrfs_free_space, offset_index);
824 if (offset < info->offset) {
825 p = &(*p)->rb_left;
826 } else if (offset > info->offset) {
827 p = &(*p)->rb_right;
828 } else {
830 * we could have a bitmap entry and an extent entry
831 * share the same offset. If this is the case, we want
832 * the extent entry to always be found first if we do a
833 * linear search through the tree, since we want to have
834 * the quickest allocation time, and allocating from an
835 * extent is faster than allocating from a bitmap. So
836 * if we're inserting a bitmap and we find an entry at
837 * this offset, we want to go right, or after this entry
838 * logically. If we are inserting an extent and we've
839 * found a bitmap, we want to go left, or before
840 * logically.
842 if (bitmap) {
843 WARN_ON(info->bitmap);
844 p = &(*p)->rb_right;
845 } else {
846 WARN_ON(!info->bitmap);
847 p = &(*p)->rb_left;
852 rb_link_node(node, parent, p);
853 rb_insert_color(node, root);
855 return 0;
859 * searches the tree for the given offset.
861 * fuzzy - If this is set, then we are trying to make an allocation, and we just
862 * want a section that has at least bytes size and comes at or after the given
863 * offset.
865 static struct btrfs_free_space *
866 tree_search_offset(struct btrfs_block_group_cache *block_group,
867 u64 offset, int bitmap_only, int fuzzy)
869 struct rb_node *n = block_group->free_space_offset.rb_node;
870 struct btrfs_free_space *entry, *prev = NULL;
872 /* find entry that is closest to the 'offset' */
873 while (1) {
874 if (!n) {
875 entry = NULL;
876 break;
879 entry = rb_entry(n, struct btrfs_free_space, offset_index);
880 prev = entry;
882 if (offset < entry->offset)
883 n = n->rb_left;
884 else if (offset > entry->offset)
885 n = n->rb_right;
886 else
887 break;
890 if (bitmap_only) {
891 if (!entry)
892 return NULL;
893 if (entry->bitmap)
894 return entry;
897 * bitmap entry and extent entry may share same offset,
898 * in that case, bitmap entry comes after extent entry.
900 n = rb_next(n);
901 if (!n)
902 return NULL;
903 entry = rb_entry(n, struct btrfs_free_space, offset_index);
904 if (entry->offset != offset)
905 return NULL;
907 WARN_ON(!entry->bitmap);
908 return entry;
909 } else if (entry) {
910 if (entry->bitmap) {
912 * if previous extent entry covers the offset,
913 * we should return it instead of the bitmap entry
915 n = &entry->offset_index;
916 while (1) {
917 n = rb_prev(n);
918 if (!n)
919 break;
920 prev = rb_entry(n, struct btrfs_free_space,
921 offset_index);
922 if (!prev->bitmap) {
923 if (prev->offset + prev->bytes > offset)
924 entry = prev;
925 break;
929 return entry;
932 if (!prev)
933 return NULL;
935 /* find last entry before the 'offset' */
936 entry = prev;
937 if (entry->offset > offset) {
938 n = rb_prev(&entry->offset_index);
939 if (n) {
940 entry = rb_entry(n, struct btrfs_free_space,
941 offset_index);
942 BUG_ON(entry->offset > offset);
943 } else {
944 if (fuzzy)
945 return entry;
946 else
947 return NULL;
951 if (entry->bitmap) {
952 n = &entry->offset_index;
953 while (1) {
954 n = rb_prev(n);
955 if (!n)
956 break;
957 prev = rb_entry(n, struct btrfs_free_space,
958 offset_index);
959 if (!prev->bitmap) {
960 if (prev->offset + prev->bytes > offset)
961 return prev;
962 break;
965 if (entry->offset + BITS_PER_BITMAP *
966 block_group->sectorsize > offset)
967 return entry;
968 } else if (entry->offset + entry->bytes > offset)
969 return entry;
971 if (!fuzzy)
972 return NULL;
974 while (1) {
975 if (entry->bitmap) {
976 if (entry->offset + BITS_PER_BITMAP *
977 block_group->sectorsize > offset)
978 break;
979 } else {
980 if (entry->offset + entry->bytes > offset)
981 break;
984 n = rb_next(&entry->offset_index);
985 if (!n)
986 return NULL;
987 entry = rb_entry(n, struct btrfs_free_space, offset_index);
989 return entry;
992 static inline void
993 __unlink_free_space(struct btrfs_block_group_cache *block_group,
994 struct btrfs_free_space *info)
996 rb_erase(&info->offset_index, &block_group->free_space_offset);
997 block_group->free_extents--;
1000 static void unlink_free_space(struct btrfs_block_group_cache *block_group,
1001 struct btrfs_free_space *info)
1003 __unlink_free_space(block_group, info);
1004 block_group->free_space -= info->bytes;
1007 static int link_free_space(struct btrfs_block_group_cache *block_group,
1008 struct btrfs_free_space *info)
1010 int ret = 0;
1012 BUG_ON(!info->bitmap && !info->bytes);
1013 ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
1014 &info->offset_index, (info->bitmap != NULL));
1015 if (ret)
1016 return ret;
1018 block_group->free_space += info->bytes;
1019 block_group->free_extents++;
1020 return ret;
1023 static void recalculate_thresholds(struct btrfs_block_group_cache *block_group)
1025 u64 max_bytes;
1026 u64 bitmap_bytes;
1027 u64 extent_bytes;
1028 u64 size = block_group->key.offset;
1031 * The goal is to keep the total amount of memory used per 1gb of space
1032 * at or below 32k, so we need to adjust how much memory we allow to be
1033 * used by extent based free space tracking
1035 if (size < 1024 * 1024 * 1024)
1036 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1037 else
1038 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1039 div64_u64(size, 1024 * 1024 * 1024);
1042 * we want to account for 1 more bitmap than what we have so we can make
1043 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1044 * we add more bitmaps.
1046 bitmap_bytes = (block_group->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1048 if (bitmap_bytes >= max_bytes) {
1049 block_group->extents_thresh = 0;
1050 return;
1054 * we want the extent entry threshold to always be at most 1/2 the maxw
1055 * bytes we can have, or whatever is less than that.
1057 extent_bytes = max_bytes - bitmap_bytes;
1058 extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1060 block_group->extents_thresh =
1061 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1064 static void bitmap_clear_bits(struct btrfs_block_group_cache *block_group,
1065 struct btrfs_free_space *info, u64 offset,
1066 u64 bytes)
1068 unsigned long start, end;
1069 unsigned long i;
1071 start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1072 end = start + bytes_to_bits(bytes, block_group->sectorsize);
1073 BUG_ON(end > BITS_PER_BITMAP);
1075 for (i = start; i < end; i++)
1076 clear_bit(i, info->bitmap);
1078 info->bytes -= bytes;
1079 block_group->free_space -= bytes;
1082 static void bitmap_set_bits(struct btrfs_block_group_cache *block_group,
1083 struct btrfs_free_space *info, u64 offset,
1084 u64 bytes)
1086 unsigned long start, end;
1087 unsigned long i;
1089 start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1090 end = start + bytes_to_bits(bytes, block_group->sectorsize);
1091 BUG_ON(end > BITS_PER_BITMAP);
1093 for (i = start; i < end; i++)
1094 set_bit(i, info->bitmap);
1096 info->bytes += bytes;
1097 block_group->free_space += bytes;
1100 static int search_bitmap(struct btrfs_block_group_cache *block_group,
1101 struct btrfs_free_space *bitmap_info, u64 *offset,
1102 u64 *bytes)
1104 unsigned long found_bits = 0;
1105 unsigned long bits, i;
1106 unsigned long next_zero;
1108 i = offset_to_bit(bitmap_info->offset, block_group->sectorsize,
1109 max_t(u64, *offset, bitmap_info->offset));
1110 bits = bytes_to_bits(*bytes, block_group->sectorsize);
1112 for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1113 i < BITS_PER_BITMAP;
1114 i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1115 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1116 BITS_PER_BITMAP, i);
1117 if ((next_zero - i) >= bits) {
1118 found_bits = next_zero - i;
1119 break;
1121 i = next_zero;
1124 if (found_bits) {
1125 *offset = (u64)(i * block_group->sectorsize) +
1126 bitmap_info->offset;
1127 *bytes = (u64)(found_bits) * block_group->sectorsize;
1128 return 0;
1131 return -1;
1134 static struct btrfs_free_space *find_free_space(struct btrfs_block_group_cache
1135 *block_group, u64 *offset,
1136 u64 *bytes, int debug)
1138 struct btrfs_free_space *entry;
1139 struct rb_node *node;
1140 int ret;
1142 if (!block_group->free_space_offset.rb_node)
1143 return NULL;
1145 entry = tree_search_offset(block_group,
1146 offset_to_bitmap(block_group, *offset),
1147 0, 1);
1148 if (!entry)
1149 return NULL;
1151 for (node = &entry->offset_index; node; node = rb_next(node)) {
1152 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1153 if (entry->bytes < *bytes)
1154 continue;
1156 if (entry->bitmap) {
1157 ret = search_bitmap(block_group, entry, offset, bytes);
1158 if (!ret)
1159 return entry;
1160 continue;
1163 *offset = entry->offset;
1164 *bytes = entry->bytes;
1165 return entry;
1168 return NULL;
1171 static void add_new_bitmap(struct btrfs_block_group_cache *block_group,
1172 struct btrfs_free_space *info, u64 offset)
1174 u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1175 int max_bitmaps = (int)div64_u64(block_group->key.offset +
1176 bytes_per_bg - 1, bytes_per_bg);
1177 BUG_ON(block_group->total_bitmaps >= max_bitmaps);
1179 info->offset = offset_to_bitmap(block_group, offset);
1180 info->bytes = 0;
1181 link_free_space(block_group, info);
1182 block_group->total_bitmaps++;
1184 recalculate_thresholds(block_group);
1187 static void free_bitmap(struct btrfs_block_group_cache *block_group,
1188 struct btrfs_free_space *bitmap_info)
1190 unlink_free_space(block_group, bitmap_info);
1191 kfree(bitmap_info->bitmap);
1192 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1193 block_group->total_bitmaps--;
1194 recalculate_thresholds(block_group);
1197 static noinline int remove_from_bitmap(struct btrfs_block_group_cache *block_group,
1198 struct btrfs_free_space *bitmap_info,
1199 u64 *offset, u64 *bytes)
1201 u64 end;
1202 u64 search_start, search_bytes;
1203 int ret;
1205 again:
1206 end = bitmap_info->offset +
1207 (u64)(BITS_PER_BITMAP * block_group->sectorsize) - 1;
1210 * XXX - this can go away after a few releases.
1212 * since the only user of btrfs_remove_free_space is the tree logging
1213 * stuff, and the only way to test that is under crash conditions, we
1214 * want to have this debug stuff here just in case somethings not
1215 * working. Search the bitmap for the space we are trying to use to
1216 * make sure its actually there. If its not there then we need to stop
1217 * because something has gone wrong.
1219 search_start = *offset;
1220 search_bytes = *bytes;
1221 search_bytes = min(search_bytes, end - search_start + 1);
1222 ret = search_bitmap(block_group, bitmap_info, &search_start,
1223 &search_bytes);
1224 BUG_ON(ret < 0 || search_start != *offset);
1226 if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1227 bitmap_clear_bits(block_group, bitmap_info, *offset,
1228 end - *offset + 1);
1229 *bytes -= end - *offset + 1;
1230 *offset = end + 1;
1231 } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1232 bitmap_clear_bits(block_group, bitmap_info, *offset, *bytes);
1233 *bytes = 0;
1236 if (*bytes) {
1237 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1238 if (!bitmap_info->bytes)
1239 free_bitmap(block_group, bitmap_info);
1242 * no entry after this bitmap, but we still have bytes to
1243 * remove, so something has gone wrong.
1245 if (!next)
1246 return -EINVAL;
1248 bitmap_info = rb_entry(next, struct btrfs_free_space,
1249 offset_index);
1252 * if the next entry isn't a bitmap we need to return to let the
1253 * extent stuff do its work.
1255 if (!bitmap_info->bitmap)
1256 return -EAGAIN;
1259 * Ok the next item is a bitmap, but it may not actually hold
1260 * the information for the rest of this free space stuff, so
1261 * look for it, and if we don't find it return so we can try
1262 * everything over again.
1264 search_start = *offset;
1265 search_bytes = *bytes;
1266 ret = search_bitmap(block_group, bitmap_info, &search_start,
1267 &search_bytes);
1268 if (ret < 0 || search_start != *offset)
1269 return -EAGAIN;
1271 goto again;
1272 } else if (!bitmap_info->bytes)
1273 free_bitmap(block_group, bitmap_info);
1275 return 0;
1278 static int insert_into_bitmap(struct btrfs_block_group_cache *block_group,
1279 struct btrfs_free_space *info)
1281 struct btrfs_free_space *bitmap_info;
1282 int added = 0;
1283 u64 bytes, offset, end;
1284 int ret;
1287 * If we are below the extents threshold then we can add this as an
1288 * extent, and don't have to deal with the bitmap
1290 if (block_group->free_extents < block_group->extents_thresh &&
1291 info->bytes > block_group->sectorsize * 4)
1292 return 0;
1295 * some block groups are so tiny they can't be enveloped by a bitmap, so
1296 * don't even bother to create a bitmap for this
1298 if (BITS_PER_BITMAP * block_group->sectorsize >
1299 block_group->key.offset)
1300 return 0;
1302 bytes = info->bytes;
1303 offset = info->offset;
1305 again:
1306 bitmap_info = tree_search_offset(block_group,
1307 offset_to_bitmap(block_group, offset),
1308 1, 0);
1309 if (!bitmap_info) {
1310 BUG_ON(added);
1311 goto new_bitmap;
1314 end = bitmap_info->offset +
1315 (u64)(BITS_PER_BITMAP * block_group->sectorsize);
1317 if (offset >= bitmap_info->offset && offset + bytes > end) {
1318 bitmap_set_bits(block_group, bitmap_info, offset,
1319 end - offset);
1320 bytes -= end - offset;
1321 offset = end;
1322 added = 0;
1323 } else if (offset >= bitmap_info->offset && offset + bytes <= end) {
1324 bitmap_set_bits(block_group, bitmap_info, offset, bytes);
1325 bytes = 0;
1326 } else {
1327 BUG();
1330 if (!bytes) {
1331 ret = 1;
1332 goto out;
1333 } else
1334 goto again;
1336 new_bitmap:
1337 if (info && info->bitmap) {
1338 add_new_bitmap(block_group, info, offset);
1339 added = 1;
1340 info = NULL;
1341 goto again;
1342 } else {
1343 spin_unlock(&block_group->tree_lock);
1345 /* no pre-allocated info, allocate a new one */
1346 if (!info) {
1347 info = kmem_cache_zalloc(btrfs_free_space_cachep,
1348 GFP_NOFS);
1349 if (!info) {
1350 spin_lock(&block_group->tree_lock);
1351 ret = -ENOMEM;
1352 goto out;
1356 /* allocate the bitmap */
1357 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1358 spin_lock(&block_group->tree_lock);
1359 if (!info->bitmap) {
1360 ret = -ENOMEM;
1361 goto out;
1363 goto again;
1366 out:
1367 if (info) {
1368 if (info->bitmap)
1369 kfree(info->bitmap);
1370 kmem_cache_free(btrfs_free_space_cachep, info);
1373 return ret;
1376 bool try_merge_free_space(struct btrfs_block_group_cache *block_group,
1377 struct btrfs_free_space *info, bool update_stat)
1379 struct btrfs_free_space *left_info;
1380 struct btrfs_free_space *right_info;
1381 bool merged = false;
1382 u64 offset = info->offset;
1383 u64 bytes = info->bytes;
1386 * first we want to see if there is free space adjacent to the range we
1387 * are adding, if there is remove that struct and add a new one to
1388 * cover the entire range
1390 right_info = tree_search_offset(block_group, offset + bytes, 0, 0);
1391 if (right_info && rb_prev(&right_info->offset_index))
1392 left_info = rb_entry(rb_prev(&right_info->offset_index),
1393 struct btrfs_free_space, offset_index);
1394 else
1395 left_info = tree_search_offset(block_group, offset - 1, 0, 0);
1397 if (right_info && !right_info->bitmap) {
1398 if (update_stat)
1399 unlink_free_space(block_group, right_info);
1400 else
1401 __unlink_free_space(block_group, right_info);
1402 info->bytes += right_info->bytes;
1403 kmem_cache_free(btrfs_free_space_cachep, right_info);
1404 merged = true;
1407 if (left_info && !left_info->bitmap &&
1408 left_info->offset + left_info->bytes == offset) {
1409 if (update_stat)
1410 unlink_free_space(block_group, left_info);
1411 else
1412 __unlink_free_space(block_group, left_info);
1413 info->offset = left_info->offset;
1414 info->bytes += left_info->bytes;
1415 kmem_cache_free(btrfs_free_space_cachep, left_info);
1416 merged = true;
1419 return merged;
1422 int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
1423 u64 offset, u64 bytes)
1425 struct btrfs_free_space *info;
1426 int ret = 0;
1428 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1429 if (!info)
1430 return -ENOMEM;
1432 info->offset = offset;
1433 info->bytes = bytes;
1435 spin_lock(&block_group->tree_lock);
1437 if (try_merge_free_space(block_group, info, true))
1438 goto link;
1441 * There was no extent directly to the left or right of this new
1442 * extent then we know we're going to have to allocate a new extent, so
1443 * before we do that see if we need to drop this into a bitmap
1445 ret = insert_into_bitmap(block_group, info);
1446 if (ret < 0) {
1447 goto out;
1448 } else if (ret) {
1449 ret = 0;
1450 goto out;
1452 link:
1453 ret = link_free_space(block_group, info);
1454 if (ret)
1455 kmem_cache_free(btrfs_free_space_cachep, info);
1456 out:
1457 spin_unlock(&block_group->tree_lock);
1459 if (ret) {
1460 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1461 BUG_ON(ret == -EEXIST);
1464 return ret;
1467 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1468 u64 offset, u64 bytes)
1470 struct btrfs_free_space *info;
1471 struct btrfs_free_space *next_info = NULL;
1472 int ret = 0;
1474 spin_lock(&block_group->tree_lock);
1476 again:
1477 info = tree_search_offset(block_group, offset, 0, 0);
1478 if (!info) {
1480 * oops didn't find an extent that matched the space we wanted
1481 * to remove, look for a bitmap instead
1483 info = tree_search_offset(block_group,
1484 offset_to_bitmap(block_group, offset),
1485 1, 0);
1486 if (!info) {
1487 WARN_ON(1);
1488 goto out_lock;
1492 if (info->bytes < bytes && rb_next(&info->offset_index)) {
1493 u64 end;
1494 next_info = rb_entry(rb_next(&info->offset_index),
1495 struct btrfs_free_space,
1496 offset_index);
1498 if (next_info->bitmap)
1499 end = next_info->offset + BITS_PER_BITMAP *
1500 block_group->sectorsize - 1;
1501 else
1502 end = next_info->offset + next_info->bytes;
1504 if (next_info->bytes < bytes ||
1505 next_info->offset > offset || offset > end) {
1506 printk(KERN_CRIT "Found free space at %llu, size %llu,"
1507 " trying to use %llu\n",
1508 (unsigned long long)info->offset,
1509 (unsigned long long)info->bytes,
1510 (unsigned long long)bytes);
1511 WARN_ON(1);
1512 ret = -EINVAL;
1513 goto out_lock;
1516 info = next_info;
1519 if (info->bytes == bytes) {
1520 unlink_free_space(block_group, info);
1521 if (info->bitmap) {
1522 kfree(info->bitmap);
1523 block_group->total_bitmaps--;
1525 kmem_cache_free(btrfs_free_space_cachep, info);
1526 goto out_lock;
1529 if (!info->bitmap && info->offset == offset) {
1530 unlink_free_space(block_group, info);
1531 info->offset += bytes;
1532 info->bytes -= bytes;
1533 link_free_space(block_group, info);
1534 goto out_lock;
1537 if (!info->bitmap && info->offset <= offset &&
1538 info->offset + info->bytes >= offset + bytes) {
1539 u64 old_start = info->offset;
1541 * we're freeing space in the middle of the info,
1542 * this can happen during tree log replay
1544 * first unlink the old info and then
1545 * insert it again after the hole we're creating
1547 unlink_free_space(block_group, info);
1548 if (offset + bytes < info->offset + info->bytes) {
1549 u64 old_end = info->offset + info->bytes;
1551 info->offset = offset + bytes;
1552 info->bytes = old_end - info->offset;
1553 ret = link_free_space(block_group, info);
1554 WARN_ON(ret);
1555 if (ret)
1556 goto out_lock;
1557 } else {
1558 /* the hole we're creating ends at the end
1559 * of the info struct, just free the info
1561 kmem_cache_free(btrfs_free_space_cachep, info);
1563 spin_unlock(&block_group->tree_lock);
1565 /* step two, insert a new info struct to cover
1566 * anything before the hole
1568 ret = btrfs_add_free_space(block_group, old_start,
1569 offset - old_start);
1570 WARN_ON(ret);
1571 goto out;
1574 ret = remove_from_bitmap(block_group, info, &offset, &bytes);
1575 if (ret == -EAGAIN)
1576 goto again;
1577 BUG_ON(ret);
1578 out_lock:
1579 spin_unlock(&block_group->tree_lock);
1580 out:
1581 return ret;
1584 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1585 u64 bytes)
1587 struct btrfs_free_space *info;
1588 struct rb_node *n;
1589 int count = 0;
1591 for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
1592 info = rb_entry(n, struct btrfs_free_space, offset_index);
1593 if (info->bytes >= bytes)
1594 count++;
1595 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1596 (unsigned long long)info->offset,
1597 (unsigned long long)info->bytes,
1598 (info->bitmap) ? "yes" : "no");
1600 printk(KERN_INFO "block group has cluster?: %s\n",
1601 list_empty(&block_group->cluster_list) ? "no" : "yes");
1602 printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1603 "\n", count);
1606 u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
1608 struct btrfs_free_space *info;
1609 struct rb_node *n;
1610 u64 ret = 0;
1612 for (n = rb_first(&block_group->free_space_offset); n;
1613 n = rb_next(n)) {
1614 info = rb_entry(n, struct btrfs_free_space, offset_index);
1615 ret += info->bytes;
1618 return ret;
1622 * for a given cluster, put all of its extents back into the free
1623 * space cache. If the block group passed doesn't match the block group
1624 * pointed to by the cluster, someone else raced in and freed the
1625 * cluster already. In that case, we just return without changing anything
1627 static int
1628 __btrfs_return_cluster_to_free_space(
1629 struct btrfs_block_group_cache *block_group,
1630 struct btrfs_free_cluster *cluster)
1632 struct btrfs_free_space *entry;
1633 struct rb_node *node;
1634 bool bitmap;
1636 spin_lock(&cluster->lock);
1637 if (cluster->block_group != block_group)
1638 goto out;
1640 bitmap = cluster->points_to_bitmap;
1641 cluster->block_group = NULL;
1642 cluster->window_start = 0;
1643 list_del_init(&cluster->block_group_list);
1644 cluster->points_to_bitmap = false;
1646 if (bitmap)
1647 goto out;
1649 node = rb_first(&cluster->root);
1650 while (node) {
1651 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1652 node = rb_next(&entry->offset_index);
1653 rb_erase(&entry->offset_index, &cluster->root);
1654 BUG_ON(entry->bitmap);
1655 try_merge_free_space(block_group, entry, false);
1656 tree_insert_offset(&block_group->free_space_offset,
1657 entry->offset, &entry->offset_index, 0);
1659 cluster->root = RB_ROOT;
1661 out:
1662 spin_unlock(&cluster->lock);
1663 btrfs_put_block_group(block_group);
1664 return 0;
1667 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1669 struct btrfs_free_space *info;
1670 struct rb_node *node;
1671 struct btrfs_free_cluster *cluster;
1672 struct list_head *head;
1674 spin_lock(&block_group->tree_lock);
1675 while ((head = block_group->cluster_list.next) !=
1676 &block_group->cluster_list) {
1677 cluster = list_entry(head, struct btrfs_free_cluster,
1678 block_group_list);
1680 WARN_ON(cluster->block_group != block_group);
1681 __btrfs_return_cluster_to_free_space(block_group, cluster);
1682 if (need_resched()) {
1683 spin_unlock(&block_group->tree_lock);
1684 cond_resched();
1685 spin_lock(&block_group->tree_lock);
1689 while ((node = rb_last(&block_group->free_space_offset)) != NULL) {
1690 info = rb_entry(node, struct btrfs_free_space, offset_index);
1691 unlink_free_space(block_group, info);
1692 if (info->bitmap)
1693 kfree(info->bitmap);
1694 kmem_cache_free(btrfs_free_space_cachep, info);
1695 if (need_resched()) {
1696 spin_unlock(&block_group->tree_lock);
1697 cond_resched();
1698 spin_lock(&block_group->tree_lock);
1702 spin_unlock(&block_group->tree_lock);
1705 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1706 u64 offset, u64 bytes, u64 empty_size)
1708 struct btrfs_free_space *entry = NULL;
1709 u64 bytes_search = bytes + empty_size;
1710 u64 ret = 0;
1712 spin_lock(&block_group->tree_lock);
1713 entry = find_free_space(block_group, &offset, &bytes_search, 0);
1714 if (!entry)
1715 goto out;
1717 ret = offset;
1718 if (entry->bitmap) {
1719 bitmap_clear_bits(block_group, entry, offset, bytes);
1720 if (!entry->bytes)
1721 free_bitmap(block_group, entry);
1722 } else {
1723 unlink_free_space(block_group, entry);
1724 entry->offset += bytes;
1725 entry->bytes -= bytes;
1726 if (!entry->bytes)
1727 kmem_cache_free(btrfs_free_space_cachep, entry);
1728 else
1729 link_free_space(block_group, entry);
1732 out:
1733 spin_unlock(&block_group->tree_lock);
1735 return ret;
1739 * given a cluster, put all of its extents back into the free space
1740 * cache. If a block group is passed, this function will only free
1741 * a cluster that belongs to the passed block group.
1743 * Otherwise, it'll get a reference on the block group pointed to by the
1744 * cluster and remove the cluster from it.
1746 int btrfs_return_cluster_to_free_space(
1747 struct btrfs_block_group_cache *block_group,
1748 struct btrfs_free_cluster *cluster)
1750 int ret;
1752 /* first, get a safe pointer to the block group */
1753 spin_lock(&cluster->lock);
1754 if (!block_group) {
1755 block_group = cluster->block_group;
1756 if (!block_group) {
1757 spin_unlock(&cluster->lock);
1758 return 0;
1760 } else if (cluster->block_group != block_group) {
1761 /* someone else has already freed it don't redo their work */
1762 spin_unlock(&cluster->lock);
1763 return 0;
1765 atomic_inc(&block_group->count);
1766 spin_unlock(&cluster->lock);
1768 /* now return any extents the cluster had on it */
1769 spin_lock(&block_group->tree_lock);
1770 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1771 spin_unlock(&block_group->tree_lock);
1773 /* finally drop our ref */
1774 btrfs_put_block_group(block_group);
1775 return ret;
1778 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1779 struct btrfs_free_cluster *cluster,
1780 u64 bytes, u64 min_start)
1782 struct btrfs_free_space *entry;
1783 int err;
1784 u64 search_start = cluster->window_start;
1785 u64 search_bytes = bytes;
1786 u64 ret = 0;
1788 spin_lock(&block_group->tree_lock);
1789 spin_lock(&cluster->lock);
1791 if (!cluster->points_to_bitmap)
1792 goto out;
1794 if (cluster->block_group != block_group)
1795 goto out;
1798 * search_start is the beginning of the bitmap, but at some point it may
1799 * be a good idea to point to the actual start of the free area in the
1800 * bitmap, so do the offset_to_bitmap trick anyway, and set bitmap_only
1801 * to 1 to make sure we get the bitmap entry
1803 entry = tree_search_offset(block_group,
1804 offset_to_bitmap(block_group, search_start),
1805 1, 0);
1806 if (!entry || !entry->bitmap)
1807 goto out;
1809 search_start = min_start;
1810 search_bytes = bytes;
1812 err = search_bitmap(block_group, entry, &search_start,
1813 &search_bytes);
1814 if (err)
1815 goto out;
1817 ret = search_start;
1818 bitmap_clear_bits(block_group, entry, ret, bytes);
1819 if (entry->bytes == 0)
1820 free_bitmap(block_group, entry);
1821 out:
1822 spin_unlock(&cluster->lock);
1823 spin_unlock(&block_group->tree_lock);
1825 return ret;
1829 * given a cluster, try to allocate 'bytes' from it, returns 0
1830 * if it couldn't find anything suitably large, or a logical disk offset
1831 * if things worked out
1833 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
1834 struct btrfs_free_cluster *cluster, u64 bytes,
1835 u64 min_start)
1837 struct btrfs_free_space *entry = NULL;
1838 struct rb_node *node;
1839 u64 ret = 0;
1841 if (cluster->points_to_bitmap)
1842 return btrfs_alloc_from_bitmap(block_group, cluster, bytes,
1843 min_start);
1845 spin_lock(&cluster->lock);
1846 if (bytes > cluster->max_size)
1847 goto out;
1849 if (cluster->block_group != block_group)
1850 goto out;
1852 node = rb_first(&cluster->root);
1853 if (!node)
1854 goto out;
1856 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1858 while(1) {
1859 if (entry->bytes < bytes || entry->offset < min_start) {
1860 struct rb_node *node;
1862 node = rb_next(&entry->offset_index);
1863 if (!node)
1864 break;
1865 entry = rb_entry(node, struct btrfs_free_space,
1866 offset_index);
1867 continue;
1869 ret = entry->offset;
1871 entry->offset += bytes;
1872 entry->bytes -= bytes;
1874 if (entry->bytes == 0)
1875 rb_erase(&entry->offset_index, &cluster->root);
1876 break;
1878 out:
1879 spin_unlock(&cluster->lock);
1881 if (!ret)
1882 return 0;
1884 spin_lock(&block_group->tree_lock);
1886 block_group->free_space -= bytes;
1887 if (entry->bytes == 0) {
1888 block_group->free_extents--;
1889 kmem_cache_free(btrfs_free_space_cachep, entry);
1892 spin_unlock(&block_group->tree_lock);
1894 return ret;
1897 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
1898 struct btrfs_free_space *entry,
1899 struct btrfs_free_cluster *cluster,
1900 u64 offset, u64 bytes, u64 min_bytes)
1902 unsigned long next_zero;
1903 unsigned long i;
1904 unsigned long search_bits;
1905 unsigned long total_bits;
1906 unsigned long found_bits;
1907 unsigned long start = 0;
1908 unsigned long total_found = 0;
1909 bool found = false;
1911 i = offset_to_bit(entry->offset, block_group->sectorsize,
1912 max_t(u64, offset, entry->offset));
1913 search_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
1914 total_bits = bytes_to_bits(bytes, block_group->sectorsize);
1916 again:
1917 found_bits = 0;
1918 for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
1919 i < BITS_PER_BITMAP;
1920 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
1921 next_zero = find_next_zero_bit(entry->bitmap,
1922 BITS_PER_BITMAP, i);
1923 if (next_zero - i >= search_bits) {
1924 found_bits = next_zero - i;
1925 break;
1927 i = next_zero;
1930 if (!found_bits)
1931 return -1;
1933 if (!found) {
1934 start = i;
1935 found = true;
1938 total_found += found_bits;
1940 if (cluster->max_size < found_bits * block_group->sectorsize)
1941 cluster->max_size = found_bits * block_group->sectorsize;
1943 if (total_found < total_bits) {
1944 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
1945 if (i - start > total_bits * 2) {
1946 total_found = 0;
1947 cluster->max_size = 0;
1948 found = false;
1950 goto again;
1953 cluster->window_start = start * block_group->sectorsize +
1954 entry->offset;
1955 cluster->points_to_bitmap = true;
1957 return 0;
1961 * here we try to find a cluster of blocks in a block group. The goal
1962 * is to find at least bytes free and up to empty_size + bytes free.
1963 * We might not find them all in one contiguous area.
1965 * returns zero and sets up cluster if things worked out, otherwise
1966 * it returns -enospc
1968 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
1969 struct btrfs_root *root,
1970 struct btrfs_block_group_cache *block_group,
1971 struct btrfs_free_cluster *cluster,
1972 u64 offset, u64 bytes, u64 empty_size)
1974 struct btrfs_free_space *entry = NULL;
1975 struct rb_node *node;
1976 struct btrfs_free_space *next;
1977 struct btrfs_free_space *last = NULL;
1978 u64 min_bytes;
1979 u64 window_start;
1980 u64 window_free;
1981 u64 max_extent = 0;
1982 bool found_bitmap = false;
1983 int ret;
1985 /* for metadata, allow allocates with more holes */
1986 if (btrfs_test_opt(root, SSD_SPREAD)) {
1987 min_bytes = bytes + empty_size;
1988 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
1990 * we want to do larger allocations when we are
1991 * flushing out the delayed refs, it helps prevent
1992 * making more work as we go along.
1994 if (trans->transaction->delayed_refs.flushing)
1995 min_bytes = max(bytes, (bytes + empty_size) >> 1);
1996 else
1997 min_bytes = max(bytes, (bytes + empty_size) >> 4);
1998 } else
1999 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2001 spin_lock(&block_group->tree_lock);
2004 * If we know we don't have enough space to make a cluster don't even
2005 * bother doing all the work to try and find one.
2007 if (block_group->free_space < min_bytes) {
2008 spin_unlock(&block_group->tree_lock);
2009 return -ENOSPC;
2012 spin_lock(&cluster->lock);
2014 /* someone already found a cluster, hooray */
2015 if (cluster->block_group) {
2016 ret = 0;
2017 goto out;
2019 again:
2020 entry = tree_search_offset(block_group, offset, found_bitmap, 1);
2021 if (!entry) {
2022 ret = -ENOSPC;
2023 goto out;
2027 * If found_bitmap is true, we exhausted our search for extent entries,
2028 * and we just want to search all of the bitmaps that we can find, and
2029 * ignore any extent entries we find.
2031 while (entry->bitmap || found_bitmap ||
2032 (!entry->bitmap && entry->bytes < min_bytes)) {
2033 struct rb_node *node = rb_next(&entry->offset_index);
2035 if (entry->bitmap && entry->bytes > bytes + empty_size) {
2036 ret = btrfs_bitmap_cluster(block_group, entry, cluster,
2037 offset, bytes + empty_size,
2038 min_bytes);
2039 if (!ret)
2040 goto got_it;
2043 if (!node) {
2044 ret = -ENOSPC;
2045 goto out;
2047 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2051 * We already searched all the extent entries from the passed in offset
2052 * to the end and didn't find enough space for the cluster, and we also
2053 * didn't find any bitmaps that met our criteria, just go ahead and exit
2055 if (found_bitmap) {
2056 ret = -ENOSPC;
2057 goto out;
2060 cluster->points_to_bitmap = false;
2061 window_start = entry->offset;
2062 window_free = entry->bytes;
2063 last = entry;
2064 max_extent = entry->bytes;
2066 while (1) {
2067 /* out window is just right, lets fill it */
2068 if (window_free >= bytes + empty_size)
2069 break;
2071 node = rb_next(&last->offset_index);
2072 if (!node) {
2073 if (found_bitmap)
2074 goto again;
2075 ret = -ENOSPC;
2076 goto out;
2078 next = rb_entry(node, struct btrfs_free_space, offset_index);
2081 * we found a bitmap, so if this search doesn't result in a
2082 * cluster, we know to go and search again for the bitmaps and
2083 * start looking for space there
2085 if (next->bitmap) {
2086 if (!found_bitmap)
2087 offset = next->offset;
2088 found_bitmap = true;
2089 last = next;
2090 continue;
2094 * we haven't filled the empty size and the window is
2095 * very large. reset and try again
2097 if (next->offset - (last->offset + last->bytes) > 128 * 1024 ||
2098 next->offset - window_start > (bytes + empty_size) * 2) {
2099 entry = next;
2100 window_start = entry->offset;
2101 window_free = entry->bytes;
2102 last = entry;
2103 max_extent = entry->bytes;
2104 } else {
2105 last = next;
2106 window_free += next->bytes;
2107 if (entry->bytes > max_extent)
2108 max_extent = entry->bytes;
2112 cluster->window_start = entry->offset;
2115 * now we've found our entries, pull them out of the free space
2116 * cache and put them into the cluster rbtree
2118 * The cluster includes an rbtree, but only uses the offset index
2119 * of each free space cache entry.
2121 while (1) {
2122 node = rb_next(&entry->offset_index);
2123 if (entry->bitmap && node) {
2124 entry = rb_entry(node, struct btrfs_free_space,
2125 offset_index);
2126 continue;
2127 } else if (entry->bitmap && !node) {
2128 break;
2131 rb_erase(&entry->offset_index, &block_group->free_space_offset);
2132 ret = tree_insert_offset(&cluster->root, entry->offset,
2133 &entry->offset_index, 0);
2134 BUG_ON(ret);
2136 if (!node || entry == last)
2137 break;
2139 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2142 cluster->max_size = max_extent;
2143 got_it:
2144 ret = 0;
2145 atomic_inc(&block_group->count);
2146 list_add_tail(&cluster->block_group_list, &block_group->cluster_list);
2147 cluster->block_group = block_group;
2148 out:
2149 spin_unlock(&cluster->lock);
2150 spin_unlock(&block_group->tree_lock);
2152 return ret;
2156 * simple code to zero out a cluster
2158 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2160 spin_lock_init(&cluster->lock);
2161 spin_lock_init(&cluster->refill_lock);
2162 cluster->root = RB_ROOT;
2163 cluster->max_size = 0;
2164 cluster->points_to_bitmap = false;
2165 INIT_LIST_HEAD(&cluster->block_group_list);
2166 cluster->block_group = NULL;