?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge remote-tracking branch 'slab/for-next'
[linux-2.6/next.git] / fs / btrfs / free-space-cache.c
blob6a265b9f85f2cb282940f127bb39d9c56cb5082f
1 /*
2 * Copyright (C) 2008 Red Hat. All rights reserved.
3 *
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.
7 *
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.
12 *
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.
17 */
18
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"
27 #include "extent_io.h"
28 #include "inode-map.h"
29
30 #define BITS_PER_BITMAP (PAGE_CACHE_SIZE * 8)
31 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
32
33 static int link_free_space(struct btrfs_free_space_ctl *ctl,
34 struct btrfs_free_space *info);
35
36 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
37 struct btrfs_path *path,
38 u64 offset)
39 {
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;
47
48 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
49 key.offset = offset;
50 key.type = 0;
51
52 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
53 if (ret < 0)
54 return ERR_PTR(ret);
55 if (ret > 0) {
56 btrfs_release_path(path);
57 return ERR_PTR(-ENOENT);
58 }
59
60 leaf = path->nodes[0];
61 header = btrfs_item_ptr(leaf, path->slots[0],
62 struct btrfs_free_space_header);
63 btrfs_free_space_key(leaf, header, &disk_key);
64 btrfs_disk_key_to_cpu(&location, &disk_key);
65 btrfs_release_path(path);
66
67 inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
68 if (!inode)
69 return ERR_PTR(-ENOENT);
70 if (IS_ERR(inode))
71 return inode;
72 if (is_bad_inode(inode)) {
73 iput(inode);
74 return ERR_PTR(-ENOENT);
75 }
76
77 inode->i_mapping->flags &= ~__GFP_FS;
78
79 return inode;
80 }
81
82 struct inode *lookup_free_space_inode(struct btrfs_root *root,
83 struct btrfs_block_group_cache
84 *block_group, struct btrfs_path *path)
85 {
86 struct inode *inode = NULL;
87
88 spin_lock(&block_group->lock);
89 if (block_group->inode)
90 inode = igrab(block_group->inode);
91 spin_unlock(&block_group->lock);
92 if (inode)
93 return inode;
94
95 inode = __lookup_free_space_inode(root, path,
96 block_group->key.objectid);
97 if (IS_ERR(inode))
98 return inode;
99
100 spin_lock(&block_group->lock);
101 if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM) {
102 printk(KERN_INFO "Old style space inode found, converting.\n");
103 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NODATASUM;
104 block_group->disk_cache_state = BTRFS_DC_CLEAR;
105 }
106
107 if (!btrfs_fs_closing(root->fs_info)) {
108 block_group->inode = igrab(inode);
109 block_group->iref = 1;
110 }
111 spin_unlock(&block_group->lock);
112
113 return inode;
114 }
115
116 int __create_free_space_inode(struct btrfs_root *root,
117 struct btrfs_trans_handle *trans,
118 struct btrfs_path *path, u64 ino, u64 offset)
119 {
120 struct btrfs_key key;
121 struct btrfs_disk_key disk_key;
122 struct btrfs_free_space_header *header;
123 struct btrfs_inode_item *inode_item;
124 struct extent_buffer *leaf;
125 int ret;
126
127 ret = btrfs_insert_empty_inode(trans, root, path, ino);
128 if (ret)
129 return ret;
130
131 leaf = path->nodes[0];
132 inode_item = btrfs_item_ptr(leaf, path->slots[0],
133 struct btrfs_inode_item);
134 btrfs_item_key(leaf, &disk_key, path->slots[0]);
135 memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
136 sizeof(*inode_item));
137 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
138 btrfs_set_inode_size(leaf, inode_item, 0);
139 btrfs_set_inode_nbytes(leaf, inode_item, 0);
140 btrfs_set_inode_uid(leaf, inode_item, 0);
141 btrfs_set_inode_gid(leaf, inode_item, 0);
142 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
143 btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
144 BTRFS_INODE_PREALLOC);
145 btrfs_set_inode_nlink(leaf, inode_item, 1);
146 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
147 btrfs_set_inode_block_group(leaf, inode_item, offset);
148 btrfs_mark_buffer_dirty(leaf);
149 btrfs_release_path(path);
150
151 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
152 key.offset = offset;
153 key.type = 0;
154
155 ret = btrfs_insert_empty_item(trans, root, path, &key,
156 sizeof(struct btrfs_free_space_header));
157 if (ret < 0) {
158 btrfs_release_path(path);
159 return ret;
160 }
161 leaf = path->nodes[0];
162 header = btrfs_item_ptr(leaf, path->slots[0],
163 struct btrfs_free_space_header);
164 memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
165 btrfs_set_free_space_key(leaf, header, &disk_key);
166 btrfs_mark_buffer_dirty(leaf);
167 btrfs_release_path(path);
168
169 return 0;
170 }
171
172 int create_free_space_inode(struct btrfs_root *root,
173 struct btrfs_trans_handle *trans,
174 struct btrfs_block_group_cache *block_group,
175 struct btrfs_path *path)
176 {
177 int ret;
178 u64 ino;
179
180 ret = btrfs_find_free_objectid(root, &ino);
181 if (ret < 0)
182 return ret;
183
184 return __create_free_space_inode(root, trans, path, ino,
185 block_group->key.objectid);
186 }
187
188 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
189 struct btrfs_trans_handle *trans,
190 struct btrfs_path *path,
191 struct inode *inode)
192 {
193 loff_t oldsize;
194 int ret = 0;
195
196 trans->block_rsv = root->orphan_block_rsv;
197 ret = btrfs_block_rsv_check(trans, root,
198 root->orphan_block_rsv,
199 0, 5);
200 if (ret)
201 return ret;
202
203 oldsize = i_size_read(inode);
204 btrfs_i_size_write(inode, 0);
205 truncate_pagecache(inode, oldsize, 0);
206
207 /*
208 * We don't need an orphan item because truncating the free space cache
209 * will never be split across transactions.
210 */
211 ret = btrfs_truncate_inode_items(trans, root, inode,
212 0, BTRFS_EXTENT_DATA_KEY);
213 if (ret) {
214 WARN_ON(1);
215 return ret;
216 }
217
218 ret = btrfs_update_inode(trans, root, inode);
219 return ret;
220 }
221
222 static int readahead_cache(struct inode *inode)
223 {
224 struct file_ra_state *ra;
225 unsigned long last_index;
226
227 ra = kzalloc(sizeof(*ra), GFP_NOFS);
228 if (!ra)
229 return -ENOMEM;
230
231 file_ra_state_init(ra, inode->i_mapping);
232 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
233
234 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
235
236 kfree(ra);
237
238 return 0;
239 }
240
241 int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
242 struct btrfs_free_space_ctl *ctl,
243 struct btrfs_path *path, u64 offset)
244 {
245 struct btrfs_free_space_header *header;
246 struct extent_buffer *leaf;
247 struct page *page;
248 struct btrfs_key key;
249 struct list_head bitmaps;
250 u64 num_entries;
251 u64 num_bitmaps;
252 u64 generation;
253 pgoff_t index = 0;
254 int ret = 0;
255
256 INIT_LIST_HEAD(&bitmaps);
257
258 /* Nothing in the space cache, goodbye */
259 if (!i_size_read(inode))
260 goto out;
261
262 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
263 key.offset = offset;
264 key.type = 0;
265
266 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
267 if (ret < 0)
268 goto out;
269 else if (ret > 0) {
270 btrfs_release_path(path);
271 ret = 0;
272 goto out;
273 }
274
275 ret = -1;
276
277 leaf = path->nodes[0];
278 header = btrfs_item_ptr(leaf, path->slots[0],
279 struct btrfs_free_space_header);
280 num_entries = btrfs_free_space_entries(leaf, header);
281 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
282 generation = btrfs_free_space_generation(leaf, header);
283 btrfs_release_path(path);
284
285 if (BTRFS_I(inode)->generation != generation) {
286 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
287 " not match free space cache generation (%llu)\n",
288 (unsigned long long)BTRFS_I(inode)->generation,
289 (unsigned long long)generation);
290 goto out;
291 }
292
293 if (!num_entries)
294 goto out;
295
296 ret = readahead_cache(inode);
297 if (ret)
298 goto out;
299
300 while (1) {
301 struct btrfs_free_space_entry *entry;
302 struct btrfs_free_space *e;
303 void *addr;
304 unsigned long offset = 0;
305 int need_loop = 0;
306
307 if (!num_entries && !num_bitmaps)
308 break;
309
310 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
311 if (!page)
312 goto free_cache;
313
314 if (!PageUptodate(page)) {
315 btrfs_readpage(NULL, page);
316 lock_page(page);
317 if (!PageUptodate(page)) {
318 unlock_page(page);
319 page_cache_release(page);
320 printk(KERN_ERR "btrfs: error reading free "
321 "space cache\n");
322 goto free_cache;
323 }
324 }
325 addr = kmap(page);
326
327 if (index == 0) {
328 u64 *gen;
329
330 /*
331 * We put a bogus crc in the front of the first page in
332 * case old kernels try to mount a fs with the new
333 * format to make sure they discard the cache.
334 */
335 addr += sizeof(u64);
336 offset += sizeof(u64);
337
338 gen = addr;
339 if (*gen != BTRFS_I(inode)->generation) {
340 printk(KERN_ERR "btrfs: space cache generation"
341 " (%llu) does not match inode (%llu)\n",
342 (unsigned long long)*gen,
343 (unsigned long long)
344 BTRFS_I(inode)->generation);
345 kunmap(page);
346 unlock_page(page);
347 page_cache_release(page);
348 goto free_cache;
349 }
350 addr += sizeof(u64);
351 offset += sizeof(u64);
352 }
353 entry = addr;
354
355 while (1) {
356 if (!num_entries)
357 break;
358
359 need_loop = 1;
360 e = kmem_cache_zalloc(btrfs_free_space_cachep,
361 GFP_NOFS);
362 if (!e) {
363 kunmap(page);
364 unlock_page(page);
365 page_cache_release(page);
366 goto free_cache;
367 }
368
369 e->offset = le64_to_cpu(entry->offset);
370 e->bytes = le64_to_cpu(entry->bytes);
371 if (!e->bytes) {
372 kunmap(page);
373 kmem_cache_free(btrfs_free_space_cachep, e);
374 unlock_page(page);
375 page_cache_release(page);
376 goto free_cache;
377 }
378
379 if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
380 spin_lock(&ctl->tree_lock);
381 ret = link_free_space(ctl, e);
382 spin_unlock(&ctl->tree_lock);
383 if (ret) {
384 printk(KERN_ERR "Duplicate entries in "
385 "free space cache, dumping\n");
386 kunmap(page);
387 unlock_page(page);
388 page_cache_release(page);
389 goto free_cache;
390 }
391 } else {
392 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
393 if (!e->bitmap) {
394 kunmap(page);
395 kmem_cache_free(
396 btrfs_free_space_cachep, e);
397 unlock_page(page);
398 page_cache_release(page);
399 goto free_cache;
400 }
401 spin_lock(&ctl->tree_lock);
402 ret = link_free_space(ctl, e);
403 ctl->total_bitmaps++;
404 ctl->op->recalc_thresholds(ctl);
405 spin_unlock(&ctl->tree_lock);
406 if (ret) {
407 printk(KERN_ERR "Duplicate entries in "
408 "free space cache, dumping\n");
409 kunmap(page);
410 unlock_page(page);
411 page_cache_release(page);
412 goto free_cache;
413 }
414 list_add_tail(&e->list, &bitmaps);
415 }
416
417 num_entries--;
418 offset += sizeof(struct btrfs_free_space_entry);
419 if (offset + sizeof(struct btrfs_free_space_entry) >=
420 PAGE_CACHE_SIZE)
421 break;
422 entry++;
423 }
424
425 /*
426 * We read an entry out of this page, we need to move on to the
427 * next page.
428 */
429 if (need_loop) {
430 kunmap(page);
431 goto next;
432 }
433
434 /*
435 * We add the bitmaps at the end of the entries in order that
436 * the bitmap entries are added to the cache.
437 */
438 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
439 list_del_init(&e->list);
440 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
441 kunmap(page);
442 num_bitmaps--;
443 next:
444 unlock_page(page);
445 page_cache_release(page);
446 index++;
447 }
448
449 ret = 1;
450 out:
451 return ret;
452 free_cache:
453 __btrfs_remove_free_space_cache(ctl);
454 goto out;
455 }
456
457 int load_free_space_cache(struct btrfs_fs_info *fs_info,
458 struct btrfs_block_group_cache *block_group)
459 {
460 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
461 struct btrfs_root *root = fs_info->tree_root;
462 struct inode *inode;
463 struct btrfs_path *path;
464 int ret;
465 bool matched;
466 u64 used = btrfs_block_group_used(&block_group->item);
467
468 /*
469 * If we're unmounting then just return, since this does a search on the
470 * normal root and not the commit root and we could deadlock.
471 */
472 if (btrfs_fs_closing(fs_info))
473 return 0;
474
475 /*
476 * If this block group has been marked to be cleared for one reason or
477 * another then we can't trust the on disk cache, so just return.
478 */
479 spin_lock(&block_group->lock);
480 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
481 spin_unlock(&block_group->lock);
482 return 0;
483 }
484 spin_unlock(&block_group->lock);
485
486 path = btrfs_alloc_path();
487 if (!path)
488 return 0;
489
490 inode = lookup_free_space_inode(root, block_group, path);
491 if (IS_ERR(inode)) {
492 btrfs_free_path(path);
493 return 0;
494 }
495
496 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
497 path, block_group->key.objectid);
498 btrfs_free_path(path);
499 if (ret <= 0)
500 goto out;
501
502 spin_lock(&ctl->tree_lock);
503 matched = (ctl->free_space == (block_group->key.offset - used -
504 block_group->bytes_super));
505 spin_unlock(&ctl->tree_lock);
506
507 if (!matched) {
508 __btrfs_remove_free_space_cache(ctl);
509 printk(KERN_ERR "block group %llu has an wrong amount of free "
510 "space\n", block_group->key.objectid);
511 ret = -1;
512 }
513 out:
514 if (ret < 0) {
515 /* This cache is bogus, make sure it gets cleared */
516 spin_lock(&block_group->lock);
517 block_group->disk_cache_state = BTRFS_DC_CLEAR;
518 spin_unlock(&block_group->lock);
519 ret = 0;
520
521 printk(KERN_ERR "btrfs: failed to load free space cache "
522 "for block group %llu\n", block_group->key.objectid);
523 }
524
525 iput(inode);
526 return ret;
527 }
528
529 int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
530 struct btrfs_free_space_ctl *ctl,
531 struct btrfs_block_group_cache *block_group,
532 struct btrfs_trans_handle *trans,
533 struct btrfs_path *path, u64 offset)
534 {
535 struct btrfs_free_space_header *header;
536 struct extent_buffer *leaf;
537 struct rb_node *node;
538 struct list_head *pos, *n;
539 struct page **pages;
540 struct page *page;
541 struct extent_state *cached_state = NULL;
542 struct btrfs_free_cluster *cluster = NULL;
543 struct extent_io_tree *unpin = NULL;
544 struct list_head bitmap_list;
545 struct btrfs_key key;
546 u64 start, end, len;
547 u64 bytes = 0;
548 u32 crc = ~(u32)0;
549 int index = 0, num_pages = 0;
550 int entries = 0;
551 int bitmaps = 0;
552 int ret = -1;
553 bool next_page = false;
554 bool out_of_space = false;
555
556 INIT_LIST_HEAD(&bitmap_list);
557
558 node = rb_first(&ctl->free_space_offset);
559 if (!node)
560 return 0;
561
562 if (!i_size_read(inode))
563 return -1;
564
565 num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
566 PAGE_CACHE_SHIFT;
567
568 filemap_write_and_wait(inode->i_mapping);
569 btrfs_wait_ordered_range(inode, inode->i_size &
570 ~(root->sectorsize - 1), (u64)-1);
571
572 pages = kzalloc(sizeof(struct page *) * num_pages, GFP_NOFS);
573 if (!pages)
574 return -1;
575
576 /* Get the cluster for this block_group if it exists */
577 if (block_group && !list_empty(&block_group->cluster_list))
578 cluster = list_entry(block_group->cluster_list.next,
579 struct btrfs_free_cluster,
580 block_group_list);
581
582 /*
583 * We shouldn't have switched the pinned extents yet so this is the
584 * right one
585 */
586 unpin = root->fs_info->pinned_extents;
587
588 /*
589 * Lock all pages first so we can lock the extent safely.
590 *
591 * NOTE: Because we hold the ref the entire time we're going to write to
592 * the page find_get_page should never fail, so we don't do a check
593 * after find_get_page at this point. Just putting this here so people
594 * know and don't freak out.
595 */
596 while (index < num_pages) {
597 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
598 if (!page) {
599 int i;
600
601 for (i = 0; i < num_pages; i++) {
602 unlock_page(pages[i]);
603 page_cache_release(pages[i]);
604 }
605 goto out;
606 }
607 pages[index] = page;
608 index++;
609 }
610
611 index = 0;
612 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
613 0, &cached_state, GFP_NOFS);
614
615 /*
616 * When searching for pinned extents, we need to start at our start
617 * offset.
618 */
619 if (block_group)
620 start = block_group->key.objectid;
621
622 /* Write out the extent entries */
623 do {
624 struct btrfs_free_space_entry *entry;
625 void *addr, *orig;
626 unsigned long offset = 0;
627
628 next_page = false;
629
630 if (index >= num_pages) {
631 out_of_space = true;
632 break;
633 }
634
635 page = pages[index];
636
637 orig = addr = kmap(page);
638 if (index == 0) {
639 u64 *gen;
640
641 /*
642 * We're going to put in a bogus crc for this page to
643 * make sure that old kernels who aren't aware of this
644 * format will be sure to discard the cache.
645 */
646 addr += sizeof(u64);
647 offset += sizeof(u64);
648
649 gen = addr;
650 *gen = trans->transid;
651 addr += sizeof(u64);
652 offset += sizeof(u64);
653 }
654 entry = addr;
655
656 memset(addr, 0, PAGE_CACHE_SIZE - offset);
657 while (node && !next_page) {
658 struct btrfs_free_space *e;
659
660 e = rb_entry(node, struct btrfs_free_space, offset_index);
661 entries++;
662
663 entry->offset = cpu_to_le64(e->offset);
664 entry->bytes = cpu_to_le64(e->bytes);
665 if (e->bitmap) {
666 entry->type = BTRFS_FREE_SPACE_BITMAP;
667 list_add_tail(&e->list, &bitmap_list);
668 bitmaps++;
669 } else {
670 entry->type = BTRFS_FREE_SPACE_EXTENT;
671 }
672 node = rb_next(node);
673 if (!node && cluster) {
674 node = rb_first(&cluster->root);
675 cluster = NULL;
676 }
677 offset += sizeof(struct btrfs_free_space_entry);
678 if (offset + sizeof(struct btrfs_free_space_entry) >=
679 PAGE_CACHE_SIZE)
680 next_page = true;
681 entry++;
682 }
683
684 /*
685 * We want to add any pinned extents to our free space cache
686 * so we don't leak the space
687 */
688 while (block_group && !next_page &&
689 (start < block_group->key.objectid +
690 block_group->key.offset)) {
691 ret = find_first_extent_bit(unpin, start, &start, &end,
692 EXTENT_DIRTY);
693 if (ret) {
694 ret = 0;
695 break;
696 }
697
698 /* This pinned extent is out of our range */
699 if (start >= block_group->key.objectid +
700 block_group->key.offset)
701 break;
702
703 len = block_group->key.objectid +
704 block_group->key.offset - start;
705 len = min(len, end + 1 - start);
706
707 entries++;
708 entry->offset = cpu_to_le64(start);
709 entry->bytes = cpu_to_le64(len);
710 entry->type = BTRFS_FREE_SPACE_EXTENT;
711
712 start = end + 1;
713 offset += sizeof(struct btrfs_free_space_entry);
714 if (offset + sizeof(struct btrfs_free_space_entry) >=
715 PAGE_CACHE_SIZE)
716 next_page = true;
717 entry++;
718 }
719
720 /* Generate bogus crc value */
721 if (index == 0) {
722 u32 *tmp;
723 crc = btrfs_csum_data(root, orig + sizeof(u64), crc,
724 PAGE_CACHE_SIZE - sizeof(u64));
725 btrfs_csum_final(crc, (char *)&crc);
726 crc++;
727 tmp = orig;
728 *tmp = crc;
729 }
730
731 kunmap(page);
732
733 bytes += PAGE_CACHE_SIZE;
734
735 index++;
736 } while (node || next_page);
737
738 /* Write out the bitmaps */
739 list_for_each_safe(pos, n, &bitmap_list) {
740 void *addr;
741 struct btrfs_free_space *entry =
742 list_entry(pos, struct btrfs_free_space, list);
743
744 if (index >= num_pages) {
745 out_of_space = true;
746 break;
747 }
748 page = pages[index];
749
750 addr = kmap(page);
751 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
752 kunmap(page);
753 bytes += PAGE_CACHE_SIZE;
754
755 list_del_init(&entry->list);
756 index++;
757 }
758
759 if (out_of_space) {
760 btrfs_drop_pages(pages, num_pages);
761 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
762 i_size_read(inode) - 1, &cached_state,
763 GFP_NOFS);
764 ret = 0;
765 goto out;
766 }
767
768 /* Zero out the rest of the pages just to make sure */
769 while (index < num_pages) {
770 void *addr;
771
772 page = pages[index];
773 addr = kmap(page);
774 memset(addr, 0, PAGE_CACHE_SIZE);
775 kunmap(page);
776 bytes += PAGE_CACHE_SIZE;
777 index++;
778 }
779
780 ret = btrfs_dirty_pages(root, inode, pages, num_pages, 0,
781 bytes, &cached_state);
782 btrfs_drop_pages(pages, num_pages);
783 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
784 i_size_read(inode) - 1, &cached_state, GFP_NOFS);
785
786 if (ret) {
787 ret = 0;
788 goto out;
789 }
790
791 BTRFS_I(inode)->generation = trans->transid;
792
793 filemap_write_and_wait(inode->i_mapping);
794
795 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
796 key.offset = offset;
797 key.type = 0;
798
799 ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
800 if (ret < 0) {
801 ret = -1;
802 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
803 EXTENT_DIRTY | EXTENT_DELALLOC |
804 EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
805 goto out;
806 }
807 leaf = path->nodes[0];
808 if (ret > 0) {
809 struct btrfs_key found_key;
810 BUG_ON(!path->slots[0]);
811 path->slots[0]--;
812 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
813 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
814 found_key.offset != offset) {
815 ret = -1;
816 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
817 EXTENT_DIRTY | EXTENT_DELALLOC |
818 EXTENT_DO_ACCOUNTING, 0, 0, NULL,
819 GFP_NOFS);
820 btrfs_release_path(path);
821 goto out;
822 }
823 }
824 header = btrfs_item_ptr(leaf, path->slots[0],
825 struct btrfs_free_space_header);
826 btrfs_set_free_space_entries(leaf, header, entries);
827 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
828 btrfs_set_free_space_generation(leaf, header, trans->transid);
829 btrfs_mark_buffer_dirty(leaf);
830 btrfs_release_path(path);
831
832 ret = 1;
833
834 out:
835 kfree(pages);
836 if (ret != 1) {
837 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
838 BTRFS_I(inode)->generation = 0;
839 }
840 btrfs_update_inode(trans, root, inode);
841 return ret;
842 }
843
844 int btrfs_write_out_cache(struct btrfs_root *root,
845 struct btrfs_trans_handle *trans,
846 struct btrfs_block_group_cache *block_group,
847 struct btrfs_path *path)
848 {
849 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
850 struct inode *inode;
851 int ret = 0;
852
853 root = root->fs_info->tree_root;
854
855 spin_lock(&block_group->lock);
856 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
857 spin_unlock(&block_group->lock);
858 return 0;
859 }
860 spin_unlock(&block_group->lock);
861
862 inode = lookup_free_space_inode(root, block_group, path);
863 if (IS_ERR(inode))
864 return 0;
865
866 ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
867 path, block_group->key.objectid);
868 if (ret < 0) {
869 spin_lock(&block_group->lock);
870 block_group->disk_cache_state = BTRFS_DC_ERROR;
871 spin_unlock(&block_group->lock);
872 ret = 0;
873
874 printk(KERN_ERR "btrfs: failed to write free space cace "
875 "for block group %llu\n", block_group->key.objectid);
876 }
877
878 iput(inode);
879 return ret;
880 }
881
882 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
883 u64 offset)
884 {
885 BUG_ON(offset < bitmap_start);
886 offset -= bitmap_start;
887 return (unsigned long)(div_u64(offset, unit));
888 }
889
890 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
891 {
892 return (unsigned long)(div_u64(bytes, unit));
893 }
894
895 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
896 u64 offset)
897 {
898 u64 bitmap_start;
899 u64 bytes_per_bitmap;
900
901 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
902 bitmap_start = offset - ctl->start;
903 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
904 bitmap_start *= bytes_per_bitmap;
905 bitmap_start += ctl->start;
906
907 return bitmap_start;
908 }
909
910 static int tree_insert_offset(struct rb_root *root, u64 offset,
911 struct rb_node *node, int bitmap)
912 {
913 struct rb_node **p = &root->rb_node;
914 struct rb_node *parent = NULL;
915 struct btrfs_free_space *info;
916
917 while (*p) {
918 parent = *p;
919 info = rb_entry(parent, struct btrfs_free_space, offset_index);
920
921 if (offset < info->offset) {
922 p = &(*p)->rb_left;
923 } else if (offset > info->offset) {
924 p = &(*p)->rb_right;
925 } else {
926 /*
927 * we could have a bitmap entry and an extent entry
928 * share the same offset. If this is the case, we want
929 * the extent entry to always be found first if we do a
930 * linear search through the tree, since we want to have
931 * the quickest allocation time, and allocating from an
932 * extent is faster than allocating from a bitmap. So
933 * if we're inserting a bitmap and we find an entry at
934 * this offset, we want to go right, or after this entry
935 * logically. If we are inserting an extent and we've
936 * found a bitmap, we want to go left, or before
937 * logically.
938 */
939 if (bitmap) {
940 if (info->bitmap) {
941 WARN_ON_ONCE(1);
942 return -EEXIST;
943 }
944 p = &(*p)->rb_right;
945 } else {
946 if (!info->bitmap) {
947 WARN_ON_ONCE(1);
948 return -EEXIST;
949 }
950 p = &(*p)->rb_left;
951 }
952 }
953 }
954
955 rb_link_node(node, parent, p);
956 rb_insert_color(node, root);
957
958 return 0;
959 }
960
961 /*
962 * searches the tree for the given offset.
963 *
964 * fuzzy - If this is set, then we are trying to make an allocation, and we just
965 * want a section that has at least bytes size and comes at or after the given
966 * offset.
967 */
968 static struct btrfs_free_space *
969 tree_search_offset(struct btrfs_free_space_ctl *ctl,
970 u64 offset, int bitmap_only, int fuzzy)
971 {
972 struct rb_node *n = ctl->free_space_offset.rb_node;
973 struct btrfs_free_space *entry, *prev = NULL;
974
975 /* find entry that is closest to the 'offset' */
976 while (1) {
977 if (!n) {
978 entry = NULL;
979 break;
980 }
981
982 entry = rb_entry(n, struct btrfs_free_space, offset_index);
983 prev = entry;
984
985 if (offset < entry->offset)
986 n = n->rb_left;
987 else if (offset > entry->offset)
988 n = n->rb_right;
989 else
990 break;
991 }
992
993 if (bitmap_only) {
994 if (!entry)
995 return NULL;
996 if (entry->bitmap)
997 return entry;
998
999 /*
1000 * bitmap entry and extent entry may share same offset,
1001 * in that case, bitmap entry comes after extent entry.
1002 */
1003 n = rb_next(n);
1004 if (!n)
1005 return NULL;
1006 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1007 if (entry->offset != offset)
1008 return NULL;
1009
1010 WARN_ON(!entry->bitmap);
1011 return entry;
1012 } else if (entry) {
1013 if (entry->bitmap) {
1014 /*
1015 * if previous extent entry covers the offset,
1016 * we should return it instead of the bitmap entry
1017 */
1018 n = &entry->offset_index;
1019 while (1) {
1020 n = rb_prev(n);
1021 if (!n)
1022 break;
1023 prev = rb_entry(n, struct btrfs_free_space,
1024 offset_index);
1025 if (!prev->bitmap) {
1026 if (prev->offset + prev->bytes > offset)
1027 entry = prev;
1028 break;
1029 }
1030 }
1031 }
1032 return entry;
1033 }
1034
1035 if (!prev)
1036 return NULL;
1037
1038 /* find last entry before the 'offset' */
1039 entry = prev;
1040 if (entry->offset > offset) {
1041 n = rb_prev(&entry->offset_index);
1042 if (n) {
1043 entry = rb_entry(n, struct btrfs_free_space,
1044 offset_index);
1045 BUG_ON(entry->offset > offset);
1046 } else {
1047 if (fuzzy)
1048 return entry;
1049 else
1050 return NULL;
1051 }
1052 }
1053
1054 if (entry->bitmap) {
1055 n = &entry->offset_index;
1056 while (1) {
1057 n = rb_prev(n);
1058 if (!n)
1059 break;
1060 prev = rb_entry(n, struct btrfs_free_space,
1061 offset_index);
1062 if (!prev->bitmap) {
1063 if (prev->offset + prev->bytes > offset)
1064 return prev;
1065 break;
1066 }
1067 }
1068 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1069 return entry;
1070 } else if (entry->offset + entry->bytes > offset)
1071 return entry;
1072
1073 if (!fuzzy)
1074 return NULL;
1075
1076 while (1) {
1077 if (entry->bitmap) {
1078 if (entry->offset + BITS_PER_BITMAP *
1079 ctl->unit > offset)
1080 break;
1081 } else {
1082 if (entry->offset + entry->bytes > offset)
1083 break;
1084 }
1085
1086 n = rb_next(&entry->offset_index);
1087 if (!n)
1088 return NULL;
1089 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1090 }
1091 return entry;
1092 }
1093
1094 static inline void
1095 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1096 struct btrfs_free_space *info)
1097 {
1098 rb_erase(&info->offset_index, &ctl->free_space_offset);
1099 ctl->free_extents--;
1100 }
1101
1102 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1103 struct btrfs_free_space *info)
1104 {
1105 __unlink_free_space(ctl, info);
1106 ctl->free_space -= info->bytes;
1107 }
1108
1109 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1110 struct btrfs_free_space *info)
1111 {
1112 int ret = 0;
1113
1114 BUG_ON(!info->bitmap && !info->bytes);
1115 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1116 &info->offset_index, (info->bitmap != NULL));
1117 if (ret)
1118 return ret;
1119
1120 ctl->free_space += info->bytes;
1121 ctl->free_extents++;
1122 return ret;
1123 }
1124
1125 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1126 {
1127 struct btrfs_block_group_cache *block_group = ctl->private;
1128 u64 max_bytes;
1129 u64 bitmap_bytes;
1130 u64 extent_bytes;
1131 u64 size = block_group->key.offset;
1132 u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1133 int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1134
1135 BUG_ON(ctl->total_bitmaps > max_bitmaps);
1136
1137 /*
1138 * The goal is to keep the total amount of memory used per 1gb of space
1139 * at or below 32k, so we need to adjust how much memory we allow to be
1140 * used by extent based free space tracking
1141 */
1142 if (size < 1024 * 1024 * 1024)
1143 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1144 else
1145 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1146 div64_u64(size, 1024 * 1024 * 1024);
1147
1148 /*
1149 * we want to account for 1 more bitmap than what we have so we can make
1150 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1151 * we add more bitmaps.
1152 */
1153 bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1154
1155 if (bitmap_bytes >= max_bytes) {
1156 ctl->extents_thresh = 0;
1157 return;
1158 }
1159
1160 /*
1161 * we want the extent entry threshold to always be at most 1/2 the maxw
1162 * bytes we can have, or whatever is less than that.
1163 */
1164 extent_bytes = max_bytes - bitmap_bytes;
1165 extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1166
1167 ctl->extents_thresh =
1168 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1169 }
1170
1171 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1172 struct btrfs_free_space *info,
1173 u64 offset, u64 bytes)
1174 {
1175 unsigned long start, count;
1176
1177 start = offset_to_bit(info->offset, ctl->unit, offset);
1178 count = bytes_to_bits(bytes, ctl->unit);
1179 BUG_ON(start + count > BITS_PER_BITMAP);
1180
1181 bitmap_clear(info->bitmap, start, count);
1182
1183 info->bytes -= bytes;
1184 }
1185
1186 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1187 struct btrfs_free_space *info, u64 offset,
1188 u64 bytes)
1189 {
1190 __bitmap_clear_bits(ctl, info, offset, bytes);
1191 ctl->free_space -= bytes;
1192 }
1193
1194 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1195 struct btrfs_free_space *info, u64 offset,
1196 u64 bytes)
1197 {
1198 unsigned long start, count;
1199
1200 start = offset_to_bit(info->offset, ctl->unit, offset);
1201 count = bytes_to_bits(bytes, ctl->unit);
1202 BUG_ON(start + count > BITS_PER_BITMAP);
1203
1204 bitmap_set(info->bitmap, start, count);
1205
1206 info->bytes += bytes;
1207 ctl->free_space += bytes;
1208 }
1209
1210 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1211 struct btrfs_free_space *bitmap_info, u64 *offset,
1212 u64 *bytes)
1213 {
1214 unsigned long found_bits = 0;
1215 unsigned long bits, i;
1216 unsigned long next_zero;
1217
1218 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1219 max_t(u64, *offset, bitmap_info->offset));
1220 bits = bytes_to_bits(*bytes, ctl->unit);
1221
1222 for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1223 i < BITS_PER_BITMAP;
1224 i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1225 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1226 BITS_PER_BITMAP, i);
1227 if ((next_zero - i) >= bits) {
1228 found_bits = next_zero - i;
1229 break;
1230 }
1231 i = next_zero;
1232 }
1233
1234 if (found_bits) {
1235 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1236 *bytes = (u64)(found_bits) * ctl->unit;
1237 return 0;
1238 }
1239
1240 return -1;
1241 }
1242
1243 static struct btrfs_free_space *
1244 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
1245 {
1246 struct btrfs_free_space *entry;
1247 struct rb_node *node;
1248 int ret;
1249
1250 if (!ctl->free_space_offset.rb_node)
1251 return NULL;
1252
1253 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1254 if (!entry)
1255 return NULL;
1256
1257 for (node = &entry->offset_index; node; node = rb_next(node)) {
1258 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1259 if (entry->bytes < *bytes)
1260 continue;
1261
1262 if (entry->bitmap) {
1263 ret = search_bitmap(ctl, entry, offset, bytes);
1264 if (!ret)
1265 return entry;
1266 continue;
1267 }
1268
1269 *offset = entry->offset;
1270 *bytes = entry->bytes;
1271 return entry;
1272 }
1273
1274 return NULL;
1275 }
1276
1277 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1278 struct btrfs_free_space *info, u64 offset)
1279 {
1280 info->offset = offset_to_bitmap(ctl, offset);
1281 info->bytes = 0;
1282 link_free_space(ctl, info);
1283 ctl->total_bitmaps++;
1284
1285 ctl->op->recalc_thresholds(ctl);
1286 }
1287
1288 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1289 struct btrfs_free_space *bitmap_info)
1290 {
1291 unlink_free_space(ctl, bitmap_info);
1292 kfree(bitmap_info->bitmap);
1293 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1294 ctl->total_bitmaps--;
1295 ctl->op->recalc_thresholds(ctl);
1296 }
1297
1298 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1299 struct btrfs_free_space *bitmap_info,
1300 u64 *offset, u64 *bytes)
1301 {
1302 u64 end;
1303 u64 search_start, search_bytes;
1304 int ret;
1305
1306 again:
1307 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1308
1309 /*
1310 * XXX - this can go away after a few releases.
1311 *
1312 * since the only user of btrfs_remove_free_space is the tree logging
1313 * stuff, and the only way to test that is under crash conditions, we
1314 * want to have this debug stuff here just in case somethings not
1315 * working. Search the bitmap for the space we are trying to use to
1316 * make sure its actually there. If its not there then we need to stop
1317 * because something has gone wrong.
1318 */
1319 search_start = *offset;
1320 search_bytes = *bytes;
1321 search_bytes = min(search_bytes, end - search_start + 1);
1322 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1323 BUG_ON(ret < 0 || search_start != *offset);
1324
1325 if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1326 bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
1327 *bytes -= end - *offset + 1;
1328 *offset = end + 1;
1329 } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1330 bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
1331 *bytes = 0;
1332 }
1333
1334 if (*bytes) {
1335 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1336 if (!bitmap_info->bytes)
1337 free_bitmap(ctl, bitmap_info);
1338
1339 /*
1340 * no entry after this bitmap, but we still have bytes to
1341 * remove, so something has gone wrong.
1342 */
1343 if (!next)
1344 return -EINVAL;
1345
1346 bitmap_info = rb_entry(next, struct btrfs_free_space,
1347 offset_index);
1348
1349 /*
1350 * if the next entry isn't a bitmap we need to return to let the
1351 * extent stuff do its work.
1352 */
1353 if (!bitmap_info->bitmap)
1354 return -EAGAIN;
1355
1356 /*
1357 * Ok the next item is a bitmap, but it may not actually hold
1358 * the information for the rest of this free space stuff, so
1359 * look for it, and if we don't find it return so we can try
1360 * everything over again.
1361 */
1362 search_start = *offset;
1363 search_bytes = *bytes;
1364 ret = search_bitmap(ctl, bitmap_info, &search_start,
1365 &search_bytes);
1366 if (ret < 0 || search_start != *offset)
1367 return -EAGAIN;
1368
1369 goto again;
1370 } else if (!bitmap_info->bytes)
1371 free_bitmap(ctl, bitmap_info);
1372
1373 return 0;
1374 }
1375
1376 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1377 struct btrfs_free_space *info, u64 offset,
1378 u64 bytes)
1379 {
1380 u64 bytes_to_set = 0;
1381 u64 end;
1382
1383 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1384
1385 bytes_to_set = min(end - offset, bytes);
1386
1387 bitmap_set_bits(ctl, info, offset, bytes_to_set);
1388
1389 return bytes_to_set;
1390
1391 }
1392
1393 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1394 struct btrfs_free_space *info)
1395 {
1396 struct btrfs_block_group_cache *block_group = ctl->private;
1397
1398 /*
1399 * If we are below the extents threshold then we can add this as an
1400 * extent, and don't have to deal with the bitmap
1401 */
1402 if (ctl->free_extents < ctl->extents_thresh) {
1403 /*
1404 * If this block group has some small extents we don't want to
1405 * use up all of our free slots in the cache with them, we want
1406 * to reserve them to larger extents, however if we have plent
1407 * of cache left then go ahead an dadd them, no sense in adding
1408 * the overhead of a bitmap if we don't have to.
1409 */
1410 if (info->bytes <= block_group->sectorsize * 4) {
1411 if (ctl->free_extents * 2 <= ctl->extents_thresh)
1412 return false;
1413 } else {
1414 return false;
1415 }
1416 }
1417
1418 /*
1419 * some block groups are so tiny they can't be enveloped by a bitmap, so
1420 * don't even bother to create a bitmap for this
1421 */
1422 if (BITS_PER_BITMAP * block_group->sectorsize >
1423 block_group->key.offset)
1424 return false;
1425
1426 return true;
1427 }
1428
1429 static struct btrfs_free_space_op free_space_op = {
1430 .recalc_thresholds = recalculate_thresholds,
1431 .use_bitmap = use_bitmap,
1432 };
1433
1434 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1435 struct btrfs_free_space *info)
1436 {
1437 struct btrfs_free_space *bitmap_info;
1438 struct btrfs_block_group_cache *block_group = NULL;
1439 int added = 0;
1440 u64 bytes, offset, bytes_added;
1441 int ret;
1442
1443 bytes = info->bytes;
1444 offset = info->offset;
1445
1446 if (!ctl->op->use_bitmap(ctl, info))
1447 return 0;
1448
1449 if (ctl->op == &free_space_op)
1450 block_group = ctl->private;
1451 again:
1452 /*
1453 * Since we link bitmaps right into the cluster we need to see if we
1454 * have a cluster here, and if so and it has our bitmap we need to add
1455 * the free space to that bitmap.
1456 */
1457 if (block_group && !list_empty(&block_group->cluster_list)) {
1458 struct btrfs_free_cluster *cluster;
1459 struct rb_node *node;
1460 struct btrfs_free_space *entry;
1461
1462 cluster = list_entry(block_group->cluster_list.next,
1463 struct btrfs_free_cluster,
1464 block_group_list);
1465 spin_lock(&cluster->lock);
1466 node = rb_first(&cluster->root);
1467 if (!node) {
1468 spin_unlock(&cluster->lock);
1469 goto no_cluster_bitmap;
1470 }
1471
1472 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1473 if (!entry->bitmap) {
1474 spin_unlock(&cluster->lock);
1475 goto no_cluster_bitmap;
1476 }
1477
1478 if (entry->offset == offset_to_bitmap(ctl, offset)) {
1479 bytes_added = add_bytes_to_bitmap(ctl, entry,
1480 offset, bytes);
1481 bytes -= bytes_added;
1482 offset += bytes_added;
1483 }
1484 spin_unlock(&cluster->lock);
1485 if (!bytes) {
1486 ret = 1;
1487 goto out;
1488 }
1489 }
1490
1491 no_cluster_bitmap:
1492 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1493 1, 0);
1494 if (!bitmap_info) {
1495 BUG_ON(added);
1496 goto new_bitmap;
1497 }
1498
1499 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1500 bytes -= bytes_added;
1501 offset += bytes_added;
1502 added = 0;
1503
1504 if (!bytes) {
1505 ret = 1;
1506 goto out;
1507 } else
1508 goto again;
1509
1510 new_bitmap:
1511 if (info && info->bitmap) {
1512 add_new_bitmap(ctl, info, offset);
1513 added = 1;
1514 info = NULL;
1515 goto again;
1516 } else {
1517 spin_unlock(&ctl->tree_lock);
1518
1519 /* no pre-allocated info, allocate a new one */
1520 if (!info) {
1521 info = kmem_cache_zalloc(btrfs_free_space_cachep,
1522 GFP_NOFS);
1523 if (!info) {
1524 spin_lock(&ctl->tree_lock);
1525 ret = -ENOMEM;
1526 goto out;
1527 }
1528 }
1529
1530 /* allocate the bitmap */
1531 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1532 spin_lock(&ctl->tree_lock);
1533 if (!info->bitmap) {
1534 ret = -ENOMEM;
1535 goto out;
1536 }
1537 goto again;
1538 }
1539
1540 out:
1541 if (info) {
1542 if (info->bitmap)
1543 kfree(info->bitmap);
1544 kmem_cache_free(btrfs_free_space_cachep, info);
1545 }
1546
1547 return ret;
1548 }
1549
1550 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1551 struct btrfs_free_space *info, bool update_stat)
1552 {
1553 struct btrfs_free_space *left_info;
1554 struct btrfs_free_space *right_info;
1555 bool merged = false;
1556 u64 offset = info->offset;
1557 u64 bytes = info->bytes;
1558
1559 /*
1560 * first we want to see if there is free space adjacent to the range we
1561 * are adding, if there is remove that struct and add a new one to
1562 * cover the entire range
1563 */
1564 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1565 if (right_info && rb_prev(&right_info->offset_index))
1566 left_info = rb_entry(rb_prev(&right_info->offset_index),
1567 struct btrfs_free_space, offset_index);
1568 else
1569 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1570
1571 if (right_info && !right_info->bitmap) {
1572 if (update_stat)
1573 unlink_free_space(ctl, right_info);
1574 else
1575 __unlink_free_space(ctl, right_info);
1576 info->bytes += right_info->bytes;
1577 kmem_cache_free(btrfs_free_space_cachep, right_info);
1578 merged = true;
1579 }
1580
1581 if (left_info && !left_info->bitmap &&
1582 left_info->offset + left_info->bytes == offset) {
1583 if (update_stat)
1584 unlink_free_space(ctl, left_info);
1585 else
1586 __unlink_free_space(ctl, left_info);
1587 info->offset = left_info->offset;
1588 info->bytes += left_info->bytes;
1589 kmem_cache_free(btrfs_free_space_cachep, left_info);
1590 merged = true;
1591 }
1592
1593 return merged;
1594 }
1595
1596 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1597 u64 offset, u64 bytes)
1598 {
1599 struct btrfs_free_space *info;
1600 int ret = 0;
1601
1602 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1603 if (!info)
1604 return -ENOMEM;
1605
1606 info->offset = offset;
1607 info->bytes = bytes;
1608
1609 spin_lock(&ctl->tree_lock);
1610
1611 if (try_merge_free_space(ctl, info, true))
1612 goto link;
1613
1614 /*
1615 * There was no extent directly to the left or right of this new
1616 * extent then we know we're going to have to allocate a new extent, so
1617 * before we do that see if we need to drop this into a bitmap
1618 */
1619 ret = insert_into_bitmap(ctl, info);
1620 if (ret < 0) {
1621 goto out;
1622 } else if (ret) {
1623 ret = 0;
1624 goto out;
1625 }
1626 link:
1627 ret = link_free_space(ctl, info);
1628 if (ret)
1629 kmem_cache_free(btrfs_free_space_cachep, info);
1630 out:
1631 spin_unlock(&ctl->tree_lock);
1632
1633 if (ret) {
1634 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1635 BUG_ON(ret == -EEXIST);
1636 }
1637
1638 return ret;
1639 }
1640
1641 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1642 u64 offset, u64 bytes)
1643 {
1644 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1645 struct btrfs_free_space *info;
1646 struct btrfs_free_space *next_info = NULL;
1647 int ret = 0;
1648
1649 spin_lock(&ctl->tree_lock);
1650
1651 again:
1652 info = tree_search_offset(ctl, offset, 0, 0);
1653 if (!info) {
1654 /*
1655 * oops didn't find an extent that matched the space we wanted
1656 * to remove, look for a bitmap instead
1657 */
1658 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1659 1, 0);
1660 if (!info) {
1661 WARN_ON(1);
1662 goto out_lock;
1663 }
1664 }
1665
1666 if (info->bytes < bytes && rb_next(&info->offset_index)) {
1667 u64 end;
1668 next_info = rb_entry(rb_next(&info->offset_index),
1669 struct btrfs_free_space,
1670 offset_index);
1671
1672 if (next_info->bitmap)
1673 end = next_info->offset +
1674 BITS_PER_BITMAP * ctl->unit - 1;
1675 else
1676 end = next_info->offset + next_info->bytes;
1677
1678 if (next_info->bytes < bytes ||
1679 next_info->offset > offset || offset > end) {
1680 printk(KERN_CRIT "Found free space at %llu, size %llu,"
1681 " trying to use %llu\n",
1682 (unsigned long long)info->offset,
1683 (unsigned long long)info->bytes,
1684 (unsigned long long)bytes);
1685 WARN_ON(1);
1686 ret = -EINVAL;
1687 goto out_lock;
1688 }
1689
1690 info = next_info;
1691 }
1692
1693 if (info->bytes == bytes) {
1694 unlink_free_space(ctl, info);
1695 if (info->bitmap) {
1696 kfree(info->bitmap);
1697 ctl->total_bitmaps--;
1698 }
1699 kmem_cache_free(btrfs_free_space_cachep, info);
1700 goto out_lock;
1701 }
1702
1703 if (!info->bitmap && info->offset == offset) {
1704 unlink_free_space(ctl, info);
1705 info->offset += bytes;
1706 info->bytes -= bytes;
1707 link_free_space(ctl, info);
1708 goto out_lock;
1709 }
1710
1711 if (!info->bitmap && info->offset <= offset &&
1712 info->offset + info->bytes >= offset + bytes) {
1713 u64 old_start = info->offset;
1714 /*
1715 * we're freeing space in the middle of the info,
1716 * this can happen during tree log replay
1717 *
1718 * first unlink the old info and then
1719 * insert it again after the hole we're creating
1720 */
1721 unlink_free_space(ctl, info);
1722 if (offset + bytes < info->offset + info->bytes) {
1723 u64 old_end = info->offset + info->bytes;
1724
1725 info->offset = offset + bytes;
1726 info->bytes = old_end - info->offset;
1727 ret = link_free_space(ctl, info);
1728 WARN_ON(ret);
1729 if (ret)
1730 goto out_lock;
1731 } else {
1732 /* the hole we're creating ends at the end
1733 * of the info struct, just free the info
1734 */
1735 kmem_cache_free(btrfs_free_space_cachep, info);
1736 }
1737 spin_unlock(&ctl->tree_lock);
1738
1739 /* step two, insert a new info struct to cover
1740 * anything before the hole
1741 */
1742 ret = btrfs_add_free_space(block_group, old_start,
1743 offset - old_start);
1744 WARN_ON(ret);
1745 goto out;
1746 }
1747
1748 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1749 if (ret == -EAGAIN)
1750 goto again;
1751 BUG_ON(ret);
1752 out_lock:
1753 spin_unlock(&ctl->tree_lock);
1754 out:
1755 return ret;
1756 }
1757
1758 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1759 u64 bytes)
1760 {
1761 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1762 struct btrfs_free_space *info;
1763 struct rb_node *n;
1764 int count = 0;
1765
1766 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1767 info = rb_entry(n, struct btrfs_free_space, offset_index);
1768 if (info->bytes >= bytes)
1769 count++;
1770 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1771 (unsigned long long)info->offset,
1772 (unsigned long long)info->bytes,
1773 (info->bitmap) ? "yes" : "no");
1774 }
1775 printk(KERN_INFO "block group has cluster?: %s\n",
1776 list_empty(&block_group->cluster_list) ? "no" : "yes");
1777 printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1778 "\n", count);
1779 }
1780
1781 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
1782 {
1783 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1784
1785 spin_lock_init(&ctl->tree_lock);
1786 ctl->unit = block_group->sectorsize;
1787 ctl->start = block_group->key.objectid;
1788 ctl->private = block_group;
1789 ctl->op = &free_space_op;
1790
1791 /*
1792 * we only want to have 32k of ram per block group for keeping
1793 * track of free space, and if we pass 1/2 of that we want to
1794 * start converting things over to using bitmaps
1795 */
1796 ctl->extents_thresh = ((1024 * 32) / 2) /
1797 sizeof(struct btrfs_free_space);
1798 }
1799
1800 /*
1801 * for a given cluster, put all of its extents back into the free
1802 * space cache. If the block group passed doesn't match the block group
1803 * pointed to by the cluster, someone else raced in and freed the
1804 * cluster already. In that case, we just return without changing anything
1805 */
1806 static int
1807 __btrfs_return_cluster_to_free_space(
1808 struct btrfs_block_group_cache *block_group,
1809 struct btrfs_free_cluster *cluster)
1810 {
1811 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1812 struct btrfs_free_space *entry;
1813 struct rb_node *node;
1814
1815 spin_lock(&cluster->lock);
1816 if (cluster->block_group != block_group)
1817 goto out;
1818
1819 cluster->block_group = NULL;
1820 cluster->window_start = 0;
1821 list_del_init(&cluster->block_group_list);
1822
1823 node = rb_first(&cluster->root);
1824 while (node) {
1825 bool bitmap;
1826
1827 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1828 node = rb_next(&entry->offset_index);
1829 rb_erase(&entry->offset_index, &cluster->root);
1830
1831 bitmap = (entry->bitmap != NULL);
1832 if (!bitmap)
1833 try_merge_free_space(ctl, entry, false);
1834 tree_insert_offset(&ctl->free_space_offset,
1835 entry->offset, &entry->offset_index, bitmap);
1836 }
1837 cluster->root = RB_ROOT;
1838
1839 out:
1840 spin_unlock(&cluster->lock);
1841 btrfs_put_block_group(block_group);
1842 return 0;
1843 }
1844
1845 void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
1846 {
1847 struct btrfs_free_space *info;
1848 struct rb_node *node;
1849
1850 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
1851 info = rb_entry(node, struct btrfs_free_space, offset_index);
1852 if (!info->bitmap) {
1853 unlink_free_space(ctl, info);
1854 kmem_cache_free(btrfs_free_space_cachep, info);
1855 } else {
1856 free_bitmap(ctl, info);
1857 }
1858 if (need_resched()) {
1859 spin_unlock(&ctl->tree_lock);
1860 cond_resched();
1861 spin_lock(&ctl->tree_lock);
1862 }
1863 }
1864 }
1865
1866 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
1867 {
1868 spin_lock(&ctl->tree_lock);
1869 __btrfs_remove_free_space_cache_locked(ctl);
1870 spin_unlock(&ctl->tree_lock);
1871 }
1872
1873 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1874 {
1875 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1876 struct btrfs_free_cluster *cluster;
1877 struct list_head *head;
1878
1879 spin_lock(&ctl->tree_lock);
1880 while ((head = block_group->cluster_list.next) !=
1881 &block_group->cluster_list) {
1882 cluster = list_entry(head, struct btrfs_free_cluster,
1883 block_group_list);
1884
1885 WARN_ON(cluster->block_group != block_group);
1886 __btrfs_return_cluster_to_free_space(block_group, cluster);
1887 if (need_resched()) {
1888 spin_unlock(&ctl->tree_lock);
1889 cond_resched();
1890 spin_lock(&ctl->tree_lock);
1891 }
1892 }
1893 __btrfs_remove_free_space_cache_locked(ctl);
1894 spin_unlock(&ctl->tree_lock);
1895
1896 }
1897
1898 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1899 u64 offset, u64 bytes, u64 empty_size)
1900 {
1901 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1902 struct btrfs_free_space *entry = NULL;
1903 u64 bytes_search = bytes + empty_size;
1904 u64 ret = 0;
1905
1906 spin_lock(&ctl->tree_lock);
1907 entry = find_free_space(ctl, &offset, &bytes_search);
1908 if (!entry)
1909 goto out;
1910
1911 ret = offset;
1912 if (entry->bitmap) {
1913 bitmap_clear_bits(ctl, entry, offset, bytes);
1914 if (!entry->bytes)
1915 free_bitmap(ctl, entry);
1916 } else {
1917 unlink_free_space(ctl, entry);
1918 entry->offset += bytes;
1919 entry->bytes -= bytes;
1920 if (!entry->bytes)
1921 kmem_cache_free(btrfs_free_space_cachep, entry);
1922 else
1923 link_free_space(ctl, entry);
1924 }
1925
1926 out:
1927 spin_unlock(&ctl->tree_lock);
1928
1929 return ret;
1930 }
1931
1932 /*
1933 * given a cluster, put all of its extents back into the free space
1934 * cache. If a block group is passed, this function will only free
1935 * a cluster that belongs to the passed block group.
1936 *
1937 * Otherwise, it'll get a reference on the block group pointed to by the
1938 * cluster and remove the cluster from it.
1939 */
1940 int btrfs_return_cluster_to_free_space(
1941 struct btrfs_block_group_cache *block_group,
1942 struct btrfs_free_cluster *cluster)
1943 {
1944 struct btrfs_free_space_ctl *ctl;
1945 int ret;
1946
1947 /* first, get a safe pointer to the block group */
1948 spin_lock(&cluster->lock);
1949 if (!block_group) {
1950 block_group = cluster->block_group;
1951 if (!block_group) {
1952 spin_unlock(&cluster->lock);
1953 return 0;
1954 }
1955 } else if (cluster->block_group != block_group) {
1956 /* someone else has already freed it don't redo their work */
1957 spin_unlock(&cluster->lock);
1958 return 0;
1959 }
1960 atomic_inc(&block_group->count);
1961 spin_unlock(&cluster->lock);
1962
1963 ctl = block_group->free_space_ctl;
1964
1965 /* now return any extents the cluster had on it */
1966 spin_lock(&ctl->tree_lock);
1967 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1968 spin_unlock(&ctl->tree_lock);
1969
1970 /* finally drop our ref */
1971 btrfs_put_block_group(block_group);
1972 return ret;
1973 }
1974
1975 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1976 struct btrfs_free_cluster *cluster,
1977 struct btrfs_free_space *entry,
1978 u64 bytes, u64 min_start)
1979 {
1980 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1981 int err;
1982 u64 search_start = cluster->window_start;
1983 u64 search_bytes = bytes;
1984 u64 ret = 0;
1985
1986 search_start = min_start;
1987 search_bytes = bytes;
1988
1989 err = search_bitmap(ctl, entry, &search_start, &search_bytes);
1990 if (err)
1991 return 0;
1992
1993 ret = search_start;
1994 __bitmap_clear_bits(ctl, entry, ret, bytes);
1995
1996 return ret;
1997 }
1998
1999 /*
2000 * given a cluster, try to allocate 'bytes' from it, returns 0
2001 * if it couldn't find anything suitably large, or a logical disk offset
2002 * if things worked out
2003 */
2004 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2005 struct btrfs_free_cluster *cluster, u64 bytes,
2006 u64 min_start)
2007 {
2008 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2009 struct btrfs_free_space *entry = NULL;
2010 struct rb_node *node;
2011 u64 ret = 0;
2012
2013 spin_lock(&cluster->lock);
2014 if (bytes > cluster->max_size)
2015 goto out;
2016
2017 if (cluster->block_group != block_group)
2018 goto out;
2019
2020 node = rb_first(&cluster->root);
2021 if (!node)
2022 goto out;
2023
2024 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2025 while(1) {
2026 if (entry->bytes < bytes ||
2027 (!entry->bitmap && entry->offset < min_start)) {
2028 node = rb_next(&entry->offset_index);
2029 if (!node)
2030 break;
2031 entry = rb_entry(node, struct btrfs_free_space,
2032 offset_index);
2033 continue;
2034 }
2035
2036 if (entry->bitmap) {
2037 ret = btrfs_alloc_from_bitmap(block_group,
2038 cluster, entry, bytes,
2039 min_start);
2040 if (ret == 0) {
2041 node = rb_next(&entry->offset_index);
2042 if (!node)
2043 break;
2044 entry = rb_entry(node, struct btrfs_free_space,
2045 offset_index);
2046 continue;
2047 }
2048 } else {
2049 ret = entry->offset;
2050
2051 entry->offset += bytes;
2052 entry->bytes -= bytes;
2053 }
2054
2055 if (entry->bytes == 0)
2056 rb_erase(&entry->offset_index, &cluster->root);
2057 break;
2058 }
2059 out:
2060 spin_unlock(&cluster->lock);
2061
2062 if (!ret)
2063 return 0;
2064
2065 spin_lock(&ctl->tree_lock);
2066
2067 ctl->free_space -= bytes;
2068 if (entry->bytes == 0) {
2069 ctl->free_extents--;
2070 if (entry->bitmap) {
2071 kfree(entry->bitmap);
2072 ctl->total_bitmaps--;
2073 ctl->op->recalc_thresholds(ctl);
2074 }
2075 kmem_cache_free(btrfs_free_space_cachep, entry);
2076 }
2077
2078 spin_unlock(&ctl->tree_lock);
2079
2080 return ret;
2081 }
2082
2083 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2084 struct btrfs_free_space *entry,
2085 struct btrfs_free_cluster *cluster,
2086 u64 offset, u64 bytes, u64 min_bytes)
2087 {
2088 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2089 unsigned long next_zero;
2090 unsigned long i;
2091 unsigned long search_bits;
2092 unsigned long total_bits;
2093 unsigned long found_bits;
2094 unsigned long start = 0;
2095 unsigned long total_found = 0;
2096 int ret;
2097 bool found = false;
2098
2099 i = offset_to_bit(entry->offset, block_group->sectorsize,
2100 max_t(u64, offset, entry->offset));
2101 search_bits = bytes_to_bits(bytes, block_group->sectorsize);
2102 total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
2103
2104 again:
2105 found_bits = 0;
2106 for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
2107 i < BITS_PER_BITMAP;
2108 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
2109 next_zero = find_next_zero_bit(entry->bitmap,
2110 BITS_PER_BITMAP, i);
2111 if (next_zero - i >= search_bits) {
2112 found_bits = next_zero - i;
2113 break;
2114 }
2115 i = next_zero;
2116 }
2117
2118 if (!found_bits)
2119 return -ENOSPC;
2120
2121 if (!found) {
2122 start = i;
2123 found = true;
2124 }
2125
2126 total_found += found_bits;
2127
2128 if (cluster->max_size < found_bits * block_group->sectorsize)
2129 cluster->max_size = found_bits * block_group->sectorsize;
2130
2131 if (total_found < total_bits) {
2132 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2133 if (i - start > total_bits * 2) {
2134 total_found = 0;
2135 cluster->max_size = 0;
2136 found = false;
2137 }
2138 goto again;
2139 }
2140
2141 cluster->window_start = start * block_group->sectorsize +
2142 entry->offset;
2143 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2144 ret = tree_insert_offset(&cluster->root, entry->offset,
2145 &entry->offset_index, 1);
2146 BUG_ON(ret);
2147
2148 return 0;
2149 }
2150
2151 /*
2152 * This searches the block group for just extents to fill the cluster with.
2153 */
2154 static noinline int
2155 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2156 struct btrfs_free_cluster *cluster,
2157 struct list_head *bitmaps, u64 offset, u64 bytes,
2158 u64 min_bytes)
2159 {
2160 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2161 struct btrfs_free_space *first = NULL;
2162 struct btrfs_free_space *entry = NULL;
2163 struct btrfs_free_space *prev = NULL;
2164 struct btrfs_free_space *last;
2165 struct rb_node *node;
2166 u64 window_start;
2167 u64 window_free;
2168 u64 max_extent;
2169 u64 max_gap = 128 * 1024;
2170
2171 entry = tree_search_offset(ctl, offset, 0, 1);
2172 if (!entry)
2173 return -ENOSPC;
2174
2175 /*
2176 * We don't want bitmaps, so just move along until we find a normal
2177 * extent entry.
2178 */
2179 while (entry->bitmap) {
2180 if (list_empty(&entry->list))
2181 list_add_tail(&entry->list, bitmaps);
2182 node = rb_next(&entry->offset_index);
2183 if (!node)
2184 return -ENOSPC;
2185 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2186 }
2187
2188 window_start = entry->offset;
2189 window_free = entry->bytes;
2190 max_extent = entry->bytes;
2191 first = entry;
2192 last = entry;
2193 prev = entry;
2194
2195 while (window_free <= min_bytes) {
2196 node = rb_next(&entry->offset_index);
2197 if (!node)
2198 return -ENOSPC;
2199 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2200
2201 if (entry->bitmap) {
2202 if (list_empty(&entry->list))
2203 list_add_tail(&entry->list, bitmaps);
2204 continue;
2205 }
2206
2207 /*
2208 * we haven't filled the empty size and the window is
2209 * very large. reset and try again
2210 */
2211 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2212 entry->offset - window_start > (min_bytes * 2)) {
2213 first = entry;
2214 window_start = entry->offset;
2215 window_free = entry->bytes;
2216 last = entry;
2217 max_extent = entry->bytes;
2218 } else {
2219 last = entry;
2220 window_free += entry->bytes;
2221 if (entry->bytes > max_extent)
2222 max_extent = entry->bytes;
2223 }
2224 prev = entry;
2225 }
2226
2227 cluster->window_start = first->offset;
2228
2229 node = &first->offset_index;
2230
2231 /*
2232 * now we've found our entries, pull them out of the free space
2233 * cache and put them into the cluster rbtree
2234 */
2235 do {
2236 int ret;
2237
2238 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2239 node = rb_next(&entry->offset_index);
2240 if (entry->bitmap)
2241 continue;
2242
2243 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2244 ret = tree_insert_offset(&cluster->root, entry->offset,
2245 &entry->offset_index, 0);
2246 BUG_ON(ret);
2247 } while (node && entry != last);
2248
2249 cluster->max_size = max_extent;
2250
2251 return 0;
2252 }
2253
2254 /*
2255 * This specifically looks for bitmaps that may work in the cluster, we assume
2256 * that we have already failed to find extents that will work.
2257 */
2258 static noinline int
2259 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2260 struct btrfs_free_cluster *cluster,
2261 struct list_head *bitmaps, u64 offset, u64 bytes,
2262 u64 min_bytes)
2263 {
2264 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2265 struct btrfs_free_space *entry;
2266 struct rb_node *node;
2267 int ret = -ENOSPC;
2268
2269 if (ctl->total_bitmaps == 0)
2270 return -ENOSPC;
2271
2272 /*
2273 * First check our cached list of bitmaps and see if there is an entry
2274 * here that will work.
2275 */
2276 list_for_each_entry(entry, bitmaps, list) {
2277 if (entry->bytes < min_bytes)
2278 continue;
2279 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2280 bytes, min_bytes);
2281 if (!ret)
2282 return 0;
2283 }
2284
2285 /*
2286 * If we do have entries on our list and we are here then we didn't find
2287 * anything, so go ahead and get the next entry after the last entry in
2288 * this list and start the search from there.
2289 */
2290 if (!list_empty(bitmaps)) {
2291 entry = list_entry(bitmaps->prev, struct btrfs_free_space,
2292 list);
2293 node = rb_next(&entry->offset_index);
2294 if (!node)
2295 return -ENOSPC;
2296 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2297 goto search;
2298 }
2299
2300 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, offset), 0, 1);
2301 if (!entry)
2302 return -ENOSPC;
2303
2304 search:
2305 node = &entry->offset_index;
2306 do {
2307 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2308 node = rb_next(&entry->offset_index);
2309 if (!entry->bitmap)
2310 continue;
2311 if (entry->bytes < min_bytes)
2312 continue;
2313 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2314 bytes, min_bytes);
2315 } while (ret && node);
2316
2317 return ret;
2318 }
2319
2320 /*
2321 * here we try to find a cluster of blocks in a block group. The goal
2322 * is to find at least bytes free and up to empty_size + bytes free.
2323 * We might not find them all in one contiguous area.
2324 *
2325 * returns zero and sets up cluster if things worked out, otherwise
2326 * it returns -enospc
2327 */
2328 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2329 struct btrfs_root *root,
2330 struct btrfs_block_group_cache *block_group,
2331 struct btrfs_free_cluster *cluster,
2332 u64 offset, u64 bytes, u64 empty_size)
2333 {
2334 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2335 struct list_head bitmaps;
2336 struct btrfs_free_space *entry, *tmp;
2337 u64 min_bytes;
2338 int ret;
2339
2340 /* for metadata, allow allocates with more holes */
2341 if (btrfs_test_opt(root, SSD_SPREAD)) {
2342 min_bytes = bytes + empty_size;
2343 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2344 /*
2345 * we want to do larger allocations when we are
2346 * flushing out the delayed refs, it helps prevent
2347 * making more work as we go along.
2348 */
2349 if (trans->transaction->delayed_refs.flushing)
2350 min_bytes = max(bytes, (bytes + empty_size) >> 1);
2351 else
2352 min_bytes = max(bytes, (bytes + empty_size) >> 4);
2353 } else
2354 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2355
2356 spin_lock(&ctl->tree_lock);
2357
2358 /*
2359 * If we know we don't have enough space to make a cluster don't even
2360 * bother doing all the work to try and find one.
2361 */
2362 if (ctl->free_space < min_bytes) {
2363 spin_unlock(&ctl->tree_lock);
2364 return -ENOSPC;
2365 }
2366
2367 spin_lock(&cluster->lock);
2368
2369 /* someone already found a cluster, hooray */
2370 if (cluster->block_group) {
2371 ret = 0;
2372 goto out;
2373 }
2374
2375 INIT_LIST_HEAD(&bitmaps);
2376 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2377 bytes, min_bytes);
2378 if (ret)
2379 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2380 offset, bytes, min_bytes);
2381
2382 /* Clear our temporary list */
2383 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2384 list_del_init(&entry->list);
2385
2386 if (!ret) {
2387 atomic_inc(&block_group->count);
2388 list_add_tail(&cluster->block_group_list,
2389 &block_group->cluster_list);
2390 cluster->block_group = block_group;
2391 }
2392 out:
2393 spin_unlock(&cluster->lock);
2394 spin_unlock(&ctl->tree_lock);
2395
2396 return ret;
2397 }
2398
2399 /*
2400 * simple code to zero out a cluster
2401 */
2402 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2403 {
2404 spin_lock_init(&cluster->lock);
2405 spin_lock_init(&cluster->refill_lock);
2406 cluster->root = RB_ROOT;
2407 cluster->max_size = 0;
2408 INIT_LIST_HEAD(&cluster->block_group_list);
2409 cluster->block_group = NULL;
2410 }
2411
2412 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2413 u64 *trimmed, u64 start, u64 end, u64 minlen)
2414 {
2415 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2416 struct btrfs_free_space *entry = NULL;
2417 struct btrfs_fs_info *fs_info = block_group->fs_info;
2418 u64 bytes = 0;
2419 u64 actually_trimmed;
2420 int ret = 0;
2421
2422 *trimmed = 0;
2423
2424 while (start < end) {
2425 spin_lock(&ctl->tree_lock);
2426
2427 if (ctl->free_space < minlen) {
2428 spin_unlock(&ctl->tree_lock);
2429 break;
2430 }
2431
2432 entry = tree_search_offset(ctl, start, 0, 1);
2433 if (!entry)
2434 entry = tree_search_offset(ctl,
2435 offset_to_bitmap(ctl, start),
2436 1, 1);
2437
2438 if (!entry || entry->offset >= end) {
2439 spin_unlock(&ctl->tree_lock);
2440 break;
2441 }
2442
2443 if (entry->bitmap) {
2444 ret = search_bitmap(ctl, entry, &start, &bytes);
2445 if (!ret) {
2446 if (start >= end) {
2447 spin_unlock(&ctl->tree_lock);
2448 break;
2449 }
2450 bytes = min(bytes, end - start);
2451 bitmap_clear_bits(ctl, entry, start, bytes);
2452 if (entry->bytes == 0)
2453 free_bitmap(ctl, entry);
2454 } else {
2455 start = entry->offset + BITS_PER_BITMAP *
2456 block_group->sectorsize;
2457 spin_unlock(&ctl->tree_lock);
2458 ret = 0;
2459 continue;
2460 }
2461 } else {
2462 start = entry->offset;
2463 bytes = min(entry->bytes, end - start);
2464 unlink_free_space(ctl, entry);
2465 kmem_cache_free(btrfs_free_space_cachep, entry);
2466 }
2467
2468 spin_unlock(&ctl->tree_lock);
2469
2470 if (bytes >= minlen) {
2471 int update_ret;
2472 update_ret = btrfs_update_reserved_bytes(block_group,
2473 bytes, 1, 1);
2474
2475 ret = btrfs_error_discard_extent(fs_info->extent_root,
2476 start,
2477 bytes,
2478 &actually_trimmed);
2479
2480 btrfs_add_free_space(block_group, start, bytes);
2481 if (!update_ret)
2482 btrfs_update_reserved_bytes(block_group,
2483 bytes, 0, 1);
2484
2485 if (ret)
2486 break;
2487 *trimmed += actually_trimmed;
2488 }
2489 start += bytes;
2490 bytes = 0;
2491
2492 if (fatal_signal_pending(current)) {
2493 ret = -ERESTARTSYS;
2494 break;
2495 }
2496
2497 cond_resched();
2498 }
2499
2500 return ret;
2501 }
2502
2503 /*
2504 * Find the left-most item in the cache tree, and then return the
2505 * smallest inode number in the item.
2506 *
2507 * Note: the returned inode number may not be the smallest one in
2508 * the tree, if the left-most item is a bitmap.
2509 */
2510 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2511 {
2512 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2513 struct btrfs_free_space *entry = NULL;
2514 u64 ino = 0;
2515
2516 spin_lock(&ctl->tree_lock);
2517
2518 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2519 goto out;
2520
2521 entry = rb_entry(rb_first(&ctl->free_space_offset),
2522 struct btrfs_free_space, offset_index);
2523
2524 if (!entry->bitmap) {
2525 ino = entry->offset;
2526
2527 unlink_free_space(ctl, entry);
2528 entry->offset++;
2529 entry->bytes--;
2530 if (!entry->bytes)
2531 kmem_cache_free(btrfs_free_space_cachep, entry);
2532 else
2533 link_free_space(ctl, entry);
2534 } else {
2535 u64 offset = 0;
2536 u64 count = 1;
2537 int ret;
2538
2539 ret = search_bitmap(ctl, entry, &offset, &count);
2540 BUG_ON(ret);
2541
2542 ino = offset;
2543 bitmap_clear_bits(ctl, entry, offset, 1);
2544 if (entry->bytes == 0)
2545 free_bitmap(ctl, entry);
2546 }
2547 out:
2548 spin_unlock(&ctl->tree_lock);
2549
2550 return ino;
2551 }
2552
2553 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2554 struct btrfs_path *path)
2555 {
2556 struct inode *inode = NULL;
2557
2558 spin_lock(&root->cache_lock);
2559 if (root->cache_inode)
2560 inode = igrab(root->cache_inode);
2561 spin_unlock(&root->cache_lock);
2562 if (inode)
2563 return inode;
2564
2565 inode = __lookup_free_space_inode(root, path, 0);
2566 if (IS_ERR(inode))
2567 return inode;
2568
2569 spin_lock(&root->cache_lock);
2570 if (!btrfs_fs_closing(root->fs_info))
2571 root->cache_inode = igrab(inode);
2572 spin_unlock(&root->cache_lock);
2573
2574 return inode;
2575 }
2576
2577 int create_free_ino_inode(struct btrfs_root *root,
2578 struct btrfs_trans_handle *trans,
2579 struct btrfs_path *path)
2580 {
2581 return __create_free_space_inode(root, trans, path,
2582 BTRFS_FREE_INO_OBJECTID, 0);
2583 }
2584
2585 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2586 {
2587 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2588 struct btrfs_path *path;
2589 struct inode *inode;
2590 int ret = 0;
2591 u64 root_gen = btrfs_root_generation(&root->root_item);
2592
2593 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2594 return 0;
2595
2596 /*
2597 * If we're unmounting then just return, since this does a search on the
2598 * normal root and not the commit root and we could deadlock.
2599 */
2600 if (btrfs_fs_closing(fs_info))
2601 return 0;
2602
2603 path = btrfs_alloc_path();
2604 if (!path)
2605 return 0;
2606
2607 inode = lookup_free_ino_inode(root, path);
2608 if (IS_ERR(inode))
2609 goto out;
2610
2611 if (root_gen != BTRFS_I(inode)->generation)
2612 goto out_put;
2613
2614 ret = __load_free_space_cache(root, inode, ctl, path, 0);
2615
2616 if (ret < 0)
2617 printk(KERN_ERR "btrfs: failed to load free ino cache for "
2618 "root %llu\n", root->root_key.objectid);
2619 out_put:
2620 iput(inode);
2621 out:
2622 btrfs_free_path(path);
2623 return ret;
2624 }
2625
2626 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2627 struct btrfs_trans_handle *trans,
2628 struct btrfs_path *path)
2629 {
2630 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2631 struct inode *inode;
2632 int ret;
2633
2634 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2635 return 0;
2636
2637 inode = lookup_free_ino_inode(root, path);
2638 if (IS_ERR(inode))
2639 return 0;
2640
2641 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2642 if (ret < 0)
2643 printk(KERN_ERR "btrfs: failed to write free ino cache "
2644 "for root %llu\n", root->root_key.objectid);
2645
2646 iput(inode);
2647 return ret;
2648 }
linux-next