Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs...
[pandora-kernel.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 #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_fs_closing(root->fs_info)) {
102                 block_group->inode = igrab(inode);
103                 block_group->iref = 1;
104         }
105         spin_unlock(&block_group->lock);
106
107         return inode;
108 }
109
110 int __create_free_space_inode(struct btrfs_root *root,
111                               struct btrfs_trans_handle *trans,
112                               struct btrfs_path *path, u64 ino, u64 offset)
113 {
114         struct btrfs_key key;
115         struct btrfs_disk_key disk_key;
116         struct btrfs_free_space_header *header;
117         struct btrfs_inode_item *inode_item;
118         struct extent_buffer *leaf;
119         int ret;
120
121         ret = btrfs_insert_empty_inode(trans, root, path, ino);
122         if (ret)
123                 return ret;
124
125         leaf = path->nodes[0];
126         inode_item = btrfs_item_ptr(leaf, path->slots[0],
127                                     struct btrfs_inode_item);
128         btrfs_item_key(leaf, &disk_key, path->slots[0]);
129         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
130                              sizeof(*inode_item));
131         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
132         btrfs_set_inode_size(leaf, inode_item, 0);
133         btrfs_set_inode_nbytes(leaf, inode_item, 0);
134         btrfs_set_inode_uid(leaf, inode_item, 0);
135         btrfs_set_inode_gid(leaf, inode_item, 0);
136         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
137         btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
138                               BTRFS_INODE_PREALLOC | BTRFS_INODE_NODATASUM);
139         btrfs_set_inode_nlink(leaf, inode_item, 1);
140         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
141         btrfs_set_inode_block_group(leaf, inode_item, offset);
142         btrfs_mark_buffer_dirty(leaf);
143         btrfs_release_path(path);
144
145         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
146         key.offset = offset;
147         key.type = 0;
148
149         ret = btrfs_insert_empty_item(trans, root, path, &key,
150                                       sizeof(struct btrfs_free_space_header));
151         if (ret < 0) {
152                 btrfs_release_path(path);
153                 return ret;
154         }
155         leaf = path->nodes[0];
156         header = btrfs_item_ptr(leaf, path->slots[0],
157                                 struct btrfs_free_space_header);
158         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
159         btrfs_set_free_space_key(leaf, header, &disk_key);
160         btrfs_mark_buffer_dirty(leaf);
161         btrfs_release_path(path);
162
163         return 0;
164 }
165
166 int create_free_space_inode(struct btrfs_root *root,
167                             struct btrfs_trans_handle *trans,
168                             struct btrfs_block_group_cache *block_group,
169                             struct btrfs_path *path)
170 {
171         int ret;
172         u64 ino;
173
174         ret = btrfs_find_free_objectid(root, &ino);
175         if (ret < 0)
176                 return ret;
177
178         return __create_free_space_inode(root, trans, path, ino,
179                                          block_group->key.objectid);
180 }
181
182 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
183                                     struct btrfs_trans_handle *trans,
184                                     struct btrfs_path *path,
185                                     struct inode *inode)
186 {
187         loff_t oldsize;
188         int ret = 0;
189
190         trans->block_rsv = root->orphan_block_rsv;
191         ret = btrfs_block_rsv_check(trans, root,
192                                     root->orphan_block_rsv,
193                                     0, 5);
194         if (ret)
195                 return ret;
196
197         oldsize = i_size_read(inode);
198         btrfs_i_size_write(inode, 0);
199         truncate_pagecache(inode, oldsize, 0);
200
201         /*
202          * We don't need an orphan item because truncating the free space cache
203          * will never be split across transactions.
204          */
205         ret = btrfs_truncate_inode_items(trans, root, inode,
206                                          0, BTRFS_EXTENT_DATA_KEY);
207         if (ret) {
208                 WARN_ON(1);
209                 return ret;
210         }
211
212         ret = btrfs_update_inode(trans, root, inode);
213         return ret;
214 }
215
216 static int readahead_cache(struct inode *inode)
217 {
218         struct file_ra_state *ra;
219         unsigned long last_index;
220
221         ra = kzalloc(sizeof(*ra), GFP_NOFS);
222         if (!ra)
223                 return -ENOMEM;
224
225         file_ra_state_init(ra, inode->i_mapping);
226         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
227
228         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
229
230         kfree(ra);
231
232         return 0;
233 }
234
235 int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
236                             struct btrfs_free_space_ctl *ctl,
237                             struct btrfs_path *path, u64 offset)
238 {
239         struct btrfs_free_space_header *header;
240         struct extent_buffer *leaf;
241         struct page *page;
242         u32 *checksums = NULL, *crc;
243         char *disk_crcs = NULL;
244         struct btrfs_key key;
245         struct list_head bitmaps;
246         u64 num_entries;
247         u64 num_bitmaps;
248         u64 generation;
249         u32 cur_crc = ~(u32)0;
250         pgoff_t index = 0;
251         unsigned long first_page_offset;
252         int num_checksums;
253         int ret = 0, ret2;
254
255         INIT_LIST_HEAD(&bitmaps);
256
257         /* Nothing in the space cache, goodbye */
258         if (!i_size_read(inode))
259                 goto out;
260
261         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
262         key.offset = offset;
263         key.type = 0;
264
265         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
266         if (ret < 0)
267                 goto out;
268         else if (ret > 0) {
269                 btrfs_release_path(path);
270                 ret = 0;
271                 goto out;
272         }
273
274         ret = -1;
275
276         leaf = path->nodes[0];
277         header = btrfs_item_ptr(leaf, path->slots[0],
278                                 struct btrfs_free_space_header);
279         num_entries = btrfs_free_space_entries(leaf, header);
280         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
281         generation = btrfs_free_space_generation(leaf, header);
282         btrfs_release_path(path);
283
284         if (BTRFS_I(inode)->generation != generation) {
285                 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
286                        " not match free space cache generation (%llu)\n",
287                        (unsigned long long)BTRFS_I(inode)->generation,
288                        (unsigned long long)generation);
289                 goto out;
290         }
291
292         if (!num_entries)
293                 goto out;
294
295         /* Setup everything for doing checksumming */
296         num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
297         checksums = crc = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
298         if (!checksums)
299                 goto out;
300         first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
301         disk_crcs = kzalloc(first_page_offset, GFP_NOFS);
302         if (!disk_crcs)
303                 goto out;
304
305         ret = readahead_cache(inode);
306         if (ret)
307                 goto out;
308
309         while (1) {
310                 struct btrfs_free_space_entry *entry;
311                 struct btrfs_free_space *e;
312                 void *addr;
313                 unsigned long offset = 0;
314                 unsigned long start_offset = 0;
315                 int need_loop = 0;
316
317                 if (!num_entries && !num_bitmaps)
318                         break;
319
320                 if (index == 0) {
321                         start_offset = first_page_offset;
322                         offset = start_offset;
323                 }
324
325                 page = grab_cache_page(inode->i_mapping, index);
326                 if (!page)
327                         goto free_cache;
328
329                 if (!PageUptodate(page)) {
330                         btrfs_readpage(NULL, page);
331                         lock_page(page);
332                         if (!PageUptodate(page)) {
333                                 unlock_page(page);
334                                 page_cache_release(page);
335                                 printk(KERN_ERR "btrfs: error reading free "
336                                        "space cache\n");
337                                 goto free_cache;
338                         }
339                 }
340                 addr = kmap(page);
341
342                 if (index == 0) {
343                         u64 *gen;
344
345                         memcpy(disk_crcs, addr, first_page_offset);
346                         gen = addr + (sizeof(u32) * num_checksums);
347                         if (*gen != BTRFS_I(inode)->generation) {
348                                 printk(KERN_ERR "btrfs: space cache generation"
349                                        " (%llu) does not match inode (%llu)\n",
350                                        (unsigned long long)*gen,
351                                        (unsigned long long)
352                                        BTRFS_I(inode)->generation);
353                                 kunmap(page);
354                                 unlock_page(page);
355                                 page_cache_release(page);
356                                 goto free_cache;
357                         }
358                         crc = (u32 *)disk_crcs;
359                 }
360                 entry = addr + start_offset;
361
362                 /* First lets check our crc before we do anything fun */
363                 cur_crc = ~(u32)0;
364                 cur_crc = btrfs_csum_data(root, addr + start_offset, cur_crc,
365                                           PAGE_CACHE_SIZE - start_offset);
366                 btrfs_csum_final(cur_crc, (char *)&cur_crc);
367                 if (cur_crc != *crc) {
368                         printk(KERN_ERR "btrfs: crc mismatch for page %lu\n",
369                                index);
370                         kunmap(page);
371                         unlock_page(page);
372                         page_cache_release(page);
373                         goto free_cache;
374                 }
375                 crc++;
376
377                 while (1) {
378                         if (!num_entries)
379                                 break;
380
381                         need_loop = 1;
382                         e = kmem_cache_zalloc(btrfs_free_space_cachep,
383                                               GFP_NOFS);
384                         if (!e) {
385                                 kunmap(page);
386                                 unlock_page(page);
387                                 page_cache_release(page);
388                                 goto free_cache;
389                         }
390
391                         e->offset = le64_to_cpu(entry->offset);
392                         e->bytes = le64_to_cpu(entry->bytes);
393                         if (!e->bytes) {
394                                 kunmap(page);
395                                 kmem_cache_free(btrfs_free_space_cachep, e);
396                                 unlock_page(page);
397                                 page_cache_release(page);
398                                 goto free_cache;
399                         }
400
401                         if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
402                                 spin_lock(&ctl->tree_lock);
403                                 ret = link_free_space(ctl, e);
404                                 spin_unlock(&ctl->tree_lock);
405                                 if (ret) {
406                                         printk(KERN_ERR "Duplicate entries in "
407                                                "free space cache, dumping\n");
408                                         kunmap(page);
409                                         unlock_page(page);
410                                         page_cache_release(page);
411                                         goto free_cache;
412                                 }
413                         } else {
414                                 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
415                                 if (!e->bitmap) {
416                                         kunmap(page);
417                                         kmem_cache_free(
418                                                 btrfs_free_space_cachep, e);
419                                         unlock_page(page);
420                                         page_cache_release(page);
421                                         goto free_cache;
422                                 }
423                                 spin_lock(&ctl->tree_lock);
424                                 ret2 = link_free_space(ctl, e);
425                                 ctl->total_bitmaps++;
426                                 ctl->op->recalc_thresholds(ctl);
427                                 spin_unlock(&ctl->tree_lock);
428                                 list_add_tail(&e->list, &bitmaps);
429                                 if (ret) {
430                                         printk(KERN_ERR "Duplicate entries in "
431                                                "free space cache, dumping\n");
432                                         kunmap(page);
433                                         unlock_page(page);
434                                         page_cache_release(page);
435                                         goto free_cache;
436                                 }
437                         }
438
439                         num_entries--;
440                         offset += sizeof(struct btrfs_free_space_entry);
441                         if (offset + sizeof(struct btrfs_free_space_entry) >=
442                             PAGE_CACHE_SIZE)
443                                 break;
444                         entry++;
445                 }
446
447                 /*
448                  * We read an entry out of this page, we need to move on to the
449                  * next page.
450                  */
451                 if (need_loop) {
452                         kunmap(page);
453                         goto next;
454                 }
455
456                 /*
457                  * We add the bitmaps at the end of the entries in order that
458                  * the bitmap entries are added to the cache.
459                  */
460                 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
461                 list_del_init(&e->list);
462                 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
463                 kunmap(page);
464                 num_bitmaps--;
465 next:
466                 unlock_page(page);
467                 page_cache_release(page);
468                 index++;
469         }
470
471         ret = 1;
472 out:
473         kfree(checksums);
474         kfree(disk_crcs);
475         return ret;
476 free_cache:
477         __btrfs_remove_free_space_cache(ctl);
478         goto out;
479 }
480
481 int load_free_space_cache(struct btrfs_fs_info *fs_info,
482                           struct btrfs_block_group_cache *block_group)
483 {
484         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
485         struct btrfs_root *root = fs_info->tree_root;
486         struct inode *inode;
487         struct btrfs_path *path;
488         int ret;
489         bool matched;
490         u64 used = btrfs_block_group_used(&block_group->item);
491
492         /*
493          * If we're unmounting then just return, since this does a search on the
494          * normal root and not the commit root and we could deadlock.
495          */
496         if (btrfs_fs_closing(fs_info))
497                 return 0;
498
499         /*
500          * If this block group has been marked to be cleared for one reason or
501          * another then we can't trust the on disk cache, so just return.
502          */
503         spin_lock(&block_group->lock);
504         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
505                 spin_unlock(&block_group->lock);
506                 return 0;
507         }
508         spin_unlock(&block_group->lock);
509
510         path = btrfs_alloc_path();
511         if (!path)
512                 return 0;
513
514         inode = lookup_free_space_inode(root, block_group, path);
515         if (IS_ERR(inode)) {
516                 btrfs_free_path(path);
517                 return 0;
518         }
519
520         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
521                                       path, block_group->key.objectid);
522         btrfs_free_path(path);
523         if (ret <= 0)
524                 goto out;
525
526         spin_lock(&ctl->tree_lock);
527         matched = (ctl->free_space == (block_group->key.offset - used -
528                                        block_group->bytes_super));
529         spin_unlock(&ctl->tree_lock);
530
531         if (!matched) {
532                 __btrfs_remove_free_space_cache(ctl);
533                 printk(KERN_ERR "block group %llu has an wrong amount of free "
534                        "space\n", block_group->key.objectid);
535                 ret = -1;
536         }
537 out:
538         if (ret < 0) {
539                 /* This cache is bogus, make sure it gets cleared */
540                 spin_lock(&block_group->lock);
541                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
542                 spin_unlock(&block_group->lock);
543                 ret = 0;
544
545                 printk(KERN_ERR "btrfs: failed to load free space cache "
546                        "for block group %llu\n", block_group->key.objectid);
547         }
548
549         iput(inode);
550         return ret;
551 }
552
553 int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
554                             struct btrfs_free_space_ctl *ctl,
555                             struct btrfs_block_group_cache *block_group,
556                             struct btrfs_trans_handle *trans,
557                             struct btrfs_path *path, u64 offset)
558 {
559         struct btrfs_free_space_header *header;
560         struct extent_buffer *leaf;
561         struct rb_node *node;
562         struct list_head *pos, *n;
563         struct page **pages;
564         struct page *page;
565         struct extent_state *cached_state = NULL;
566         struct btrfs_free_cluster *cluster = NULL;
567         struct extent_io_tree *unpin = NULL;
568         struct list_head bitmap_list;
569         struct btrfs_key key;
570         u64 start, end, len;
571         u64 bytes = 0;
572         u32 *crc, *checksums;
573         unsigned long first_page_offset;
574         int index = 0, num_pages = 0;
575         int entries = 0;
576         int bitmaps = 0;
577         int ret = -1;
578         bool next_page = false;
579         bool out_of_space = false;
580
581         INIT_LIST_HEAD(&bitmap_list);
582
583         node = rb_first(&ctl->free_space_offset);
584         if (!node)
585                 return 0;
586
587         if (!i_size_read(inode))
588                 return -1;
589
590         num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
591                 PAGE_CACHE_SHIFT;
592
593         /* Since the first page has all of our checksums and our generation we
594          * need to calculate the offset into the page that we can start writing
595          * our entries.
596          */
597         first_page_offset = (sizeof(u32) * num_pages) + sizeof(u64);
598
599         filemap_write_and_wait(inode->i_mapping);
600         btrfs_wait_ordered_range(inode, inode->i_size &
601                                  ~(root->sectorsize - 1), (u64)-1);
602
603         /* make sure we don't overflow that first page */
604         if (first_page_offset + sizeof(struct btrfs_free_space_entry) >= PAGE_CACHE_SIZE) {
605                 /* this is really the same as running out of space, where we also return 0 */
606                 printk(KERN_CRIT "Btrfs: free space cache was too big for the crc page\n");
607                 ret = 0;
608                 goto out_update;
609         }
610
611         /* We need a checksum per page. */
612         crc = checksums = kzalloc(sizeof(u32) * num_pages, GFP_NOFS);
613         if (!crc)
614                 return -1;
615
616         pages = kzalloc(sizeof(struct page *) * num_pages, GFP_NOFS);
617         if (!pages) {
618                 kfree(crc);
619                 return -1;
620         }
621
622         /* Get the cluster for this block_group if it exists */
623         if (block_group && !list_empty(&block_group->cluster_list))
624                 cluster = list_entry(block_group->cluster_list.next,
625                                      struct btrfs_free_cluster,
626                                      block_group_list);
627
628         /*
629          * We shouldn't have switched the pinned extents yet so this is the
630          * right one
631          */
632         unpin = root->fs_info->pinned_extents;
633
634         /*
635          * Lock all pages first so we can lock the extent safely.
636          *
637          * NOTE: Because we hold the ref the entire time we're going to write to
638          * the page find_get_page should never fail, so we don't do a check
639          * after find_get_page at this point.  Just putting this here so people
640          * know and don't freak out.
641          */
642         while (index < num_pages) {
643                 page = grab_cache_page(inode->i_mapping, index);
644                 if (!page) {
645                         int i;
646
647                         for (i = 0; i < num_pages; i++) {
648                                 unlock_page(pages[i]);
649                                 page_cache_release(pages[i]);
650                         }
651                         goto out_free;
652                 }
653                 pages[index] = page;
654                 index++;
655         }
656
657         index = 0;
658         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
659                          0, &cached_state, GFP_NOFS);
660
661         /*
662          * When searching for pinned extents, we need to start at our start
663          * offset.
664          */
665         if (block_group)
666                 start = block_group->key.objectid;
667
668         /* Write out the extent entries */
669         do {
670                 struct btrfs_free_space_entry *entry;
671                 void *addr;
672                 unsigned long offset = 0;
673                 unsigned long start_offset = 0;
674
675                 next_page = false;
676
677                 if (index == 0) {
678                         start_offset = first_page_offset;
679                         offset = start_offset;
680                 }
681
682                 if (index >= num_pages) {
683                         out_of_space = true;
684                         break;
685                 }
686
687                 page = pages[index];
688
689                 addr = kmap(page);
690                 entry = addr + start_offset;
691
692                 memset(addr, 0, PAGE_CACHE_SIZE);
693                 while (node && !next_page) {
694                         struct btrfs_free_space *e;
695
696                         e = rb_entry(node, struct btrfs_free_space, offset_index);
697                         entries++;
698
699                         entry->offset = cpu_to_le64(e->offset);
700                         entry->bytes = cpu_to_le64(e->bytes);
701                         if (e->bitmap) {
702                                 entry->type = BTRFS_FREE_SPACE_BITMAP;
703                                 list_add_tail(&e->list, &bitmap_list);
704                                 bitmaps++;
705                         } else {
706                                 entry->type = BTRFS_FREE_SPACE_EXTENT;
707                         }
708                         node = rb_next(node);
709                         if (!node && cluster) {
710                                 node = rb_first(&cluster->root);
711                                 cluster = NULL;
712                         }
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                 /*
721                  * We want to add any pinned extents to our free space cache
722                  * so we don't leak the space
723                  */
724                 while (block_group && !next_page &&
725                        (start < block_group->key.objectid +
726                         block_group->key.offset)) {
727                         ret = find_first_extent_bit(unpin, start, &start, &end,
728                                                     EXTENT_DIRTY);
729                         if (ret) {
730                                 ret = 0;
731                                 break;
732                         }
733
734                         /* This pinned extent is out of our range */
735                         if (start >= block_group->key.objectid +
736                             block_group->key.offset)
737                                 break;
738
739                         len = block_group->key.objectid +
740                                 block_group->key.offset - start;
741                         len = min(len, end + 1 - start);
742
743                         entries++;
744                         entry->offset = cpu_to_le64(start);
745                         entry->bytes = cpu_to_le64(len);
746                         entry->type = BTRFS_FREE_SPACE_EXTENT;
747
748                         start = end + 1;
749                         offset += sizeof(struct btrfs_free_space_entry);
750                         if (offset + sizeof(struct btrfs_free_space_entry) >=
751                             PAGE_CACHE_SIZE)
752                                 next_page = true;
753                         entry++;
754                 }
755                 *crc = ~(u32)0;
756                 *crc = btrfs_csum_data(root, addr + start_offset, *crc,
757                                        PAGE_CACHE_SIZE - start_offset);
758                 kunmap(page);
759
760                 btrfs_csum_final(*crc, (char *)crc);
761                 crc++;
762
763                 bytes += PAGE_CACHE_SIZE;
764
765                 index++;
766         } while (node || next_page);
767
768         /* Write out the bitmaps */
769         list_for_each_safe(pos, n, &bitmap_list) {
770                 void *addr;
771                 struct btrfs_free_space *entry =
772                         list_entry(pos, struct btrfs_free_space, list);
773
774                 if (index >= num_pages) {
775                         out_of_space = true;
776                         break;
777                 }
778                 page = pages[index];
779
780                 addr = kmap(page);
781                 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
782                 *crc = ~(u32)0;
783                 *crc = btrfs_csum_data(root, addr, *crc, PAGE_CACHE_SIZE);
784                 kunmap(page);
785                 btrfs_csum_final(*crc, (char *)crc);
786                 crc++;
787                 bytes += PAGE_CACHE_SIZE;
788
789                 list_del_init(&entry->list);
790                 index++;
791         }
792
793         if (out_of_space) {
794                 btrfs_drop_pages(pages, num_pages);
795                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
796                                      i_size_read(inode) - 1, &cached_state,
797                                      GFP_NOFS);
798                 ret = 0;
799                 goto out_free;
800         }
801
802         /* Zero out the rest of the pages just to make sure */
803         while (index < num_pages) {
804                 void *addr;
805
806                 page = pages[index];
807                 addr = kmap(page);
808                 memset(addr, 0, PAGE_CACHE_SIZE);
809                 kunmap(page);
810                 bytes += PAGE_CACHE_SIZE;
811                 index++;
812         }
813
814         /* Write the checksums and trans id to the first page */
815         {
816                 void *addr;
817                 u64 *gen;
818
819                 page = pages[0];
820
821                 addr = kmap(page);
822                 memcpy(addr, checksums, sizeof(u32) * num_pages);
823                 gen = addr + (sizeof(u32) * num_pages);
824                 *gen = trans->transid;
825                 kunmap(page);
826         }
827
828         ret = btrfs_dirty_pages(root, inode, pages, num_pages, 0,
829                                             bytes, &cached_state);
830         btrfs_drop_pages(pages, num_pages);
831         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
832                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
833
834         if (ret) {
835                 ret = 0;
836                 goto out_free;
837         }
838
839         BTRFS_I(inode)->generation = trans->transid;
840
841         filemap_write_and_wait(inode->i_mapping);
842
843         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
844         key.offset = offset;
845         key.type = 0;
846
847         ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
848         if (ret < 0) {
849                 ret = -1;
850                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
851                                  EXTENT_DIRTY | EXTENT_DELALLOC |
852                                  EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
853                 goto out_free;
854         }
855         leaf = path->nodes[0];
856         if (ret > 0) {
857                 struct btrfs_key found_key;
858                 BUG_ON(!path->slots[0]);
859                 path->slots[0]--;
860                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
861                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
862                     found_key.offset != offset) {
863                         ret = -1;
864                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
865                                          EXTENT_DIRTY | EXTENT_DELALLOC |
866                                          EXTENT_DO_ACCOUNTING, 0, 0, NULL,
867                                          GFP_NOFS);
868                         btrfs_release_path(path);
869                         goto out_free;
870                 }
871         }
872         header = btrfs_item_ptr(leaf, path->slots[0],
873                                 struct btrfs_free_space_header);
874         btrfs_set_free_space_entries(leaf, header, entries);
875         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
876         btrfs_set_free_space_generation(leaf, header, trans->transid);
877         btrfs_mark_buffer_dirty(leaf);
878         btrfs_release_path(path);
879
880         ret = 1;
881
882 out_free:
883         kfree(checksums);
884         kfree(pages);
885
886 out_update:
887         if (ret != 1) {
888                 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
889                 BTRFS_I(inode)->generation = 0;
890         }
891         btrfs_update_inode(trans, root, inode);
892         return ret;
893 }
894
895 int btrfs_write_out_cache(struct btrfs_root *root,
896                           struct btrfs_trans_handle *trans,
897                           struct btrfs_block_group_cache *block_group,
898                           struct btrfs_path *path)
899 {
900         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
901         struct inode *inode;
902         int ret = 0;
903
904         root = root->fs_info->tree_root;
905
906         spin_lock(&block_group->lock);
907         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
908                 spin_unlock(&block_group->lock);
909                 return 0;
910         }
911         spin_unlock(&block_group->lock);
912
913         inode = lookup_free_space_inode(root, block_group, path);
914         if (IS_ERR(inode))
915                 return 0;
916
917         ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
918                                       path, block_group->key.objectid);
919         if (ret < 0) {
920                 spin_lock(&block_group->lock);
921                 block_group->disk_cache_state = BTRFS_DC_ERROR;
922                 spin_unlock(&block_group->lock);
923                 ret = 0;
924
925                 printk(KERN_ERR "btrfs: failed to write free space cace "
926                        "for block group %llu\n", block_group->key.objectid);
927         }
928
929         iput(inode);
930         return ret;
931 }
932
933 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
934                                           u64 offset)
935 {
936         BUG_ON(offset < bitmap_start);
937         offset -= bitmap_start;
938         return (unsigned long)(div_u64(offset, unit));
939 }
940
941 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
942 {
943         return (unsigned long)(div_u64(bytes, unit));
944 }
945
946 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
947                                    u64 offset)
948 {
949         u64 bitmap_start;
950         u64 bytes_per_bitmap;
951
952         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
953         bitmap_start = offset - ctl->start;
954         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
955         bitmap_start *= bytes_per_bitmap;
956         bitmap_start += ctl->start;
957
958         return bitmap_start;
959 }
960
961 static int tree_insert_offset(struct rb_root *root, u64 offset,
962                               struct rb_node *node, int bitmap)
963 {
964         struct rb_node **p = &root->rb_node;
965         struct rb_node *parent = NULL;
966         struct btrfs_free_space *info;
967
968         while (*p) {
969                 parent = *p;
970                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
971
972                 if (offset < info->offset) {
973                         p = &(*p)->rb_left;
974                 } else if (offset > info->offset) {
975                         p = &(*p)->rb_right;
976                 } else {
977                         /*
978                          * we could have a bitmap entry and an extent entry
979                          * share the same offset.  If this is the case, we want
980                          * the extent entry to always be found first if we do a
981                          * linear search through the tree, since we want to have
982                          * the quickest allocation time, and allocating from an
983                          * extent is faster than allocating from a bitmap.  So
984                          * if we're inserting a bitmap and we find an entry at
985                          * this offset, we want to go right, or after this entry
986                          * logically.  If we are inserting an extent and we've
987                          * found a bitmap, we want to go left, or before
988                          * logically.
989                          */
990                         if (bitmap) {
991                                 if (info->bitmap) {
992                                         WARN_ON_ONCE(1);
993                                         return -EEXIST;
994                                 }
995                                 p = &(*p)->rb_right;
996                         } else {
997                                 if (!info->bitmap) {
998                                         WARN_ON_ONCE(1);
999                                         return -EEXIST;
1000                                 }
1001                                 p = &(*p)->rb_left;
1002                         }
1003                 }
1004         }
1005
1006         rb_link_node(node, parent, p);
1007         rb_insert_color(node, root);
1008
1009         return 0;
1010 }
1011
1012 /*
1013  * searches the tree for the given offset.
1014  *
1015  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1016  * want a section that has at least bytes size and comes at or after the given
1017  * offset.
1018  */
1019 static struct btrfs_free_space *
1020 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1021                    u64 offset, int bitmap_only, int fuzzy)
1022 {
1023         struct rb_node *n = ctl->free_space_offset.rb_node;
1024         struct btrfs_free_space *entry, *prev = NULL;
1025
1026         /* find entry that is closest to the 'offset' */
1027         while (1) {
1028                 if (!n) {
1029                         entry = NULL;
1030                         break;
1031                 }
1032
1033                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1034                 prev = entry;
1035
1036                 if (offset < entry->offset)
1037                         n = n->rb_left;
1038                 else if (offset > entry->offset)
1039                         n = n->rb_right;
1040                 else
1041                         break;
1042         }
1043
1044         if (bitmap_only) {
1045                 if (!entry)
1046                         return NULL;
1047                 if (entry->bitmap)
1048                         return entry;
1049
1050                 /*
1051                  * bitmap entry and extent entry may share same offset,
1052                  * in that case, bitmap entry comes after extent entry.
1053                  */
1054                 n = rb_next(n);
1055                 if (!n)
1056                         return NULL;
1057                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1058                 if (entry->offset != offset)
1059                         return NULL;
1060
1061                 WARN_ON(!entry->bitmap);
1062                 return entry;
1063         } else if (entry) {
1064                 if (entry->bitmap) {
1065                         /*
1066                          * if previous extent entry covers the offset,
1067                          * we should return it instead of the bitmap entry
1068                          */
1069                         n = &entry->offset_index;
1070                         while (1) {
1071                                 n = rb_prev(n);
1072                                 if (!n)
1073                                         break;
1074                                 prev = rb_entry(n, struct btrfs_free_space,
1075                                                 offset_index);
1076                                 if (!prev->bitmap) {
1077                                         if (prev->offset + prev->bytes > offset)
1078                                                 entry = prev;
1079                                         break;
1080                                 }
1081                         }
1082                 }
1083                 return entry;
1084         }
1085
1086         if (!prev)
1087                 return NULL;
1088
1089         /* find last entry before the 'offset' */
1090         entry = prev;
1091         if (entry->offset > offset) {
1092                 n = rb_prev(&entry->offset_index);
1093                 if (n) {
1094                         entry = rb_entry(n, struct btrfs_free_space,
1095                                         offset_index);
1096                         BUG_ON(entry->offset > offset);
1097                 } else {
1098                         if (fuzzy)
1099                                 return entry;
1100                         else
1101                                 return NULL;
1102                 }
1103         }
1104
1105         if (entry->bitmap) {
1106                 n = &entry->offset_index;
1107                 while (1) {
1108                         n = rb_prev(n);
1109                         if (!n)
1110                                 break;
1111                         prev = rb_entry(n, struct btrfs_free_space,
1112                                         offset_index);
1113                         if (!prev->bitmap) {
1114                                 if (prev->offset + prev->bytes > offset)
1115                                         return prev;
1116                                 break;
1117                         }
1118                 }
1119                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1120                         return entry;
1121         } else if (entry->offset + entry->bytes > offset)
1122                 return entry;
1123
1124         if (!fuzzy)
1125                 return NULL;
1126
1127         while (1) {
1128                 if (entry->bitmap) {
1129                         if (entry->offset + BITS_PER_BITMAP *
1130                             ctl->unit > offset)
1131                                 break;
1132                 } else {
1133                         if (entry->offset + entry->bytes > offset)
1134                                 break;
1135                 }
1136
1137                 n = rb_next(&entry->offset_index);
1138                 if (!n)
1139                         return NULL;
1140                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1141         }
1142         return entry;
1143 }
1144
1145 static inline void
1146 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1147                     struct btrfs_free_space *info)
1148 {
1149         rb_erase(&info->offset_index, &ctl->free_space_offset);
1150         ctl->free_extents--;
1151 }
1152
1153 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1154                               struct btrfs_free_space *info)
1155 {
1156         __unlink_free_space(ctl, info);
1157         ctl->free_space -= info->bytes;
1158 }
1159
1160 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1161                            struct btrfs_free_space *info)
1162 {
1163         int ret = 0;
1164
1165         BUG_ON(!info->bitmap && !info->bytes);
1166         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1167                                  &info->offset_index, (info->bitmap != NULL));
1168         if (ret)
1169                 return ret;
1170
1171         ctl->free_space += info->bytes;
1172         ctl->free_extents++;
1173         return ret;
1174 }
1175
1176 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1177 {
1178         struct btrfs_block_group_cache *block_group = ctl->private;
1179         u64 max_bytes;
1180         u64 bitmap_bytes;
1181         u64 extent_bytes;
1182         u64 size = block_group->key.offset;
1183         u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1184         int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1185
1186         BUG_ON(ctl->total_bitmaps > max_bitmaps);
1187
1188         /*
1189          * The goal is to keep the total amount of memory used per 1gb of space
1190          * at or below 32k, so we need to adjust how much memory we allow to be
1191          * used by extent based free space tracking
1192          */
1193         if (size < 1024 * 1024 * 1024)
1194                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1195         else
1196                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1197                         div64_u64(size, 1024 * 1024 * 1024);
1198
1199         /*
1200          * we want to account for 1 more bitmap than what we have so we can make
1201          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1202          * we add more bitmaps.
1203          */
1204         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1205
1206         if (bitmap_bytes >= max_bytes) {
1207                 ctl->extents_thresh = 0;
1208                 return;
1209         }
1210
1211         /*
1212          * we want the extent entry threshold to always be at most 1/2 the maxw
1213          * bytes we can have, or whatever is less than that.
1214          */
1215         extent_bytes = max_bytes - bitmap_bytes;
1216         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1217
1218         ctl->extents_thresh =
1219                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1220 }
1221
1222 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1223                               struct btrfs_free_space *info, u64 offset,
1224                               u64 bytes)
1225 {
1226         unsigned long start, count;
1227
1228         start = offset_to_bit(info->offset, ctl->unit, offset);
1229         count = bytes_to_bits(bytes, ctl->unit);
1230         BUG_ON(start + count > BITS_PER_BITMAP);
1231
1232         bitmap_clear(info->bitmap, start, count);
1233
1234         info->bytes -= bytes;
1235         ctl->free_space -= bytes;
1236 }
1237
1238 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1239                             struct btrfs_free_space *info, u64 offset,
1240                             u64 bytes)
1241 {
1242         unsigned long start, count;
1243
1244         start = offset_to_bit(info->offset, ctl->unit, offset);
1245         count = bytes_to_bits(bytes, ctl->unit);
1246         BUG_ON(start + count > BITS_PER_BITMAP);
1247
1248         bitmap_set(info->bitmap, start, count);
1249
1250         info->bytes += bytes;
1251         ctl->free_space += bytes;
1252 }
1253
1254 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1255                          struct btrfs_free_space *bitmap_info, u64 *offset,
1256                          u64 *bytes)
1257 {
1258         unsigned long found_bits = 0;
1259         unsigned long bits, i;
1260         unsigned long next_zero;
1261
1262         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1263                           max_t(u64, *offset, bitmap_info->offset));
1264         bits = bytes_to_bits(*bytes, ctl->unit);
1265
1266         for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1267              i < BITS_PER_BITMAP;
1268              i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1269                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1270                                                BITS_PER_BITMAP, i);
1271                 if ((next_zero - i) >= bits) {
1272                         found_bits = next_zero - i;
1273                         break;
1274                 }
1275                 i = next_zero;
1276         }
1277
1278         if (found_bits) {
1279                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1280                 *bytes = (u64)(found_bits) * ctl->unit;
1281                 return 0;
1282         }
1283
1284         return -1;
1285 }
1286
1287 static struct btrfs_free_space *
1288 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
1289 {
1290         struct btrfs_free_space *entry;
1291         struct rb_node *node;
1292         int ret;
1293
1294         if (!ctl->free_space_offset.rb_node)
1295                 return NULL;
1296
1297         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1298         if (!entry)
1299                 return NULL;
1300
1301         for (node = &entry->offset_index; node; node = rb_next(node)) {
1302                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1303                 if (entry->bytes < *bytes)
1304                         continue;
1305
1306                 if (entry->bitmap) {
1307                         ret = search_bitmap(ctl, entry, offset, bytes);
1308                         if (!ret)
1309                                 return entry;
1310                         continue;
1311                 }
1312
1313                 *offset = entry->offset;
1314                 *bytes = entry->bytes;
1315                 return entry;
1316         }
1317
1318         return NULL;
1319 }
1320
1321 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1322                            struct btrfs_free_space *info, u64 offset)
1323 {
1324         info->offset = offset_to_bitmap(ctl, offset);
1325         info->bytes = 0;
1326         link_free_space(ctl, info);
1327         ctl->total_bitmaps++;
1328
1329         ctl->op->recalc_thresholds(ctl);
1330 }
1331
1332 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1333                         struct btrfs_free_space *bitmap_info)
1334 {
1335         unlink_free_space(ctl, bitmap_info);
1336         kfree(bitmap_info->bitmap);
1337         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1338         ctl->total_bitmaps--;
1339         ctl->op->recalc_thresholds(ctl);
1340 }
1341
1342 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1343                               struct btrfs_free_space *bitmap_info,
1344                               u64 *offset, u64 *bytes)
1345 {
1346         u64 end;
1347         u64 search_start, search_bytes;
1348         int ret;
1349
1350 again:
1351         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1352
1353         /*
1354          * XXX - this can go away after a few releases.
1355          *
1356          * since the only user of btrfs_remove_free_space is the tree logging
1357          * stuff, and the only way to test that is under crash conditions, we
1358          * want to have this debug stuff here just in case somethings not
1359          * working.  Search the bitmap for the space we are trying to use to
1360          * make sure its actually there.  If its not there then we need to stop
1361          * because something has gone wrong.
1362          */
1363         search_start = *offset;
1364         search_bytes = *bytes;
1365         search_bytes = min(search_bytes, end - search_start + 1);
1366         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1367         BUG_ON(ret < 0 || search_start != *offset);
1368
1369         if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1370                 bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
1371                 *bytes -= end - *offset + 1;
1372                 *offset = end + 1;
1373         } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1374                 bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
1375                 *bytes = 0;
1376         }
1377
1378         if (*bytes) {
1379                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1380                 if (!bitmap_info->bytes)
1381                         free_bitmap(ctl, bitmap_info);
1382
1383                 /*
1384                  * no entry after this bitmap, but we still have bytes to
1385                  * remove, so something has gone wrong.
1386                  */
1387                 if (!next)
1388                         return -EINVAL;
1389
1390                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1391                                        offset_index);
1392
1393                 /*
1394                  * if the next entry isn't a bitmap we need to return to let the
1395                  * extent stuff do its work.
1396                  */
1397                 if (!bitmap_info->bitmap)
1398                         return -EAGAIN;
1399
1400                 /*
1401                  * Ok the next item is a bitmap, but it may not actually hold
1402                  * the information for the rest of this free space stuff, so
1403                  * look for it, and if we don't find it return so we can try
1404                  * everything over again.
1405                  */
1406                 search_start = *offset;
1407                 search_bytes = *bytes;
1408                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1409                                     &search_bytes);
1410                 if (ret < 0 || search_start != *offset)
1411                         return -EAGAIN;
1412
1413                 goto again;
1414         } else if (!bitmap_info->bytes)
1415                 free_bitmap(ctl, bitmap_info);
1416
1417         return 0;
1418 }
1419
1420 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1421                       struct btrfs_free_space *info)
1422 {
1423         struct btrfs_block_group_cache *block_group = ctl->private;
1424
1425         /*
1426          * If we are below the extents threshold then we can add this as an
1427          * extent, and don't have to deal with the bitmap
1428          */
1429         if (ctl->free_extents < ctl->extents_thresh) {
1430                 /*
1431                  * If this block group has some small extents we don't want to
1432                  * use up all of our free slots in the cache with them, we want
1433                  * to reserve them to larger extents, however if we have plent
1434                  * of cache left then go ahead an dadd them, no sense in adding
1435                  * the overhead of a bitmap if we don't have to.
1436                  */
1437                 if (info->bytes <= block_group->sectorsize * 4) {
1438                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1439                                 return false;
1440                 } else {
1441                         return false;
1442                 }
1443         }
1444
1445         /*
1446          * some block groups are so tiny they can't be enveloped by a bitmap, so
1447          * don't even bother to create a bitmap for this
1448          */
1449         if (BITS_PER_BITMAP * block_group->sectorsize >
1450             block_group->key.offset)
1451                 return false;
1452
1453         return true;
1454 }
1455
1456 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1457                               struct btrfs_free_space *info)
1458 {
1459         struct btrfs_free_space *bitmap_info;
1460         int added = 0;
1461         u64 bytes, offset, end;
1462         int ret;
1463
1464         bytes = info->bytes;
1465         offset = info->offset;
1466
1467         if (!ctl->op->use_bitmap(ctl, info))
1468                 return 0;
1469
1470 again:
1471         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1472                                          1, 0);
1473         if (!bitmap_info) {
1474                 BUG_ON(added);
1475                 goto new_bitmap;
1476         }
1477
1478         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1479
1480         if (offset >= bitmap_info->offset && offset + bytes > end) {
1481                 bitmap_set_bits(ctl, bitmap_info, offset, end - offset);
1482                 bytes -= end - offset;
1483                 offset = end;
1484                 added = 0;
1485         } else if (offset >= bitmap_info->offset && offset + bytes <= end) {
1486                 bitmap_set_bits(ctl, bitmap_info, offset, bytes);
1487                 bytes = 0;
1488         } else {
1489                 BUG();
1490         }
1491
1492         if (!bytes) {
1493                 ret = 1;
1494                 goto out;
1495         } else
1496                 goto again;
1497
1498 new_bitmap:
1499         if (info && info->bitmap) {
1500                 add_new_bitmap(ctl, info, offset);
1501                 added = 1;
1502                 info = NULL;
1503                 goto again;
1504         } else {
1505                 spin_unlock(&ctl->tree_lock);
1506
1507                 /* no pre-allocated info, allocate a new one */
1508                 if (!info) {
1509                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
1510                                                  GFP_NOFS);
1511                         if (!info) {
1512                                 spin_lock(&ctl->tree_lock);
1513                                 ret = -ENOMEM;
1514                                 goto out;
1515                         }
1516                 }
1517
1518                 /* allocate the bitmap */
1519                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1520                 spin_lock(&ctl->tree_lock);
1521                 if (!info->bitmap) {
1522                         ret = -ENOMEM;
1523                         goto out;
1524                 }
1525                 goto again;
1526         }
1527
1528 out:
1529         if (info) {
1530                 if (info->bitmap)
1531                         kfree(info->bitmap);
1532                 kmem_cache_free(btrfs_free_space_cachep, info);
1533         }
1534
1535         return ret;
1536 }
1537
1538 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1539                           struct btrfs_free_space *info, bool update_stat)
1540 {
1541         struct btrfs_free_space *left_info;
1542         struct btrfs_free_space *right_info;
1543         bool merged = false;
1544         u64 offset = info->offset;
1545         u64 bytes = info->bytes;
1546
1547         /*
1548          * first we want to see if there is free space adjacent to the range we
1549          * are adding, if there is remove that struct and add a new one to
1550          * cover the entire range
1551          */
1552         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1553         if (right_info && rb_prev(&right_info->offset_index))
1554                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1555                                      struct btrfs_free_space, offset_index);
1556         else
1557                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1558
1559         if (right_info && !right_info->bitmap) {
1560                 if (update_stat)
1561                         unlink_free_space(ctl, right_info);
1562                 else
1563                         __unlink_free_space(ctl, right_info);
1564                 info->bytes += right_info->bytes;
1565                 kmem_cache_free(btrfs_free_space_cachep, right_info);
1566                 merged = true;
1567         }
1568
1569         if (left_info && !left_info->bitmap &&
1570             left_info->offset + left_info->bytes == offset) {
1571                 if (update_stat)
1572                         unlink_free_space(ctl, left_info);
1573                 else
1574                         __unlink_free_space(ctl, left_info);
1575                 info->offset = left_info->offset;
1576                 info->bytes += left_info->bytes;
1577                 kmem_cache_free(btrfs_free_space_cachep, left_info);
1578                 merged = true;
1579         }
1580
1581         return merged;
1582 }
1583
1584 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1585                            u64 offset, u64 bytes)
1586 {
1587         struct btrfs_free_space *info;
1588         int ret = 0;
1589
1590         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1591         if (!info)
1592                 return -ENOMEM;
1593
1594         info->offset = offset;
1595         info->bytes = bytes;
1596
1597         spin_lock(&ctl->tree_lock);
1598
1599         if (try_merge_free_space(ctl, info, true))
1600                 goto link;
1601
1602         /*
1603          * There was no extent directly to the left or right of this new
1604          * extent then we know we're going to have to allocate a new extent, so
1605          * before we do that see if we need to drop this into a bitmap
1606          */
1607         ret = insert_into_bitmap(ctl, info);
1608         if (ret < 0) {
1609                 goto out;
1610         } else if (ret) {
1611                 ret = 0;
1612                 goto out;
1613         }
1614 link:
1615         ret = link_free_space(ctl, info);
1616         if (ret)
1617                 kmem_cache_free(btrfs_free_space_cachep, info);
1618 out:
1619         spin_unlock(&ctl->tree_lock);
1620
1621         if (ret) {
1622                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1623                 BUG_ON(ret == -EEXIST);
1624         }
1625
1626         return ret;
1627 }
1628
1629 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1630                             u64 offset, u64 bytes)
1631 {
1632         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1633         struct btrfs_free_space *info;
1634         struct btrfs_free_space *next_info = NULL;
1635         int ret = 0;
1636
1637         spin_lock(&ctl->tree_lock);
1638
1639 again:
1640         info = tree_search_offset(ctl, offset, 0, 0);
1641         if (!info) {
1642                 /*
1643                  * oops didn't find an extent that matched the space we wanted
1644                  * to remove, look for a bitmap instead
1645                  */
1646                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1647                                           1, 0);
1648                 if (!info) {
1649                         WARN_ON(1);
1650                         goto out_lock;
1651                 }
1652         }
1653
1654         if (info->bytes < bytes && rb_next(&info->offset_index)) {
1655                 u64 end;
1656                 next_info = rb_entry(rb_next(&info->offset_index),
1657                                              struct btrfs_free_space,
1658                                              offset_index);
1659
1660                 if (next_info->bitmap)
1661                         end = next_info->offset +
1662                               BITS_PER_BITMAP * ctl->unit - 1;
1663                 else
1664                         end = next_info->offset + next_info->bytes;
1665
1666                 if (next_info->bytes < bytes ||
1667                     next_info->offset > offset || offset > end) {
1668                         printk(KERN_CRIT "Found free space at %llu, size %llu,"
1669                               " trying to use %llu\n",
1670                               (unsigned long long)info->offset,
1671                               (unsigned long long)info->bytes,
1672                               (unsigned long long)bytes);
1673                         WARN_ON(1);
1674                         ret = -EINVAL;
1675                         goto out_lock;
1676                 }
1677
1678                 info = next_info;
1679         }
1680
1681         if (info->bytes == bytes) {
1682                 unlink_free_space(ctl, info);
1683                 if (info->bitmap) {
1684                         kfree(info->bitmap);
1685                         ctl->total_bitmaps--;
1686                 }
1687                 kmem_cache_free(btrfs_free_space_cachep, info);
1688                 goto out_lock;
1689         }
1690
1691         if (!info->bitmap && info->offset == offset) {
1692                 unlink_free_space(ctl, info);
1693                 info->offset += bytes;
1694                 info->bytes -= bytes;
1695                 link_free_space(ctl, info);
1696                 goto out_lock;
1697         }
1698
1699         if (!info->bitmap && info->offset <= offset &&
1700             info->offset + info->bytes >= offset + bytes) {
1701                 u64 old_start = info->offset;
1702                 /*
1703                  * we're freeing space in the middle of the info,
1704                  * this can happen during tree log replay
1705                  *
1706                  * first unlink the old info and then
1707                  * insert it again after the hole we're creating
1708                  */
1709                 unlink_free_space(ctl, info);
1710                 if (offset + bytes < info->offset + info->bytes) {
1711                         u64 old_end = info->offset + info->bytes;
1712
1713                         info->offset = offset + bytes;
1714                         info->bytes = old_end - info->offset;
1715                         ret = link_free_space(ctl, info);
1716                         WARN_ON(ret);
1717                         if (ret)
1718                                 goto out_lock;
1719                 } else {
1720                         /* the hole we're creating ends at the end
1721                          * of the info struct, just free the info
1722                          */
1723                         kmem_cache_free(btrfs_free_space_cachep, info);
1724                 }
1725                 spin_unlock(&ctl->tree_lock);
1726
1727                 /* step two, insert a new info struct to cover
1728                  * anything before the hole
1729                  */
1730                 ret = btrfs_add_free_space(block_group, old_start,
1731                                            offset - old_start);
1732                 WARN_ON(ret);
1733                 goto out;
1734         }
1735
1736         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1737         if (ret == -EAGAIN)
1738                 goto again;
1739         BUG_ON(ret);
1740 out_lock:
1741         spin_unlock(&ctl->tree_lock);
1742 out:
1743         return ret;
1744 }
1745
1746 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1747                            u64 bytes)
1748 {
1749         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1750         struct btrfs_free_space *info;
1751         struct rb_node *n;
1752         int count = 0;
1753
1754         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1755                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1756                 if (info->bytes >= bytes)
1757                         count++;
1758                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1759                        (unsigned long long)info->offset,
1760                        (unsigned long long)info->bytes,
1761                        (info->bitmap) ? "yes" : "no");
1762         }
1763         printk(KERN_INFO "block group has cluster?: %s\n",
1764                list_empty(&block_group->cluster_list) ? "no" : "yes");
1765         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1766                "\n", count);
1767 }
1768
1769 static struct btrfs_free_space_op free_space_op = {
1770         .recalc_thresholds      = recalculate_thresholds,
1771         .use_bitmap             = use_bitmap,
1772 };
1773
1774 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
1775 {
1776         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1777
1778         spin_lock_init(&ctl->tree_lock);
1779         ctl->unit = block_group->sectorsize;
1780         ctl->start = block_group->key.objectid;
1781         ctl->private = block_group;
1782         ctl->op = &free_space_op;
1783
1784         /*
1785          * we only want to have 32k of ram per block group for keeping
1786          * track of free space, and if we pass 1/2 of that we want to
1787          * start converting things over to using bitmaps
1788          */
1789         ctl->extents_thresh = ((1024 * 32) / 2) /
1790                                 sizeof(struct btrfs_free_space);
1791 }
1792
1793 /*
1794  * for a given cluster, put all of its extents back into the free
1795  * space cache.  If the block group passed doesn't match the block group
1796  * pointed to by the cluster, someone else raced in and freed the
1797  * cluster already.  In that case, we just return without changing anything
1798  */
1799 static int
1800 __btrfs_return_cluster_to_free_space(
1801                              struct btrfs_block_group_cache *block_group,
1802                              struct btrfs_free_cluster *cluster)
1803 {
1804         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1805         struct btrfs_free_space *entry;
1806         struct rb_node *node;
1807
1808         spin_lock(&cluster->lock);
1809         if (cluster->block_group != block_group)
1810                 goto out;
1811
1812         cluster->block_group = NULL;
1813         cluster->window_start = 0;
1814         list_del_init(&cluster->block_group_list);
1815
1816         node = rb_first(&cluster->root);
1817         while (node) {
1818                 bool bitmap;
1819
1820                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1821                 node = rb_next(&entry->offset_index);
1822                 rb_erase(&entry->offset_index, &cluster->root);
1823
1824                 bitmap = (entry->bitmap != NULL);
1825                 if (!bitmap)
1826                         try_merge_free_space(ctl, entry, false);
1827                 tree_insert_offset(&ctl->free_space_offset,
1828                                    entry->offset, &entry->offset_index, bitmap);
1829         }
1830         cluster->root = RB_ROOT;
1831
1832 out:
1833         spin_unlock(&cluster->lock);
1834         btrfs_put_block_group(block_group);
1835         return 0;
1836 }
1837
1838 void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
1839 {
1840         struct btrfs_free_space *info;
1841         struct rb_node *node;
1842
1843         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
1844                 info = rb_entry(node, struct btrfs_free_space, offset_index);
1845                 unlink_free_space(ctl, info);
1846                 kfree(info->bitmap);
1847                 kmem_cache_free(btrfs_free_space_cachep, info);
1848                 if (need_resched()) {
1849                         spin_unlock(&ctl->tree_lock);
1850                         cond_resched();
1851                         spin_lock(&ctl->tree_lock);
1852                 }
1853         }
1854 }
1855
1856 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
1857 {
1858         spin_lock(&ctl->tree_lock);
1859         __btrfs_remove_free_space_cache_locked(ctl);
1860         spin_unlock(&ctl->tree_lock);
1861 }
1862
1863 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1864 {
1865         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1866         struct btrfs_free_cluster *cluster;
1867         struct list_head *head;
1868
1869         spin_lock(&ctl->tree_lock);
1870         while ((head = block_group->cluster_list.next) !=
1871                &block_group->cluster_list) {
1872                 cluster = list_entry(head, struct btrfs_free_cluster,
1873                                      block_group_list);
1874
1875                 WARN_ON(cluster->block_group != block_group);
1876                 __btrfs_return_cluster_to_free_space(block_group, cluster);
1877                 if (need_resched()) {
1878                         spin_unlock(&ctl->tree_lock);
1879                         cond_resched();
1880                         spin_lock(&ctl->tree_lock);
1881                 }
1882         }
1883         __btrfs_remove_free_space_cache_locked(ctl);
1884         spin_unlock(&ctl->tree_lock);
1885
1886 }
1887
1888 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1889                                u64 offset, u64 bytes, u64 empty_size)
1890 {
1891         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1892         struct btrfs_free_space *entry = NULL;
1893         u64 bytes_search = bytes + empty_size;
1894         u64 ret = 0;
1895
1896         spin_lock(&ctl->tree_lock);
1897         entry = find_free_space(ctl, &offset, &bytes_search);
1898         if (!entry)
1899                 goto out;
1900
1901         ret = offset;
1902         if (entry->bitmap) {
1903                 bitmap_clear_bits(ctl, entry, offset, bytes);
1904                 if (!entry->bytes)
1905                         free_bitmap(ctl, entry);
1906         } else {
1907                 unlink_free_space(ctl, entry);
1908                 entry->offset += bytes;
1909                 entry->bytes -= bytes;
1910                 if (!entry->bytes)
1911                         kmem_cache_free(btrfs_free_space_cachep, entry);
1912                 else
1913                         link_free_space(ctl, entry);
1914         }
1915
1916 out:
1917         spin_unlock(&ctl->tree_lock);
1918
1919         return ret;
1920 }
1921
1922 /*
1923  * given a cluster, put all of its extents back into the free space
1924  * cache.  If a block group is passed, this function will only free
1925  * a cluster that belongs to the passed block group.
1926  *
1927  * Otherwise, it'll get a reference on the block group pointed to by the
1928  * cluster and remove the cluster from it.
1929  */
1930 int btrfs_return_cluster_to_free_space(
1931                                struct btrfs_block_group_cache *block_group,
1932                                struct btrfs_free_cluster *cluster)
1933 {
1934         struct btrfs_free_space_ctl *ctl;
1935         int ret;
1936
1937         /* first, get a safe pointer to the block group */
1938         spin_lock(&cluster->lock);
1939         if (!block_group) {
1940                 block_group = cluster->block_group;
1941                 if (!block_group) {
1942                         spin_unlock(&cluster->lock);
1943                         return 0;
1944                 }
1945         } else if (cluster->block_group != block_group) {
1946                 /* someone else has already freed it don't redo their work */
1947                 spin_unlock(&cluster->lock);
1948                 return 0;
1949         }
1950         atomic_inc(&block_group->count);
1951         spin_unlock(&cluster->lock);
1952
1953         ctl = block_group->free_space_ctl;
1954
1955         /* now return any extents the cluster had on it */
1956         spin_lock(&ctl->tree_lock);
1957         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1958         spin_unlock(&ctl->tree_lock);
1959
1960         /* finally drop our ref */
1961         btrfs_put_block_group(block_group);
1962         return ret;
1963 }
1964
1965 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1966                                    struct btrfs_free_cluster *cluster,
1967                                    struct btrfs_free_space *entry,
1968                                    u64 bytes, u64 min_start)
1969 {
1970         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1971         int err;
1972         u64 search_start = cluster->window_start;
1973         u64 search_bytes = bytes;
1974         u64 ret = 0;
1975
1976         search_start = min_start;
1977         search_bytes = bytes;
1978
1979         err = search_bitmap(ctl, entry, &search_start, &search_bytes);
1980         if (err)
1981                 return 0;
1982
1983         ret = search_start;
1984         bitmap_clear_bits(ctl, entry, ret, bytes);
1985
1986         return ret;
1987 }
1988
1989 /*
1990  * given a cluster, try to allocate 'bytes' from it, returns 0
1991  * if it couldn't find anything suitably large, or a logical disk offset
1992  * if things worked out
1993  */
1994 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
1995                              struct btrfs_free_cluster *cluster, u64 bytes,
1996                              u64 min_start)
1997 {
1998         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1999         struct btrfs_free_space *entry = NULL;
2000         struct rb_node *node;
2001         u64 ret = 0;
2002
2003         spin_lock(&cluster->lock);
2004         if (bytes > cluster->max_size)
2005                 goto out;
2006
2007         if (cluster->block_group != block_group)
2008                 goto out;
2009
2010         node = rb_first(&cluster->root);
2011         if (!node)
2012                 goto out;
2013
2014         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2015         while(1) {
2016                 if (entry->bytes < bytes ||
2017                     (!entry->bitmap && entry->offset < min_start)) {
2018                         node = rb_next(&entry->offset_index);
2019                         if (!node)
2020                                 break;
2021                         entry = rb_entry(node, struct btrfs_free_space,
2022                                          offset_index);
2023                         continue;
2024                 }
2025
2026                 if (entry->bitmap) {
2027                         ret = btrfs_alloc_from_bitmap(block_group,
2028                                                       cluster, entry, bytes,
2029                                                       min_start);
2030                         if (ret == 0) {
2031                                 node = rb_next(&entry->offset_index);
2032                                 if (!node)
2033                                         break;
2034                                 entry = rb_entry(node, struct btrfs_free_space,
2035                                                  offset_index);
2036                                 continue;
2037                         }
2038                 } else {
2039
2040                         ret = entry->offset;
2041
2042                         entry->offset += bytes;
2043                         entry->bytes -= bytes;
2044                 }
2045
2046                 if (entry->bytes == 0)
2047                         rb_erase(&entry->offset_index, &cluster->root);
2048                 break;
2049         }
2050 out:
2051         spin_unlock(&cluster->lock);
2052
2053         if (!ret)
2054                 return 0;
2055
2056         spin_lock(&ctl->tree_lock);
2057
2058         ctl->free_space -= bytes;
2059         if (entry->bytes == 0) {
2060                 ctl->free_extents--;
2061                 if (entry->bitmap) {
2062                         kfree(entry->bitmap);
2063                         ctl->total_bitmaps--;
2064                         ctl->op->recalc_thresholds(ctl);
2065                 }
2066                 kmem_cache_free(btrfs_free_space_cachep, entry);
2067         }
2068
2069         spin_unlock(&ctl->tree_lock);
2070
2071         return ret;
2072 }
2073
2074 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2075                                 struct btrfs_free_space *entry,
2076                                 struct btrfs_free_cluster *cluster,
2077                                 u64 offset, u64 bytes, u64 min_bytes)
2078 {
2079         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2080         unsigned long next_zero;
2081         unsigned long i;
2082         unsigned long search_bits;
2083         unsigned long total_bits;
2084         unsigned long found_bits;
2085         unsigned long start = 0;
2086         unsigned long total_found = 0;
2087         int ret;
2088         bool found = false;
2089
2090         i = offset_to_bit(entry->offset, block_group->sectorsize,
2091                           max_t(u64, offset, entry->offset));
2092         search_bits = bytes_to_bits(bytes, block_group->sectorsize);
2093         total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
2094
2095 again:
2096         found_bits = 0;
2097         for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
2098              i < BITS_PER_BITMAP;
2099              i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
2100                 next_zero = find_next_zero_bit(entry->bitmap,
2101                                                BITS_PER_BITMAP, i);
2102                 if (next_zero - i >= search_bits) {
2103                         found_bits = next_zero - i;
2104                         break;
2105                 }
2106                 i = next_zero;
2107         }
2108
2109         if (!found_bits)
2110                 return -ENOSPC;
2111
2112         if (!found) {
2113                 start = i;
2114                 found = true;
2115         }
2116
2117         total_found += found_bits;
2118
2119         if (cluster->max_size < found_bits * block_group->sectorsize)
2120                 cluster->max_size = found_bits * block_group->sectorsize;
2121
2122         if (total_found < total_bits) {
2123                 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2124                 if (i - start > total_bits * 2) {
2125                         total_found = 0;
2126                         cluster->max_size = 0;
2127                         found = false;
2128                 }
2129                 goto again;
2130         }
2131
2132         cluster->window_start = start * block_group->sectorsize +
2133                 entry->offset;
2134         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2135         ret = tree_insert_offset(&cluster->root, entry->offset,
2136                                  &entry->offset_index, 1);
2137         BUG_ON(ret);
2138
2139         return 0;
2140 }
2141
2142 /*
2143  * This searches the block group for just extents to fill the cluster with.
2144  */
2145 static int setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2146                                    struct btrfs_free_cluster *cluster,
2147                                    u64 offset, u64 bytes, u64 min_bytes)
2148 {
2149         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2150         struct btrfs_free_space *first = NULL;
2151         struct btrfs_free_space *entry = NULL;
2152         struct btrfs_free_space *prev = NULL;
2153         struct btrfs_free_space *last;
2154         struct rb_node *node;
2155         u64 window_start;
2156         u64 window_free;
2157         u64 max_extent;
2158         u64 max_gap = 128 * 1024;
2159
2160         entry = tree_search_offset(ctl, offset, 0, 1);
2161         if (!entry)
2162                 return -ENOSPC;
2163
2164         /*
2165          * We don't want bitmaps, so just move along until we find a normal
2166          * extent entry.
2167          */
2168         while (entry->bitmap) {
2169                 node = rb_next(&entry->offset_index);
2170                 if (!node)
2171                         return -ENOSPC;
2172                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2173         }
2174
2175         window_start = entry->offset;
2176         window_free = entry->bytes;
2177         max_extent = entry->bytes;
2178         first = entry;
2179         last = entry;
2180         prev = entry;
2181
2182         while (window_free <= min_bytes) {
2183                 node = rb_next(&entry->offset_index);
2184                 if (!node)
2185                         return -ENOSPC;
2186                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2187
2188                 if (entry->bitmap)
2189                         continue;
2190                 /*
2191                  * we haven't filled the empty size and the window is
2192                  * very large.  reset and try again
2193                  */
2194                 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2195                     entry->offset - window_start > (min_bytes * 2)) {
2196                         first = entry;
2197                         window_start = entry->offset;
2198                         window_free = entry->bytes;
2199                         last = entry;
2200                         max_extent = entry->bytes;
2201                 } else {
2202                         last = entry;
2203                         window_free += entry->bytes;
2204                         if (entry->bytes > max_extent)
2205                                 max_extent = entry->bytes;
2206                 }
2207                 prev = entry;
2208         }
2209
2210         cluster->window_start = first->offset;
2211
2212         node = &first->offset_index;
2213
2214         /*
2215          * now we've found our entries, pull them out of the free space
2216          * cache and put them into the cluster rbtree
2217          */
2218         do {
2219                 int ret;
2220
2221                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2222                 node = rb_next(&entry->offset_index);
2223                 if (entry->bitmap)
2224                         continue;
2225
2226                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2227                 ret = tree_insert_offset(&cluster->root, entry->offset,
2228                                          &entry->offset_index, 0);
2229                 BUG_ON(ret);
2230         } while (node && entry != last);
2231
2232         cluster->max_size = max_extent;
2233
2234         return 0;
2235 }
2236
2237 /*
2238  * This specifically looks for bitmaps that may work in the cluster, we assume
2239  * that we have already failed to find extents that will work.
2240  */
2241 static int setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2242                                 struct btrfs_free_cluster *cluster,
2243                                 u64 offset, u64 bytes, u64 min_bytes)
2244 {
2245         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2246         struct btrfs_free_space *entry;
2247         struct rb_node *node;
2248         int ret = -ENOSPC;
2249
2250         if (ctl->total_bitmaps == 0)
2251                 return -ENOSPC;
2252
2253         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, offset), 0, 1);
2254         if (!entry)
2255                 return -ENOSPC;
2256
2257         node = &entry->offset_index;
2258         do {
2259                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2260                 node = rb_next(&entry->offset_index);
2261                 if (!entry->bitmap)
2262                         continue;
2263                 if (entry->bytes < min_bytes)
2264                         continue;
2265                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2266                                            bytes, min_bytes);
2267         } while (ret && node);
2268
2269         return ret;
2270 }
2271
2272 /*
2273  * here we try to find a cluster of blocks in a block group.  The goal
2274  * is to find at least bytes free and up to empty_size + bytes free.
2275  * We might not find them all in one contiguous area.
2276  *
2277  * returns zero and sets up cluster if things worked out, otherwise
2278  * it returns -enospc
2279  */
2280 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2281                              struct btrfs_root *root,
2282                              struct btrfs_block_group_cache *block_group,
2283                              struct btrfs_free_cluster *cluster,
2284                              u64 offset, u64 bytes, u64 empty_size)
2285 {
2286         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2287         u64 min_bytes;
2288         int ret;
2289
2290         /* for metadata, allow allocates with more holes */
2291         if (btrfs_test_opt(root, SSD_SPREAD)) {
2292                 min_bytes = bytes + empty_size;
2293         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2294                 /*
2295                  * we want to do larger allocations when we are
2296                  * flushing out the delayed refs, it helps prevent
2297                  * making more work as we go along.
2298                  */
2299                 if (trans->transaction->delayed_refs.flushing)
2300                         min_bytes = max(bytes, (bytes + empty_size) >> 1);
2301                 else
2302                         min_bytes = max(bytes, (bytes + empty_size) >> 4);
2303         } else
2304                 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2305
2306         spin_lock(&ctl->tree_lock);
2307
2308         /*
2309          * If we know we don't have enough space to make a cluster don't even
2310          * bother doing all the work to try and find one.
2311          */
2312         if (ctl->free_space < min_bytes) {
2313                 spin_unlock(&ctl->tree_lock);
2314                 return -ENOSPC;
2315         }
2316
2317         spin_lock(&cluster->lock);
2318
2319         /* someone already found a cluster, hooray */
2320         if (cluster->block_group) {
2321                 ret = 0;
2322                 goto out;
2323         }
2324
2325         ret = setup_cluster_no_bitmap(block_group, cluster, offset, bytes,
2326                                       min_bytes);
2327         if (ret)
2328                 ret = setup_cluster_bitmap(block_group, cluster, offset,
2329                                            bytes, min_bytes);
2330
2331         if (!ret) {
2332                 atomic_inc(&block_group->count);
2333                 list_add_tail(&cluster->block_group_list,
2334                               &block_group->cluster_list);
2335                 cluster->block_group = block_group;
2336         }
2337 out:
2338         spin_unlock(&cluster->lock);
2339         spin_unlock(&ctl->tree_lock);
2340
2341         return ret;
2342 }
2343
2344 /*
2345  * simple code to zero out a cluster
2346  */
2347 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2348 {
2349         spin_lock_init(&cluster->lock);
2350         spin_lock_init(&cluster->refill_lock);
2351         cluster->root = RB_ROOT;
2352         cluster->max_size = 0;
2353         INIT_LIST_HEAD(&cluster->block_group_list);
2354         cluster->block_group = NULL;
2355 }
2356
2357 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2358                            u64 *trimmed, u64 start, u64 end, u64 minlen)
2359 {
2360         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2361         struct btrfs_free_space *entry = NULL;
2362         struct btrfs_fs_info *fs_info = block_group->fs_info;
2363         u64 bytes = 0;
2364         u64 actually_trimmed;
2365         int ret = 0;
2366
2367         *trimmed = 0;
2368
2369         while (start < end) {
2370                 spin_lock(&ctl->tree_lock);
2371
2372                 if (ctl->free_space < minlen) {
2373                         spin_unlock(&ctl->tree_lock);
2374                         break;
2375                 }
2376
2377                 entry = tree_search_offset(ctl, start, 0, 1);
2378                 if (!entry)
2379                         entry = tree_search_offset(ctl,
2380                                                    offset_to_bitmap(ctl, start),
2381                                                    1, 1);
2382
2383                 if (!entry || entry->offset >= end) {
2384                         spin_unlock(&ctl->tree_lock);
2385                         break;
2386                 }
2387
2388                 if (entry->bitmap) {
2389                         ret = search_bitmap(ctl, entry, &start, &bytes);
2390                         if (!ret) {
2391                                 if (start >= end) {
2392                                         spin_unlock(&ctl->tree_lock);
2393                                         break;
2394                                 }
2395                                 bytes = min(bytes, end - start);
2396                                 bitmap_clear_bits(ctl, entry, start, bytes);
2397                                 if (entry->bytes == 0)
2398                                         free_bitmap(ctl, entry);
2399                         } else {
2400                                 start = entry->offset + BITS_PER_BITMAP *
2401                                         block_group->sectorsize;
2402                                 spin_unlock(&ctl->tree_lock);
2403                                 ret = 0;
2404                                 continue;
2405                         }
2406                 } else {
2407                         start = entry->offset;
2408                         bytes = min(entry->bytes, end - start);
2409                         unlink_free_space(ctl, entry);
2410                         kmem_cache_free(btrfs_free_space_cachep, entry);
2411                 }
2412
2413                 spin_unlock(&ctl->tree_lock);
2414
2415                 if (bytes >= minlen) {
2416                         int update_ret;
2417                         update_ret = btrfs_update_reserved_bytes(block_group,
2418                                                                  bytes, 1, 1);
2419
2420                         ret = btrfs_error_discard_extent(fs_info->extent_root,
2421                                                          start,
2422                                                          bytes,
2423                                                          &actually_trimmed);
2424
2425                         btrfs_add_free_space(block_group, start, bytes);
2426                         if (!update_ret)
2427                                 btrfs_update_reserved_bytes(block_group,
2428                                                             bytes, 0, 1);
2429
2430                         if (ret)
2431                                 break;
2432                         *trimmed += actually_trimmed;
2433                 }
2434                 start += bytes;
2435                 bytes = 0;
2436
2437                 if (fatal_signal_pending(current)) {
2438                         ret = -ERESTARTSYS;
2439                         break;
2440                 }
2441
2442                 cond_resched();
2443         }
2444
2445         return ret;
2446 }
2447
2448 /*
2449  * Find the left-most item in the cache tree, and then return the
2450  * smallest inode number in the item.
2451  *
2452  * Note: the returned inode number may not be the smallest one in
2453  * the tree, if the left-most item is a bitmap.
2454  */
2455 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2456 {
2457         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2458         struct btrfs_free_space *entry = NULL;
2459         u64 ino = 0;
2460
2461         spin_lock(&ctl->tree_lock);
2462
2463         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2464                 goto out;
2465
2466         entry = rb_entry(rb_first(&ctl->free_space_offset),
2467                          struct btrfs_free_space, offset_index);
2468
2469         if (!entry->bitmap) {
2470                 ino = entry->offset;
2471
2472                 unlink_free_space(ctl, entry);
2473                 entry->offset++;
2474                 entry->bytes--;
2475                 if (!entry->bytes)
2476                         kmem_cache_free(btrfs_free_space_cachep, entry);
2477                 else
2478                         link_free_space(ctl, entry);
2479         } else {
2480                 u64 offset = 0;
2481                 u64 count = 1;
2482                 int ret;
2483
2484                 ret = search_bitmap(ctl, entry, &offset, &count);
2485                 BUG_ON(ret);
2486
2487                 ino = offset;
2488                 bitmap_clear_bits(ctl, entry, offset, 1);
2489                 if (entry->bytes == 0)
2490                         free_bitmap(ctl, entry);
2491         }
2492 out:
2493         spin_unlock(&ctl->tree_lock);
2494
2495         return ino;
2496 }
2497
2498 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2499                                     struct btrfs_path *path)
2500 {
2501         struct inode *inode = NULL;
2502
2503         spin_lock(&root->cache_lock);
2504         if (root->cache_inode)
2505                 inode = igrab(root->cache_inode);
2506         spin_unlock(&root->cache_lock);
2507         if (inode)
2508                 return inode;
2509
2510         inode = __lookup_free_space_inode(root, path, 0);
2511         if (IS_ERR(inode))
2512                 return inode;
2513
2514         spin_lock(&root->cache_lock);
2515         if (!btrfs_fs_closing(root->fs_info))
2516                 root->cache_inode = igrab(inode);
2517         spin_unlock(&root->cache_lock);
2518
2519         return inode;
2520 }
2521
2522 int create_free_ino_inode(struct btrfs_root *root,
2523                           struct btrfs_trans_handle *trans,
2524                           struct btrfs_path *path)
2525 {
2526         return __create_free_space_inode(root, trans, path,
2527                                          BTRFS_FREE_INO_OBJECTID, 0);
2528 }
2529
2530 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2531 {
2532         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2533         struct btrfs_path *path;
2534         struct inode *inode;
2535         int ret = 0;
2536         u64 root_gen = btrfs_root_generation(&root->root_item);
2537
2538         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2539                 return 0;
2540
2541         /*
2542          * If we're unmounting then just return, since this does a search on the
2543          * normal root and not the commit root and we could deadlock.
2544          */
2545         if (btrfs_fs_closing(fs_info))
2546                 return 0;
2547
2548         path = btrfs_alloc_path();
2549         if (!path)
2550                 return 0;
2551
2552         inode = lookup_free_ino_inode(root, path);
2553         if (IS_ERR(inode))
2554                 goto out;
2555
2556         if (root_gen != BTRFS_I(inode)->generation)
2557                 goto out_put;
2558
2559         ret = __load_free_space_cache(root, inode, ctl, path, 0);
2560
2561         if (ret < 0)
2562                 printk(KERN_ERR "btrfs: failed to load free ino cache for "
2563                        "root %llu\n", root->root_key.objectid);
2564 out_put:
2565         iput(inode);
2566 out:
2567         btrfs_free_path(path);
2568         return ret;
2569 }
2570
2571 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2572                               struct btrfs_trans_handle *trans,
2573                               struct btrfs_path *path)
2574 {
2575         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2576         struct inode *inode;
2577         int ret;
2578
2579         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2580                 return 0;
2581
2582         inode = lookup_free_ino_inode(root, path);
2583         if (IS_ERR(inode))
2584                 return 0;
2585
2586         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2587         if (ret < 0)
2588                 printk(KERN_ERR "btrfs: failed to write free ino cache "
2589                        "for root %llu\n", root->root_key.objectid);
2590
2591         iput(inode);
2592         return ret;
2593 }