GNU Linux-libre 4.4.297-gnu1
[releases.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 <linux/ratelimit.h>
24 #include "ctree.h"
25 #include "free-space-cache.h"
26 #include "transaction.h"
27 #include "disk-io.h"
28 #include "extent_io.h"
29 #include "inode-map.h"
30 #include "volumes.h"
31
32 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
33 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
34
35 struct btrfs_trim_range {
36         u64 start;
37         u64 bytes;
38         struct list_head list;
39 };
40
41 static int link_free_space(struct btrfs_free_space_ctl *ctl,
42                            struct btrfs_free_space *info);
43 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
44                               struct btrfs_free_space *info);
45
46 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
47                                                struct btrfs_path *path,
48                                                u64 offset)
49 {
50         struct btrfs_key key;
51         struct btrfs_key location;
52         struct btrfs_disk_key disk_key;
53         struct btrfs_free_space_header *header;
54         struct extent_buffer *leaf;
55         struct inode *inode = NULL;
56         int ret;
57
58         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
59         key.offset = offset;
60         key.type = 0;
61
62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
63         if (ret < 0)
64                 return ERR_PTR(ret);
65         if (ret > 0) {
66                 btrfs_release_path(path);
67                 return ERR_PTR(-ENOENT);
68         }
69
70         leaf = path->nodes[0];
71         header = btrfs_item_ptr(leaf, path->slots[0],
72                                 struct btrfs_free_space_header);
73         btrfs_free_space_key(leaf, header, &disk_key);
74         btrfs_disk_key_to_cpu(&location, &disk_key);
75         btrfs_release_path(path);
76
77         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
78         if (!inode)
79                 return ERR_PTR(-ENOENT);
80         if (IS_ERR(inode))
81                 return inode;
82         if (is_bad_inode(inode)) {
83                 iput(inode);
84                 return ERR_PTR(-ENOENT);
85         }
86
87         mapping_set_gfp_mask(inode->i_mapping,
88                         mapping_gfp_constraint(inode->i_mapping,
89                         ~(__GFP_FS | __GFP_HIGHMEM)));
90
91         return inode;
92 }
93
94 struct inode *lookup_free_space_inode(struct btrfs_root *root,
95                                       struct btrfs_block_group_cache
96                                       *block_group, struct btrfs_path *path)
97 {
98         struct inode *inode = NULL;
99         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
100
101         spin_lock(&block_group->lock);
102         if (block_group->inode)
103                 inode = igrab(block_group->inode);
104         spin_unlock(&block_group->lock);
105         if (inode)
106                 return inode;
107
108         inode = __lookup_free_space_inode(root, path,
109                                           block_group->key.objectid);
110         if (IS_ERR(inode))
111                 return inode;
112
113         spin_lock(&block_group->lock);
114         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
115                 btrfs_info(root->fs_info,
116                         "Old style space inode found, converting.");
117                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
118                         BTRFS_INODE_NODATACOW;
119                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
120         }
121
122         if (!block_group->iref) {
123                 block_group->inode = igrab(inode);
124                 block_group->iref = 1;
125         }
126         spin_unlock(&block_group->lock);
127
128         return inode;
129 }
130
131 static int __create_free_space_inode(struct btrfs_root *root,
132                                      struct btrfs_trans_handle *trans,
133                                      struct btrfs_path *path,
134                                      u64 ino, u64 offset)
135 {
136         struct btrfs_key key;
137         struct btrfs_disk_key disk_key;
138         struct btrfs_free_space_header *header;
139         struct btrfs_inode_item *inode_item;
140         struct extent_buffer *leaf;
141         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
142         int ret;
143
144         ret = btrfs_insert_empty_inode(trans, root, path, ino);
145         if (ret)
146                 return ret;
147
148         /* We inline crc's for the free disk space cache */
149         if (ino != BTRFS_FREE_INO_OBJECTID)
150                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
151
152         leaf = path->nodes[0];
153         inode_item = btrfs_item_ptr(leaf, path->slots[0],
154                                     struct btrfs_inode_item);
155         btrfs_item_key(leaf, &disk_key, path->slots[0]);
156         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
157                              sizeof(*inode_item));
158         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
159         btrfs_set_inode_size(leaf, inode_item, 0);
160         btrfs_set_inode_nbytes(leaf, inode_item, 0);
161         btrfs_set_inode_uid(leaf, inode_item, 0);
162         btrfs_set_inode_gid(leaf, inode_item, 0);
163         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
164         btrfs_set_inode_flags(leaf, inode_item, flags);
165         btrfs_set_inode_nlink(leaf, inode_item, 1);
166         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
167         btrfs_set_inode_block_group(leaf, inode_item, offset);
168         btrfs_mark_buffer_dirty(leaf);
169         btrfs_release_path(path);
170
171         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
172         key.offset = offset;
173         key.type = 0;
174         ret = btrfs_insert_empty_item(trans, root, path, &key,
175                                       sizeof(struct btrfs_free_space_header));
176         if (ret < 0) {
177                 btrfs_release_path(path);
178                 return ret;
179         }
180
181         leaf = path->nodes[0];
182         header = btrfs_item_ptr(leaf, path->slots[0],
183                                 struct btrfs_free_space_header);
184         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
185         btrfs_set_free_space_key(leaf, header, &disk_key);
186         btrfs_mark_buffer_dirty(leaf);
187         btrfs_release_path(path);
188
189         return 0;
190 }
191
192 int create_free_space_inode(struct btrfs_root *root,
193                             struct btrfs_trans_handle *trans,
194                             struct btrfs_block_group_cache *block_group,
195                             struct btrfs_path *path)
196 {
197         int ret;
198         u64 ino;
199
200         ret = btrfs_find_free_objectid(root, &ino);
201         if (ret < 0)
202                 return ret;
203
204         return __create_free_space_inode(root, trans, path, ino,
205                                          block_group->key.objectid);
206 }
207
208 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
209                                        struct btrfs_block_rsv *rsv)
210 {
211         u64 needed_bytes;
212         int ret;
213
214         /* 1 for slack space, 1 for updating the inode */
215         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
216                 btrfs_calc_trans_metadata_size(root, 1);
217
218         spin_lock(&rsv->lock);
219         if (rsv->reserved < needed_bytes)
220                 ret = -ENOSPC;
221         else
222                 ret = 0;
223         spin_unlock(&rsv->lock);
224         return ret;
225 }
226
227 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
228                                     struct btrfs_trans_handle *trans,
229                                     struct btrfs_block_group_cache *block_group,
230                                     struct inode *inode)
231 {
232         int ret = 0;
233         struct btrfs_path *path = btrfs_alloc_path();
234         bool locked = false;
235
236         if (!path) {
237                 ret = -ENOMEM;
238                 goto fail;
239         }
240
241         if (block_group) {
242                 locked = true;
243                 mutex_lock(&trans->transaction->cache_write_mutex);
244                 if (!list_empty(&block_group->io_list)) {
245                         list_del_init(&block_group->io_list);
246
247                         btrfs_wait_cache_io(root, trans, block_group,
248                                             &block_group->io_ctl, path,
249                                             block_group->key.objectid);
250                         btrfs_put_block_group(block_group);
251                 }
252
253                 /*
254                  * now that we've truncated the cache away, its no longer
255                  * setup or written
256                  */
257                 spin_lock(&block_group->lock);
258                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
259                 spin_unlock(&block_group->lock);
260         }
261         btrfs_free_path(path);
262
263         btrfs_i_size_write(inode, 0);
264         truncate_pagecache(inode, 0);
265
266         /*
267          * We don't need an orphan item because truncating the free space cache
268          * will never be split across transactions.
269          * We don't need to check for -EAGAIN because we're a free space
270          * cache inode
271          */
272         ret = btrfs_truncate_inode_items(trans, root, inode,
273                                          0, BTRFS_EXTENT_DATA_KEY);
274         if (ret)
275                 goto fail;
276
277         ret = btrfs_update_inode(trans, root, inode);
278
279 fail:
280         if (locked)
281                 mutex_unlock(&trans->transaction->cache_write_mutex);
282         if (ret)
283                 btrfs_abort_transaction(trans, root, ret);
284
285         return ret;
286 }
287
288 static int readahead_cache(struct inode *inode)
289 {
290         struct file_ra_state *ra;
291         unsigned long last_index;
292
293         ra = kzalloc(sizeof(*ra), GFP_NOFS);
294         if (!ra)
295                 return -ENOMEM;
296
297         file_ra_state_init(ra, inode->i_mapping);
298         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
299
300         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
301
302         kfree(ra);
303
304         return 0;
305 }
306
307 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
308                        struct btrfs_root *root, int write)
309 {
310         int num_pages;
311         int check_crcs = 0;
312
313         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_CACHE_SIZE);
314
315         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
316                 check_crcs = 1;
317
318         /* Make sure we can fit our crcs into the first page */
319         if (write && check_crcs &&
320             (num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE)
321                 return -ENOSPC;
322
323         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
324
325         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
326         if (!io_ctl->pages)
327                 return -ENOMEM;
328
329         io_ctl->num_pages = num_pages;
330         io_ctl->root = root;
331         io_ctl->check_crcs = check_crcs;
332         io_ctl->inode = inode;
333
334         return 0;
335 }
336
337 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
338 {
339         kfree(io_ctl->pages);
340         io_ctl->pages = NULL;
341 }
342
343 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
344 {
345         if (io_ctl->cur) {
346                 io_ctl->cur = NULL;
347                 io_ctl->orig = NULL;
348         }
349 }
350
351 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
352 {
353         ASSERT(io_ctl->index < io_ctl->num_pages);
354         io_ctl->page = io_ctl->pages[io_ctl->index++];
355         io_ctl->cur = page_address(io_ctl->page);
356         io_ctl->orig = io_ctl->cur;
357         io_ctl->size = PAGE_CACHE_SIZE;
358         if (clear)
359                 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
360 }
361
362 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
363 {
364         int i;
365
366         io_ctl_unmap_page(io_ctl);
367
368         for (i = 0; i < io_ctl->num_pages; i++) {
369                 if (io_ctl->pages[i]) {
370                         ClearPageChecked(io_ctl->pages[i]);
371                         unlock_page(io_ctl->pages[i]);
372                         page_cache_release(io_ctl->pages[i]);
373                 }
374         }
375 }
376
377 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
378                                 int uptodate)
379 {
380         struct page *page;
381         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
382         int i;
383
384         for (i = 0; i < io_ctl->num_pages; i++) {
385                 page = find_or_create_page(inode->i_mapping, i, mask);
386                 if (!page) {
387                         io_ctl_drop_pages(io_ctl);
388                         return -ENOMEM;
389                 }
390                 io_ctl->pages[i] = page;
391                 if (uptodate && !PageUptodate(page)) {
392                         btrfs_readpage(NULL, page);
393                         lock_page(page);
394                         if (page->mapping != inode->i_mapping) {
395                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
396                                           "free space cache page truncated");
397                                 io_ctl_drop_pages(io_ctl);
398                                 return -EIO;
399                         }
400                         if (!PageUptodate(page)) {
401                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
402                                            "error reading free space cache");
403                                 io_ctl_drop_pages(io_ctl);
404                                 return -EIO;
405                         }
406                 }
407         }
408
409         for (i = 0; i < io_ctl->num_pages; i++) {
410                 clear_page_dirty_for_io(io_ctl->pages[i]);
411                 set_page_extent_mapped(io_ctl->pages[i]);
412         }
413
414         return 0;
415 }
416
417 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
418 {
419         __le64 *val;
420
421         io_ctl_map_page(io_ctl, 1);
422
423         /*
424          * Skip the csum areas.  If we don't check crcs then we just have a
425          * 64bit chunk at the front of the first page.
426          */
427         if (io_ctl->check_crcs) {
428                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
429                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
430         } else {
431                 io_ctl->cur += sizeof(u64);
432                 io_ctl->size -= sizeof(u64) * 2;
433         }
434
435         val = io_ctl->cur;
436         *val = cpu_to_le64(generation);
437         io_ctl->cur += sizeof(u64);
438 }
439
440 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
441 {
442         __le64 *gen;
443
444         /*
445          * Skip the crc area.  If we don't check crcs then we just have a 64bit
446          * chunk at the front of the first page.
447          */
448         if (io_ctl->check_crcs) {
449                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
450                 io_ctl->size -= sizeof(u64) +
451                         (sizeof(u32) * io_ctl->num_pages);
452         } else {
453                 io_ctl->cur += sizeof(u64);
454                 io_ctl->size -= sizeof(u64) * 2;
455         }
456
457         gen = io_ctl->cur;
458         if (le64_to_cpu(*gen) != generation) {
459                 btrfs_err_rl(io_ctl->root->fs_info,
460                         "space cache generation (%llu) does not match inode (%llu)",
461                                 *gen, generation);
462                 io_ctl_unmap_page(io_ctl);
463                 return -EIO;
464         }
465         io_ctl->cur += sizeof(u64);
466         return 0;
467 }
468
469 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
470 {
471         u32 *tmp;
472         u32 crc = ~(u32)0;
473         unsigned offset = 0;
474
475         if (!io_ctl->check_crcs) {
476                 io_ctl_unmap_page(io_ctl);
477                 return;
478         }
479
480         if (index == 0)
481                 offset = sizeof(u32) * io_ctl->num_pages;
482
483         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
484                               PAGE_CACHE_SIZE - offset);
485         btrfs_csum_final(crc, (char *)&crc);
486         io_ctl_unmap_page(io_ctl);
487         tmp = page_address(io_ctl->pages[0]);
488         tmp += index;
489         *tmp = crc;
490 }
491
492 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
493 {
494         u32 *tmp, val;
495         u32 crc = ~(u32)0;
496         unsigned offset = 0;
497
498         if (!io_ctl->check_crcs) {
499                 io_ctl_map_page(io_ctl, 0);
500                 return 0;
501         }
502
503         if (index == 0)
504                 offset = sizeof(u32) * io_ctl->num_pages;
505
506         tmp = page_address(io_ctl->pages[0]);
507         tmp += index;
508         val = *tmp;
509
510         io_ctl_map_page(io_ctl, 0);
511         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
512                               PAGE_CACHE_SIZE - offset);
513         btrfs_csum_final(crc, (char *)&crc);
514         if (val != crc) {
515                 btrfs_err_rl(io_ctl->root->fs_info,
516                         "csum mismatch on free space cache");
517                 io_ctl_unmap_page(io_ctl);
518                 return -EIO;
519         }
520
521         return 0;
522 }
523
524 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
525                             void *bitmap)
526 {
527         struct btrfs_free_space_entry *entry;
528
529         if (!io_ctl->cur)
530                 return -ENOSPC;
531
532         entry = io_ctl->cur;
533         entry->offset = cpu_to_le64(offset);
534         entry->bytes = cpu_to_le64(bytes);
535         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
536                 BTRFS_FREE_SPACE_EXTENT;
537         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
538         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
539
540         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
541                 return 0;
542
543         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
544
545         /* No more pages to map */
546         if (io_ctl->index >= io_ctl->num_pages)
547                 return 0;
548
549         /* map the next page */
550         io_ctl_map_page(io_ctl, 1);
551         return 0;
552 }
553
554 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
555 {
556         if (!io_ctl->cur)
557                 return -ENOSPC;
558
559         /*
560          * If we aren't at the start of the current page, unmap this one and
561          * map the next one if there is any left.
562          */
563         if (io_ctl->cur != io_ctl->orig) {
564                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
565                 if (io_ctl->index >= io_ctl->num_pages)
566                         return -ENOSPC;
567                 io_ctl_map_page(io_ctl, 0);
568         }
569
570         memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
571         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
572         if (io_ctl->index < io_ctl->num_pages)
573                 io_ctl_map_page(io_ctl, 0);
574         return 0;
575 }
576
577 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
578 {
579         /*
580          * If we're not on the boundary we know we've modified the page and we
581          * need to crc the page.
582          */
583         if (io_ctl->cur != io_ctl->orig)
584                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
585         else
586                 io_ctl_unmap_page(io_ctl);
587
588         while (io_ctl->index < io_ctl->num_pages) {
589                 io_ctl_map_page(io_ctl, 1);
590                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
591         }
592 }
593
594 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
595                             struct btrfs_free_space *entry, u8 *type)
596 {
597         struct btrfs_free_space_entry *e;
598         int ret;
599
600         if (!io_ctl->cur) {
601                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
602                 if (ret)
603                         return ret;
604         }
605
606         e = io_ctl->cur;
607         entry->offset = le64_to_cpu(e->offset);
608         entry->bytes = le64_to_cpu(e->bytes);
609         *type = e->type;
610         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
611         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
612
613         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
614                 return 0;
615
616         io_ctl_unmap_page(io_ctl);
617
618         return 0;
619 }
620
621 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
622                               struct btrfs_free_space *entry)
623 {
624         int ret;
625
626         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
627         if (ret)
628                 return ret;
629
630         memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
631         io_ctl_unmap_page(io_ctl);
632
633         return 0;
634 }
635
636 /*
637  * Since we attach pinned extents after the fact we can have contiguous sections
638  * of free space that are split up in entries.  This poses a problem with the
639  * tree logging stuff since it could have allocated across what appears to be 2
640  * entries since we would have merged the entries when adding the pinned extents
641  * back to the free space cache.  So run through the space cache that we just
642  * loaded and merge contiguous entries.  This will make the log replay stuff not
643  * blow up and it will make for nicer allocator behavior.
644  */
645 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
646 {
647         struct btrfs_free_space *e, *prev = NULL;
648         struct rb_node *n;
649
650 again:
651         spin_lock(&ctl->tree_lock);
652         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
653                 e = rb_entry(n, struct btrfs_free_space, offset_index);
654                 if (!prev)
655                         goto next;
656                 if (e->bitmap || prev->bitmap)
657                         goto next;
658                 if (prev->offset + prev->bytes == e->offset) {
659                         unlink_free_space(ctl, prev);
660                         unlink_free_space(ctl, e);
661                         prev->bytes += e->bytes;
662                         kmem_cache_free(btrfs_free_space_cachep, e);
663                         link_free_space(ctl, prev);
664                         prev = NULL;
665                         spin_unlock(&ctl->tree_lock);
666                         goto again;
667                 }
668 next:
669                 prev = e;
670         }
671         spin_unlock(&ctl->tree_lock);
672 }
673
674 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
675                                    struct btrfs_free_space_ctl *ctl,
676                                    struct btrfs_path *path, u64 offset)
677 {
678         struct btrfs_free_space_header *header;
679         struct extent_buffer *leaf;
680         struct btrfs_io_ctl io_ctl;
681         struct btrfs_key key;
682         struct btrfs_free_space *e, *n;
683         LIST_HEAD(bitmaps);
684         u64 num_entries;
685         u64 num_bitmaps;
686         u64 generation;
687         u8 type;
688         int ret = 0;
689
690         /* Nothing in the space cache, goodbye */
691         if (!i_size_read(inode))
692                 return 0;
693
694         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
695         key.offset = offset;
696         key.type = 0;
697
698         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
699         if (ret < 0)
700                 return 0;
701         else if (ret > 0) {
702                 btrfs_release_path(path);
703                 return 0;
704         }
705
706         ret = -1;
707
708         leaf = path->nodes[0];
709         header = btrfs_item_ptr(leaf, path->slots[0],
710                                 struct btrfs_free_space_header);
711         num_entries = btrfs_free_space_entries(leaf, header);
712         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
713         generation = btrfs_free_space_generation(leaf, header);
714         btrfs_release_path(path);
715
716         if (!BTRFS_I(inode)->generation) {
717                 btrfs_info(root->fs_info,
718                            "The free space cache file (%llu) is invalid. skip it\n",
719                            offset);
720                 return 0;
721         }
722
723         if (BTRFS_I(inode)->generation != generation) {
724                 btrfs_err(root->fs_info,
725                         "free space inode generation (%llu) "
726                         "did not match free space cache generation (%llu)",
727                         BTRFS_I(inode)->generation, generation);
728                 return 0;
729         }
730
731         if (!num_entries)
732                 return 0;
733
734         ret = io_ctl_init(&io_ctl, inode, root, 0);
735         if (ret)
736                 return ret;
737
738         ret = readahead_cache(inode);
739         if (ret)
740                 goto out;
741
742         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
743         if (ret)
744                 goto out;
745
746         ret = io_ctl_check_crc(&io_ctl, 0);
747         if (ret)
748                 goto free_cache;
749
750         ret = io_ctl_check_generation(&io_ctl, generation);
751         if (ret)
752                 goto free_cache;
753
754         while (num_entries) {
755                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
756                                       GFP_NOFS);
757                 if (!e) {
758                         ret = -ENOMEM;
759                         goto free_cache;
760                 }
761
762                 ret = io_ctl_read_entry(&io_ctl, e, &type);
763                 if (ret) {
764                         kmem_cache_free(btrfs_free_space_cachep, e);
765                         goto free_cache;
766                 }
767
768                 if (!e->bytes) {
769                         ret = -1;
770                         kmem_cache_free(btrfs_free_space_cachep, e);
771                         goto free_cache;
772                 }
773
774                 if (type == BTRFS_FREE_SPACE_EXTENT) {
775                         spin_lock(&ctl->tree_lock);
776                         ret = link_free_space(ctl, e);
777                         spin_unlock(&ctl->tree_lock);
778                         if (ret) {
779                                 btrfs_err(root->fs_info,
780                                         "Duplicate entries in free space cache, dumping");
781                                 kmem_cache_free(btrfs_free_space_cachep, e);
782                                 goto free_cache;
783                         }
784                 } else {
785                         ASSERT(num_bitmaps);
786                         num_bitmaps--;
787                         e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
788                         if (!e->bitmap) {
789                                 ret = -ENOMEM;
790                                 kmem_cache_free(
791                                         btrfs_free_space_cachep, e);
792                                 goto free_cache;
793                         }
794                         spin_lock(&ctl->tree_lock);
795                         ret = link_free_space(ctl, e);
796                         ctl->total_bitmaps++;
797                         ctl->op->recalc_thresholds(ctl);
798                         spin_unlock(&ctl->tree_lock);
799                         if (ret) {
800                                 btrfs_err(root->fs_info,
801                                         "Duplicate entries in free space cache, dumping");
802                                 kmem_cache_free(btrfs_free_space_cachep, e);
803                                 goto free_cache;
804                         }
805                         list_add_tail(&e->list, &bitmaps);
806                 }
807
808                 num_entries--;
809         }
810
811         io_ctl_unmap_page(&io_ctl);
812
813         /*
814          * We add the bitmaps at the end of the entries in order that
815          * the bitmap entries are added to the cache.
816          */
817         list_for_each_entry_safe(e, n, &bitmaps, list) {
818                 list_del_init(&e->list);
819                 ret = io_ctl_read_bitmap(&io_ctl, e);
820                 if (ret)
821                         goto free_cache;
822         }
823
824         io_ctl_drop_pages(&io_ctl);
825         merge_space_tree(ctl);
826         ret = 1;
827 out:
828         io_ctl_free(&io_ctl);
829         return ret;
830 free_cache:
831         io_ctl_drop_pages(&io_ctl);
832         __btrfs_remove_free_space_cache(ctl);
833         goto out;
834 }
835
836 int load_free_space_cache(struct btrfs_fs_info *fs_info,
837                           struct btrfs_block_group_cache *block_group)
838 {
839         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
840         struct btrfs_root *root = fs_info->tree_root;
841         struct inode *inode;
842         struct btrfs_path *path;
843         int ret = 0;
844         bool matched;
845         u64 used = btrfs_block_group_used(&block_group->item);
846
847         /*
848          * If this block group has been marked to be cleared for one reason or
849          * another then we can't trust the on disk cache, so just return.
850          */
851         spin_lock(&block_group->lock);
852         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
853                 spin_unlock(&block_group->lock);
854                 return 0;
855         }
856         spin_unlock(&block_group->lock);
857
858         path = btrfs_alloc_path();
859         if (!path)
860                 return 0;
861         path->search_commit_root = 1;
862         path->skip_locking = 1;
863
864         inode = lookup_free_space_inode(root, block_group, path);
865         if (IS_ERR(inode)) {
866                 btrfs_free_path(path);
867                 return 0;
868         }
869
870         /* We may have converted the inode and made the cache invalid. */
871         spin_lock(&block_group->lock);
872         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
873                 spin_unlock(&block_group->lock);
874                 btrfs_free_path(path);
875                 goto out;
876         }
877         spin_unlock(&block_group->lock);
878
879         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
880                                       path, block_group->key.objectid);
881         btrfs_free_path(path);
882         if (ret <= 0)
883                 goto out;
884
885         spin_lock(&ctl->tree_lock);
886         matched = (ctl->free_space == (block_group->key.offset - used -
887                                        block_group->bytes_super));
888         spin_unlock(&ctl->tree_lock);
889
890         if (!matched) {
891                 __btrfs_remove_free_space_cache(ctl);
892                 btrfs_warn(fs_info, "block group %llu has wrong amount of free space",
893                         block_group->key.objectid);
894                 ret = -1;
895         }
896 out:
897         if (ret < 0) {
898                 /* This cache is bogus, make sure it gets cleared */
899                 spin_lock(&block_group->lock);
900                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
901                 spin_unlock(&block_group->lock);
902                 ret = 0;
903
904                 btrfs_warn(fs_info, "failed to load free space cache for block group %llu, rebuilding it now",
905                         block_group->key.objectid);
906         }
907
908         iput(inode);
909         return ret;
910 }
911
912 static noinline_for_stack
913 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
914                               struct btrfs_free_space_ctl *ctl,
915                               struct btrfs_block_group_cache *block_group,
916                               int *entries, int *bitmaps,
917                               struct list_head *bitmap_list)
918 {
919         int ret;
920         struct btrfs_free_cluster *cluster = NULL;
921         struct btrfs_free_cluster *cluster_locked = NULL;
922         struct rb_node *node = rb_first(&ctl->free_space_offset);
923         struct btrfs_trim_range *trim_entry;
924
925         /* Get the cluster for this block_group if it exists */
926         if (block_group && !list_empty(&block_group->cluster_list)) {
927                 cluster = list_entry(block_group->cluster_list.next,
928                                      struct btrfs_free_cluster,
929                                      block_group_list);
930         }
931
932         if (!node && cluster) {
933                 cluster_locked = cluster;
934                 spin_lock(&cluster_locked->lock);
935                 node = rb_first(&cluster->root);
936                 cluster = NULL;
937         }
938
939         /* Write out the extent entries */
940         while (node) {
941                 struct btrfs_free_space *e;
942
943                 e = rb_entry(node, struct btrfs_free_space, offset_index);
944                 *entries += 1;
945
946                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
947                                        e->bitmap);
948                 if (ret)
949                         goto fail;
950
951                 if (e->bitmap) {
952                         list_add_tail(&e->list, bitmap_list);
953                         *bitmaps += 1;
954                 }
955                 node = rb_next(node);
956                 if (!node && cluster) {
957                         node = rb_first(&cluster->root);
958                         cluster_locked = cluster;
959                         spin_lock(&cluster_locked->lock);
960                         cluster = NULL;
961                 }
962         }
963         if (cluster_locked) {
964                 spin_unlock(&cluster_locked->lock);
965                 cluster_locked = NULL;
966         }
967
968         /*
969          * Make sure we don't miss any range that was removed from our rbtree
970          * because trimming is running. Otherwise after a umount+mount (or crash
971          * after committing the transaction) we would leak free space and get
972          * an inconsistent free space cache report from fsck.
973          */
974         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
975                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
976                                        trim_entry->bytes, NULL);
977                 if (ret)
978                         goto fail;
979                 *entries += 1;
980         }
981
982         return 0;
983 fail:
984         if (cluster_locked)
985                 spin_unlock(&cluster_locked->lock);
986         return -ENOSPC;
987 }
988
989 static noinline_for_stack int
990 update_cache_item(struct btrfs_trans_handle *trans,
991                   struct btrfs_root *root,
992                   struct inode *inode,
993                   struct btrfs_path *path, u64 offset,
994                   int entries, int bitmaps)
995 {
996         struct btrfs_key key;
997         struct btrfs_free_space_header *header;
998         struct extent_buffer *leaf;
999         int ret;
1000
1001         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1002         key.offset = offset;
1003         key.type = 0;
1004
1005         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1006         if (ret < 0) {
1007                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1008                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1009                                  GFP_NOFS);
1010                 goto fail;
1011         }
1012         leaf = path->nodes[0];
1013         if (ret > 0) {
1014                 struct btrfs_key found_key;
1015                 ASSERT(path->slots[0]);
1016                 path->slots[0]--;
1017                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1018                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1019                     found_key.offset != offset) {
1020                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1021                                          inode->i_size - 1,
1022                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1023                                          NULL, GFP_NOFS);
1024                         btrfs_release_path(path);
1025                         goto fail;
1026                 }
1027         }
1028
1029         BTRFS_I(inode)->generation = trans->transid;
1030         header = btrfs_item_ptr(leaf, path->slots[0],
1031                                 struct btrfs_free_space_header);
1032         btrfs_set_free_space_entries(leaf, header, entries);
1033         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1034         btrfs_set_free_space_generation(leaf, header, trans->transid);
1035         btrfs_mark_buffer_dirty(leaf);
1036         btrfs_release_path(path);
1037
1038         return 0;
1039
1040 fail:
1041         return -1;
1042 }
1043
1044 static noinline_for_stack int
1045 write_pinned_extent_entries(struct btrfs_root *root,
1046                             struct btrfs_block_group_cache *block_group,
1047                             struct btrfs_io_ctl *io_ctl,
1048                             int *entries)
1049 {
1050         u64 start, extent_start, extent_end, len;
1051         struct extent_io_tree *unpin = NULL;
1052         int ret;
1053
1054         if (!block_group)
1055                 return 0;
1056
1057         /*
1058          * We want to add any pinned extents to our free space cache
1059          * so we don't leak the space
1060          *
1061          * We shouldn't have switched the pinned extents yet so this is the
1062          * right one
1063          */
1064         unpin = root->fs_info->pinned_extents;
1065
1066         start = block_group->key.objectid;
1067
1068         while (start < block_group->key.objectid + block_group->key.offset) {
1069                 ret = find_first_extent_bit(unpin, start,
1070                                             &extent_start, &extent_end,
1071                                             EXTENT_DIRTY, NULL);
1072                 if (ret)
1073                         return 0;
1074
1075                 /* This pinned extent is out of our range */
1076                 if (extent_start >= block_group->key.objectid +
1077                     block_group->key.offset)
1078                         return 0;
1079
1080                 extent_start = max(extent_start, start);
1081                 extent_end = min(block_group->key.objectid +
1082                                  block_group->key.offset, extent_end + 1);
1083                 len = extent_end - extent_start;
1084
1085                 *entries += 1;
1086                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1087                 if (ret)
1088                         return -ENOSPC;
1089
1090                 start = extent_end;
1091         }
1092
1093         return 0;
1094 }
1095
1096 static noinline_for_stack int
1097 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1098 {
1099         struct list_head *pos, *n;
1100         int ret;
1101
1102         /* Write out the bitmaps */
1103         list_for_each_safe(pos, n, bitmap_list) {
1104                 struct btrfs_free_space *entry =
1105                         list_entry(pos, struct btrfs_free_space, list);
1106
1107                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1108                 if (ret)
1109                         return -ENOSPC;
1110                 list_del_init(&entry->list);
1111         }
1112
1113         return 0;
1114 }
1115
1116 static int flush_dirty_cache(struct inode *inode)
1117 {
1118         int ret;
1119
1120         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1121         if (ret)
1122                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1123                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1124                                  GFP_NOFS);
1125
1126         return ret;
1127 }
1128
1129 static void noinline_for_stack
1130 cleanup_bitmap_list(struct list_head *bitmap_list)
1131 {
1132         struct list_head *pos, *n;
1133
1134         list_for_each_safe(pos, n, bitmap_list) {
1135                 struct btrfs_free_space *entry =
1136                         list_entry(pos, struct btrfs_free_space, list);
1137                 list_del_init(&entry->list);
1138         }
1139 }
1140
1141 static void noinline_for_stack
1142 cleanup_write_cache_enospc(struct inode *inode,
1143                            struct btrfs_io_ctl *io_ctl,
1144                            struct extent_state **cached_state,
1145                            struct list_head *bitmap_list)
1146 {
1147         io_ctl_drop_pages(io_ctl);
1148         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1149                              i_size_read(inode) - 1, cached_state,
1150                              GFP_NOFS);
1151 }
1152
1153 int btrfs_wait_cache_io(struct btrfs_root *root,
1154                         struct btrfs_trans_handle *trans,
1155                         struct btrfs_block_group_cache *block_group,
1156                         struct btrfs_io_ctl *io_ctl,
1157                         struct btrfs_path *path, u64 offset)
1158 {
1159         int ret;
1160         struct inode *inode = io_ctl->inode;
1161
1162         if (!inode)
1163                 return 0;
1164
1165         if (block_group)
1166                 root = root->fs_info->tree_root;
1167
1168         /* Flush the dirty pages in the cache file. */
1169         ret = flush_dirty_cache(inode);
1170         if (ret)
1171                 goto out;
1172
1173         /* Update the cache item to tell everyone this cache file is valid. */
1174         ret = update_cache_item(trans, root, inode, path, offset,
1175                                 io_ctl->entries, io_ctl->bitmaps);
1176 out:
1177         io_ctl_free(io_ctl);
1178         if (ret) {
1179                 invalidate_inode_pages2(inode->i_mapping);
1180                 BTRFS_I(inode)->generation = 0;
1181                 if (block_group) {
1182 #ifdef DEBUG
1183                         btrfs_err(root->fs_info,
1184                                 "failed to write free space cache for block group %llu",
1185                                 block_group->key.objectid);
1186 #endif
1187                 }
1188         }
1189         btrfs_update_inode(trans, root, inode);
1190
1191         if (block_group) {
1192                 /* the dirty list is protected by the dirty_bgs_lock */
1193                 spin_lock(&trans->transaction->dirty_bgs_lock);
1194
1195                 /* the disk_cache_state is protected by the block group lock */
1196                 spin_lock(&block_group->lock);
1197
1198                 /*
1199                  * only mark this as written if we didn't get put back on
1200                  * the dirty list while waiting for IO.   Otherwise our
1201                  * cache state won't be right, and we won't get written again
1202                  */
1203                 if (!ret && list_empty(&block_group->dirty_list))
1204                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1205                 else if (ret)
1206                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1207
1208                 spin_unlock(&block_group->lock);
1209                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1210                 io_ctl->inode = NULL;
1211                 iput(inode);
1212         }
1213
1214         return ret;
1215
1216 }
1217
1218 /**
1219  * __btrfs_write_out_cache - write out cached info to an inode
1220  * @root - the root the inode belongs to
1221  * @ctl - the free space cache we are going to write out
1222  * @block_group - the block_group for this cache if it belongs to a block_group
1223  * @trans - the trans handle
1224  * @path - the path to use
1225  * @offset - the offset for the key we'll insert
1226  *
1227  * This function writes out a free space cache struct to disk for quick recovery
1228  * on mount.  This will return 0 if it was successful in writing the cache out,
1229  * or an errno if it was not.
1230  */
1231 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1232                                    struct btrfs_free_space_ctl *ctl,
1233                                    struct btrfs_block_group_cache *block_group,
1234                                    struct btrfs_io_ctl *io_ctl,
1235                                    struct btrfs_trans_handle *trans,
1236                                    struct btrfs_path *path, u64 offset)
1237 {
1238         struct extent_state *cached_state = NULL;
1239         LIST_HEAD(bitmap_list);
1240         int entries = 0;
1241         int bitmaps = 0;
1242         int ret;
1243         int must_iput = 0;
1244
1245         if (!i_size_read(inode))
1246                 return -EIO;
1247
1248         WARN_ON(io_ctl->pages);
1249         ret = io_ctl_init(io_ctl, inode, root, 1);
1250         if (ret)
1251                 return ret;
1252
1253         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1254                 down_write(&block_group->data_rwsem);
1255                 spin_lock(&block_group->lock);
1256                 if (block_group->delalloc_bytes) {
1257                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1258                         spin_unlock(&block_group->lock);
1259                         up_write(&block_group->data_rwsem);
1260                         BTRFS_I(inode)->generation = 0;
1261                         ret = 0;
1262                         must_iput = 1;
1263                         goto out;
1264                 }
1265                 spin_unlock(&block_group->lock);
1266         }
1267
1268         /* Lock all pages first so we can lock the extent safely. */
1269         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1270         if (ret)
1271                 goto out_unlock;
1272
1273         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1274                          0, &cached_state);
1275
1276         io_ctl_set_generation(io_ctl, trans->transid);
1277
1278         mutex_lock(&ctl->cache_writeout_mutex);
1279         /* Write out the extent entries in the free space cache */
1280         spin_lock(&ctl->tree_lock);
1281         ret = write_cache_extent_entries(io_ctl, ctl,
1282                                          block_group, &entries, &bitmaps,
1283                                          &bitmap_list);
1284         if (ret)
1285                 goto out_nospc_locked;
1286
1287         /*
1288          * Some spaces that are freed in the current transaction are pinned,
1289          * they will be added into free space cache after the transaction is
1290          * committed, we shouldn't lose them.
1291          *
1292          * If this changes while we are working we'll get added back to
1293          * the dirty list and redo it.  No locking needed
1294          */
1295         ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1296         if (ret)
1297                 goto out_nospc_locked;
1298
1299         /*
1300          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1301          * locked while doing it because a concurrent trim can be manipulating
1302          * or freeing the bitmap.
1303          */
1304         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1305         spin_unlock(&ctl->tree_lock);
1306         mutex_unlock(&ctl->cache_writeout_mutex);
1307         if (ret)
1308                 goto out_nospc;
1309
1310         /* Zero out the rest of the pages just to make sure */
1311         io_ctl_zero_remaining_pages(io_ctl);
1312
1313         /* Everything is written out, now we dirty the pages in the file. */
1314         ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1315                                 0, i_size_read(inode), &cached_state);
1316         if (ret)
1317                 goto out_nospc;
1318
1319         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1320                 up_write(&block_group->data_rwsem);
1321         /*
1322          * Release the pages and unlock the extent, we will flush
1323          * them out later
1324          */
1325         io_ctl_drop_pages(io_ctl);
1326
1327         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1328                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1329
1330         /*
1331          * at this point the pages are under IO and we're happy,
1332          * The caller is responsible for waiting on them and updating the
1333          * the cache and the inode
1334          */
1335         io_ctl->entries = entries;
1336         io_ctl->bitmaps = bitmaps;
1337
1338         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1339         if (ret)
1340                 goto out;
1341
1342         return 0;
1343
1344 out:
1345         io_ctl->inode = NULL;
1346         io_ctl_free(io_ctl);
1347         if (ret) {
1348                 invalidate_inode_pages2(inode->i_mapping);
1349                 BTRFS_I(inode)->generation = 0;
1350         }
1351         btrfs_update_inode(trans, root, inode);
1352         if (must_iput)
1353                 iput(inode);
1354         return ret;
1355
1356 out_nospc_locked:
1357         cleanup_bitmap_list(&bitmap_list);
1358         spin_unlock(&ctl->tree_lock);
1359         mutex_unlock(&ctl->cache_writeout_mutex);
1360
1361 out_nospc:
1362         cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1363
1364 out_unlock:
1365         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1366                 up_write(&block_group->data_rwsem);
1367
1368         goto out;
1369 }
1370
1371 int btrfs_write_out_cache(struct btrfs_root *root,
1372                           struct btrfs_trans_handle *trans,
1373                           struct btrfs_block_group_cache *block_group,
1374                           struct btrfs_path *path)
1375 {
1376         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1377         struct inode *inode;
1378         int ret = 0;
1379
1380         root = root->fs_info->tree_root;
1381
1382         spin_lock(&block_group->lock);
1383         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1384                 spin_unlock(&block_group->lock);
1385                 return 0;
1386         }
1387         spin_unlock(&block_group->lock);
1388
1389         inode = lookup_free_space_inode(root, block_group, path);
1390         if (IS_ERR(inode))
1391                 return 0;
1392
1393         ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1394                                       &block_group->io_ctl, trans,
1395                                       path, block_group->key.objectid);
1396         if (ret) {
1397 #ifdef DEBUG
1398                 btrfs_err(root->fs_info,
1399                         "failed to write free space cache for block group %llu",
1400                         block_group->key.objectid);
1401 #endif
1402                 spin_lock(&block_group->lock);
1403                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1404                 spin_unlock(&block_group->lock);
1405
1406                 block_group->io_ctl.inode = NULL;
1407                 iput(inode);
1408         }
1409
1410         /*
1411          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1412          * to wait for IO and put the inode
1413          */
1414
1415         return ret;
1416 }
1417
1418 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1419                                           u64 offset)
1420 {
1421         ASSERT(offset >= bitmap_start);
1422         offset -= bitmap_start;
1423         return (unsigned long)(div_u64(offset, unit));
1424 }
1425
1426 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1427 {
1428         return (unsigned long)(div_u64(bytes, unit));
1429 }
1430
1431 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1432                                    u64 offset)
1433 {
1434         u64 bitmap_start;
1435         u32 bytes_per_bitmap;
1436
1437         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1438         bitmap_start = offset - ctl->start;
1439         bitmap_start = div_u64(bitmap_start, bytes_per_bitmap);
1440         bitmap_start *= bytes_per_bitmap;
1441         bitmap_start += ctl->start;
1442
1443         return bitmap_start;
1444 }
1445
1446 static int tree_insert_offset(struct rb_root *root, u64 offset,
1447                               struct rb_node *node, int bitmap)
1448 {
1449         struct rb_node **p = &root->rb_node;
1450         struct rb_node *parent = NULL;
1451         struct btrfs_free_space *info;
1452
1453         while (*p) {
1454                 parent = *p;
1455                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1456
1457                 if (offset < info->offset) {
1458                         p = &(*p)->rb_left;
1459                 } else if (offset > info->offset) {
1460                         p = &(*p)->rb_right;
1461                 } else {
1462                         /*
1463                          * we could have a bitmap entry and an extent entry
1464                          * share the same offset.  If this is the case, we want
1465                          * the extent entry to always be found first if we do a
1466                          * linear search through the tree, since we want to have
1467                          * the quickest allocation time, and allocating from an
1468                          * extent is faster than allocating from a bitmap.  So
1469                          * if we're inserting a bitmap and we find an entry at
1470                          * this offset, we want to go right, or after this entry
1471                          * logically.  If we are inserting an extent and we've
1472                          * found a bitmap, we want to go left, or before
1473                          * logically.
1474                          */
1475                         if (bitmap) {
1476                                 if (info->bitmap) {
1477                                         WARN_ON_ONCE(1);
1478                                         return -EEXIST;
1479                                 }
1480                                 p = &(*p)->rb_right;
1481                         } else {
1482                                 if (!info->bitmap) {
1483                                         WARN_ON_ONCE(1);
1484                                         return -EEXIST;
1485                                 }
1486                                 p = &(*p)->rb_left;
1487                         }
1488                 }
1489         }
1490
1491         rb_link_node(node, parent, p);
1492         rb_insert_color(node, root);
1493
1494         return 0;
1495 }
1496
1497 /*
1498  * searches the tree for the given offset.
1499  *
1500  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1501  * want a section that has at least bytes size and comes at or after the given
1502  * offset.
1503  */
1504 static struct btrfs_free_space *
1505 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1506                    u64 offset, int bitmap_only, int fuzzy)
1507 {
1508         struct rb_node *n = ctl->free_space_offset.rb_node;
1509         struct btrfs_free_space *entry, *prev = NULL;
1510
1511         /* find entry that is closest to the 'offset' */
1512         while (1) {
1513                 if (!n) {
1514                         entry = NULL;
1515                         break;
1516                 }
1517
1518                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1519                 prev = entry;
1520
1521                 if (offset < entry->offset)
1522                         n = n->rb_left;
1523                 else if (offset > entry->offset)
1524                         n = n->rb_right;
1525                 else
1526                         break;
1527         }
1528
1529         if (bitmap_only) {
1530                 if (!entry)
1531                         return NULL;
1532                 if (entry->bitmap)
1533                         return entry;
1534
1535                 /*
1536                  * bitmap entry and extent entry may share same offset,
1537                  * in that case, bitmap entry comes after extent entry.
1538                  */
1539                 n = rb_next(n);
1540                 if (!n)
1541                         return NULL;
1542                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1543                 if (entry->offset != offset)
1544                         return NULL;
1545
1546                 WARN_ON(!entry->bitmap);
1547                 return entry;
1548         } else if (entry) {
1549                 if (entry->bitmap) {
1550                         /*
1551                          * if previous extent entry covers the offset,
1552                          * we should return it instead of the bitmap entry
1553                          */
1554                         n = rb_prev(&entry->offset_index);
1555                         if (n) {
1556                                 prev = rb_entry(n, struct btrfs_free_space,
1557                                                 offset_index);
1558                                 if (!prev->bitmap &&
1559                                     prev->offset + prev->bytes > offset)
1560                                         entry = prev;
1561                         }
1562                 }
1563                 return entry;
1564         }
1565
1566         if (!prev)
1567                 return NULL;
1568
1569         /* find last entry before the 'offset' */
1570         entry = prev;
1571         if (entry->offset > offset) {
1572                 n = rb_prev(&entry->offset_index);
1573                 if (n) {
1574                         entry = rb_entry(n, struct btrfs_free_space,
1575                                         offset_index);
1576                         ASSERT(entry->offset <= offset);
1577                 } else {
1578                         if (fuzzy)
1579                                 return entry;
1580                         else
1581                                 return NULL;
1582                 }
1583         }
1584
1585         if (entry->bitmap) {
1586                 n = rb_prev(&entry->offset_index);
1587                 if (n) {
1588                         prev = rb_entry(n, struct btrfs_free_space,
1589                                         offset_index);
1590                         if (!prev->bitmap &&
1591                             prev->offset + prev->bytes > offset)
1592                                 return prev;
1593                 }
1594                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1595                         return entry;
1596         } else if (entry->offset + entry->bytes > offset)
1597                 return entry;
1598
1599         if (!fuzzy)
1600                 return NULL;
1601
1602         while (1) {
1603                 if (entry->bitmap) {
1604                         if (entry->offset + BITS_PER_BITMAP *
1605                             ctl->unit > offset)
1606                                 break;
1607                 } else {
1608                         if (entry->offset + entry->bytes > offset)
1609                                 break;
1610                 }
1611
1612                 n = rb_next(&entry->offset_index);
1613                 if (!n)
1614                         return NULL;
1615                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1616         }
1617         return entry;
1618 }
1619
1620 static inline void
1621 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1622                     struct btrfs_free_space *info)
1623 {
1624         rb_erase(&info->offset_index, &ctl->free_space_offset);
1625         ctl->free_extents--;
1626 }
1627
1628 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1629                               struct btrfs_free_space *info)
1630 {
1631         __unlink_free_space(ctl, info);
1632         ctl->free_space -= info->bytes;
1633 }
1634
1635 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1636                            struct btrfs_free_space *info)
1637 {
1638         int ret = 0;
1639
1640         ASSERT(info->bytes || info->bitmap);
1641         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1642                                  &info->offset_index, (info->bitmap != NULL));
1643         if (ret)
1644                 return ret;
1645
1646         ctl->free_space += info->bytes;
1647         ctl->free_extents++;
1648         return ret;
1649 }
1650
1651 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1652 {
1653         struct btrfs_block_group_cache *block_group = ctl->private;
1654         u64 max_bytes;
1655         u64 bitmap_bytes;
1656         u64 extent_bytes;
1657         u64 size = block_group->key.offset;
1658         u32 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1659         u32 max_bitmaps = div_u64(size + bytes_per_bg - 1, bytes_per_bg);
1660
1661         max_bitmaps = max_t(u32, max_bitmaps, 1);
1662
1663         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1664
1665         /*
1666          * The goal is to keep the total amount of memory used per 1gb of space
1667          * at or below 32k, so we need to adjust how much memory we allow to be
1668          * used by extent based free space tracking
1669          */
1670         if (size < 1024 * 1024 * 1024)
1671                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1672         else
1673                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1674                         div_u64(size, 1024 * 1024 * 1024);
1675
1676         /*
1677          * we want to account for 1 more bitmap than what we have so we can make
1678          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1679          * we add more bitmaps.
1680          */
1681         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1682
1683         if (bitmap_bytes >= max_bytes) {
1684                 ctl->extents_thresh = 0;
1685                 return;
1686         }
1687
1688         /*
1689          * we want the extent entry threshold to always be at most 1/2 the max
1690          * bytes we can have, or whatever is less than that.
1691          */
1692         extent_bytes = max_bytes - bitmap_bytes;
1693         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1694
1695         ctl->extents_thresh =
1696                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1697 }
1698
1699 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1700                                        struct btrfs_free_space *info,
1701                                        u64 offset, u64 bytes)
1702 {
1703         unsigned long start, count;
1704
1705         start = offset_to_bit(info->offset, ctl->unit, offset);
1706         count = bytes_to_bits(bytes, ctl->unit);
1707         ASSERT(start + count <= BITS_PER_BITMAP);
1708
1709         bitmap_clear(info->bitmap, start, count);
1710
1711         info->bytes -= bytes;
1712         if (info->max_extent_size > ctl->unit)
1713                 info->max_extent_size = 0;
1714 }
1715
1716 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1717                               struct btrfs_free_space *info, u64 offset,
1718                               u64 bytes)
1719 {
1720         __bitmap_clear_bits(ctl, info, offset, bytes);
1721         ctl->free_space -= bytes;
1722 }
1723
1724 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1725                             struct btrfs_free_space *info, u64 offset,
1726                             u64 bytes)
1727 {
1728         unsigned long start, count;
1729
1730         start = offset_to_bit(info->offset, ctl->unit, offset);
1731         count = bytes_to_bits(bytes, ctl->unit);
1732         ASSERT(start + count <= BITS_PER_BITMAP);
1733
1734         bitmap_set(info->bitmap, start, count);
1735
1736         info->bytes += bytes;
1737         ctl->free_space += bytes;
1738 }
1739
1740 /*
1741  * If we can not find suitable extent, we will use bytes to record
1742  * the size of the max extent.
1743  */
1744 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1745                          struct btrfs_free_space *bitmap_info, u64 *offset,
1746                          u64 *bytes, bool for_alloc)
1747 {
1748         unsigned long found_bits = 0;
1749         unsigned long max_bits = 0;
1750         unsigned long bits, i;
1751         unsigned long next_zero;
1752         unsigned long extent_bits;
1753
1754         /*
1755          * Skip searching the bitmap if we don't have a contiguous section that
1756          * is large enough for this allocation.
1757          */
1758         if (for_alloc &&
1759             bitmap_info->max_extent_size &&
1760             bitmap_info->max_extent_size < *bytes) {
1761                 *bytes = bitmap_info->max_extent_size;
1762                 return -1;
1763         }
1764
1765         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1766                           max_t(u64, *offset, bitmap_info->offset));
1767         bits = bytes_to_bits(*bytes, ctl->unit);
1768
1769         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1770                 if (for_alloc && bits == 1) {
1771                         found_bits = 1;
1772                         break;
1773                 }
1774                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1775                                                BITS_PER_BITMAP, i);
1776                 extent_bits = next_zero - i;
1777                 if (extent_bits >= bits) {
1778                         found_bits = extent_bits;
1779                         break;
1780                 } else if (extent_bits > max_bits) {
1781                         max_bits = extent_bits;
1782                 }
1783                 i = next_zero;
1784         }
1785
1786         if (found_bits) {
1787                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1788                 *bytes = (u64)(found_bits) * ctl->unit;
1789                 return 0;
1790         }
1791
1792         *bytes = (u64)(max_bits) * ctl->unit;
1793         bitmap_info->max_extent_size = *bytes;
1794         return -1;
1795 }
1796
1797 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1798 {
1799         if (entry->bitmap)
1800                 return entry->max_extent_size;
1801         return entry->bytes;
1802 }
1803
1804 /* Cache the size of the max extent in bytes */
1805 static struct btrfs_free_space *
1806 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1807                 unsigned long align, u64 *max_extent_size)
1808 {
1809         struct btrfs_free_space *entry;
1810         struct rb_node *node;
1811         u64 tmp;
1812         u64 align_off;
1813         int ret;
1814
1815         if (!ctl->free_space_offset.rb_node)
1816                 goto out;
1817
1818         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1819         if (!entry)
1820                 goto out;
1821
1822         for (node = &entry->offset_index; node; node = rb_next(node)) {
1823                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1824                 if (entry->bytes < *bytes) {
1825                         *max_extent_size = max(get_max_extent_size(entry),
1826                                                *max_extent_size);
1827                         continue;
1828                 }
1829
1830                 /* make sure the space returned is big enough
1831                  * to match our requested alignment
1832                  */
1833                 if (*bytes >= align) {
1834                         tmp = entry->offset - ctl->start + align - 1;
1835                         tmp = div64_u64(tmp, align);
1836                         tmp = tmp * align + ctl->start;
1837                         align_off = tmp - entry->offset;
1838                 } else {
1839                         align_off = 0;
1840                         tmp = entry->offset;
1841                 }
1842
1843                 if (entry->bytes < *bytes + align_off) {
1844                         *max_extent_size = max(get_max_extent_size(entry),
1845                                                *max_extent_size);
1846                         continue;
1847                 }
1848
1849                 if (entry->bitmap) {
1850                         u64 size = *bytes;
1851
1852                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1853                         if (!ret) {
1854                                 *offset = tmp;
1855                                 *bytes = size;
1856                                 return entry;
1857                         } else {
1858                                 *max_extent_size =
1859                                         max(get_max_extent_size(entry),
1860                                             *max_extent_size);
1861                         }
1862                         continue;
1863                 }
1864
1865                 *offset = tmp;
1866                 *bytes = entry->bytes - align_off;
1867                 return entry;
1868         }
1869 out:
1870         return NULL;
1871 }
1872
1873 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1874                            struct btrfs_free_space *info, u64 offset)
1875 {
1876         info->offset = offset_to_bitmap(ctl, offset);
1877         info->bytes = 0;
1878         INIT_LIST_HEAD(&info->list);
1879         link_free_space(ctl, info);
1880         ctl->total_bitmaps++;
1881
1882         ctl->op->recalc_thresholds(ctl);
1883 }
1884
1885 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1886                         struct btrfs_free_space *bitmap_info)
1887 {
1888         unlink_free_space(ctl, bitmap_info);
1889         kfree(bitmap_info->bitmap);
1890         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1891         ctl->total_bitmaps--;
1892         ctl->op->recalc_thresholds(ctl);
1893 }
1894
1895 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1896                               struct btrfs_free_space *bitmap_info,
1897                               u64 *offset, u64 *bytes)
1898 {
1899         u64 end;
1900         u64 search_start, search_bytes;
1901         int ret;
1902
1903 again:
1904         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1905
1906         /*
1907          * We need to search for bits in this bitmap.  We could only cover some
1908          * of the extent in this bitmap thanks to how we add space, so we need
1909          * to search for as much as it as we can and clear that amount, and then
1910          * go searching for the next bit.
1911          */
1912         search_start = *offset;
1913         search_bytes = ctl->unit;
1914         search_bytes = min(search_bytes, end - search_start + 1);
1915         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1916                             false);
1917         if (ret < 0 || search_start != *offset)
1918                 return -EINVAL;
1919
1920         /* We may have found more bits than what we need */
1921         search_bytes = min(search_bytes, *bytes);
1922
1923         /* Cannot clear past the end of the bitmap */
1924         search_bytes = min(search_bytes, end - search_start + 1);
1925
1926         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1927         *offset += search_bytes;
1928         *bytes -= search_bytes;
1929
1930         if (*bytes) {
1931                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1932                 if (!bitmap_info->bytes)
1933                         free_bitmap(ctl, bitmap_info);
1934
1935                 /*
1936                  * no entry after this bitmap, but we still have bytes to
1937                  * remove, so something has gone wrong.
1938                  */
1939                 if (!next)
1940                         return -EINVAL;
1941
1942                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1943                                        offset_index);
1944
1945                 /*
1946                  * if the next entry isn't a bitmap we need to return to let the
1947                  * extent stuff do its work.
1948                  */
1949                 if (!bitmap_info->bitmap)
1950                         return -EAGAIN;
1951
1952                 /*
1953                  * Ok the next item is a bitmap, but it may not actually hold
1954                  * the information for the rest of this free space stuff, so
1955                  * look for it, and if we don't find it return so we can try
1956                  * everything over again.
1957                  */
1958                 search_start = *offset;
1959                 search_bytes = ctl->unit;
1960                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1961                                     &search_bytes, false);
1962                 if (ret < 0 || search_start != *offset)
1963                         return -EAGAIN;
1964
1965                 goto again;
1966         } else if (!bitmap_info->bytes)
1967                 free_bitmap(ctl, bitmap_info);
1968
1969         return 0;
1970 }
1971
1972 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1973                                struct btrfs_free_space *info, u64 offset,
1974                                u64 bytes)
1975 {
1976         u64 bytes_to_set = 0;
1977         u64 end;
1978
1979         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1980
1981         bytes_to_set = min(end - offset, bytes);
1982
1983         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1984
1985         /*
1986          * We set some bytes, we have no idea what the max extent size is
1987          * anymore.
1988          */
1989         info->max_extent_size = 0;
1990
1991         return bytes_to_set;
1992
1993 }
1994
1995 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1996                       struct btrfs_free_space *info)
1997 {
1998         struct btrfs_block_group_cache *block_group = ctl->private;
1999         bool forced = false;
2000
2001 #ifdef CONFIG_BTRFS_DEBUG
2002         if (btrfs_should_fragment_free_space(block_group->fs_info->extent_root,
2003                                              block_group))
2004                 forced = true;
2005 #endif
2006
2007         /*
2008          * If we are below the extents threshold then we can add this as an
2009          * extent, and don't have to deal with the bitmap
2010          */
2011         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2012                 /*
2013                  * If this block group has some small extents we don't want to
2014                  * use up all of our free slots in the cache with them, we want
2015                  * to reserve them to larger extents, however if we have plent
2016                  * of cache left then go ahead an dadd them, no sense in adding
2017                  * the overhead of a bitmap if we don't have to.
2018                  */
2019                 if (info->bytes <= block_group->sectorsize * 4) {
2020                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
2021                                 return false;
2022                 } else {
2023                         return false;
2024                 }
2025         }
2026
2027         /*
2028          * The original block groups from mkfs can be really small, like 8
2029          * megabytes, so don't bother with a bitmap for those entries.  However
2030          * some block groups can be smaller than what a bitmap would cover but
2031          * are still large enough that they could overflow the 32k memory limit,
2032          * so allow those block groups to still be allowed to have a bitmap
2033          * entry.
2034          */
2035         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2036                 return false;
2037
2038         return true;
2039 }
2040
2041 static struct btrfs_free_space_op free_space_op = {
2042         .recalc_thresholds      = recalculate_thresholds,
2043         .use_bitmap             = use_bitmap,
2044 };
2045
2046 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2047                               struct btrfs_free_space *info)
2048 {
2049         struct btrfs_free_space *bitmap_info;
2050         struct btrfs_block_group_cache *block_group = NULL;
2051         int added = 0;
2052         u64 bytes, offset, bytes_added;
2053         int ret;
2054
2055         bytes = info->bytes;
2056         offset = info->offset;
2057
2058         if (!ctl->op->use_bitmap(ctl, info))
2059                 return 0;
2060
2061         if (ctl->op == &free_space_op)
2062                 block_group = ctl->private;
2063 again:
2064         /*
2065          * Since we link bitmaps right into the cluster we need to see if we
2066          * have a cluster here, and if so and it has our bitmap we need to add
2067          * the free space to that bitmap.
2068          */
2069         if (block_group && !list_empty(&block_group->cluster_list)) {
2070                 struct btrfs_free_cluster *cluster;
2071                 struct rb_node *node;
2072                 struct btrfs_free_space *entry;
2073
2074                 cluster = list_entry(block_group->cluster_list.next,
2075                                      struct btrfs_free_cluster,
2076                                      block_group_list);
2077                 spin_lock(&cluster->lock);
2078                 node = rb_first(&cluster->root);
2079                 if (!node) {
2080                         spin_unlock(&cluster->lock);
2081                         goto no_cluster_bitmap;
2082                 }
2083
2084                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2085                 if (!entry->bitmap) {
2086                         spin_unlock(&cluster->lock);
2087                         goto no_cluster_bitmap;
2088                 }
2089
2090                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2091                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2092                                                           offset, bytes);
2093                         bytes -= bytes_added;
2094                         offset += bytes_added;
2095                 }
2096                 spin_unlock(&cluster->lock);
2097                 if (!bytes) {
2098                         ret = 1;
2099                         goto out;
2100                 }
2101         }
2102
2103 no_cluster_bitmap:
2104         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2105                                          1, 0);
2106         if (!bitmap_info) {
2107                 ASSERT(added == 0);
2108                 goto new_bitmap;
2109         }
2110
2111         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2112         bytes -= bytes_added;
2113         offset += bytes_added;
2114         added = 0;
2115
2116         if (!bytes) {
2117                 ret = 1;
2118                 goto out;
2119         } else
2120                 goto again;
2121
2122 new_bitmap:
2123         if (info && info->bitmap) {
2124                 add_new_bitmap(ctl, info, offset);
2125                 added = 1;
2126                 info = NULL;
2127                 goto again;
2128         } else {
2129                 spin_unlock(&ctl->tree_lock);
2130
2131                 /* no pre-allocated info, allocate a new one */
2132                 if (!info) {
2133                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2134                                                  GFP_NOFS);
2135                         if (!info) {
2136                                 spin_lock(&ctl->tree_lock);
2137                                 ret = -ENOMEM;
2138                                 goto out;
2139                         }
2140                 }
2141
2142                 /* allocate the bitmap */
2143                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
2144                 spin_lock(&ctl->tree_lock);
2145                 if (!info->bitmap) {
2146                         ret = -ENOMEM;
2147                         goto out;
2148                 }
2149                 goto again;
2150         }
2151
2152 out:
2153         if (info) {
2154                 if (info->bitmap)
2155                         kfree(info->bitmap);
2156                 kmem_cache_free(btrfs_free_space_cachep, info);
2157         }
2158
2159         return ret;
2160 }
2161
2162 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2163                           struct btrfs_free_space *info, bool update_stat)
2164 {
2165         struct btrfs_free_space *left_info = NULL;
2166         struct btrfs_free_space *right_info;
2167         bool merged = false;
2168         u64 offset = info->offset;
2169         u64 bytes = info->bytes;
2170
2171         /*
2172          * first we want to see if there is free space adjacent to the range we
2173          * are adding, if there is remove that struct and add a new one to
2174          * cover the entire range
2175          */
2176         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2177         if (right_info && rb_prev(&right_info->offset_index))
2178                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2179                                      struct btrfs_free_space, offset_index);
2180         else if (!right_info)
2181                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2182
2183         if (right_info && !right_info->bitmap) {
2184                 if (update_stat)
2185                         unlink_free_space(ctl, right_info);
2186                 else
2187                         __unlink_free_space(ctl, right_info);
2188                 info->bytes += right_info->bytes;
2189                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2190                 merged = true;
2191         }
2192
2193         if (left_info && !left_info->bitmap &&
2194             left_info->offset + left_info->bytes == offset) {
2195                 if (update_stat)
2196                         unlink_free_space(ctl, left_info);
2197                 else
2198                         __unlink_free_space(ctl, left_info);
2199                 info->offset = left_info->offset;
2200                 info->bytes += left_info->bytes;
2201                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2202                 merged = true;
2203         }
2204
2205         return merged;
2206 }
2207
2208 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2209                                      struct btrfs_free_space *info,
2210                                      bool update_stat)
2211 {
2212         struct btrfs_free_space *bitmap;
2213         unsigned long i;
2214         unsigned long j;
2215         const u64 end = info->offset + info->bytes;
2216         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2217         u64 bytes;
2218
2219         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2220         if (!bitmap)
2221                 return false;
2222
2223         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2224         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2225         if (j == i)
2226                 return false;
2227         bytes = (j - i) * ctl->unit;
2228         info->bytes += bytes;
2229
2230         if (update_stat)
2231                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2232         else
2233                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2234
2235         if (!bitmap->bytes)
2236                 free_bitmap(ctl, bitmap);
2237
2238         return true;
2239 }
2240
2241 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2242                                        struct btrfs_free_space *info,
2243                                        bool update_stat)
2244 {
2245         struct btrfs_free_space *bitmap;
2246         u64 bitmap_offset;
2247         unsigned long i;
2248         unsigned long j;
2249         unsigned long prev_j;
2250         u64 bytes;
2251
2252         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2253         /* If we're on a boundary, try the previous logical bitmap. */
2254         if (bitmap_offset == info->offset) {
2255                 if (info->offset == 0)
2256                         return false;
2257                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2258         }
2259
2260         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2261         if (!bitmap)
2262                 return false;
2263
2264         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2265         j = 0;
2266         prev_j = (unsigned long)-1;
2267         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2268                 if (j > i)
2269                         break;
2270                 prev_j = j;
2271         }
2272         if (prev_j == i)
2273                 return false;
2274
2275         if (prev_j == (unsigned long)-1)
2276                 bytes = (i + 1) * ctl->unit;
2277         else
2278                 bytes = (i - prev_j) * ctl->unit;
2279
2280         info->offset -= bytes;
2281         info->bytes += bytes;
2282
2283         if (update_stat)
2284                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2285         else
2286                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2287
2288         if (!bitmap->bytes)
2289                 free_bitmap(ctl, bitmap);
2290
2291         return true;
2292 }
2293
2294 /*
2295  * We prefer always to allocate from extent entries, both for clustered and
2296  * non-clustered allocation requests. So when attempting to add a new extent
2297  * entry, try to see if there's adjacent free space in bitmap entries, and if
2298  * there is, migrate that space from the bitmaps to the extent.
2299  * Like this we get better chances of satisfying space allocation requests
2300  * because we attempt to satisfy them based on a single cache entry, and never
2301  * on 2 or more entries - even if the entries represent a contiguous free space
2302  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2303  * ends).
2304  */
2305 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2306                               struct btrfs_free_space *info,
2307                               bool update_stat)
2308 {
2309         /*
2310          * Only work with disconnected entries, as we can change their offset,
2311          * and must be extent entries.
2312          */
2313         ASSERT(!info->bitmap);
2314         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2315
2316         if (ctl->total_bitmaps > 0) {
2317                 bool stole_end;
2318                 bool stole_front = false;
2319
2320                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2321                 if (ctl->total_bitmaps > 0)
2322                         stole_front = steal_from_bitmap_to_front(ctl, info,
2323                                                                  update_stat);
2324
2325                 if (stole_end || stole_front)
2326                         try_merge_free_space(ctl, info, update_stat);
2327         }
2328 }
2329
2330 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2331                            u64 offset, u64 bytes)
2332 {
2333         struct btrfs_free_space *info;
2334         int ret = 0;
2335
2336         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2337         if (!info)
2338                 return -ENOMEM;
2339
2340         info->offset = offset;
2341         info->bytes = bytes;
2342         RB_CLEAR_NODE(&info->offset_index);
2343
2344         spin_lock(&ctl->tree_lock);
2345
2346         if (try_merge_free_space(ctl, info, true))
2347                 goto link;
2348
2349         /*
2350          * There was no extent directly to the left or right of this new
2351          * extent then we know we're going to have to allocate a new extent, so
2352          * before we do that see if we need to drop this into a bitmap
2353          */
2354         ret = insert_into_bitmap(ctl, info);
2355         if (ret < 0) {
2356                 goto out;
2357         } else if (ret) {
2358                 ret = 0;
2359                 goto out;
2360         }
2361 link:
2362         /*
2363          * Only steal free space from adjacent bitmaps if we're sure we're not
2364          * going to add the new free space to existing bitmap entries - because
2365          * that would mean unnecessary work that would be reverted. Therefore
2366          * attempt to steal space from bitmaps if we're adding an extent entry.
2367          */
2368         steal_from_bitmap(ctl, info, true);
2369
2370         ret = link_free_space(ctl, info);
2371         if (ret)
2372                 kmem_cache_free(btrfs_free_space_cachep, info);
2373 out:
2374         spin_unlock(&ctl->tree_lock);
2375
2376         if (ret) {
2377                 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2378                 ASSERT(ret != -EEXIST);
2379         }
2380
2381         return ret;
2382 }
2383
2384 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2385                             u64 offset, u64 bytes)
2386 {
2387         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2388         struct btrfs_free_space *info;
2389         int ret;
2390         bool re_search = false;
2391
2392         spin_lock(&ctl->tree_lock);
2393
2394 again:
2395         ret = 0;
2396         if (!bytes)
2397                 goto out_lock;
2398
2399         info = tree_search_offset(ctl, offset, 0, 0);
2400         if (!info) {
2401                 /*
2402                  * oops didn't find an extent that matched the space we wanted
2403                  * to remove, look for a bitmap instead
2404                  */
2405                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2406                                           1, 0);
2407                 if (!info) {
2408                         /*
2409                          * If we found a partial bit of our free space in a
2410                          * bitmap but then couldn't find the other part this may
2411                          * be a problem, so WARN about it.
2412                          */
2413                         WARN_ON(re_search);
2414                         goto out_lock;
2415                 }
2416         }
2417
2418         re_search = false;
2419         if (!info->bitmap) {
2420                 unlink_free_space(ctl, info);
2421                 if (offset == info->offset) {
2422                         u64 to_free = min(bytes, info->bytes);
2423
2424                         info->bytes -= to_free;
2425                         info->offset += to_free;
2426                         if (info->bytes) {
2427                                 ret = link_free_space(ctl, info);
2428                                 WARN_ON(ret);
2429                         } else {
2430                                 kmem_cache_free(btrfs_free_space_cachep, info);
2431                         }
2432
2433                         offset += to_free;
2434                         bytes -= to_free;
2435                         goto again;
2436                 } else {
2437                         u64 old_end = info->bytes + info->offset;
2438
2439                         info->bytes = offset - info->offset;
2440                         ret = link_free_space(ctl, info);
2441                         WARN_ON(ret);
2442                         if (ret)
2443                                 goto out_lock;
2444
2445                         /* Not enough bytes in this entry to satisfy us */
2446                         if (old_end < offset + bytes) {
2447                                 bytes -= old_end - offset;
2448                                 offset = old_end;
2449                                 goto again;
2450                         } else if (old_end == offset + bytes) {
2451                                 /* all done */
2452                                 goto out_lock;
2453                         }
2454                         spin_unlock(&ctl->tree_lock);
2455
2456                         ret = btrfs_add_free_space(block_group, offset + bytes,
2457                                                    old_end - (offset + bytes));
2458                         WARN_ON(ret);
2459                         goto out;
2460                 }
2461         }
2462
2463         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2464         if (ret == -EAGAIN) {
2465                 re_search = true;
2466                 goto again;
2467         }
2468 out_lock:
2469         spin_unlock(&ctl->tree_lock);
2470 out:
2471         return ret;
2472 }
2473
2474 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2475                            u64 bytes)
2476 {
2477         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2478         struct btrfs_free_space *info;
2479         struct rb_node *n;
2480         int count = 0;
2481
2482         spin_lock(&ctl->tree_lock);
2483         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2484                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2485                 if (info->bytes >= bytes && !block_group->ro)
2486                         count++;
2487                 btrfs_crit(block_group->fs_info,
2488                            "entry offset %llu, bytes %llu, bitmap %s",
2489                            info->offset, info->bytes,
2490                        (info->bitmap) ? "yes" : "no");
2491         }
2492         spin_unlock(&ctl->tree_lock);
2493         btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2494                list_empty(&block_group->cluster_list) ? "no" : "yes");
2495         btrfs_info(block_group->fs_info,
2496                    "%d blocks of free space at or bigger than bytes is", count);
2497 }
2498
2499 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2500 {
2501         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2502
2503         spin_lock_init(&ctl->tree_lock);
2504         ctl->unit = block_group->sectorsize;
2505         ctl->start = block_group->key.objectid;
2506         ctl->private = block_group;
2507         ctl->op = &free_space_op;
2508         INIT_LIST_HEAD(&ctl->trimming_ranges);
2509         mutex_init(&ctl->cache_writeout_mutex);
2510
2511         /*
2512          * we only want to have 32k of ram per block group for keeping
2513          * track of free space, and if we pass 1/2 of that we want to
2514          * start converting things over to using bitmaps
2515          */
2516         ctl->extents_thresh = ((1024 * 32) / 2) /
2517                                 sizeof(struct btrfs_free_space);
2518 }
2519
2520 /*
2521  * for a given cluster, put all of its extents back into the free
2522  * space cache.  If the block group passed doesn't match the block group
2523  * pointed to by the cluster, someone else raced in and freed the
2524  * cluster already.  In that case, we just return without changing anything
2525  */
2526 static int
2527 __btrfs_return_cluster_to_free_space(
2528                              struct btrfs_block_group_cache *block_group,
2529                              struct btrfs_free_cluster *cluster)
2530 {
2531         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2532         struct btrfs_free_space *entry;
2533         struct rb_node *node;
2534
2535         spin_lock(&cluster->lock);
2536         if (cluster->block_group != block_group)
2537                 goto out;
2538
2539         cluster->block_group = NULL;
2540         cluster->window_start = 0;
2541         list_del_init(&cluster->block_group_list);
2542
2543         node = rb_first(&cluster->root);
2544         while (node) {
2545                 bool bitmap;
2546
2547                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2548                 node = rb_next(&entry->offset_index);
2549                 rb_erase(&entry->offset_index, &cluster->root);
2550                 RB_CLEAR_NODE(&entry->offset_index);
2551
2552                 bitmap = (entry->bitmap != NULL);
2553                 if (!bitmap) {
2554                         try_merge_free_space(ctl, entry, false);
2555                         steal_from_bitmap(ctl, entry, false);
2556                 }
2557                 tree_insert_offset(&ctl->free_space_offset,
2558                                    entry->offset, &entry->offset_index, bitmap);
2559         }
2560         cluster->root = RB_ROOT;
2561
2562 out:
2563         spin_unlock(&cluster->lock);
2564         btrfs_put_block_group(block_group);
2565         return 0;
2566 }
2567
2568 static void __btrfs_remove_free_space_cache_locked(
2569                                 struct btrfs_free_space_ctl *ctl)
2570 {
2571         struct btrfs_free_space *info;
2572         struct rb_node *node;
2573
2574         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2575                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2576                 if (!info->bitmap) {
2577                         unlink_free_space(ctl, info);
2578                         kmem_cache_free(btrfs_free_space_cachep, info);
2579                 } else {
2580                         free_bitmap(ctl, info);
2581                 }
2582
2583                 cond_resched_lock(&ctl->tree_lock);
2584         }
2585 }
2586
2587 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2588 {
2589         spin_lock(&ctl->tree_lock);
2590         __btrfs_remove_free_space_cache_locked(ctl);
2591         spin_unlock(&ctl->tree_lock);
2592 }
2593
2594 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2595 {
2596         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2597         struct btrfs_free_cluster *cluster;
2598         struct list_head *head;
2599
2600         spin_lock(&ctl->tree_lock);
2601         while ((head = block_group->cluster_list.next) !=
2602                &block_group->cluster_list) {
2603                 cluster = list_entry(head, struct btrfs_free_cluster,
2604                                      block_group_list);
2605
2606                 WARN_ON(cluster->block_group != block_group);
2607                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2608
2609                 cond_resched_lock(&ctl->tree_lock);
2610         }
2611         __btrfs_remove_free_space_cache_locked(ctl);
2612         spin_unlock(&ctl->tree_lock);
2613
2614 }
2615
2616 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2617                                u64 offset, u64 bytes, u64 empty_size,
2618                                u64 *max_extent_size)
2619 {
2620         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2621         struct btrfs_free_space *entry = NULL;
2622         u64 bytes_search = bytes + empty_size;
2623         u64 ret = 0;
2624         u64 align_gap = 0;
2625         u64 align_gap_len = 0;
2626
2627         spin_lock(&ctl->tree_lock);
2628         entry = find_free_space(ctl, &offset, &bytes_search,
2629                                 block_group->full_stripe_len, max_extent_size);
2630         if (!entry)
2631                 goto out;
2632
2633         ret = offset;
2634         if (entry->bitmap) {
2635                 bitmap_clear_bits(ctl, entry, offset, bytes);
2636                 if (!entry->bytes)
2637                         free_bitmap(ctl, entry);
2638         } else {
2639                 unlink_free_space(ctl, entry);
2640                 align_gap_len = offset - entry->offset;
2641                 align_gap = entry->offset;
2642
2643                 entry->offset = offset + bytes;
2644                 WARN_ON(entry->bytes < bytes + align_gap_len);
2645
2646                 entry->bytes -= bytes + align_gap_len;
2647                 if (!entry->bytes)
2648                         kmem_cache_free(btrfs_free_space_cachep, entry);
2649                 else
2650                         link_free_space(ctl, entry);
2651         }
2652 out:
2653         spin_unlock(&ctl->tree_lock);
2654
2655         if (align_gap_len)
2656                 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2657         return ret;
2658 }
2659
2660 /*
2661  * given a cluster, put all of its extents back into the free space
2662  * cache.  If a block group is passed, this function will only free
2663  * a cluster that belongs to the passed block group.
2664  *
2665  * Otherwise, it'll get a reference on the block group pointed to by the
2666  * cluster and remove the cluster from it.
2667  */
2668 int btrfs_return_cluster_to_free_space(
2669                                struct btrfs_block_group_cache *block_group,
2670                                struct btrfs_free_cluster *cluster)
2671 {
2672         struct btrfs_free_space_ctl *ctl;
2673         int ret;
2674
2675         /* first, get a safe pointer to the block group */
2676         spin_lock(&cluster->lock);
2677         if (!block_group) {
2678                 block_group = cluster->block_group;
2679                 if (!block_group) {
2680                         spin_unlock(&cluster->lock);
2681                         return 0;
2682                 }
2683         } else if (cluster->block_group != block_group) {
2684                 /* someone else has already freed it don't redo their work */
2685                 spin_unlock(&cluster->lock);
2686                 return 0;
2687         }
2688         atomic_inc(&block_group->count);
2689         spin_unlock(&cluster->lock);
2690
2691         ctl = block_group->free_space_ctl;
2692
2693         /* now return any extents the cluster had on it */
2694         spin_lock(&ctl->tree_lock);
2695         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2696         spin_unlock(&ctl->tree_lock);
2697
2698         /* finally drop our ref */
2699         btrfs_put_block_group(block_group);
2700         return ret;
2701 }
2702
2703 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2704                                    struct btrfs_free_cluster *cluster,
2705                                    struct btrfs_free_space *entry,
2706                                    u64 bytes, u64 min_start,
2707                                    u64 *max_extent_size)
2708 {
2709         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2710         int err;
2711         u64 search_start = cluster->window_start;
2712         u64 search_bytes = bytes;
2713         u64 ret = 0;
2714
2715         search_start = min_start;
2716         search_bytes = bytes;
2717
2718         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2719         if (err) {
2720                 *max_extent_size = max(get_max_extent_size(entry),
2721                                        *max_extent_size);
2722                 return 0;
2723         }
2724
2725         ret = search_start;
2726         __bitmap_clear_bits(ctl, entry, ret, bytes);
2727
2728         return ret;
2729 }
2730
2731 /*
2732  * given a cluster, try to allocate 'bytes' from it, returns 0
2733  * if it couldn't find anything suitably large, or a logical disk offset
2734  * if things worked out
2735  */
2736 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2737                              struct btrfs_free_cluster *cluster, u64 bytes,
2738                              u64 min_start, u64 *max_extent_size)
2739 {
2740         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2741         struct btrfs_free_space *entry = NULL;
2742         struct rb_node *node;
2743         u64 ret = 0;
2744
2745         spin_lock(&cluster->lock);
2746         if (bytes > cluster->max_size)
2747                 goto out;
2748
2749         if (cluster->block_group != block_group)
2750                 goto out;
2751
2752         node = rb_first(&cluster->root);
2753         if (!node)
2754                 goto out;
2755
2756         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2757         while (1) {
2758                 if (entry->bytes < bytes)
2759                         *max_extent_size = max(get_max_extent_size(entry),
2760                                                *max_extent_size);
2761
2762                 if (entry->bytes < bytes ||
2763                     (!entry->bitmap && entry->offset < min_start)) {
2764                         node = rb_next(&entry->offset_index);
2765                         if (!node)
2766                                 break;
2767                         entry = rb_entry(node, struct btrfs_free_space,
2768                                          offset_index);
2769                         continue;
2770                 }
2771
2772                 if (entry->bitmap) {
2773                         ret = btrfs_alloc_from_bitmap(block_group,
2774                                                       cluster, entry, bytes,
2775                                                       cluster->window_start,
2776                                                       max_extent_size);
2777                         if (ret == 0) {
2778                                 node = rb_next(&entry->offset_index);
2779                                 if (!node)
2780                                         break;
2781                                 entry = rb_entry(node, struct btrfs_free_space,
2782                                                  offset_index);
2783                                 continue;
2784                         }
2785                         cluster->window_start += bytes;
2786                 } else {
2787                         ret = entry->offset;
2788
2789                         entry->offset += bytes;
2790                         entry->bytes -= bytes;
2791                 }
2792
2793                 if (entry->bytes == 0)
2794                         rb_erase(&entry->offset_index, &cluster->root);
2795                 break;
2796         }
2797 out:
2798         spin_unlock(&cluster->lock);
2799
2800         if (!ret)
2801                 return 0;
2802
2803         spin_lock(&ctl->tree_lock);
2804
2805         ctl->free_space -= bytes;
2806         if (entry->bytes == 0) {
2807                 ctl->free_extents--;
2808                 if (entry->bitmap) {
2809                         kfree(entry->bitmap);
2810                         ctl->total_bitmaps--;
2811                         ctl->op->recalc_thresholds(ctl);
2812                 }
2813                 kmem_cache_free(btrfs_free_space_cachep, entry);
2814         }
2815
2816         spin_unlock(&ctl->tree_lock);
2817
2818         return ret;
2819 }
2820
2821 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2822                                 struct btrfs_free_space *entry,
2823                                 struct btrfs_free_cluster *cluster,
2824                                 u64 offset, u64 bytes,
2825                                 u64 cont1_bytes, u64 min_bytes)
2826 {
2827         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2828         unsigned long next_zero;
2829         unsigned long i;
2830         unsigned long want_bits;
2831         unsigned long min_bits;
2832         unsigned long found_bits;
2833         unsigned long max_bits = 0;
2834         unsigned long start = 0;
2835         unsigned long total_found = 0;
2836         int ret;
2837
2838         i = offset_to_bit(entry->offset, ctl->unit,
2839                           max_t(u64, offset, entry->offset));
2840         want_bits = bytes_to_bits(bytes, ctl->unit);
2841         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2842
2843         /*
2844          * Don't bother looking for a cluster in this bitmap if it's heavily
2845          * fragmented.
2846          */
2847         if (entry->max_extent_size &&
2848             entry->max_extent_size < cont1_bytes)
2849                 return -ENOSPC;
2850 again:
2851         found_bits = 0;
2852         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2853                 next_zero = find_next_zero_bit(entry->bitmap,
2854                                                BITS_PER_BITMAP, i);
2855                 if (next_zero - i >= min_bits) {
2856                         found_bits = next_zero - i;
2857                         if (found_bits > max_bits)
2858                                 max_bits = found_bits;
2859                         break;
2860                 }
2861                 if (next_zero - i > max_bits)
2862                         max_bits = next_zero - i;
2863                 i = next_zero;
2864         }
2865
2866         if (!found_bits) {
2867                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2868                 return -ENOSPC;
2869         }
2870
2871         if (!total_found) {
2872                 start = i;
2873                 cluster->max_size = 0;
2874         }
2875
2876         total_found += found_bits;
2877
2878         if (cluster->max_size < found_bits * ctl->unit)
2879                 cluster->max_size = found_bits * ctl->unit;
2880
2881         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2882                 i = next_zero + 1;
2883                 goto again;
2884         }
2885
2886         cluster->window_start = start * ctl->unit + entry->offset;
2887         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2888         ret = tree_insert_offset(&cluster->root, entry->offset,
2889                                  &entry->offset_index, 1);
2890         ASSERT(!ret); /* -EEXIST; Logic error */
2891
2892         trace_btrfs_setup_cluster(block_group, cluster,
2893                                   total_found * ctl->unit, 1);
2894         return 0;
2895 }
2896
2897 /*
2898  * This searches the block group for just extents to fill the cluster with.
2899  * Try to find a cluster with at least bytes total bytes, at least one
2900  * extent of cont1_bytes, and other clusters of at least min_bytes.
2901  */
2902 static noinline int
2903 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2904                         struct btrfs_free_cluster *cluster,
2905                         struct list_head *bitmaps, u64 offset, u64 bytes,
2906                         u64 cont1_bytes, u64 min_bytes)
2907 {
2908         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2909         struct btrfs_free_space *first = NULL;
2910         struct btrfs_free_space *entry = NULL;
2911         struct btrfs_free_space *last;
2912         struct rb_node *node;
2913         u64 window_free;
2914         u64 max_extent;
2915         u64 total_size = 0;
2916
2917         entry = tree_search_offset(ctl, offset, 0, 1);
2918         if (!entry)
2919                 return -ENOSPC;
2920
2921         /*
2922          * We don't want bitmaps, so just move along until we find a normal
2923          * extent entry.
2924          */
2925         while (entry->bitmap || entry->bytes < min_bytes) {
2926                 if (entry->bitmap && list_empty(&entry->list))
2927                         list_add_tail(&entry->list, bitmaps);
2928                 node = rb_next(&entry->offset_index);
2929                 if (!node)
2930                         return -ENOSPC;
2931                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2932         }
2933
2934         window_free = entry->bytes;
2935         max_extent = entry->bytes;
2936         first = entry;
2937         last = entry;
2938
2939         for (node = rb_next(&entry->offset_index); node;
2940              node = rb_next(&entry->offset_index)) {
2941                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2942
2943                 if (entry->bitmap) {
2944                         if (list_empty(&entry->list))
2945                                 list_add_tail(&entry->list, bitmaps);
2946                         continue;
2947                 }
2948
2949                 if (entry->bytes < min_bytes)
2950                         continue;
2951
2952                 last = entry;
2953                 window_free += entry->bytes;
2954                 if (entry->bytes > max_extent)
2955                         max_extent = entry->bytes;
2956         }
2957
2958         if (window_free < bytes || max_extent < cont1_bytes)
2959                 return -ENOSPC;
2960
2961         cluster->window_start = first->offset;
2962
2963         node = &first->offset_index;
2964
2965         /*
2966          * now we've found our entries, pull them out of the free space
2967          * cache and put them into the cluster rbtree
2968          */
2969         do {
2970                 int ret;
2971
2972                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2973                 node = rb_next(&entry->offset_index);
2974                 if (entry->bitmap || entry->bytes < min_bytes)
2975                         continue;
2976
2977                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2978                 ret = tree_insert_offset(&cluster->root, entry->offset,
2979                                          &entry->offset_index, 0);
2980                 total_size += entry->bytes;
2981                 ASSERT(!ret); /* -EEXIST; Logic error */
2982         } while (node && entry != last);
2983
2984         cluster->max_size = max_extent;
2985         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2986         return 0;
2987 }
2988
2989 /*
2990  * This specifically looks for bitmaps that may work in the cluster, we assume
2991  * that we have already failed to find extents that will work.
2992  */
2993 static noinline int
2994 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2995                      struct btrfs_free_cluster *cluster,
2996                      struct list_head *bitmaps, u64 offset, u64 bytes,
2997                      u64 cont1_bytes, u64 min_bytes)
2998 {
2999         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3000         struct btrfs_free_space *entry = NULL;
3001         int ret = -ENOSPC;
3002         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3003
3004         if (ctl->total_bitmaps == 0)
3005                 return -ENOSPC;
3006
3007         /*
3008          * The bitmap that covers offset won't be in the list unless offset
3009          * is just its start offset.
3010          */
3011         if (!list_empty(bitmaps))
3012                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3013
3014         if (!entry || entry->offset != bitmap_offset) {
3015                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3016                 if (entry && list_empty(&entry->list))
3017                         list_add(&entry->list, bitmaps);
3018         }
3019
3020         list_for_each_entry(entry, bitmaps, list) {
3021                 if (entry->bytes < bytes)
3022                         continue;
3023                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3024                                            bytes, cont1_bytes, min_bytes);
3025                 if (!ret)
3026                         return 0;
3027         }
3028
3029         /*
3030          * The bitmaps list has all the bitmaps that record free space
3031          * starting after offset, so no more search is required.
3032          */
3033         return -ENOSPC;
3034 }
3035
3036 /*
3037  * here we try to find a cluster of blocks in a block group.  The goal
3038  * is to find at least bytes+empty_size.
3039  * We might not find them all in one contiguous area.
3040  *
3041  * returns zero and sets up cluster if things worked out, otherwise
3042  * it returns -enospc
3043  */
3044 int btrfs_find_space_cluster(struct btrfs_root *root,
3045                              struct btrfs_block_group_cache *block_group,
3046                              struct btrfs_free_cluster *cluster,
3047                              u64 offset, u64 bytes, u64 empty_size)
3048 {
3049         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3050         struct btrfs_free_space *entry, *tmp;
3051         LIST_HEAD(bitmaps);
3052         u64 min_bytes;
3053         u64 cont1_bytes;
3054         int ret;
3055
3056         /*
3057          * Choose the minimum extent size we'll require for this
3058          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3059          * For metadata, allow allocates with smaller extents.  For
3060          * data, keep it dense.
3061          */
3062         if (btrfs_test_opt(root, SSD_SPREAD)) {
3063                 cont1_bytes = min_bytes = bytes + empty_size;
3064         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3065                 cont1_bytes = bytes;
3066                 min_bytes = block_group->sectorsize;
3067         } else {
3068                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3069                 min_bytes = block_group->sectorsize;
3070         }
3071
3072         spin_lock(&ctl->tree_lock);
3073
3074         /*
3075          * If we know we don't have enough space to make a cluster don't even
3076          * bother doing all the work to try and find one.
3077          */
3078         if (ctl->free_space < bytes) {
3079                 spin_unlock(&ctl->tree_lock);
3080                 return -ENOSPC;
3081         }
3082
3083         spin_lock(&cluster->lock);
3084
3085         /* someone already found a cluster, hooray */
3086         if (cluster->block_group) {
3087                 ret = 0;
3088                 goto out;
3089         }
3090
3091         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3092                                  min_bytes);
3093
3094         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3095                                       bytes + empty_size,
3096                                       cont1_bytes, min_bytes);
3097         if (ret)
3098                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3099                                            offset, bytes + empty_size,
3100                                            cont1_bytes, min_bytes);
3101
3102         /* Clear our temporary list */
3103         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3104                 list_del_init(&entry->list);
3105
3106         if (!ret) {
3107                 atomic_inc(&block_group->count);
3108                 list_add_tail(&cluster->block_group_list,
3109                               &block_group->cluster_list);
3110                 cluster->block_group = block_group;
3111         } else {
3112                 trace_btrfs_failed_cluster_setup(block_group);
3113         }
3114 out:
3115         spin_unlock(&cluster->lock);
3116         spin_unlock(&ctl->tree_lock);
3117
3118         return ret;
3119 }
3120
3121 /*
3122  * simple code to zero out a cluster
3123  */
3124 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3125 {
3126         spin_lock_init(&cluster->lock);
3127         spin_lock_init(&cluster->refill_lock);
3128         cluster->root = RB_ROOT;
3129         cluster->max_size = 0;
3130         cluster->fragmented = false;
3131         INIT_LIST_HEAD(&cluster->block_group_list);
3132         cluster->block_group = NULL;
3133 }
3134
3135 static int do_trimming(struct btrfs_block_group_cache *block_group,
3136                        u64 *total_trimmed, u64 start, u64 bytes,
3137                        u64 reserved_start, u64 reserved_bytes,
3138                        struct btrfs_trim_range *trim_entry)
3139 {
3140         struct btrfs_space_info *space_info = block_group->space_info;
3141         struct btrfs_fs_info *fs_info = block_group->fs_info;
3142         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3143         int ret;
3144         int update = 0;
3145         u64 trimmed = 0;
3146
3147         spin_lock(&space_info->lock);
3148         spin_lock(&block_group->lock);
3149         if (!block_group->ro) {
3150                 block_group->reserved += reserved_bytes;
3151                 space_info->bytes_reserved += reserved_bytes;
3152                 update = 1;
3153         }
3154         spin_unlock(&block_group->lock);
3155         spin_unlock(&space_info->lock);
3156
3157         ret = btrfs_discard_extent(fs_info->extent_root,
3158                                    start, bytes, &trimmed);
3159         if (!ret)
3160                 *total_trimmed += trimmed;
3161
3162         mutex_lock(&ctl->cache_writeout_mutex);
3163         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3164         list_del(&trim_entry->list);
3165         mutex_unlock(&ctl->cache_writeout_mutex);
3166
3167         if (update) {
3168                 spin_lock(&space_info->lock);
3169                 spin_lock(&block_group->lock);
3170                 if (block_group->ro)
3171                         space_info->bytes_readonly += reserved_bytes;
3172                 block_group->reserved -= reserved_bytes;
3173                 space_info->bytes_reserved -= reserved_bytes;
3174                 spin_unlock(&space_info->lock);
3175                 spin_unlock(&block_group->lock);
3176         }
3177
3178         return ret;
3179 }
3180
3181 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3182                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3183 {
3184         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3185         struct btrfs_free_space *entry;
3186         struct rb_node *node;
3187         int ret = 0;
3188         u64 extent_start;
3189         u64 extent_bytes;
3190         u64 bytes;
3191
3192         while (start < end) {
3193                 struct btrfs_trim_range trim_entry;
3194
3195                 mutex_lock(&ctl->cache_writeout_mutex);
3196                 spin_lock(&ctl->tree_lock);
3197
3198                 if (ctl->free_space < minlen) {
3199                         spin_unlock(&ctl->tree_lock);
3200                         mutex_unlock(&ctl->cache_writeout_mutex);
3201                         break;
3202                 }
3203
3204                 entry = tree_search_offset(ctl, start, 0, 1);
3205                 if (!entry) {
3206                         spin_unlock(&ctl->tree_lock);
3207                         mutex_unlock(&ctl->cache_writeout_mutex);
3208                         break;
3209                 }
3210
3211                 /* skip bitmaps */
3212                 while (entry->bitmap) {
3213                         node = rb_next(&entry->offset_index);
3214                         if (!node) {
3215                                 spin_unlock(&ctl->tree_lock);
3216                                 mutex_unlock(&ctl->cache_writeout_mutex);
3217                                 goto out;
3218                         }
3219                         entry = rb_entry(node, struct btrfs_free_space,
3220                                          offset_index);
3221                 }
3222
3223                 if (entry->offset >= end) {
3224                         spin_unlock(&ctl->tree_lock);
3225                         mutex_unlock(&ctl->cache_writeout_mutex);
3226                         break;
3227                 }
3228
3229                 extent_start = entry->offset;
3230                 extent_bytes = entry->bytes;
3231                 start = max(start, extent_start);
3232                 bytes = min(extent_start + extent_bytes, end) - start;
3233                 if (bytes < minlen) {
3234                         spin_unlock(&ctl->tree_lock);
3235                         mutex_unlock(&ctl->cache_writeout_mutex);
3236                         goto next;
3237                 }
3238
3239                 unlink_free_space(ctl, entry);
3240                 kmem_cache_free(btrfs_free_space_cachep, entry);
3241
3242                 spin_unlock(&ctl->tree_lock);
3243                 trim_entry.start = extent_start;
3244                 trim_entry.bytes = extent_bytes;
3245                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3246                 mutex_unlock(&ctl->cache_writeout_mutex);
3247
3248                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3249                                   extent_start, extent_bytes, &trim_entry);
3250                 if (ret)
3251                         break;
3252 next:
3253                 start += bytes;
3254
3255                 if (fatal_signal_pending(current)) {
3256                         ret = -ERESTARTSYS;
3257                         break;
3258                 }
3259
3260                 cond_resched();
3261         }
3262 out:
3263         return ret;
3264 }
3265
3266 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3267                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3268 {
3269         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3270         struct btrfs_free_space *entry;
3271         int ret = 0;
3272         int ret2;
3273         u64 bytes;
3274         u64 offset = offset_to_bitmap(ctl, start);
3275
3276         while (offset < end) {
3277                 bool next_bitmap = false;
3278                 struct btrfs_trim_range trim_entry;
3279
3280                 mutex_lock(&ctl->cache_writeout_mutex);
3281                 spin_lock(&ctl->tree_lock);
3282
3283                 if (ctl->free_space < minlen) {
3284                         spin_unlock(&ctl->tree_lock);
3285                         mutex_unlock(&ctl->cache_writeout_mutex);
3286                         break;
3287                 }
3288
3289                 entry = tree_search_offset(ctl, offset, 1, 0);
3290                 if (!entry) {
3291                         spin_unlock(&ctl->tree_lock);
3292                         mutex_unlock(&ctl->cache_writeout_mutex);
3293                         next_bitmap = true;
3294                         goto next;
3295                 }
3296
3297                 bytes = minlen;
3298                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3299                 if (ret2 || start >= end) {
3300                         spin_unlock(&ctl->tree_lock);
3301                         mutex_unlock(&ctl->cache_writeout_mutex);
3302                         next_bitmap = true;
3303                         goto next;
3304                 }
3305
3306                 bytes = min(bytes, end - start);
3307                 if (bytes < minlen) {
3308                         spin_unlock(&ctl->tree_lock);
3309                         mutex_unlock(&ctl->cache_writeout_mutex);
3310                         goto next;
3311                 }
3312
3313                 bitmap_clear_bits(ctl, entry, start, bytes);
3314                 if (entry->bytes == 0)
3315                         free_bitmap(ctl, entry);
3316
3317                 spin_unlock(&ctl->tree_lock);
3318                 trim_entry.start = start;
3319                 trim_entry.bytes = bytes;
3320                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3321                 mutex_unlock(&ctl->cache_writeout_mutex);
3322
3323                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3324                                   start, bytes, &trim_entry);
3325                 if (ret)
3326                         break;
3327 next:
3328                 if (next_bitmap) {
3329                         offset += BITS_PER_BITMAP * ctl->unit;
3330                 } else {
3331                         start += bytes;
3332                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3333                                 offset += BITS_PER_BITMAP * ctl->unit;
3334                 }
3335
3336                 if (fatal_signal_pending(current)) {
3337                         ret = -ERESTARTSYS;
3338                         break;
3339                 }
3340
3341                 cond_resched();
3342         }
3343
3344         return ret;
3345 }
3346
3347 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3348 {
3349         atomic_inc(&cache->trimming);
3350 }
3351
3352 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3353 {
3354         struct extent_map_tree *em_tree;
3355         struct extent_map *em;
3356         bool cleanup;
3357
3358         spin_lock(&block_group->lock);
3359         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3360                    block_group->removed);
3361         spin_unlock(&block_group->lock);
3362
3363         if (cleanup) {
3364                 lock_chunks(block_group->fs_info->chunk_root);
3365                 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3366                 write_lock(&em_tree->lock);
3367                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3368                                            1);
3369                 BUG_ON(!em); /* logic error, can't happen */
3370                 /*
3371                  * remove_extent_mapping() will delete us from the pinned_chunks
3372                  * list, which is protected by the chunk mutex.
3373                  */
3374                 remove_extent_mapping(em_tree, em);
3375                 write_unlock(&em_tree->lock);
3376                 unlock_chunks(block_group->fs_info->chunk_root);
3377
3378                 /* once for us and once for the tree */
3379                 free_extent_map(em);
3380                 free_extent_map(em);
3381
3382                 /*
3383                  * We've left one free space entry and other tasks trimming
3384                  * this block group have left 1 entry each one. Free them.
3385                  */
3386                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3387         }
3388 }
3389
3390 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3391                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3392 {
3393         int ret;
3394
3395         *trimmed = 0;
3396
3397         spin_lock(&block_group->lock);
3398         if (block_group->removed) {
3399                 spin_unlock(&block_group->lock);
3400                 return 0;
3401         }
3402         btrfs_get_block_group_trimming(block_group);
3403         spin_unlock(&block_group->lock);
3404
3405         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3406         if (ret)
3407                 goto out;
3408
3409         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3410 out:
3411         btrfs_put_block_group_trimming(block_group);
3412         return ret;
3413 }
3414
3415 /*
3416  * Find the left-most item in the cache tree, and then return the
3417  * smallest inode number in the item.
3418  *
3419  * Note: the returned inode number may not be the smallest one in
3420  * the tree, if the left-most item is a bitmap.
3421  */
3422 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3423 {
3424         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3425         struct btrfs_free_space *entry = NULL;
3426         u64 ino = 0;
3427
3428         spin_lock(&ctl->tree_lock);
3429
3430         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3431                 goto out;
3432
3433         entry = rb_entry(rb_first(&ctl->free_space_offset),
3434                          struct btrfs_free_space, offset_index);
3435
3436         if (!entry->bitmap) {
3437                 ino = entry->offset;
3438
3439                 unlink_free_space(ctl, entry);
3440                 entry->offset++;
3441                 entry->bytes--;
3442                 if (!entry->bytes)
3443                         kmem_cache_free(btrfs_free_space_cachep, entry);
3444                 else
3445                         link_free_space(ctl, entry);
3446         } else {
3447                 u64 offset = 0;
3448                 u64 count = 1;
3449                 int ret;
3450
3451                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3452                 /* Logic error; Should be empty if it can't find anything */
3453                 ASSERT(!ret);
3454
3455                 ino = offset;
3456                 bitmap_clear_bits(ctl, entry, offset, 1);
3457                 if (entry->bytes == 0)
3458                         free_bitmap(ctl, entry);
3459         }
3460 out:
3461         spin_unlock(&ctl->tree_lock);
3462
3463         return ino;
3464 }
3465
3466 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3467                                     struct btrfs_path *path)
3468 {
3469         struct inode *inode = NULL;
3470
3471         spin_lock(&root->ino_cache_lock);
3472         if (root->ino_cache_inode)
3473                 inode = igrab(root->ino_cache_inode);
3474         spin_unlock(&root->ino_cache_lock);
3475         if (inode)
3476                 return inode;
3477
3478         inode = __lookup_free_space_inode(root, path, 0);
3479         if (IS_ERR(inode))
3480                 return inode;
3481
3482         spin_lock(&root->ino_cache_lock);
3483         if (!btrfs_fs_closing(root->fs_info))
3484                 root->ino_cache_inode = igrab(inode);
3485         spin_unlock(&root->ino_cache_lock);
3486
3487         return inode;
3488 }
3489
3490 int create_free_ino_inode(struct btrfs_root *root,
3491                           struct btrfs_trans_handle *trans,
3492                           struct btrfs_path *path)
3493 {
3494         return __create_free_space_inode(root, trans, path,
3495                                          BTRFS_FREE_INO_OBJECTID, 0);
3496 }
3497
3498 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3499 {
3500         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3501         struct btrfs_path *path;
3502         struct inode *inode;
3503         int ret = 0;
3504         u64 root_gen = btrfs_root_generation(&root->root_item);
3505
3506         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3507                 return 0;
3508
3509         /*
3510          * If we're unmounting then just return, since this does a search on the
3511          * normal root and not the commit root and we could deadlock.
3512          */
3513         if (btrfs_fs_closing(fs_info))
3514                 return 0;
3515
3516         path = btrfs_alloc_path();
3517         if (!path)
3518                 return 0;
3519
3520         inode = lookup_free_ino_inode(root, path);
3521         if (IS_ERR(inode))
3522                 goto out;
3523
3524         if (root_gen != BTRFS_I(inode)->generation)
3525                 goto out_put;
3526
3527         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3528
3529         if (ret < 0)
3530                 btrfs_err(fs_info,
3531                         "failed to load free ino cache for root %llu",
3532                         root->root_key.objectid);
3533 out_put:
3534         iput(inode);
3535 out:
3536         btrfs_free_path(path);
3537         return ret;
3538 }
3539
3540 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3541                               struct btrfs_trans_handle *trans,
3542                               struct btrfs_path *path,
3543                               struct inode *inode)
3544 {
3545         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3546         int ret;
3547         struct btrfs_io_ctl io_ctl;
3548         bool release_metadata = true;
3549
3550         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3551                 return 0;
3552
3553         memset(&io_ctl, 0, sizeof(io_ctl));
3554         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3555                                       trans, path, 0);
3556         if (!ret) {
3557                 /*
3558                  * At this point writepages() didn't error out, so our metadata
3559                  * reservation is released when the writeback finishes, at
3560                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3561                  * with or without an error.
3562                  */
3563                 release_metadata = false;
3564                 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3565         }
3566
3567         if (ret) {
3568                 if (release_metadata)
3569                         btrfs_delalloc_release_metadata(inode, inode->i_size);
3570 #ifdef DEBUG
3571                 btrfs_err(root->fs_info,
3572                         "failed to write free ino cache for root %llu",
3573                         root->root_key.objectid);
3574 #endif
3575         }
3576
3577         return ret;
3578 }
3579
3580 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3581 /*
3582  * Use this if you need to make a bitmap or extent entry specifically, it
3583  * doesn't do any of the merging that add_free_space does, this acts a lot like
3584  * how the free space cache loading stuff works, so you can get really weird
3585  * configurations.
3586  */
3587 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3588                               u64 offset, u64 bytes, bool bitmap)
3589 {
3590         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3591         struct btrfs_free_space *info = NULL, *bitmap_info;
3592         void *map = NULL;
3593         u64 bytes_added;
3594         int ret;
3595
3596 again:
3597         if (!info) {
3598                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3599                 if (!info)
3600                         return -ENOMEM;
3601         }
3602
3603         if (!bitmap) {
3604                 spin_lock(&ctl->tree_lock);
3605                 info->offset = offset;
3606                 info->bytes = bytes;
3607                 info->max_extent_size = 0;
3608                 ret = link_free_space(ctl, info);
3609                 spin_unlock(&ctl->tree_lock);
3610                 if (ret)
3611                         kmem_cache_free(btrfs_free_space_cachep, info);
3612                 return ret;
3613         }
3614
3615         if (!map) {
3616                 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3617                 if (!map) {
3618                         kmem_cache_free(btrfs_free_space_cachep, info);
3619                         return -ENOMEM;
3620                 }
3621         }
3622
3623         spin_lock(&ctl->tree_lock);
3624         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3625                                          1, 0);
3626         if (!bitmap_info) {
3627                 info->bitmap = map;
3628                 map = NULL;
3629                 add_new_bitmap(ctl, info, offset);
3630                 bitmap_info = info;
3631                 info = NULL;
3632         }
3633
3634         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3635
3636         bytes -= bytes_added;
3637         offset += bytes_added;
3638         spin_unlock(&ctl->tree_lock);
3639
3640         if (bytes)
3641                 goto again;
3642
3643         if (info)
3644                 kmem_cache_free(btrfs_free_space_cachep, info);
3645         if (map)
3646                 kfree(map);
3647         return 0;
3648 }
3649
3650 /*
3651  * Checks to see if the given range is in the free space cache.  This is really
3652  * just used to check the absence of space, so if there is free space in the
3653  * range at all we will return 1.
3654  */
3655 int test_check_exists(struct btrfs_block_group_cache *cache,
3656                       u64 offset, u64 bytes)
3657 {
3658         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3659         struct btrfs_free_space *info;
3660         int ret = 0;
3661
3662         spin_lock(&ctl->tree_lock);
3663         info = tree_search_offset(ctl, offset, 0, 0);
3664         if (!info) {
3665                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3666                                           1, 0);
3667                 if (!info)
3668                         goto out;
3669         }
3670
3671 have_info:
3672         if (info->bitmap) {
3673                 u64 bit_off, bit_bytes;
3674                 struct rb_node *n;
3675                 struct btrfs_free_space *tmp;
3676
3677                 bit_off = offset;
3678                 bit_bytes = ctl->unit;
3679                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3680                 if (!ret) {
3681                         if (bit_off == offset) {
3682                                 ret = 1;
3683                                 goto out;
3684                         } else if (bit_off > offset &&
3685                                    offset + bytes > bit_off) {
3686                                 ret = 1;
3687                                 goto out;
3688                         }
3689                 }
3690
3691                 n = rb_prev(&info->offset_index);
3692                 while (n) {
3693                         tmp = rb_entry(n, struct btrfs_free_space,
3694                                        offset_index);
3695                         if (tmp->offset + tmp->bytes < offset)
3696                                 break;
3697                         if (offset + bytes < tmp->offset) {
3698                                 n = rb_prev(&info->offset_index);
3699                                 continue;
3700                         }
3701                         info = tmp;
3702                         goto have_info;
3703                 }
3704
3705                 n = rb_next(&info->offset_index);
3706                 while (n) {
3707                         tmp = rb_entry(n, struct btrfs_free_space,
3708                                        offset_index);
3709                         if (offset + bytes < tmp->offset)
3710                                 break;
3711                         if (tmp->offset + tmp->bytes < offset) {
3712                                 n = rb_next(&info->offset_index);
3713                                 continue;
3714                         }
3715                         info = tmp;
3716                         goto have_info;
3717                 }
3718
3719                 ret = 0;
3720                 goto out;
3721         }
3722
3723         if (info->offset == offset) {
3724                 ret = 1;
3725                 goto out;
3726         }
3727
3728         if (offset > info->offset && offset < info->offset + info->bytes)
3729                 ret = 1;
3730 out:
3731         spin_unlock(&ctl->tree_lock);
3732         return ret;
3733 }
3734 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */