Btrfs: load free space cache if it exists
[linux-2.6.git] / fs / btrfs / free-space-cache.c
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
28 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
29 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
30
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);
35
36 struct inode *lookup_free_space_inode(struct btrfs_root *root,
37                                       struct btrfs_block_group_cache
38                                       *block_group, struct btrfs_path *path)
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         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;
54
55         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
56         key.offset = block_group->key.objectid;
57         key.type = 0;
58
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);
65         }
66
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);
73
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);
82         }
83
84         spin_lock(&block_group->lock);
85         if (!root->fs_info->closing) {
86                 block_group->inode = igrab(inode);
87                 block_group->iref = 1;
88         }
89         spin_unlock(&block_group->lock);
90
91         return inode;
92 }
93
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)
98 {
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;
106
107         ret = btrfs_find_free_objectid(trans, root, 0, &objectid);
108         if (ret < 0)
109                 return ret;
110
111         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
112         if (ret)
113                 return ret;
114
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);
135
136         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
137         key.offset = block_group->key.objectid;
138         key.type = 0;
139
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;
145         }
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);
153
154         return 0;
155 }
156
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)
161 {
162         loff_t oldsize;
163         int ret = 0;
164
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;
171
172         oldsize = i_size_read(inode);
173         btrfs_i_size_write(inode, 0);
174         truncate_pagecache(inode, oldsize, 0);
175
176         /*
177          * We don't need an orphan item because truncating the free space cache
178          * will never be split across transactions.
179          */
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;
185         }
186
187         return btrfs_update_inode(trans, root, inode);
188 }
189
190 static int readahead_cache(struct inode *inode)
191 {
192         struct file_ra_state *ra;
193         unsigned long last_index;
194
195         ra = kzalloc(sizeof(*ra), GFP_NOFS);
196         if (!ra)
197                 return -ENOMEM;
198
199         file_ra_state_init(ra, inode->i_mapping);
200         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
201
202         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
203
204         kfree(ra);
205
206         return 0;
207 }
208
209 int load_free_space_cache(struct btrfs_fs_info *fs_info,
210                           struct btrfs_block_group_cache *block_group)
211 {
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;
230
231         /*
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.
234          */
235         smp_mb();
236         if (fs_info->closing)
237                 return 0;
238
239         /*
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.
242          */
243         spin_lock(&block_group->lock);
244         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
245                 printk(KERN_ERR "not reading block group %llu, dcs is %d\n", block_group->key.objectid,
246                        block_group->disk_cache_state);
247                 spin_unlock(&block_group->lock);
248                 return 0;
249         }
250         spin_unlock(&block_group->lock);
251
252         INIT_LIST_HEAD(&bitmaps);
253
254         path = btrfs_alloc_path();
255         if (!path)
256                 return 0;
257
258         inode = lookup_free_space_inode(root, block_group, path);
259         if (IS_ERR(inode)) {
260                 btrfs_free_path(path);
261                 return 0;
262         }
263
264         /* Nothing in the space cache, goodbye */
265         if (!i_size_read(inode)) {
266                 btrfs_free_path(path);
267                 goto out;
268         }
269
270         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
271         key.offset = block_group->key.objectid;
272         key.type = 0;
273
274         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
275         if (ret) {
276                 btrfs_free_path(path);
277                 goto out;
278         }
279
280         leaf = path->nodes[0];
281         header = btrfs_item_ptr(leaf, path->slots[0],
282                                 struct btrfs_free_space_header);
283         num_entries = btrfs_free_space_entries(leaf, header);
284         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
285         generation = btrfs_free_space_generation(leaf, header);
286         btrfs_free_path(path);
287
288         if (BTRFS_I(inode)->generation != generation) {
289                 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
290                        " not match free space cache generation (%llu) for "
291                        "block group %llu\n",
292                        (unsigned long long)BTRFS_I(inode)->generation,
293                        (unsigned long long)generation,
294                        (unsigned long long)block_group->key.objectid);
295                 goto out;
296         }
297
298         if (!num_entries)
299                 goto out;
300
301         /* Setup everything for doing checksumming */
302         num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
303         checksums = crc = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
304         if (!checksums)
305                 goto out;
306         first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
307         disk_crcs = kzalloc(first_page_offset, GFP_NOFS);
308         if (!disk_crcs)
309                 goto out;
310
311         ret = readahead_cache(inode);
312         if (ret) {
313                 ret = 0;
314                 goto out;
315         }
316
317         while (1) {
318                 struct btrfs_free_space_entry *entry;
319                 struct btrfs_free_space *e;
320                 void *addr;
321                 unsigned long offset = 0;
322                 unsigned long start_offset = 0;
323                 int need_loop = 0;
324
325                 if (!num_entries && !num_bitmaps)
326                         break;
327
328                 if (index == 0) {
329                         start_offset = first_page_offset;
330                         offset = start_offset;
331                 }
332
333                 page = grab_cache_page(inode->i_mapping, index);
334                 if (!page) {
335                         ret = 0;
336                         goto free_cache;
337                 }
338
339                 if (!PageUptodate(page)) {
340                         btrfs_readpage(NULL, page);
341                         lock_page(page);
342                         if (!PageUptodate(page)) {
343                                 unlock_page(page);
344                                 page_cache_release(page);
345                                 printk(KERN_ERR "btrfs: error reading free "
346                                        "space cache: %llu\n",
347                                        (unsigned long long)
348                                        block_group->key.objectid);
349                                 goto free_cache;
350                         }
351                 }
352                 addr = kmap(page);
353
354                 if (index == 0) {
355                         u64 *gen;
356
357                         memcpy(disk_crcs, addr, first_page_offset);
358                         gen = addr + (sizeof(u32) * num_checksums);
359                         if (*gen != BTRFS_I(inode)->generation) {
360                                 printk(KERN_ERR "btrfs: space cache generation"
361                                        " (%llu) does not match inode (%llu) "
362                                        "for block group %llu\n",
363                                        (unsigned long long)*gen,
364                                        (unsigned long long)
365                                        BTRFS_I(inode)->generation,
366                                        (unsigned long long)
367                                        block_group->key.objectid);
368                                 kunmap(page);
369                                 unlock_page(page);
370                                 page_cache_release(page);
371                                 goto free_cache;
372                         }
373                         crc = (u32 *)disk_crcs;
374                 }
375                 entry = addr + start_offset;
376
377                 /* First lets check our crc before we do anything fun */
378                 cur_crc = ~(u32)0;
379                 cur_crc = btrfs_csum_data(root, addr + start_offset, cur_crc,
380                                           PAGE_CACHE_SIZE - start_offset);
381                 btrfs_csum_final(cur_crc, (char *)&cur_crc);
382                 if (cur_crc != *crc) {
383                         printk(KERN_ERR "btrfs: crc mismatch for page %lu in "
384                                "block group %llu\n", index,
385                                (unsigned long long)block_group->key.objectid);
386                         kunmap(page);
387                         unlock_page(page);
388                         page_cache_release(page);
389                         goto free_cache;
390                 }
391                 crc++;
392
393                 while (1) {
394                         if (!num_entries)
395                                 break;
396
397                         need_loop = 1;
398                         e = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
399                         if (!e) {
400                                 kunmap(page);
401                                 unlock_page(page);
402                                 page_cache_release(page);
403                                 goto free_cache;
404                         }
405
406                         e->offset = le64_to_cpu(entry->offset);
407                         e->bytes = le64_to_cpu(entry->bytes);
408                         if (!e->bytes) {
409                                 kunmap(page);
410                                 kfree(e);
411                                 unlock_page(page);
412                                 page_cache_release(page);
413                                 goto free_cache;
414                         }
415
416                         if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
417                                 spin_lock(&block_group->tree_lock);
418                                 ret = link_free_space(block_group, e);
419                                 spin_unlock(&block_group->tree_lock);
420                                 BUG_ON(ret);
421                         } else {
422                                 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
423                                 if (!e->bitmap) {
424                                         kunmap(page);
425                                         kfree(e);
426                                         unlock_page(page);
427                                         page_cache_release(page);
428                                         goto free_cache;
429                                 }
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);
436                         }
437
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++;
444                 }
445
446                 /*
447                  * We read an entry out of this page, we need to move on to the
448                  * next page.
449                  */
450                 if (need_loop) {
451                         kunmap(page);
452                         goto next;
453                 }
454
455                 /*
456                  * We add the bitmaps at the end of the entries in order that
457                  * the bitmap entries are added to the cache.
458                  */
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++;
468         }
469
470         ret = 1;
471 out:
472         kfree(checksums);
473         kfree(disk_crcs);
474         iput(inode);
475         return ret;
476
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;
484 }
485
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)
490 {
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;
508
509         root = root->fs_info->tree_root;
510
511         INIT_LIST_HEAD(&bitmap_list);
512
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;
517         }
518         spin_unlock(&block_group->lock);
519
520         inode = lookup_free_space_inode(root, block_group, path);
521         if (IS_ERR(inode))
522                 return 0;
523
524         if (!i_size_read(inode)) {
525                 iput(inode);
526                 return 0;
527         }
528
529         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
530         filemap_write_and_wait(inode->i_mapping);
531         btrfs_wait_ordered_range(inode, inode->i_size &
532                                  ~(root->sectorsize - 1), (u64)-1);
533
534         /* We need a checksum per page. */
535         num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
536         crc = checksums  = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
537         if (!crc) {
538                 iput(inode);
539                 return 0;
540         }
541
542         /* Since the first page has all of our checksums and our generation we
543          * need to calculate the offset into the page that we can start writing
544          * our entries.
545          */
546         first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
547
548         node = rb_first(&block_group->free_space_offset);
549         if (!node)
550                 goto out_free;
551
552         /*
553          * Lock all pages first so we can lock the extent safely.
554          *
555          * NOTE: Because we hold the ref the entire time we're going to write to
556          * the page find_get_page should never fail, so we don't do a check
557          * after find_get_page at this point.  Just putting this here so people
558          * know and don't freak out.
559          */
560         while (index <= last_index) {
561                 page = grab_cache_page(inode->i_mapping, index);
562                 if (!page) {
563                         pgoff_t i = 0;
564
565                         while (i < index) {
566                                 page = find_get_page(inode->i_mapping, i);
567                                 unlock_page(page);
568                                 page_cache_release(page);
569                                 page_cache_release(page);
570                                 i++;
571                         }
572                         goto out_free;
573                 }
574                 index++;
575         }
576
577         index = 0;
578         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
579                          0, &cached_state, GFP_NOFS);
580
581         /* Write out the extent entries */
582         do {
583                 struct btrfs_free_space_entry *entry;
584                 void *addr;
585                 unsigned long offset = 0;
586                 unsigned long start_offset = 0;
587
588                 if (index == 0) {
589                         start_offset = first_page_offset;
590                         offset = start_offset;
591                 }
592
593                 page = find_get_page(inode->i_mapping, index);
594
595                 addr = kmap(page);
596                 entry = addr + start_offset;
597
598                 memset(addr, 0, PAGE_CACHE_SIZE);
599                 while (1) {
600                         struct btrfs_free_space *e;
601
602                         e = rb_entry(node, struct btrfs_free_space, offset_index);
603                         entries++;
604
605                         entry->offset = cpu_to_le64(e->offset);
606                         entry->bytes = cpu_to_le64(e->bytes);
607                         if (e->bitmap) {
608                                 entry->type = BTRFS_FREE_SPACE_BITMAP;
609                                 list_add_tail(&e->list, &bitmap_list);
610                                 bitmaps++;
611                         } else {
612                                 entry->type = BTRFS_FREE_SPACE_EXTENT;
613                         }
614                         node = rb_next(node);
615                         if (!node)
616                                 break;
617                         offset += sizeof(struct btrfs_free_space_entry);
618                         if (offset + sizeof(struct btrfs_free_space_entry) >=
619                             PAGE_CACHE_SIZE)
620                                 break;
621                         entry++;
622                 }
623                 *crc = ~(u32)0;
624                 *crc = btrfs_csum_data(root, addr + start_offset, *crc,
625                                        PAGE_CACHE_SIZE - start_offset);
626                 kunmap(page);
627
628                 btrfs_csum_final(*crc, (char *)crc);
629                 crc++;
630
631                 bytes += PAGE_CACHE_SIZE;
632
633                 ClearPageChecked(page);
634                 set_page_extent_mapped(page);
635                 SetPageUptodate(page);
636                 set_page_dirty(page);
637
638                 /*
639                  * We need to release our reference we got for grab_cache_page,
640                  * except for the first page which will hold our checksums, we
641                  * do that below.
642                  */
643                 if (index != 0) {
644                         unlock_page(page);
645                         page_cache_release(page);
646                 }
647
648                 page_cache_release(page);
649
650                 index++;
651         } while (node);
652
653         /* Write out the bitmaps */
654         list_for_each_safe(pos, n, &bitmap_list) {
655                 void *addr;
656                 struct btrfs_free_space *entry =
657                         list_entry(pos, struct btrfs_free_space, list);
658
659                 page = find_get_page(inode->i_mapping, index);
660
661                 addr = kmap(page);
662                 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
663                 *crc = ~(u32)0;
664                 *crc = btrfs_csum_data(root, addr, *crc, PAGE_CACHE_SIZE);
665                 kunmap(page);
666                 btrfs_csum_final(*crc, (char *)crc);
667                 crc++;
668                 bytes += PAGE_CACHE_SIZE;
669
670                 ClearPageChecked(page);
671                 set_page_extent_mapped(page);
672                 SetPageUptodate(page);
673                 set_page_dirty(page);
674                 unlock_page(page);
675                 page_cache_release(page);
676                 page_cache_release(page);
677                 list_del_init(&entry->list);
678                 index++;
679         }
680
681         /* Zero out the rest of the pages just to make sure */
682         while (index <= last_index) {
683                 void *addr;
684
685                 page = find_get_page(inode->i_mapping, index);
686
687                 addr = kmap(page);
688                 memset(addr, 0, PAGE_CACHE_SIZE);
689                 kunmap(page);
690                 ClearPageChecked(page);
691                 set_page_extent_mapped(page);
692                 SetPageUptodate(page);
693                 set_page_dirty(page);
694                 unlock_page(page);
695                 page_cache_release(page);
696                 page_cache_release(page);
697                 bytes += PAGE_CACHE_SIZE;
698                 index++;
699         }
700
701         btrfs_set_extent_delalloc(inode, 0, bytes - 1, &cached_state);
702
703         /* Write the checksums and trans id to the first page */
704         {
705                 void *addr;
706                 u64 *gen;
707
708                 page = find_get_page(inode->i_mapping, 0);
709
710                 addr = kmap(page);
711                 memcpy(addr, checksums, sizeof(u32) * num_checksums);
712                 gen = addr + (sizeof(u32) * num_checksums);
713                 *gen = trans->transid;
714                 kunmap(page);
715                 ClearPageChecked(page);
716                 set_page_extent_mapped(page);
717                 SetPageUptodate(page);
718                 set_page_dirty(page);
719                 unlock_page(page);
720                 page_cache_release(page);
721                 page_cache_release(page);
722         }
723         BTRFS_I(inode)->generation = trans->transid;
724
725         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
726                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
727
728         filemap_write_and_wait(inode->i_mapping);
729
730         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
731         key.offset = block_group->key.objectid;
732         key.type = 0;
733
734         ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
735         if (ret < 0) {
736                 ret = 0;
737                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
738                                  EXTENT_DIRTY | EXTENT_DELALLOC |
739                                  EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
740                 goto out_free;
741         }
742         leaf = path->nodes[0];
743         if (ret > 0) {
744                 struct btrfs_key found_key;
745                 BUG_ON(!path->slots[0]);
746                 path->slots[0]--;
747                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
748                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
749                     found_key.offset != block_group->key.objectid) {
750                         ret = 0;
751                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
752                                          EXTENT_DIRTY | EXTENT_DELALLOC |
753                                          EXTENT_DO_ACCOUNTING, 0, 0, NULL,
754                                          GFP_NOFS);
755                         btrfs_release_path(root, path);
756                         goto out_free;
757                 }
758         }
759         header = btrfs_item_ptr(leaf, path->slots[0],
760                                 struct btrfs_free_space_header);
761         btrfs_set_free_space_entries(leaf, header, entries);
762         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
763         btrfs_set_free_space_generation(leaf, header, trans->transid);
764         btrfs_mark_buffer_dirty(leaf);
765         btrfs_release_path(root, path);
766
767         ret = 1;
768
769 out_free:
770         if (ret == 0) {
771                 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
772                 spin_lock(&block_group->lock);
773                 block_group->disk_cache_state = BTRFS_DC_ERROR;
774                 spin_unlock(&block_group->lock);
775                 BTRFS_I(inode)->generation = 0;
776         }
777         kfree(checksums);
778         btrfs_update_inode(trans, root, inode);
779         iput(inode);
780         return ret;
781 }
782
783 static inline unsigned long offset_to_bit(u64 bitmap_start, u64 sectorsize,
784                                           u64 offset)
785 {
786         BUG_ON(offset < bitmap_start);
787         offset -= bitmap_start;
788         return (unsigned long)(div64_u64(offset, sectorsize));
789 }
790
791 static inline unsigned long bytes_to_bits(u64 bytes, u64 sectorsize)
792 {
793         return (unsigned long)(div64_u64(bytes, sectorsize));
794 }
795
796 static inline u64 offset_to_bitmap(struct btrfs_block_group_cache *block_group,
797                                    u64 offset)
798 {
799         u64 bitmap_start;
800         u64 bytes_per_bitmap;
801
802         bytes_per_bitmap = BITS_PER_BITMAP * block_group->sectorsize;
803         bitmap_start = offset - block_group->key.objectid;
804         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
805         bitmap_start *= bytes_per_bitmap;
806         bitmap_start += block_group->key.objectid;
807
808         return bitmap_start;
809 }
810
811 static int tree_insert_offset(struct rb_root *root, u64 offset,
812                               struct rb_node *node, int bitmap)
813 {
814         struct rb_node **p = &root->rb_node;
815         struct rb_node *parent = NULL;
816         struct btrfs_free_space *info;
817
818         while (*p) {
819                 parent = *p;
820                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
821
822                 if (offset < info->offset) {
823                         p = &(*p)->rb_left;
824                 } else if (offset > info->offset) {
825                         p = &(*p)->rb_right;
826                 } else {
827                         /*
828                          * we could have a bitmap entry and an extent entry
829                          * share the same offset.  If this is the case, we want
830                          * the extent entry to always be found first if we do a
831                          * linear search through the tree, since we want to have
832                          * the quickest allocation time, and allocating from an
833                          * extent is faster than allocating from a bitmap.  So
834                          * if we're inserting a bitmap and we find an entry at
835                          * this offset, we want to go right, or after this entry
836                          * logically.  If we are inserting an extent and we've
837                          * found a bitmap, we want to go left, or before
838                          * logically.
839                          */
840                         if (bitmap) {
841                                 WARN_ON(info->bitmap);
842                                 p = &(*p)->rb_right;
843                         } else {
844                                 WARN_ON(!info->bitmap);
845                                 p = &(*p)->rb_left;
846                         }
847                 }
848         }
849
850         rb_link_node(node, parent, p);
851         rb_insert_color(node, root);
852
853         return 0;
854 }
855
856 /*
857  * searches the tree for the given offset.
858  *
859  * fuzzy - If this is set, then we are trying to make an allocation, and we just
860  * want a section that has at least bytes size and comes at or after the given
861  * offset.
862  */
863 static struct btrfs_free_space *
864 tree_search_offset(struct btrfs_block_group_cache *block_group,
865                    u64 offset, int bitmap_only, int fuzzy)
866 {
867         struct rb_node *n = block_group->free_space_offset.rb_node;
868         struct btrfs_free_space *entry, *prev = NULL;
869
870         /* find entry that is closest to the 'offset' */
871         while (1) {
872                 if (!n) {
873                         entry = NULL;
874                         break;
875                 }
876
877                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
878                 prev = entry;
879
880                 if (offset < entry->offset)
881                         n = n->rb_left;
882                 else if (offset > entry->offset)
883                         n = n->rb_right;
884                 else
885                         break;
886         }
887
888         if (bitmap_only) {
889                 if (!entry)
890                         return NULL;
891                 if (entry->bitmap)
892                         return entry;
893
894                 /*
895                  * bitmap entry and extent entry may share same offset,
896                  * in that case, bitmap entry comes after extent entry.
897                  */
898                 n = rb_next(n);
899                 if (!n)
900                         return NULL;
901                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
902                 if (entry->offset != offset)
903                         return NULL;
904
905                 WARN_ON(!entry->bitmap);
906                 return entry;
907         } else if (entry) {
908                 if (entry->bitmap) {
909                         /*
910                          * if previous extent entry covers the offset,
911                          * we should return it instead of the bitmap entry
912                          */
913                         n = &entry->offset_index;
914                         while (1) {
915                                 n = rb_prev(n);
916                                 if (!n)
917                                         break;
918                                 prev = rb_entry(n, struct btrfs_free_space,
919                                                 offset_index);
920                                 if (!prev->bitmap) {
921                                         if (prev->offset + prev->bytes > offset)
922                                                 entry = prev;
923                                         break;
924                                 }
925                         }
926                 }
927                 return entry;
928         }
929
930         if (!prev)
931                 return NULL;
932
933         /* find last entry before the 'offset' */
934         entry = prev;
935         if (entry->offset > offset) {
936                 n = rb_prev(&entry->offset_index);
937                 if (n) {
938                         entry = rb_entry(n, struct btrfs_free_space,
939                                         offset_index);
940                         BUG_ON(entry->offset > offset);
941                 } else {
942                         if (fuzzy)
943                                 return entry;
944                         else
945                                 return NULL;
946                 }
947         }
948
949         if (entry->bitmap) {
950                 n = &entry->offset_index;
951                 while (1) {
952                         n = rb_prev(n);
953                         if (!n)
954                                 break;
955                         prev = rb_entry(n, struct btrfs_free_space,
956                                         offset_index);
957                         if (!prev->bitmap) {
958                                 if (prev->offset + prev->bytes > offset)
959                                         return prev;
960                                 break;
961                         }
962                 }
963                 if (entry->offset + BITS_PER_BITMAP *
964                     block_group->sectorsize > offset)
965                         return entry;
966         } else if (entry->offset + entry->bytes > offset)
967                 return entry;
968
969         if (!fuzzy)
970                 return NULL;
971
972         while (1) {
973                 if (entry->bitmap) {
974                         if (entry->offset + BITS_PER_BITMAP *
975                             block_group->sectorsize > offset)
976                                 break;
977                 } else {
978                         if (entry->offset + entry->bytes > offset)
979                                 break;
980                 }
981
982                 n = rb_next(&entry->offset_index);
983                 if (!n)
984                         return NULL;
985                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
986         }
987         return entry;
988 }
989
990 static void unlink_free_space(struct btrfs_block_group_cache *block_group,
991                               struct btrfs_free_space *info)
992 {
993         rb_erase(&info->offset_index, &block_group->free_space_offset);
994         block_group->free_extents--;
995         block_group->free_space -= info->bytes;
996 }
997
998 static int link_free_space(struct btrfs_block_group_cache *block_group,
999                            struct btrfs_free_space *info)
1000 {
1001         int ret = 0;
1002
1003         BUG_ON(!info->bitmap && !info->bytes);
1004         ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
1005                                  &info->offset_index, (info->bitmap != NULL));
1006         if (ret)
1007                 return ret;
1008
1009         block_group->free_space += info->bytes;
1010         block_group->free_extents++;
1011         return ret;
1012 }
1013
1014 static void recalculate_thresholds(struct btrfs_block_group_cache *block_group)
1015 {
1016         u64 max_bytes;
1017         u64 bitmap_bytes;
1018         u64 extent_bytes;
1019
1020         /*
1021          * The goal is to keep the total amount of memory used per 1gb of space
1022          * at or below 32k, so we need to adjust how much memory we allow to be
1023          * used by extent based free space tracking
1024          */
1025         max_bytes = MAX_CACHE_BYTES_PER_GIG *
1026                 (div64_u64(block_group->key.offset, 1024 * 1024 * 1024));
1027
1028         /*
1029          * we want to account for 1 more bitmap than what we have so we can make
1030          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1031          * we add more bitmaps.
1032          */
1033         bitmap_bytes = (block_group->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1034
1035         if (bitmap_bytes >= max_bytes) {
1036                 block_group->extents_thresh = 0;
1037                 return;
1038         }
1039
1040         /*
1041          * we want the extent entry threshold to always be at most 1/2 the maxw
1042          * bytes we can have, or whatever is less than that.
1043          */
1044         extent_bytes = max_bytes - bitmap_bytes;
1045         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1046
1047         block_group->extents_thresh =
1048                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1049 }
1050
1051 static void bitmap_clear_bits(struct btrfs_block_group_cache *block_group,
1052                               struct btrfs_free_space *info, u64 offset,
1053                               u64 bytes)
1054 {
1055         unsigned long start, end;
1056         unsigned long i;
1057
1058         start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1059         end = start + bytes_to_bits(bytes, block_group->sectorsize);
1060         BUG_ON(end > BITS_PER_BITMAP);
1061
1062         for (i = start; i < end; i++)
1063                 clear_bit(i, info->bitmap);
1064
1065         info->bytes -= bytes;
1066         block_group->free_space -= bytes;
1067 }
1068
1069 static void bitmap_set_bits(struct btrfs_block_group_cache *block_group,
1070                             struct btrfs_free_space *info, u64 offset,
1071                             u64 bytes)
1072 {
1073         unsigned long start, end;
1074         unsigned long i;
1075
1076         start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1077         end = start + bytes_to_bits(bytes, block_group->sectorsize);
1078         BUG_ON(end > BITS_PER_BITMAP);
1079
1080         for (i = start; i < end; i++)
1081                 set_bit(i, info->bitmap);
1082
1083         info->bytes += bytes;
1084         block_group->free_space += bytes;
1085 }
1086
1087 static int search_bitmap(struct btrfs_block_group_cache *block_group,
1088                          struct btrfs_free_space *bitmap_info, u64 *offset,
1089                          u64 *bytes)
1090 {
1091         unsigned long found_bits = 0;
1092         unsigned long bits, i;
1093         unsigned long next_zero;
1094
1095         i = offset_to_bit(bitmap_info->offset, block_group->sectorsize,
1096                           max_t(u64, *offset, bitmap_info->offset));
1097         bits = bytes_to_bits(*bytes, block_group->sectorsize);
1098
1099         for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1100              i < BITS_PER_BITMAP;
1101              i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1102                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1103                                                BITS_PER_BITMAP, i);
1104                 if ((next_zero - i) >= bits) {
1105                         found_bits = next_zero - i;
1106                         break;
1107                 }
1108                 i = next_zero;
1109         }
1110
1111         if (found_bits) {
1112                 *offset = (u64)(i * block_group->sectorsize) +
1113                         bitmap_info->offset;
1114                 *bytes = (u64)(found_bits) * block_group->sectorsize;
1115                 return 0;
1116         }
1117
1118         return -1;
1119 }
1120
1121 static struct btrfs_free_space *find_free_space(struct btrfs_block_group_cache
1122                                                 *block_group, u64 *offset,
1123                                                 u64 *bytes, int debug)
1124 {
1125         struct btrfs_free_space *entry;
1126         struct rb_node *node;
1127         int ret;
1128
1129         if (!block_group->free_space_offset.rb_node)
1130                 return NULL;
1131
1132         entry = tree_search_offset(block_group,
1133                                    offset_to_bitmap(block_group, *offset),
1134                                    0, 1);
1135         if (!entry)
1136                 return NULL;
1137
1138         for (node = &entry->offset_index; node; node = rb_next(node)) {
1139                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1140                 if (entry->bytes < *bytes)
1141                         continue;
1142
1143                 if (entry->bitmap) {
1144                         ret = search_bitmap(block_group, entry, offset, bytes);
1145                         if (!ret)
1146                                 return entry;
1147                         continue;
1148                 }
1149
1150                 *offset = entry->offset;
1151                 *bytes = entry->bytes;
1152                 return entry;
1153         }
1154
1155         return NULL;
1156 }
1157
1158 static void add_new_bitmap(struct btrfs_block_group_cache *block_group,
1159                            struct btrfs_free_space *info, u64 offset)
1160 {
1161         u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1162         int max_bitmaps = (int)div64_u64(block_group->key.offset +
1163                                          bytes_per_bg - 1, bytes_per_bg);
1164         BUG_ON(block_group->total_bitmaps >= max_bitmaps);
1165
1166         info->offset = offset_to_bitmap(block_group, offset);
1167         info->bytes = 0;
1168         link_free_space(block_group, info);
1169         block_group->total_bitmaps++;
1170
1171         recalculate_thresholds(block_group);
1172 }
1173
1174 static noinline int remove_from_bitmap(struct btrfs_block_group_cache *block_group,
1175                               struct btrfs_free_space *bitmap_info,
1176                               u64 *offset, u64 *bytes)
1177 {
1178         u64 end;
1179         u64 search_start, search_bytes;
1180         int ret;
1181
1182 again:
1183         end = bitmap_info->offset +
1184                 (u64)(BITS_PER_BITMAP * block_group->sectorsize) - 1;
1185
1186         /*
1187          * XXX - this can go away after a few releases.
1188          *
1189          * since the only user of btrfs_remove_free_space is the tree logging
1190          * stuff, and the only way to test that is under crash conditions, we
1191          * want to have this debug stuff here just in case somethings not
1192          * working.  Search the bitmap for the space we are trying to use to
1193          * make sure its actually there.  If its not there then we need to stop
1194          * because something has gone wrong.
1195          */
1196         search_start = *offset;
1197         search_bytes = *bytes;
1198         ret = search_bitmap(block_group, bitmap_info, &search_start,
1199                             &search_bytes);
1200         BUG_ON(ret < 0 || search_start != *offset);
1201
1202         if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1203                 bitmap_clear_bits(block_group, bitmap_info, *offset,
1204                                   end - *offset + 1);
1205                 *bytes -= end - *offset + 1;
1206                 *offset = end + 1;
1207         } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1208                 bitmap_clear_bits(block_group, bitmap_info, *offset, *bytes);
1209                 *bytes = 0;
1210         }
1211
1212         if (*bytes) {
1213                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1214                 if (!bitmap_info->bytes) {
1215                         unlink_free_space(block_group, bitmap_info);
1216                         kfree(bitmap_info->bitmap);
1217                         kfree(bitmap_info);
1218                         block_group->total_bitmaps--;
1219                         recalculate_thresholds(block_group);
1220                 }
1221
1222                 /*
1223                  * no entry after this bitmap, but we still have bytes to
1224                  * remove, so something has gone wrong.
1225                  */
1226                 if (!next)
1227                         return -EINVAL;
1228
1229                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1230                                        offset_index);
1231
1232                 /*
1233                  * if the next entry isn't a bitmap we need to return to let the
1234                  * extent stuff do its work.
1235                  */
1236                 if (!bitmap_info->bitmap)
1237                         return -EAGAIN;
1238
1239                 /*
1240                  * Ok the next item is a bitmap, but it may not actually hold
1241                  * the information for the rest of this free space stuff, so
1242                  * look for it, and if we don't find it return so we can try
1243                  * everything over again.
1244                  */
1245                 search_start = *offset;
1246                 search_bytes = *bytes;
1247                 ret = search_bitmap(block_group, bitmap_info, &search_start,
1248                                     &search_bytes);
1249                 if (ret < 0 || search_start != *offset)
1250                         return -EAGAIN;
1251
1252                 goto again;
1253         } else if (!bitmap_info->bytes) {
1254                 unlink_free_space(block_group, bitmap_info);
1255                 kfree(bitmap_info->bitmap);
1256                 kfree(bitmap_info);
1257                 block_group->total_bitmaps--;
1258                 recalculate_thresholds(block_group);
1259         }
1260
1261         return 0;
1262 }
1263
1264 static int insert_into_bitmap(struct btrfs_block_group_cache *block_group,
1265                               struct btrfs_free_space *info)
1266 {
1267         struct btrfs_free_space *bitmap_info;
1268         int added = 0;
1269         u64 bytes, offset, end;
1270         int ret;
1271
1272         /*
1273          * If we are below the extents threshold then we can add this as an
1274          * extent, and don't have to deal with the bitmap
1275          */
1276         if (block_group->free_extents < block_group->extents_thresh &&
1277             info->bytes > block_group->sectorsize * 4)
1278                 return 0;
1279
1280         /*
1281          * some block groups are so tiny they can't be enveloped by a bitmap, so
1282          * don't even bother to create a bitmap for this
1283          */
1284         if (BITS_PER_BITMAP * block_group->sectorsize >
1285             block_group->key.offset)
1286                 return 0;
1287
1288         bytes = info->bytes;
1289         offset = info->offset;
1290
1291 again:
1292         bitmap_info = tree_search_offset(block_group,
1293                                          offset_to_bitmap(block_group, offset),
1294                                          1, 0);
1295         if (!bitmap_info) {
1296                 BUG_ON(added);
1297                 goto new_bitmap;
1298         }
1299
1300         end = bitmap_info->offset +
1301                 (u64)(BITS_PER_BITMAP * block_group->sectorsize);
1302
1303         if (offset >= bitmap_info->offset && offset + bytes > end) {
1304                 bitmap_set_bits(block_group, bitmap_info, offset,
1305                                 end - offset);
1306                 bytes -= end - offset;
1307                 offset = end;
1308                 added = 0;
1309         } else if (offset >= bitmap_info->offset && offset + bytes <= end) {
1310                 bitmap_set_bits(block_group, bitmap_info, offset, bytes);
1311                 bytes = 0;
1312         } else {
1313                 BUG();
1314         }
1315
1316         if (!bytes) {
1317                 ret = 1;
1318                 goto out;
1319         } else
1320                 goto again;
1321
1322 new_bitmap:
1323         if (info && info->bitmap) {
1324                 add_new_bitmap(block_group, info, offset);
1325                 added = 1;
1326                 info = NULL;
1327                 goto again;
1328         } else {
1329                 spin_unlock(&block_group->tree_lock);
1330
1331                 /* no pre-allocated info, allocate a new one */
1332                 if (!info) {
1333                         info = kzalloc(sizeof(struct btrfs_free_space),
1334                                        GFP_NOFS);
1335                         if (!info) {
1336                                 spin_lock(&block_group->tree_lock);
1337                                 ret = -ENOMEM;
1338                                 goto out;
1339                         }
1340                 }
1341
1342                 /* allocate the bitmap */
1343                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1344                 spin_lock(&block_group->tree_lock);
1345                 if (!info->bitmap) {
1346                         ret = -ENOMEM;
1347                         goto out;
1348                 }
1349                 goto again;
1350         }
1351
1352 out:
1353         if (info) {
1354                 if (info->bitmap)
1355                         kfree(info->bitmap);
1356                 kfree(info);
1357         }
1358
1359         return ret;
1360 }
1361
1362 int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
1363                          u64 offset, u64 bytes)
1364 {
1365         struct btrfs_free_space *right_info = NULL;
1366         struct btrfs_free_space *left_info = NULL;
1367         struct btrfs_free_space *info = NULL;
1368         int ret = 0;
1369
1370         info = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
1371         if (!info)
1372                 return -ENOMEM;
1373
1374         info->offset = offset;
1375         info->bytes = bytes;
1376
1377         spin_lock(&block_group->tree_lock);
1378
1379         /*
1380          * first we want to see if there is free space adjacent to the range we
1381          * are adding, if there is remove that struct and add a new one to
1382          * cover the entire range
1383          */
1384         right_info = tree_search_offset(block_group, offset + bytes, 0, 0);
1385         if (right_info && rb_prev(&right_info->offset_index))
1386                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1387                                      struct btrfs_free_space, offset_index);
1388         else
1389                 left_info = tree_search_offset(block_group, offset - 1, 0, 0);
1390
1391         /*
1392          * If there was no extent directly to the left or right of this new
1393          * extent then we know we're going to have to allocate a new extent, so
1394          * before we do that see if we need to drop this into a bitmap
1395          */
1396         if ((!left_info || left_info->bitmap) &&
1397             (!right_info || right_info->bitmap)) {
1398                 ret = insert_into_bitmap(block_group, info);
1399
1400                 if (ret < 0) {
1401                         goto out;
1402                 } else if (ret) {
1403                         ret = 0;
1404                         goto out;
1405                 }
1406         }
1407
1408         if (right_info && !right_info->bitmap) {
1409                 unlink_free_space(block_group, right_info);
1410                 info->bytes += right_info->bytes;
1411                 kfree(right_info);
1412         }
1413
1414         if (left_info && !left_info->bitmap &&
1415             left_info->offset + left_info->bytes == offset) {
1416                 unlink_free_space(block_group, left_info);
1417                 info->offset = left_info->offset;
1418                 info->bytes += left_info->bytes;
1419                 kfree(left_info);
1420         }
1421
1422         ret = link_free_space(block_group, info);
1423         if (ret)
1424                 kfree(info);
1425 out:
1426         spin_unlock(&block_group->tree_lock);
1427
1428         if (ret) {
1429                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1430                 BUG_ON(ret == -EEXIST);
1431         }
1432
1433         return ret;
1434 }
1435
1436 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1437                             u64 offset, u64 bytes)
1438 {
1439         struct btrfs_free_space *info;
1440         struct btrfs_free_space *next_info = NULL;
1441         int ret = 0;
1442
1443         spin_lock(&block_group->tree_lock);
1444
1445 again:
1446         info = tree_search_offset(block_group, offset, 0, 0);
1447         if (!info) {
1448                 /*
1449                  * oops didn't find an extent that matched the space we wanted
1450                  * to remove, look for a bitmap instead
1451                  */
1452                 info = tree_search_offset(block_group,
1453                                           offset_to_bitmap(block_group, offset),
1454                                           1, 0);
1455                 if (!info) {
1456                         WARN_ON(1);
1457                         goto out_lock;
1458                 }
1459         }
1460
1461         if (info->bytes < bytes && rb_next(&info->offset_index)) {
1462                 u64 end;
1463                 next_info = rb_entry(rb_next(&info->offset_index),
1464                                              struct btrfs_free_space,
1465                                              offset_index);
1466
1467                 if (next_info->bitmap)
1468                         end = next_info->offset + BITS_PER_BITMAP *
1469                                 block_group->sectorsize - 1;
1470                 else
1471                         end = next_info->offset + next_info->bytes;
1472
1473                 if (next_info->bytes < bytes ||
1474                     next_info->offset > offset || offset > end) {
1475                         printk(KERN_CRIT "Found free space at %llu, size %llu,"
1476                               " trying to use %llu\n",
1477                               (unsigned long long)info->offset,
1478                               (unsigned long long)info->bytes,
1479                               (unsigned long long)bytes);
1480                         WARN_ON(1);
1481                         ret = -EINVAL;
1482                         goto out_lock;
1483                 }
1484
1485                 info = next_info;
1486         }
1487
1488         if (info->bytes == bytes) {
1489                 unlink_free_space(block_group, info);
1490                 if (info->bitmap) {
1491                         kfree(info->bitmap);
1492                         block_group->total_bitmaps--;
1493                 }
1494                 kfree(info);
1495                 goto out_lock;
1496         }
1497
1498         if (!info->bitmap && info->offset == offset) {
1499                 unlink_free_space(block_group, info);
1500                 info->offset += bytes;
1501                 info->bytes -= bytes;
1502                 link_free_space(block_group, info);
1503                 goto out_lock;
1504         }
1505
1506         if (!info->bitmap && info->offset <= offset &&
1507             info->offset + info->bytes >= offset + bytes) {
1508                 u64 old_start = info->offset;
1509                 /*
1510                  * we're freeing space in the middle of the info,
1511                  * this can happen during tree log replay
1512                  *
1513                  * first unlink the old info and then
1514                  * insert it again after the hole we're creating
1515                  */
1516                 unlink_free_space(block_group, info);
1517                 if (offset + bytes < info->offset + info->bytes) {
1518                         u64 old_end = info->offset + info->bytes;
1519
1520                         info->offset = offset + bytes;
1521                         info->bytes = old_end - info->offset;
1522                         ret = link_free_space(block_group, info);
1523                         WARN_ON(ret);
1524                         if (ret)
1525                                 goto out_lock;
1526                 } else {
1527                         /* the hole we're creating ends at the end
1528                          * of the info struct, just free the info
1529                          */
1530                         kfree(info);
1531                 }
1532                 spin_unlock(&block_group->tree_lock);
1533
1534                 /* step two, insert a new info struct to cover
1535                  * anything before the hole
1536                  */
1537                 ret = btrfs_add_free_space(block_group, old_start,
1538                                            offset - old_start);
1539                 WARN_ON(ret);
1540                 goto out;
1541         }
1542
1543         ret = remove_from_bitmap(block_group, info, &offset, &bytes);
1544         if (ret == -EAGAIN)
1545                 goto again;
1546         BUG_ON(ret);
1547 out_lock:
1548         spin_unlock(&block_group->tree_lock);
1549 out:
1550         return ret;
1551 }
1552
1553 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1554                            u64 bytes)
1555 {
1556         struct btrfs_free_space *info;
1557         struct rb_node *n;
1558         int count = 0;
1559
1560         for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
1561                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1562                 if (info->bytes >= bytes)
1563                         count++;
1564                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1565                        (unsigned long long)info->offset,
1566                        (unsigned long long)info->bytes,
1567                        (info->bitmap) ? "yes" : "no");
1568         }
1569         printk(KERN_INFO "block group has cluster?: %s\n",
1570                list_empty(&block_group->cluster_list) ? "no" : "yes");
1571         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1572                "\n", count);
1573 }
1574
1575 u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
1576 {
1577         struct btrfs_free_space *info;
1578         struct rb_node *n;
1579         u64 ret = 0;
1580
1581         for (n = rb_first(&block_group->free_space_offset); n;
1582              n = rb_next(n)) {
1583                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1584                 ret += info->bytes;
1585         }
1586
1587         return ret;
1588 }
1589
1590 /*
1591  * for a given cluster, put all of its extents back into the free
1592  * space cache.  If the block group passed doesn't match the block group
1593  * pointed to by the cluster, someone else raced in and freed the
1594  * cluster already.  In that case, we just return without changing anything
1595  */
1596 static int
1597 __btrfs_return_cluster_to_free_space(
1598                              struct btrfs_block_group_cache *block_group,
1599                              struct btrfs_free_cluster *cluster)
1600 {
1601         struct btrfs_free_space *entry;
1602         struct rb_node *node;
1603         bool bitmap;
1604
1605         spin_lock(&cluster->lock);
1606         if (cluster->block_group != block_group)
1607                 goto out;
1608
1609         bitmap = cluster->points_to_bitmap;
1610         cluster->block_group = NULL;
1611         cluster->window_start = 0;
1612         list_del_init(&cluster->block_group_list);
1613         cluster->points_to_bitmap = false;
1614
1615         if (bitmap)
1616                 goto out;
1617
1618         node = rb_first(&cluster->root);
1619         while (node) {
1620                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1621                 node = rb_next(&entry->offset_index);
1622                 rb_erase(&entry->offset_index, &cluster->root);
1623                 BUG_ON(entry->bitmap);
1624                 tree_insert_offset(&block_group->free_space_offset,
1625                                    entry->offset, &entry->offset_index, 0);
1626         }
1627         cluster->root = RB_ROOT;
1628
1629 out:
1630         spin_unlock(&cluster->lock);
1631         btrfs_put_block_group(block_group);
1632         return 0;
1633 }
1634
1635 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1636 {
1637         struct btrfs_free_space *info;
1638         struct rb_node *node;
1639         struct btrfs_free_cluster *cluster;
1640         struct list_head *head;
1641
1642         spin_lock(&block_group->tree_lock);
1643         while ((head = block_group->cluster_list.next) !=
1644                &block_group->cluster_list) {
1645                 cluster = list_entry(head, struct btrfs_free_cluster,
1646                                      block_group_list);
1647
1648                 WARN_ON(cluster->block_group != block_group);
1649                 __btrfs_return_cluster_to_free_space(block_group, cluster);
1650                 if (need_resched()) {
1651                         spin_unlock(&block_group->tree_lock);
1652                         cond_resched();
1653                         spin_lock(&block_group->tree_lock);
1654                 }
1655         }
1656
1657         while ((node = rb_last(&block_group->free_space_offset)) != NULL) {
1658                 info = rb_entry(node, struct btrfs_free_space, offset_index);
1659                 unlink_free_space(block_group, info);
1660                 if (info->bitmap)
1661                         kfree(info->bitmap);
1662                 kfree(info);
1663                 if (need_resched()) {
1664                         spin_unlock(&block_group->tree_lock);
1665                         cond_resched();
1666                         spin_lock(&block_group->tree_lock);
1667                 }
1668         }
1669
1670         spin_unlock(&block_group->tree_lock);
1671 }
1672
1673 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1674                                u64 offset, u64 bytes, u64 empty_size)
1675 {
1676         struct btrfs_free_space *entry = NULL;
1677         u64 bytes_search = bytes + empty_size;
1678         u64 ret = 0;
1679
1680         spin_lock(&block_group->tree_lock);
1681         entry = find_free_space(block_group, &offset, &bytes_search, 0);
1682         if (!entry)
1683                 goto out;
1684
1685         ret = offset;
1686         if (entry->bitmap) {
1687                 bitmap_clear_bits(block_group, entry, offset, bytes);
1688                 if (!entry->bytes) {
1689                         unlink_free_space(block_group, entry);
1690                         kfree(entry->bitmap);
1691                         kfree(entry);
1692                         block_group->total_bitmaps--;
1693                         recalculate_thresholds(block_group);
1694                 }
1695         } else {
1696                 unlink_free_space(block_group, entry);
1697                 entry->offset += bytes;
1698                 entry->bytes -= bytes;
1699                 if (!entry->bytes)
1700                         kfree(entry);
1701                 else
1702                         link_free_space(block_group, entry);
1703         }
1704
1705 out:
1706         spin_unlock(&block_group->tree_lock);
1707
1708         return ret;
1709 }
1710
1711 /*
1712  * given a cluster, put all of its extents back into the free space
1713  * cache.  If a block group is passed, this function will only free
1714  * a cluster that belongs to the passed block group.
1715  *
1716  * Otherwise, it'll get a reference on the block group pointed to by the
1717  * cluster and remove the cluster from it.
1718  */
1719 int btrfs_return_cluster_to_free_space(
1720                                struct btrfs_block_group_cache *block_group,
1721                                struct btrfs_free_cluster *cluster)
1722 {
1723         int ret;
1724
1725         /* first, get a safe pointer to the block group */
1726         spin_lock(&cluster->lock);
1727         if (!block_group) {
1728                 block_group = cluster->block_group;
1729                 if (!block_group) {
1730                         spin_unlock(&cluster->lock);
1731                         return 0;
1732                 }
1733         } else if (cluster->block_group != block_group) {
1734                 /* someone else has already freed it don't redo their work */
1735                 spin_unlock(&cluster->lock);
1736                 return 0;
1737         }
1738         atomic_inc(&block_group->count);
1739         spin_unlock(&cluster->lock);
1740
1741         /* now return any extents the cluster had on it */
1742         spin_lock(&block_group->tree_lock);
1743         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1744         spin_unlock(&block_group->tree_lock);
1745
1746         /* finally drop our ref */
1747         btrfs_put_block_group(block_group);
1748         return ret;
1749 }
1750
1751 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1752                                    struct btrfs_free_cluster *cluster,
1753                                    u64 bytes, u64 min_start)
1754 {
1755         struct btrfs_free_space *entry;
1756         int err;
1757         u64 search_start = cluster->window_start;
1758         u64 search_bytes = bytes;
1759         u64 ret = 0;
1760
1761         spin_lock(&block_group->tree_lock);
1762         spin_lock(&cluster->lock);
1763
1764         if (!cluster->points_to_bitmap)
1765                 goto out;
1766
1767         if (cluster->block_group != block_group)
1768                 goto out;
1769
1770         /*
1771          * search_start is the beginning of the bitmap, but at some point it may
1772          * be a good idea to point to the actual start of the free area in the
1773          * bitmap, so do the offset_to_bitmap trick anyway, and set bitmap_only
1774          * to 1 to make sure we get the bitmap entry
1775          */
1776         entry = tree_search_offset(block_group,
1777                                    offset_to_bitmap(block_group, search_start),
1778                                    1, 0);
1779         if (!entry || !entry->bitmap)
1780                 goto out;
1781
1782         search_start = min_start;
1783         search_bytes = bytes;
1784
1785         err = search_bitmap(block_group, entry, &search_start,
1786                             &search_bytes);
1787         if (err)
1788                 goto out;
1789
1790         ret = search_start;
1791         bitmap_clear_bits(block_group, entry, ret, bytes);
1792 out:
1793         spin_unlock(&cluster->lock);
1794         spin_unlock(&block_group->tree_lock);
1795
1796         return ret;
1797 }
1798
1799 /*
1800  * given a cluster, try to allocate 'bytes' from it, returns 0
1801  * if it couldn't find anything suitably large, or a logical disk offset
1802  * if things worked out
1803  */
1804 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
1805                              struct btrfs_free_cluster *cluster, u64 bytes,
1806                              u64 min_start)
1807 {
1808         struct btrfs_free_space *entry = NULL;
1809         struct rb_node *node;
1810         u64 ret = 0;
1811
1812         if (cluster->points_to_bitmap)
1813                 return btrfs_alloc_from_bitmap(block_group, cluster, bytes,
1814                                                min_start);
1815
1816         spin_lock(&cluster->lock);
1817         if (bytes > cluster->max_size)
1818                 goto out;
1819
1820         if (cluster->block_group != block_group)
1821                 goto out;
1822
1823         node = rb_first(&cluster->root);
1824         if (!node)
1825                 goto out;
1826
1827         entry = rb_entry(node, struct btrfs_free_space, offset_index);
1828
1829         while(1) {
1830                 if (entry->bytes < bytes || entry->offset < min_start) {
1831                         struct rb_node *node;
1832
1833                         node = rb_next(&entry->offset_index);
1834                         if (!node)
1835                                 break;
1836                         entry = rb_entry(node, struct btrfs_free_space,
1837                                          offset_index);
1838                         continue;
1839                 }
1840                 ret = entry->offset;
1841
1842                 entry->offset += bytes;
1843                 entry->bytes -= bytes;
1844
1845                 if (entry->bytes == 0) {
1846                         rb_erase(&entry->offset_index, &cluster->root);
1847                         kfree(entry);
1848                 }
1849                 break;
1850         }
1851 out:
1852         spin_unlock(&cluster->lock);
1853
1854         return ret;
1855 }
1856
1857 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
1858                                 struct btrfs_free_space *entry,
1859                                 struct btrfs_free_cluster *cluster,
1860                                 u64 offset, u64 bytes, u64 min_bytes)
1861 {
1862         unsigned long next_zero;
1863         unsigned long i;
1864         unsigned long search_bits;
1865         unsigned long total_bits;
1866         unsigned long found_bits;
1867         unsigned long start = 0;
1868         unsigned long total_found = 0;
1869         bool found = false;
1870
1871         i = offset_to_bit(entry->offset, block_group->sectorsize,
1872                           max_t(u64, offset, entry->offset));
1873         search_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
1874         total_bits = bytes_to_bits(bytes, block_group->sectorsize);
1875
1876 again:
1877         found_bits = 0;
1878         for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
1879              i < BITS_PER_BITMAP;
1880              i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
1881                 next_zero = find_next_zero_bit(entry->bitmap,
1882                                                BITS_PER_BITMAP, i);
1883                 if (next_zero - i >= search_bits) {
1884                         found_bits = next_zero - i;
1885                         break;
1886                 }
1887                 i = next_zero;
1888         }
1889
1890         if (!found_bits)
1891                 return -1;
1892
1893         if (!found) {
1894                 start = i;
1895                 found = true;
1896         }
1897
1898         total_found += found_bits;
1899
1900         if (cluster->max_size < found_bits * block_group->sectorsize)
1901                 cluster->max_size = found_bits * block_group->sectorsize;
1902
1903         if (total_found < total_bits) {
1904                 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
1905                 if (i - start > total_bits * 2) {
1906                         total_found = 0;
1907                         cluster->max_size = 0;
1908                         found = false;
1909                 }
1910                 goto again;
1911         }
1912
1913         cluster->window_start = start * block_group->sectorsize +
1914                 entry->offset;
1915         cluster->points_to_bitmap = true;
1916
1917         return 0;
1918 }
1919
1920 /*
1921  * here we try to find a cluster of blocks in a block group.  The goal
1922  * is to find at least bytes free and up to empty_size + bytes free.
1923  * We might not find them all in one contiguous area.
1924  *
1925  * returns zero and sets up cluster if things worked out, otherwise
1926  * it returns -enospc
1927  */
1928 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
1929                              struct btrfs_root *root,
1930                              struct btrfs_block_group_cache *block_group,
1931                              struct btrfs_free_cluster *cluster,
1932                              u64 offset, u64 bytes, u64 empty_size)
1933 {
1934         struct btrfs_free_space *entry = NULL;
1935         struct rb_node *node;
1936         struct btrfs_free_space *next;
1937         struct btrfs_free_space *last = NULL;
1938         u64 min_bytes;
1939         u64 window_start;
1940         u64 window_free;
1941         u64 max_extent = 0;
1942         bool found_bitmap = false;
1943         int ret;
1944
1945         /* for metadata, allow allocates with more holes */
1946         if (btrfs_test_opt(root, SSD_SPREAD)) {
1947                 min_bytes = bytes + empty_size;
1948         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
1949                 /*
1950                  * we want to do larger allocations when we are
1951                  * flushing out the delayed refs, it helps prevent
1952                  * making more work as we go along.
1953                  */
1954                 if (trans->transaction->delayed_refs.flushing)
1955                         min_bytes = max(bytes, (bytes + empty_size) >> 1);
1956                 else
1957                         min_bytes = max(bytes, (bytes + empty_size) >> 4);
1958         } else
1959                 min_bytes = max(bytes, (bytes + empty_size) >> 2);
1960
1961         spin_lock(&block_group->tree_lock);
1962         spin_lock(&cluster->lock);
1963
1964         /* someone already found a cluster, hooray */
1965         if (cluster->block_group) {
1966                 ret = 0;
1967                 goto out;
1968         }
1969 again:
1970         entry = tree_search_offset(block_group, offset, found_bitmap, 1);
1971         if (!entry) {
1972                 ret = -ENOSPC;
1973                 goto out;
1974         }
1975
1976         /*
1977          * If found_bitmap is true, we exhausted our search for extent entries,
1978          * and we just want to search all of the bitmaps that we can find, and
1979          * ignore any extent entries we find.
1980          */
1981         while (entry->bitmap || found_bitmap ||
1982                (!entry->bitmap && entry->bytes < min_bytes)) {
1983                 struct rb_node *node = rb_next(&entry->offset_index);
1984
1985                 if (entry->bitmap && entry->bytes > bytes + empty_size) {
1986                         ret = btrfs_bitmap_cluster(block_group, entry, cluster,
1987                                                    offset, bytes + empty_size,
1988                                                    min_bytes);
1989                         if (!ret)
1990                                 goto got_it;
1991                 }
1992
1993                 if (!node) {
1994                         ret = -ENOSPC;
1995                         goto out;
1996                 }
1997                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1998         }
1999
2000         /*
2001          * We already searched all the extent entries from the passed in offset
2002          * to the end and didn't find enough space for the cluster, and we also
2003          * didn't find any bitmaps that met our criteria, just go ahead and exit
2004          */
2005         if (found_bitmap) {
2006                 ret = -ENOSPC;
2007                 goto out;
2008         }
2009
2010         cluster->points_to_bitmap = false;
2011         window_start = entry->offset;
2012         window_free = entry->bytes;
2013         last = entry;
2014         max_extent = entry->bytes;
2015
2016         while (1) {
2017                 /* out window is just right, lets fill it */
2018                 if (window_free >= bytes + empty_size)
2019                         break;
2020
2021                 node = rb_next(&last->offset_index);
2022                 if (!node) {
2023                         if (found_bitmap)
2024                                 goto again;
2025                         ret = -ENOSPC;
2026                         goto out;
2027                 }
2028                 next = rb_entry(node, struct btrfs_free_space, offset_index);
2029
2030                 /*
2031                  * we found a bitmap, so if this search doesn't result in a
2032                  * cluster, we know to go and search again for the bitmaps and
2033                  * start looking for space there
2034                  */
2035                 if (next->bitmap) {
2036                         if (!found_bitmap)
2037                                 offset = next->offset;
2038                         found_bitmap = true;
2039                         last = next;
2040                         continue;
2041                 }
2042
2043                 /*
2044                  * we haven't filled the empty size and the window is
2045                  * very large.  reset and try again
2046                  */
2047                 if (next->offset - (last->offset + last->bytes) > 128 * 1024 ||
2048                     next->offset - window_start > (bytes + empty_size) * 2) {
2049                         entry = next;
2050                         window_start = entry->offset;
2051                         window_free = entry->bytes;
2052                         last = entry;
2053                         max_extent = entry->bytes;
2054                 } else {
2055                         last = next;
2056                         window_free += next->bytes;
2057                         if (entry->bytes > max_extent)
2058                                 max_extent = entry->bytes;
2059                 }
2060         }
2061
2062         cluster->window_start = entry->offset;
2063
2064         /*
2065          * now we've found our entries, pull them out of the free space
2066          * cache and put them into the cluster rbtree
2067          *
2068          * The cluster includes an rbtree, but only uses the offset index
2069          * of each free space cache entry.
2070          */
2071         while (1) {
2072                 node = rb_next(&entry->offset_index);
2073                 if (entry->bitmap && node) {
2074                         entry = rb_entry(node, struct btrfs_free_space,
2075                                          offset_index);
2076                         continue;
2077                 } else if (entry->bitmap && !node) {
2078                         break;
2079                 }
2080
2081                 rb_erase(&entry->offset_index, &block_group->free_space_offset);
2082                 ret = tree_insert_offset(&cluster->root, entry->offset,
2083                                          &entry->offset_index, 0);
2084                 BUG_ON(ret);
2085
2086                 if (!node || entry == last)
2087                         break;
2088
2089                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2090         }
2091
2092         cluster->max_size = max_extent;
2093 got_it:
2094         ret = 0;
2095         atomic_inc(&block_group->count);
2096         list_add_tail(&cluster->block_group_list, &block_group->cluster_list);
2097         cluster->block_group = block_group;
2098 out:
2099         spin_unlock(&cluster->lock);
2100         spin_unlock(&block_group->tree_lock);
2101
2102         return ret;
2103 }
2104
2105 /*
2106  * simple code to zero out a cluster
2107  */
2108 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2109 {
2110         spin_lock_init(&cluster->lock);
2111         spin_lock_init(&cluster->refill_lock);
2112         cluster->root = RB_ROOT;
2113         cluster->max_size = 0;
2114         cluster->points_to_bitmap = false;
2115         INIT_LIST_HEAD(&cluster->block_group_list);
2116         cluster->block_group = NULL;
2117 }
2118