btrfs: protect space cache inode alloc with GFP_NOFS
[sfrench/cifs-2.6.git] / fs / btrfs / free-space-cache.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2008 Red Hat.  All rights reserved.
4  */
5
6 #include <linux/pagemap.h>
7 #include <linux/sched.h>
8 #include <linux/sched/signal.h>
9 #include <linux/slab.h>
10 #include <linux/math64.h>
11 #include <linux/ratelimit.h>
12 #include <linux/error-injection.h>
13 #include <linux/sched/mm.h>
14 #include "ctree.h"
15 #include "free-space-cache.h"
16 #include "transaction.h"
17 #include "disk-io.h"
18 #include "extent_io.h"
19 #include "inode-map.h"
20 #include "volumes.h"
21
22 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
23 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
24
25 struct btrfs_trim_range {
26         u64 start;
27         u64 bytes;
28         struct list_head list;
29 };
30
31 static int link_free_space(struct btrfs_free_space_ctl *ctl,
32                            struct btrfs_free_space *info);
33 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
34                               struct btrfs_free_space *info);
35 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
36                              struct btrfs_trans_handle *trans,
37                              struct btrfs_io_ctl *io_ctl,
38                              struct btrfs_path *path);
39
40 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
41                                                struct btrfs_path *path,
42                                                u64 offset)
43 {
44         struct btrfs_fs_info *fs_info = root->fs_info;
45         struct btrfs_key key;
46         struct btrfs_key location;
47         struct btrfs_disk_key disk_key;
48         struct btrfs_free_space_header *header;
49         struct extent_buffer *leaf;
50         struct inode *inode = NULL;
51         unsigned nofs_flag;
52         int ret;
53
54         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
55         key.offset = offset;
56         key.type = 0;
57
58         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
59         if (ret < 0)
60                 return ERR_PTR(ret);
61         if (ret > 0) {
62                 btrfs_release_path(path);
63                 return ERR_PTR(-ENOENT);
64         }
65
66         leaf = path->nodes[0];
67         header = btrfs_item_ptr(leaf, path->slots[0],
68                                 struct btrfs_free_space_header);
69         btrfs_free_space_key(leaf, header, &disk_key);
70         btrfs_disk_key_to_cpu(&location, &disk_key);
71         btrfs_release_path(path);
72
73         /*
74          * We are often under a trans handle at this point, so we need to make
75          * sure NOFS is set to keep us from deadlocking.
76          */
77         nofs_flag = memalloc_nofs_save();
78         inode = btrfs_iget(fs_info->sb, &location, root, NULL);
79         memalloc_nofs_restore(nofs_flag);
80         if (IS_ERR(inode))
81                 return inode;
82
83         mapping_set_gfp_mask(inode->i_mapping,
84                         mapping_gfp_constraint(inode->i_mapping,
85                         ~(__GFP_FS | __GFP_HIGHMEM)));
86
87         return inode;
88 }
89
90 struct inode *lookup_free_space_inode(struct btrfs_fs_info *fs_info,
91                                       struct btrfs_block_group_cache
92                                       *block_group, struct btrfs_path *path)
93 {
94         struct inode *inode = NULL;
95         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
96
97         spin_lock(&block_group->lock);
98         if (block_group->inode)
99                 inode = igrab(block_group->inode);
100         spin_unlock(&block_group->lock);
101         if (inode)
102                 return inode;
103
104         inode = __lookup_free_space_inode(fs_info->tree_root, path,
105                                           block_group->key.objectid);
106         if (IS_ERR(inode))
107                 return inode;
108
109         spin_lock(&block_group->lock);
110         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
111                 btrfs_info(fs_info, "Old style space inode found, converting.");
112                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
113                         BTRFS_INODE_NODATACOW;
114                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
115         }
116
117         if (!block_group->iref) {
118                 block_group->inode = igrab(inode);
119                 block_group->iref = 1;
120         }
121         spin_unlock(&block_group->lock);
122
123         return inode;
124 }
125
126 static int __create_free_space_inode(struct btrfs_root *root,
127                                      struct btrfs_trans_handle *trans,
128                                      struct btrfs_path *path,
129                                      u64 ino, u64 offset)
130 {
131         struct btrfs_key key;
132         struct btrfs_disk_key disk_key;
133         struct btrfs_free_space_header *header;
134         struct btrfs_inode_item *inode_item;
135         struct extent_buffer *leaf;
136         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
137         int ret;
138
139         ret = btrfs_insert_empty_inode(trans, root, path, ino);
140         if (ret)
141                 return ret;
142
143         /* We inline crc's for the free disk space cache */
144         if (ino != BTRFS_FREE_INO_OBJECTID)
145                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
146
147         leaf = path->nodes[0];
148         inode_item = btrfs_item_ptr(leaf, path->slots[0],
149                                     struct btrfs_inode_item);
150         btrfs_item_key(leaf, &disk_key, path->slots[0]);
151         memzero_extent_buffer(leaf, (unsigned long)inode_item,
152                              sizeof(*inode_item));
153         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
154         btrfs_set_inode_size(leaf, inode_item, 0);
155         btrfs_set_inode_nbytes(leaf, inode_item, 0);
156         btrfs_set_inode_uid(leaf, inode_item, 0);
157         btrfs_set_inode_gid(leaf, inode_item, 0);
158         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
159         btrfs_set_inode_flags(leaf, inode_item, flags);
160         btrfs_set_inode_nlink(leaf, inode_item, 1);
161         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
162         btrfs_set_inode_block_group(leaf, inode_item, offset);
163         btrfs_mark_buffer_dirty(leaf);
164         btrfs_release_path(path);
165
166         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
167         key.offset = offset;
168         key.type = 0;
169         ret = btrfs_insert_empty_item(trans, root, path, &key,
170                                       sizeof(struct btrfs_free_space_header));
171         if (ret < 0) {
172                 btrfs_release_path(path);
173                 return ret;
174         }
175
176         leaf = path->nodes[0];
177         header = btrfs_item_ptr(leaf, path->slots[0],
178                                 struct btrfs_free_space_header);
179         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
180         btrfs_set_free_space_key(leaf, header, &disk_key);
181         btrfs_mark_buffer_dirty(leaf);
182         btrfs_release_path(path);
183
184         return 0;
185 }
186
187 int create_free_space_inode(struct btrfs_fs_info *fs_info,
188                             struct btrfs_trans_handle *trans,
189                             struct btrfs_block_group_cache *block_group,
190                             struct btrfs_path *path)
191 {
192         int ret;
193         u64 ino;
194
195         ret = btrfs_find_free_objectid(fs_info->tree_root, &ino);
196         if (ret < 0)
197                 return ret;
198
199         return __create_free_space_inode(fs_info->tree_root, trans, path, ino,
200                                          block_group->key.objectid);
201 }
202
203 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
204                                        struct btrfs_block_rsv *rsv)
205 {
206         u64 needed_bytes;
207         int ret;
208
209         /* 1 for slack space, 1 for updating the inode */
210         needed_bytes = btrfs_calc_trunc_metadata_size(fs_info, 1) +
211                 btrfs_calc_trans_metadata_size(fs_info, 1);
212
213         spin_lock(&rsv->lock);
214         if (rsv->reserved < needed_bytes)
215                 ret = -ENOSPC;
216         else
217                 ret = 0;
218         spin_unlock(&rsv->lock);
219         return ret;
220 }
221
222 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
223                                     struct btrfs_block_group_cache *block_group,
224                                     struct inode *inode)
225 {
226         struct btrfs_root *root = BTRFS_I(inode)->root;
227         int ret = 0;
228         bool locked = false;
229
230         if (block_group) {
231                 struct btrfs_path *path = btrfs_alloc_path();
232
233                 if (!path) {
234                         ret = -ENOMEM;
235                         goto fail;
236                 }
237                 locked = true;
238                 mutex_lock(&trans->transaction->cache_write_mutex);
239                 if (!list_empty(&block_group->io_list)) {
240                         list_del_init(&block_group->io_list);
241
242                         btrfs_wait_cache_io(trans, block_group, path);
243                         btrfs_put_block_group(block_group);
244                 }
245
246                 /*
247                  * now that we've truncated the cache away, its no longer
248                  * setup or written
249                  */
250                 spin_lock(&block_group->lock);
251                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
252                 spin_unlock(&block_group->lock);
253                 btrfs_free_path(path);
254         }
255
256         btrfs_i_size_write(BTRFS_I(inode), 0);
257         truncate_pagecache(inode, 0);
258
259         /*
260          * We skip the throttling logic for free space cache inodes, so we don't
261          * need to check for -EAGAIN.
262          */
263         ret = btrfs_truncate_inode_items(trans, root, inode,
264                                          0, BTRFS_EXTENT_DATA_KEY);
265         if (ret)
266                 goto fail;
267
268         ret = btrfs_update_inode(trans, root, inode);
269
270 fail:
271         if (locked)
272                 mutex_unlock(&trans->transaction->cache_write_mutex);
273         if (ret)
274                 btrfs_abort_transaction(trans, ret);
275
276         return ret;
277 }
278
279 static void readahead_cache(struct inode *inode)
280 {
281         struct file_ra_state *ra;
282         unsigned long last_index;
283
284         ra = kzalloc(sizeof(*ra), GFP_NOFS);
285         if (!ra)
286                 return;
287
288         file_ra_state_init(ra, inode->i_mapping);
289         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
290
291         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
292
293         kfree(ra);
294 }
295
296 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
297                        int write)
298 {
299         int num_pages;
300         int check_crcs = 0;
301
302         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
303
304         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
305                 check_crcs = 1;
306
307         /* Make sure we can fit our crcs and generation into the first page */
308         if (write && check_crcs &&
309             (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
310                 return -ENOSPC;
311
312         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
313
314         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
315         if (!io_ctl->pages)
316                 return -ENOMEM;
317
318         io_ctl->num_pages = num_pages;
319         io_ctl->fs_info = btrfs_sb(inode->i_sb);
320         io_ctl->check_crcs = check_crcs;
321         io_ctl->inode = inode;
322
323         return 0;
324 }
325 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
326
327 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
328 {
329         kfree(io_ctl->pages);
330         io_ctl->pages = NULL;
331 }
332
333 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
334 {
335         if (io_ctl->cur) {
336                 io_ctl->cur = NULL;
337                 io_ctl->orig = NULL;
338         }
339 }
340
341 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
342 {
343         ASSERT(io_ctl->index < io_ctl->num_pages);
344         io_ctl->page = io_ctl->pages[io_ctl->index++];
345         io_ctl->cur = page_address(io_ctl->page);
346         io_ctl->orig = io_ctl->cur;
347         io_ctl->size = PAGE_SIZE;
348         if (clear)
349                 clear_page(io_ctl->cur);
350 }
351
352 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
353 {
354         int i;
355
356         io_ctl_unmap_page(io_ctl);
357
358         for (i = 0; i < io_ctl->num_pages; i++) {
359                 if (io_ctl->pages[i]) {
360                         ClearPageChecked(io_ctl->pages[i]);
361                         unlock_page(io_ctl->pages[i]);
362                         put_page(io_ctl->pages[i]);
363                 }
364         }
365 }
366
367 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
368                                 int uptodate)
369 {
370         struct page *page;
371         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
372         int i;
373
374         for (i = 0; i < io_ctl->num_pages; i++) {
375                 page = find_or_create_page(inode->i_mapping, i, mask);
376                 if (!page) {
377                         io_ctl_drop_pages(io_ctl);
378                         return -ENOMEM;
379                 }
380                 io_ctl->pages[i] = page;
381                 if (uptodate && !PageUptodate(page)) {
382                         btrfs_readpage(NULL, page);
383                         lock_page(page);
384                         if (!PageUptodate(page)) {
385                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
386                                            "error reading free space cache");
387                                 io_ctl_drop_pages(io_ctl);
388                                 return -EIO;
389                         }
390                 }
391         }
392
393         for (i = 0; i < io_ctl->num_pages; i++) {
394                 clear_page_dirty_for_io(io_ctl->pages[i]);
395                 set_page_extent_mapped(io_ctl->pages[i]);
396         }
397
398         return 0;
399 }
400
401 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
402 {
403         __le64 *val;
404
405         io_ctl_map_page(io_ctl, 1);
406
407         /*
408          * Skip the csum areas.  If we don't check crcs then we just have a
409          * 64bit chunk at the front of the first page.
410          */
411         if (io_ctl->check_crcs) {
412                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
413                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
414         } else {
415                 io_ctl->cur += sizeof(u64);
416                 io_ctl->size -= sizeof(u64) * 2;
417         }
418
419         val = io_ctl->cur;
420         *val = cpu_to_le64(generation);
421         io_ctl->cur += sizeof(u64);
422 }
423
424 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
425 {
426         __le64 *gen;
427
428         /*
429          * Skip the crc area.  If we don't check crcs then we just have a 64bit
430          * chunk at the front of the first page.
431          */
432         if (io_ctl->check_crcs) {
433                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
434                 io_ctl->size -= sizeof(u64) +
435                         (sizeof(u32) * io_ctl->num_pages);
436         } else {
437                 io_ctl->cur += sizeof(u64);
438                 io_ctl->size -= sizeof(u64) * 2;
439         }
440
441         gen = io_ctl->cur;
442         if (le64_to_cpu(*gen) != generation) {
443                 btrfs_err_rl(io_ctl->fs_info,
444                         "space cache generation (%llu) does not match inode (%llu)",
445                                 *gen, generation);
446                 io_ctl_unmap_page(io_ctl);
447                 return -EIO;
448         }
449         io_ctl->cur += sizeof(u64);
450         return 0;
451 }
452
453 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
454 {
455         u32 *tmp;
456         u32 crc = ~(u32)0;
457         unsigned offset = 0;
458
459         if (!io_ctl->check_crcs) {
460                 io_ctl_unmap_page(io_ctl);
461                 return;
462         }
463
464         if (index == 0)
465                 offset = sizeof(u32) * io_ctl->num_pages;
466
467         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
468                               PAGE_SIZE - offset);
469         btrfs_csum_final(crc, (u8 *)&crc);
470         io_ctl_unmap_page(io_ctl);
471         tmp = page_address(io_ctl->pages[0]);
472         tmp += index;
473         *tmp = crc;
474 }
475
476 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
477 {
478         u32 *tmp, val;
479         u32 crc = ~(u32)0;
480         unsigned offset = 0;
481
482         if (!io_ctl->check_crcs) {
483                 io_ctl_map_page(io_ctl, 0);
484                 return 0;
485         }
486
487         if (index == 0)
488                 offset = sizeof(u32) * io_ctl->num_pages;
489
490         tmp = page_address(io_ctl->pages[0]);
491         tmp += index;
492         val = *tmp;
493
494         io_ctl_map_page(io_ctl, 0);
495         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
496                               PAGE_SIZE - offset);
497         btrfs_csum_final(crc, (u8 *)&crc);
498         if (val != crc) {
499                 btrfs_err_rl(io_ctl->fs_info,
500                         "csum mismatch on free space cache");
501                 io_ctl_unmap_page(io_ctl);
502                 return -EIO;
503         }
504
505         return 0;
506 }
507
508 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
509                             void *bitmap)
510 {
511         struct btrfs_free_space_entry *entry;
512
513         if (!io_ctl->cur)
514                 return -ENOSPC;
515
516         entry = io_ctl->cur;
517         entry->offset = cpu_to_le64(offset);
518         entry->bytes = cpu_to_le64(bytes);
519         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
520                 BTRFS_FREE_SPACE_EXTENT;
521         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
522         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
523
524         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
525                 return 0;
526
527         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
528
529         /* No more pages to map */
530         if (io_ctl->index >= io_ctl->num_pages)
531                 return 0;
532
533         /* map the next page */
534         io_ctl_map_page(io_ctl, 1);
535         return 0;
536 }
537
538 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
539 {
540         if (!io_ctl->cur)
541                 return -ENOSPC;
542
543         /*
544          * If we aren't at the start of the current page, unmap this one and
545          * map the next one if there is any left.
546          */
547         if (io_ctl->cur != io_ctl->orig) {
548                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
549                 if (io_ctl->index >= io_ctl->num_pages)
550                         return -ENOSPC;
551                 io_ctl_map_page(io_ctl, 0);
552         }
553
554         copy_page(io_ctl->cur, bitmap);
555         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
556         if (io_ctl->index < io_ctl->num_pages)
557                 io_ctl_map_page(io_ctl, 0);
558         return 0;
559 }
560
561 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
562 {
563         /*
564          * If we're not on the boundary we know we've modified the page and we
565          * need to crc the page.
566          */
567         if (io_ctl->cur != io_ctl->orig)
568                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
569         else
570                 io_ctl_unmap_page(io_ctl);
571
572         while (io_ctl->index < io_ctl->num_pages) {
573                 io_ctl_map_page(io_ctl, 1);
574                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
575         }
576 }
577
578 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
579                             struct btrfs_free_space *entry, u8 *type)
580 {
581         struct btrfs_free_space_entry *e;
582         int ret;
583
584         if (!io_ctl->cur) {
585                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
586                 if (ret)
587                         return ret;
588         }
589
590         e = io_ctl->cur;
591         entry->offset = le64_to_cpu(e->offset);
592         entry->bytes = le64_to_cpu(e->bytes);
593         *type = e->type;
594         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
595         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
596
597         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
598                 return 0;
599
600         io_ctl_unmap_page(io_ctl);
601
602         return 0;
603 }
604
605 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
606                               struct btrfs_free_space *entry)
607 {
608         int ret;
609
610         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
611         if (ret)
612                 return ret;
613
614         copy_page(entry->bitmap, io_ctl->cur);
615         io_ctl_unmap_page(io_ctl);
616
617         return 0;
618 }
619
620 /*
621  * Since we attach pinned extents after the fact we can have contiguous sections
622  * of free space that are split up in entries.  This poses a problem with the
623  * tree logging stuff since it could have allocated across what appears to be 2
624  * entries since we would have merged the entries when adding the pinned extents
625  * back to the free space cache.  So run through the space cache that we just
626  * loaded and merge contiguous entries.  This will make the log replay stuff not
627  * blow up and it will make for nicer allocator behavior.
628  */
629 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
630 {
631         struct btrfs_free_space *e, *prev = NULL;
632         struct rb_node *n;
633
634 again:
635         spin_lock(&ctl->tree_lock);
636         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
637                 e = rb_entry(n, struct btrfs_free_space, offset_index);
638                 if (!prev)
639                         goto next;
640                 if (e->bitmap || prev->bitmap)
641                         goto next;
642                 if (prev->offset + prev->bytes == e->offset) {
643                         unlink_free_space(ctl, prev);
644                         unlink_free_space(ctl, e);
645                         prev->bytes += e->bytes;
646                         kmem_cache_free(btrfs_free_space_cachep, e);
647                         link_free_space(ctl, prev);
648                         prev = NULL;
649                         spin_unlock(&ctl->tree_lock);
650                         goto again;
651                 }
652 next:
653                 prev = e;
654         }
655         spin_unlock(&ctl->tree_lock);
656 }
657
658 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
659                                    struct btrfs_free_space_ctl *ctl,
660                                    struct btrfs_path *path, u64 offset)
661 {
662         struct btrfs_fs_info *fs_info = root->fs_info;
663         struct btrfs_free_space_header *header;
664         struct extent_buffer *leaf;
665         struct btrfs_io_ctl io_ctl;
666         struct btrfs_key key;
667         struct btrfs_free_space *e, *n;
668         LIST_HEAD(bitmaps);
669         u64 num_entries;
670         u64 num_bitmaps;
671         u64 generation;
672         u8 type;
673         int ret = 0;
674
675         /* Nothing in the space cache, goodbye */
676         if (!i_size_read(inode))
677                 return 0;
678
679         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
680         key.offset = offset;
681         key.type = 0;
682
683         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
684         if (ret < 0)
685                 return 0;
686         else if (ret > 0) {
687                 btrfs_release_path(path);
688                 return 0;
689         }
690
691         ret = -1;
692
693         leaf = path->nodes[0];
694         header = btrfs_item_ptr(leaf, path->slots[0],
695                                 struct btrfs_free_space_header);
696         num_entries = btrfs_free_space_entries(leaf, header);
697         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
698         generation = btrfs_free_space_generation(leaf, header);
699         btrfs_release_path(path);
700
701         if (!BTRFS_I(inode)->generation) {
702                 btrfs_info(fs_info,
703                            "the free space cache file (%llu) is invalid, skip it",
704                            offset);
705                 return 0;
706         }
707
708         if (BTRFS_I(inode)->generation != generation) {
709                 btrfs_err(fs_info,
710                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
711                           BTRFS_I(inode)->generation, generation);
712                 return 0;
713         }
714
715         if (!num_entries)
716                 return 0;
717
718         ret = io_ctl_init(&io_ctl, inode, 0);
719         if (ret)
720                 return ret;
721
722         readahead_cache(inode);
723
724         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
725         if (ret)
726                 goto out;
727
728         ret = io_ctl_check_crc(&io_ctl, 0);
729         if (ret)
730                 goto free_cache;
731
732         ret = io_ctl_check_generation(&io_ctl, generation);
733         if (ret)
734                 goto free_cache;
735
736         while (num_entries) {
737                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
738                                       GFP_NOFS);
739                 if (!e)
740                         goto free_cache;
741
742                 ret = io_ctl_read_entry(&io_ctl, e, &type);
743                 if (ret) {
744                         kmem_cache_free(btrfs_free_space_cachep, e);
745                         goto free_cache;
746                 }
747
748                 if (!e->bytes) {
749                         kmem_cache_free(btrfs_free_space_cachep, e);
750                         goto free_cache;
751                 }
752
753                 if (type == BTRFS_FREE_SPACE_EXTENT) {
754                         spin_lock(&ctl->tree_lock);
755                         ret = link_free_space(ctl, e);
756                         spin_unlock(&ctl->tree_lock);
757                         if (ret) {
758                                 btrfs_err(fs_info,
759                                         "Duplicate entries in free space cache, dumping");
760                                 kmem_cache_free(btrfs_free_space_cachep, e);
761                                 goto free_cache;
762                         }
763                 } else {
764                         ASSERT(num_bitmaps);
765                         num_bitmaps--;
766                         e->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
767                         if (!e->bitmap) {
768                                 kmem_cache_free(
769                                         btrfs_free_space_cachep, e);
770                                 goto free_cache;
771                         }
772                         spin_lock(&ctl->tree_lock);
773                         ret = link_free_space(ctl, e);
774                         ctl->total_bitmaps++;
775                         ctl->op->recalc_thresholds(ctl);
776                         spin_unlock(&ctl->tree_lock);
777                         if (ret) {
778                                 btrfs_err(fs_info,
779                                         "Duplicate entries in free space cache, dumping");
780                                 kmem_cache_free(btrfs_free_space_cachep, e);
781                                 goto free_cache;
782                         }
783                         list_add_tail(&e->list, &bitmaps);
784                 }
785
786                 num_entries--;
787         }
788
789         io_ctl_unmap_page(&io_ctl);
790
791         /*
792          * We add the bitmaps at the end of the entries in order that
793          * the bitmap entries are added to the cache.
794          */
795         list_for_each_entry_safe(e, n, &bitmaps, list) {
796                 list_del_init(&e->list);
797                 ret = io_ctl_read_bitmap(&io_ctl, e);
798                 if (ret)
799                         goto free_cache;
800         }
801
802         io_ctl_drop_pages(&io_ctl);
803         merge_space_tree(ctl);
804         ret = 1;
805 out:
806         io_ctl_free(&io_ctl);
807         return ret;
808 free_cache:
809         io_ctl_drop_pages(&io_ctl);
810         __btrfs_remove_free_space_cache(ctl);
811         goto out;
812 }
813
814 int load_free_space_cache(struct btrfs_fs_info *fs_info,
815                           struct btrfs_block_group_cache *block_group)
816 {
817         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
818         struct inode *inode;
819         struct btrfs_path *path;
820         int ret = 0;
821         bool matched;
822         u64 used = btrfs_block_group_used(&block_group->item);
823
824         /*
825          * If this block group has been marked to be cleared for one reason or
826          * another then we can't trust the on disk cache, so just return.
827          */
828         spin_lock(&block_group->lock);
829         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
830                 spin_unlock(&block_group->lock);
831                 return 0;
832         }
833         spin_unlock(&block_group->lock);
834
835         path = btrfs_alloc_path();
836         if (!path)
837                 return 0;
838         path->search_commit_root = 1;
839         path->skip_locking = 1;
840
841         inode = lookup_free_space_inode(fs_info, block_group, path);
842         if (IS_ERR(inode)) {
843                 btrfs_free_path(path);
844                 return 0;
845         }
846
847         /* We may have converted the inode and made the cache invalid. */
848         spin_lock(&block_group->lock);
849         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
850                 spin_unlock(&block_group->lock);
851                 btrfs_free_path(path);
852                 goto out;
853         }
854         spin_unlock(&block_group->lock);
855
856         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
857                                       path, block_group->key.objectid);
858         btrfs_free_path(path);
859         if (ret <= 0)
860                 goto out;
861
862         spin_lock(&ctl->tree_lock);
863         matched = (ctl->free_space == (block_group->key.offset - used -
864                                        block_group->bytes_super));
865         spin_unlock(&ctl->tree_lock);
866
867         if (!matched) {
868                 __btrfs_remove_free_space_cache(ctl);
869                 btrfs_warn(fs_info,
870                            "block group %llu has wrong amount of free space",
871                            block_group->key.objectid);
872                 ret = -1;
873         }
874 out:
875         if (ret < 0) {
876                 /* This cache is bogus, make sure it gets cleared */
877                 spin_lock(&block_group->lock);
878                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
879                 spin_unlock(&block_group->lock);
880                 ret = 0;
881
882                 btrfs_warn(fs_info,
883                            "failed to load free space cache for block group %llu, rebuilding it now",
884                            block_group->key.objectid);
885         }
886
887         iput(inode);
888         return ret;
889 }
890
891 static noinline_for_stack
892 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
893                               struct btrfs_free_space_ctl *ctl,
894                               struct btrfs_block_group_cache *block_group,
895                               int *entries, int *bitmaps,
896                               struct list_head *bitmap_list)
897 {
898         int ret;
899         struct btrfs_free_cluster *cluster = NULL;
900         struct btrfs_free_cluster *cluster_locked = NULL;
901         struct rb_node *node = rb_first(&ctl->free_space_offset);
902         struct btrfs_trim_range *trim_entry;
903
904         /* Get the cluster for this block_group if it exists */
905         if (block_group && !list_empty(&block_group->cluster_list)) {
906                 cluster = list_entry(block_group->cluster_list.next,
907                                      struct btrfs_free_cluster,
908                                      block_group_list);
909         }
910
911         if (!node && cluster) {
912                 cluster_locked = cluster;
913                 spin_lock(&cluster_locked->lock);
914                 node = rb_first(&cluster->root);
915                 cluster = NULL;
916         }
917
918         /* Write out the extent entries */
919         while (node) {
920                 struct btrfs_free_space *e;
921
922                 e = rb_entry(node, struct btrfs_free_space, offset_index);
923                 *entries += 1;
924
925                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
926                                        e->bitmap);
927                 if (ret)
928                         goto fail;
929
930                 if (e->bitmap) {
931                         list_add_tail(&e->list, bitmap_list);
932                         *bitmaps += 1;
933                 }
934                 node = rb_next(node);
935                 if (!node && cluster) {
936                         node = rb_first(&cluster->root);
937                         cluster_locked = cluster;
938                         spin_lock(&cluster_locked->lock);
939                         cluster = NULL;
940                 }
941         }
942         if (cluster_locked) {
943                 spin_unlock(&cluster_locked->lock);
944                 cluster_locked = NULL;
945         }
946
947         /*
948          * Make sure we don't miss any range that was removed from our rbtree
949          * because trimming is running. Otherwise after a umount+mount (or crash
950          * after committing the transaction) we would leak free space and get
951          * an inconsistent free space cache report from fsck.
952          */
953         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
954                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
955                                        trim_entry->bytes, NULL);
956                 if (ret)
957                         goto fail;
958                 *entries += 1;
959         }
960
961         return 0;
962 fail:
963         if (cluster_locked)
964                 spin_unlock(&cluster_locked->lock);
965         return -ENOSPC;
966 }
967
968 static noinline_for_stack int
969 update_cache_item(struct btrfs_trans_handle *trans,
970                   struct btrfs_root *root,
971                   struct inode *inode,
972                   struct btrfs_path *path, u64 offset,
973                   int entries, int bitmaps)
974 {
975         struct btrfs_key key;
976         struct btrfs_free_space_header *header;
977         struct extent_buffer *leaf;
978         int ret;
979
980         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
981         key.offset = offset;
982         key.type = 0;
983
984         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
985         if (ret < 0) {
986                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
987                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
988                 goto fail;
989         }
990         leaf = path->nodes[0];
991         if (ret > 0) {
992                 struct btrfs_key found_key;
993                 ASSERT(path->slots[0]);
994                 path->slots[0]--;
995                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
996                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
997                     found_key.offset != offset) {
998                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
999                                          inode->i_size - 1,
1000                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1001                                          NULL);
1002                         btrfs_release_path(path);
1003                         goto fail;
1004                 }
1005         }
1006
1007         BTRFS_I(inode)->generation = trans->transid;
1008         header = btrfs_item_ptr(leaf, path->slots[0],
1009                                 struct btrfs_free_space_header);
1010         btrfs_set_free_space_entries(leaf, header, entries);
1011         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1012         btrfs_set_free_space_generation(leaf, header, trans->transid);
1013         btrfs_mark_buffer_dirty(leaf);
1014         btrfs_release_path(path);
1015
1016         return 0;
1017
1018 fail:
1019         return -1;
1020 }
1021
1022 static noinline_for_stack int
1023 write_pinned_extent_entries(struct btrfs_fs_info *fs_info,
1024                             struct btrfs_block_group_cache *block_group,
1025                             struct btrfs_io_ctl *io_ctl,
1026                             int *entries)
1027 {
1028         u64 start, extent_start, extent_end, len;
1029         struct extent_io_tree *unpin = NULL;
1030         int ret;
1031
1032         if (!block_group)
1033                 return 0;
1034
1035         /*
1036          * We want to add any pinned extents to our free space cache
1037          * so we don't leak the space
1038          *
1039          * We shouldn't have switched the pinned extents yet so this is the
1040          * right one
1041          */
1042         unpin = fs_info->pinned_extents;
1043
1044         start = block_group->key.objectid;
1045
1046         while (start < block_group->key.objectid + block_group->key.offset) {
1047                 ret = find_first_extent_bit(unpin, start,
1048                                             &extent_start, &extent_end,
1049                                             EXTENT_DIRTY, NULL);
1050                 if (ret)
1051                         return 0;
1052
1053                 /* This pinned extent is out of our range */
1054                 if (extent_start >= block_group->key.objectid +
1055                     block_group->key.offset)
1056                         return 0;
1057
1058                 extent_start = max(extent_start, start);
1059                 extent_end = min(block_group->key.objectid +
1060                                  block_group->key.offset, extent_end + 1);
1061                 len = extent_end - extent_start;
1062
1063                 *entries += 1;
1064                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1065                 if (ret)
1066                         return -ENOSPC;
1067
1068                 start = extent_end;
1069         }
1070
1071         return 0;
1072 }
1073
1074 static noinline_for_stack int
1075 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1076 {
1077         struct btrfs_free_space *entry, *next;
1078         int ret;
1079
1080         /* Write out the bitmaps */
1081         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1082                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1083                 if (ret)
1084                         return -ENOSPC;
1085                 list_del_init(&entry->list);
1086         }
1087
1088         return 0;
1089 }
1090
1091 static int flush_dirty_cache(struct inode *inode)
1092 {
1093         int ret;
1094
1095         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1096         if (ret)
1097                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1098                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
1099
1100         return ret;
1101 }
1102
1103 static void noinline_for_stack
1104 cleanup_bitmap_list(struct list_head *bitmap_list)
1105 {
1106         struct btrfs_free_space *entry, *next;
1107
1108         list_for_each_entry_safe(entry, next, bitmap_list, list)
1109                 list_del_init(&entry->list);
1110 }
1111
1112 static void noinline_for_stack
1113 cleanup_write_cache_enospc(struct inode *inode,
1114                            struct btrfs_io_ctl *io_ctl,
1115                            struct extent_state **cached_state)
1116 {
1117         io_ctl_drop_pages(io_ctl);
1118         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1119                              i_size_read(inode) - 1, cached_state);
1120 }
1121
1122 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1123                                  struct btrfs_trans_handle *trans,
1124                                  struct btrfs_block_group_cache *block_group,
1125                                  struct btrfs_io_ctl *io_ctl,
1126                                  struct btrfs_path *path, u64 offset)
1127 {
1128         int ret;
1129         struct inode *inode = io_ctl->inode;
1130
1131         if (!inode)
1132                 return 0;
1133
1134         /* Flush the dirty pages in the cache file. */
1135         ret = flush_dirty_cache(inode);
1136         if (ret)
1137                 goto out;
1138
1139         /* Update the cache item to tell everyone this cache file is valid. */
1140         ret = update_cache_item(trans, root, inode, path, offset,
1141                                 io_ctl->entries, io_ctl->bitmaps);
1142 out:
1143         io_ctl_free(io_ctl);
1144         if (ret) {
1145                 invalidate_inode_pages2(inode->i_mapping);
1146                 BTRFS_I(inode)->generation = 0;
1147                 if (block_group) {
1148 #ifdef DEBUG
1149                         btrfs_err(root->fs_info,
1150                                   "failed to write free space cache for block group %llu",
1151                                   block_group->key.objectid);
1152 #endif
1153                 }
1154         }
1155         btrfs_update_inode(trans, root, inode);
1156
1157         if (block_group) {
1158                 /* the dirty list is protected by the dirty_bgs_lock */
1159                 spin_lock(&trans->transaction->dirty_bgs_lock);
1160
1161                 /* the disk_cache_state is protected by the block group lock */
1162                 spin_lock(&block_group->lock);
1163
1164                 /*
1165                  * only mark this as written if we didn't get put back on
1166                  * the dirty list while waiting for IO.   Otherwise our
1167                  * cache state won't be right, and we won't get written again
1168                  */
1169                 if (!ret && list_empty(&block_group->dirty_list))
1170                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1171                 else if (ret)
1172                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1173
1174                 spin_unlock(&block_group->lock);
1175                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1176                 io_ctl->inode = NULL;
1177                 iput(inode);
1178         }
1179
1180         return ret;
1181
1182 }
1183
1184 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1185                                     struct btrfs_trans_handle *trans,
1186                                     struct btrfs_io_ctl *io_ctl,
1187                                     struct btrfs_path *path)
1188 {
1189         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1190 }
1191
1192 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1193                         struct btrfs_block_group_cache *block_group,
1194                         struct btrfs_path *path)
1195 {
1196         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1197                                      block_group, &block_group->io_ctl,
1198                                      path, block_group->key.objectid);
1199 }
1200
1201 /**
1202  * __btrfs_write_out_cache - write out cached info to an inode
1203  * @root - the root the inode belongs to
1204  * @ctl - the free space cache we are going to write out
1205  * @block_group - the block_group for this cache if it belongs to a block_group
1206  * @trans - the trans handle
1207  *
1208  * This function writes out a free space cache struct to disk for quick recovery
1209  * on mount.  This will return 0 if it was successful in writing the cache out,
1210  * or an errno if it was not.
1211  */
1212 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1213                                    struct btrfs_free_space_ctl *ctl,
1214                                    struct btrfs_block_group_cache *block_group,
1215                                    struct btrfs_io_ctl *io_ctl,
1216                                    struct btrfs_trans_handle *trans)
1217 {
1218         struct btrfs_fs_info *fs_info = root->fs_info;
1219         struct extent_state *cached_state = NULL;
1220         LIST_HEAD(bitmap_list);
1221         int entries = 0;
1222         int bitmaps = 0;
1223         int ret;
1224         int must_iput = 0;
1225
1226         if (!i_size_read(inode))
1227                 return -EIO;
1228
1229         WARN_ON(io_ctl->pages);
1230         ret = io_ctl_init(io_ctl, inode, 1);
1231         if (ret)
1232                 return ret;
1233
1234         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1235                 down_write(&block_group->data_rwsem);
1236                 spin_lock(&block_group->lock);
1237                 if (block_group->delalloc_bytes) {
1238                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1239                         spin_unlock(&block_group->lock);
1240                         up_write(&block_group->data_rwsem);
1241                         BTRFS_I(inode)->generation = 0;
1242                         ret = 0;
1243                         must_iput = 1;
1244                         goto out;
1245                 }
1246                 spin_unlock(&block_group->lock);
1247         }
1248
1249         /* Lock all pages first so we can lock the extent safely. */
1250         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1251         if (ret)
1252                 goto out_unlock;
1253
1254         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1255                          &cached_state);
1256
1257         io_ctl_set_generation(io_ctl, trans->transid);
1258
1259         mutex_lock(&ctl->cache_writeout_mutex);
1260         /* Write out the extent entries in the free space cache */
1261         spin_lock(&ctl->tree_lock);
1262         ret = write_cache_extent_entries(io_ctl, ctl,
1263                                          block_group, &entries, &bitmaps,
1264                                          &bitmap_list);
1265         if (ret)
1266                 goto out_nospc_locked;
1267
1268         /*
1269          * Some spaces that are freed in the current transaction are pinned,
1270          * they will be added into free space cache after the transaction is
1271          * committed, we shouldn't lose them.
1272          *
1273          * If this changes while we are working we'll get added back to
1274          * the dirty list and redo it.  No locking needed
1275          */
1276         ret = write_pinned_extent_entries(fs_info, block_group,
1277                                           io_ctl, &entries);
1278         if (ret)
1279                 goto out_nospc_locked;
1280
1281         /*
1282          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1283          * locked while doing it because a concurrent trim can be manipulating
1284          * or freeing the bitmap.
1285          */
1286         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1287         spin_unlock(&ctl->tree_lock);
1288         mutex_unlock(&ctl->cache_writeout_mutex);
1289         if (ret)
1290                 goto out_nospc;
1291
1292         /* Zero out the rest of the pages just to make sure */
1293         io_ctl_zero_remaining_pages(io_ctl);
1294
1295         /* Everything is written out, now we dirty the pages in the file. */
1296         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1297                                 i_size_read(inode), &cached_state);
1298         if (ret)
1299                 goto out_nospc;
1300
1301         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1302                 up_write(&block_group->data_rwsem);
1303         /*
1304          * Release the pages and unlock the extent, we will flush
1305          * them out later
1306          */
1307         io_ctl_drop_pages(io_ctl);
1308
1309         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1310                              i_size_read(inode) - 1, &cached_state);
1311
1312         /*
1313          * at this point the pages are under IO and we're happy,
1314          * The caller is responsible for waiting on them and updating the
1315          * the cache and the inode
1316          */
1317         io_ctl->entries = entries;
1318         io_ctl->bitmaps = bitmaps;
1319
1320         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1321         if (ret)
1322                 goto out;
1323
1324         return 0;
1325
1326 out:
1327         io_ctl->inode = NULL;
1328         io_ctl_free(io_ctl);
1329         if (ret) {
1330                 invalidate_inode_pages2(inode->i_mapping);
1331                 BTRFS_I(inode)->generation = 0;
1332         }
1333         btrfs_update_inode(trans, root, inode);
1334         if (must_iput)
1335                 iput(inode);
1336         return ret;
1337
1338 out_nospc_locked:
1339         cleanup_bitmap_list(&bitmap_list);
1340         spin_unlock(&ctl->tree_lock);
1341         mutex_unlock(&ctl->cache_writeout_mutex);
1342
1343 out_nospc:
1344         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1345
1346 out_unlock:
1347         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1348                 up_write(&block_group->data_rwsem);
1349
1350         goto out;
1351 }
1352
1353 int btrfs_write_out_cache(struct btrfs_fs_info *fs_info,
1354                           struct btrfs_trans_handle *trans,
1355                           struct btrfs_block_group_cache *block_group,
1356                           struct btrfs_path *path)
1357 {
1358         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1359         struct inode *inode;
1360         int ret = 0;
1361
1362         spin_lock(&block_group->lock);
1363         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1364                 spin_unlock(&block_group->lock);
1365                 return 0;
1366         }
1367         spin_unlock(&block_group->lock);
1368
1369         inode = lookup_free_space_inode(fs_info, block_group, path);
1370         if (IS_ERR(inode))
1371                 return 0;
1372
1373         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1374                                 block_group, &block_group->io_ctl, trans);
1375         if (ret) {
1376 #ifdef DEBUG
1377                 btrfs_err(fs_info,
1378                           "failed to write free space cache for block group %llu",
1379                           block_group->key.objectid);
1380 #endif
1381                 spin_lock(&block_group->lock);
1382                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1383                 spin_unlock(&block_group->lock);
1384
1385                 block_group->io_ctl.inode = NULL;
1386                 iput(inode);
1387         }
1388
1389         /*
1390          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1391          * to wait for IO and put the inode
1392          */
1393
1394         return ret;
1395 }
1396
1397 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1398                                           u64 offset)
1399 {
1400         ASSERT(offset >= bitmap_start);
1401         offset -= bitmap_start;
1402         return (unsigned long)(div_u64(offset, unit));
1403 }
1404
1405 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1406 {
1407         return (unsigned long)(div_u64(bytes, unit));
1408 }
1409
1410 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1411                                    u64 offset)
1412 {
1413         u64 bitmap_start;
1414         u64 bytes_per_bitmap;
1415
1416         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1417         bitmap_start = offset - ctl->start;
1418         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1419         bitmap_start *= bytes_per_bitmap;
1420         bitmap_start += ctl->start;
1421
1422         return bitmap_start;
1423 }
1424
1425 static int tree_insert_offset(struct rb_root *root, u64 offset,
1426                               struct rb_node *node, int bitmap)
1427 {
1428         struct rb_node **p = &root->rb_node;
1429         struct rb_node *parent = NULL;
1430         struct btrfs_free_space *info;
1431
1432         while (*p) {
1433                 parent = *p;
1434                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1435
1436                 if (offset < info->offset) {
1437                         p = &(*p)->rb_left;
1438                 } else if (offset > info->offset) {
1439                         p = &(*p)->rb_right;
1440                 } else {
1441                         /*
1442                          * we could have a bitmap entry and an extent entry
1443                          * share the same offset.  If this is the case, we want
1444                          * the extent entry to always be found first if we do a
1445                          * linear search through the tree, since we want to have
1446                          * the quickest allocation time, and allocating from an
1447                          * extent is faster than allocating from a bitmap.  So
1448                          * if we're inserting a bitmap and we find an entry at
1449                          * this offset, we want to go right, or after this entry
1450                          * logically.  If we are inserting an extent and we've
1451                          * found a bitmap, we want to go left, or before
1452                          * logically.
1453                          */
1454                         if (bitmap) {
1455                                 if (info->bitmap) {
1456                                         WARN_ON_ONCE(1);
1457                                         return -EEXIST;
1458                                 }
1459                                 p = &(*p)->rb_right;
1460                         } else {
1461                                 if (!info->bitmap) {
1462                                         WARN_ON_ONCE(1);
1463                                         return -EEXIST;
1464                                 }
1465                                 p = &(*p)->rb_left;
1466                         }
1467                 }
1468         }
1469
1470         rb_link_node(node, parent, p);
1471         rb_insert_color(node, root);
1472
1473         return 0;
1474 }
1475
1476 /*
1477  * searches the tree for the given offset.
1478  *
1479  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1480  * want a section that has at least bytes size and comes at or after the given
1481  * offset.
1482  */
1483 static struct btrfs_free_space *
1484 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1485                    u64 offset, int bitmap_only, int fuzzy)
1486 {
1487         struct rb_node *n = ctl->free_space_offset.rb_node;
1488         struct btrfs_free_space *entry, *prev = NULL;
1489
1490         /* find entry that is closest to the 'offset' */
1491         while (1) {
1492                 if (!n) {
1493                         entry = NULL;
1494                         break;
1495                 }
1496
1497                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1498                 prev = entry;
1499
1500                 if (offset < entry->offset)
1501                         n = n->rb_left;
1502                 else if (offset > entry->offset)
1503                         n = n->rb_right;
1504                 else
1505                         break;
1506         }
1507
1508         if (bitmap_only) {
1509                 if (!entry)
1510                         return NULL;
1511                 if (entry->bitmap)
1512                         return entry;
1513
1514                 /*
1515                  * bitmap entry and extent entry may share same offset,
1516                  * in that case, bitmap entry comes after extent entry.
1517                  */
1518                 n = rb_next(n);
1519                 if (!n)
1520                         return NULL;
1521                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1522                 if (entry->offset != offset)
1523                         return NULL;
1524
1525                 WARN_ON(!entry->bitmap);
1526                 return entry;
1527         } else if (entry) {
1528                 if (entry->bitmap) {
1529                         /*
1530                          * if previous extent entry covers the offset,
1531                          * we should return it instead of the bitmap entry
1532                          */
1533                         n = rb_prev(&entry->offset_index);
1534                         if (n) {
1535                                 prev = rb_entry(n, struct btrfs_free_space,
1536                                                 offset_index);
1537                                 if (!prev->bitmap &&
1538                                     prev->offset + prev->bytes > offset)
1539                                         entry = prev;
1540                         }
1541                 }
1542                 return entry;
1543         }
1544
1545         if (!prev)
1546                 return NULL;
1547
1548         /* find last entry before the 'offset' */
1549         entry = prev;
1550         if (entry->offset > offset) {
1551                 n = rb_prev(&entry->offset_index);
1552                 if (n) {
1553                         entry = rb_entry(n, struct btrfs_free_space,
1554                                         offset_index);
1555                         ASSERT(entry->offset <= offset);
1556                 } else {
1557                         if (fuzzy)
1558                                 return entry;
1559                         else
1560                                 return NULL;
1561                 }
1562         }
1563
1564         if (entry->bitmap) {
1565                 n = rb_prev(&entry->offset_index);
1566                 if (n) {
1567                         prev = rb_entry(n, struct btrfs_free_space,
1568                                         offset_index);
1569                         if (!prev->bitmap &&
1570                             prev->offset + prev->bytes > offset)
1571                                 return prev;
1572                 }
1573                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1574                         return entry;
1575         } else if (entry->offset + entry->bytes > offset)
1576                 return entry;
1577
1578         if (!fuzzy)
1579                 return NULL;
1580
1581         while (1) {
1582                 if (entry->bitmap) {
1583                         if (entry->offset + BITS_PER_BITMAP *
1584                             ctl->unit > offset)
1585                                 break;
1586                 } else {
1587                         if (entry->offset + entry->bytes > offset)
1588                                 break;
1589                 }
1590
1591                 n = rb_next(&entry->offset_index);
1592                 if (!n)
1593                         return NULL;
1594                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1595         }
1596         return entry;
1597 }
1598
1599 static inline void
1600 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1601                     struct btrfs_free_space *info)
1602 {
1603         rb_erase(&info->offset_index, &ctl->free_space_offset);
1604         ctl->free_extents--;
1605 }
1606
1607 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1608                               struct btrfs_free_space *info)
1609 {
1610         __unlink_free_space(ctl, info);
1611         ctl->free_space -= info->bytes;
1612 }
1613
1614 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1615                            struct btrfs_free_space *info)
1616 {
1617         int ret = 0;
1618
1619         ASSERT(info->bytes || info->bitmap);
1620         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1621                                  &info->offset_index, (info->bitmap != NULL));
1622         if (ret)
1623                 return ret;
1624
1625         ctl->free_space += info->bytes;
1626         ctl->free_extents++;
1627         return ret;
1628 }
1629
1630 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1631 {
1632         struct btrfs_block_group_cache *block_group = ctl->private;
1633         u64 max_bytes;
1634         u64 bitmap_bytes;
1635         u64 extent_bytes;
1636         u64 size = block_group->key.offset;
1637         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1638         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1639
1640         max_bitmaps = max_t(u64, max_bitmaps, 1);
1641
1642         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1643
1644         /*
1645          * The goal is to keep the total amount of memory used per 1gb of space
1646          * at or below 32k, so we need to adjust how much memory we allow to be
1647          * used by extent based free space tracking
1648          */
1649         if (size < SZ_1G)
1650                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1651         else
1652                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1653
1654         /*
1655          * we want to account for 1 more bitmap than what we have so we can make
1656          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1657          * we add more bitmaps.
1658          */
1659         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1660
1661         if (bitmap_bytes >= max_bytes) {
1662                 ctl->extents_thresh = 0;
1663                 return;
1664         }
1665
1666         /*
1667          * we want the extent entry threshold to always be at most 1/2 the max
1668          * bytes we can have, or whatever is less than that.
1669          */
1670         extent_bytes = max_bytes - bitmap_bytes;
1671         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1672
1673         ctl->extents_thresh =
1674                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1675 }
1676
1677 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1678                                        struct btrfs_free_space *info,
1679                                        u64 offset, u64 bytes)
1680 {
1681         unsigned long start, count;
1682
1683         start = offset_to_bit(info->offset, ctl->unit, offset);
1684         count = bytes_to_bits(bytes, ctl->unit);
1685         ASSERT(start + count <= BITS_PER_BITMAP);
1686
1687         bitmap_clear(info->bitmap, start, count);
1688
1689         info->bytes -= bytes;
1690 }
1691
1692 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1693                               struct btrfs_free_space *info, u64 offset,
1694                               u64 bytes)
1695 {
1696         __bitmap_clear_bits(ctl, info, offset, bytes);
1697         ctl->free_space -= bytes;
1698 }
1699
1700 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1701                             struct btrfs_free_space *info, u64 offset,
1702                             u64 bytes)
1703 {
1704         unsigned long start, count;
1705
1706         start = offset_to_bit(info->offset, ctl->unit, offset);
1707         count = bytes_to_bits(bytes, ctl->unit);
1708         ASSERT(start + count <= BITS_PER_BITMAP);
1709
1710         bitmap_set(info->bitmap, start, count);
1711
1712         info->bytes += bytes;
1713         ctl->free_space += bytes;
1714 }
1715
1716 /*
1717  * If we can not find suitable extent, we will use bytes to record
1718  * the size of the max extent.
1719  */
1720 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1721                          struct btrfs_free_space *bitmap_info, u64 *offset,
1722                          u64 *bytes, bool for_alloc)
1723 {
1724         unsigned long found_bits = 0;
1725         unsigned long max_bits = 0;
1726         unsigned long bits, i;
1727         unsigned long next_zero;
1728         unsigned long extent_bits;
1729
1730         /*
1731          * Skip searching the bitmap if we don't have a contiguous section that
1732          * is large enough for this allocation.
1733          */
1734         if (for_alloc &&
1735             bitmap_info->max_extent_size &&
1736             bitmap_info->max_extent_size < *bytes) {
1737                 *bytes = bitmap_info->max_extent_size;
1738                 return -1;
1739         }
1740
1741         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1742                           max_t(u64, *offset, bitmap_info->offset));
1743         bits = bytes_to_bits(*bytes, ctl->unit);
1744
1745         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1746                 if (for_alloc && bits == 1) {
1747                         found_bits = 1;
1748                         break;
1749                 }
1750                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1751                                                BITS_PER_BITMAP, i);
1752                 extent_bits = next_zero - i;
1753                 if (extent_bits >= bits) {
1754                         found_bits = extent_bits;
1755                         break;
1756                 } else if (extent_bits > max_bits) {
1757                         max_bits = extent_bits;
1758                 }
1759                 i = next_zero;
1760         }
1761
1762         if (found_bits) {
1763                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1764                 *bytes = (u64)(found_bits) * ctl->unit;
1765                 return 0;
1766         }
1767
1768         *bytes = (u64)(max_bits) * ctl->unit;
1769         bitmap_info->max_extent_size = *bytes;
1770         return -1;
1771 }
1772
1773 /* Cache the size of the max extent in bytes */
1774 static struct btrfs_free_space *
1775 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1776                 unsigned long align, u64 *max_extent_size)
1777 {
1778         struct btrfs_free_space *entry;
1779         struct rb_node *node;
1780         u64 tmp;
1781         u64 align_off;
1782         int ret;
1783
1784         if (!ctl->free_space_offset.rb_node)
1785                 goto out;
1786
1787         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1788         if (!entry)
1789                 goto out;
1790
1791         for (node = &entry->offset_index; node; node = rb_next(node)) {
1792                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1793                 if (entry->bytes < *bytes) {
1794                         if (entry->bytes > *max_extent_size)
1795                                 *max_extent_size = entry->bytes;
1796                         continue;
1797                 }
1798
1799                 /* make sure the space returned is big enough
1800                  * to match our requested alignment
1801                  */
1802                 if (*bytes >= align) {
1803                         tmp = entry->offset - ctl->start + align - 1;
1804                         tmp = div64_u64(tmp, align);
1805                         tmp = tmp * align + ctl->start;
1806                         align_off = tmp - entry->offset;
1807                 } else {
1808                         align_off = 0;
1809                         tmp = entry->offset;
1810                 }
1811
1812                 if (entry->bytes < *bytes + align_off) {
1813                         if (entry->bytes > *max_extent_size)
1814                                 *max_extent_size = entry->bytes;
1815                         continue;
1816                 }
1817
1818                 if (entry->bitmap) {
1819                         u64 size = *bytes;
1820
1821                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1822                         if (!ret) {
1823                                 *offset = tmp;
1824                                 *bytes = size;
1825                                 return entry;
1826                         } else if (size > *max_extent_size) {
1827                                 *max_extent_size = size;
1828                         }
1829                         continue;
1830                 }
1831
1832                 *offset = tmp;
1833                 *bytes = entry->bytes - align_off;
1834                 return entry;
1835         }
1836 out:
1837         return NULL;
1838 }
1839
1840 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1841                            struct btrfs_free_space *info, u64 offset)
1842 {
1843         info->offset = offset_to_bitmap(ctl, offset);
1844         info->bytes = 0;
1845         INIT_LIST_HEAD(&info->list);
1846         link_free_space(ctl, info);
1847         ctl->total_bitmaps++;
1848
1849         ctl->op->recalc_thresholds(ctl);
1850 }
1851
1852 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1853                         struct btrfs_free_space *bitmap_info)
1854 {
1855         unlink_free_space(ctl, bitmap_info);
1856         kfree(bitmap_info->bitmap);
1857         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1858         ctl->total_bitmaps--;
1859         ctl->op->recalc_thresholds(ctl);
1860 }
1861
1862 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1863                               struct btrfs_free_space *bitmap_info,
1864                               u64 *offset, u64 *bytes)
1865 {
1866         u64 end;
1867         u64 search_start, search_bytes;
1868         int ret;
1869
1870 again:
1871         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1872
1873         /*
1874          * We need to search for bits in this bitmap.  We could only cover some
1875          * of the extent in this bitmap thanks to how we add space, so we need
1876          * to search for as much as it as we can and clear that amount, and then
1877          * go searching for the next bit.
1878          */
1879         search_start = *offset;
1880         search_bytes = ctl->unit;
1881         search_bytes = min(search_bytes, end - search_start + 1);
1882         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1883                             false);
1884         if (ret < 0 || search_start != *offset)
1885                 return -EINVAL;
1886
1887         /* We may have found more bits than what we need */
1888         search_bytes = min(search_bytes, *bytes);
1889
1890         /* Cannot clear past the end of the bitmap */
1891         search_bytes = min(search_bytes, end - search_start + 1);
1892
1893         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1894         *offset += search_bytes;
1895         *bytes -= search_bytes;
1896
1897         if (*bytes) {
1898                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1899                 if (!bitmap_info->bytes)
1900                         free_bitmap(ctl, bitmap_info);
1901
1902                 /*
1903                  * no entry after this bitmap, but we still have bytes to
1904                  * remove, so something has gone wrong.
1905                  */
1906                 if (!next)
1907                         return -EINVAL;
1908
1909                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1910                                        offset_index);
1911
1912                 /*
1913                  * if the next entry isn't a bitmap we need to return to let the
1914                  * extent stuff do its work.
1915                  */
1916                 if (!bitmap_info->bitmap)
1917                         return -EAGAIN;
1918
1919                 /*
1920                  * Ok the next item is a bitmap, but it may not actually hold
1921                  * the information for the rest of this free space stuff, so
1922                  * look for it, and if we don't find it return so we can try
1923                  * everything over again.
1924                  */
1925                 search_start = *offset;
1926                 search_bytes = ctl->unit;
1927                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1928                                     &search_bytes, false);
1929                 if (ret < 0 || search_start != *offset)
1930                         return -EAGAIN;
1931
1932                 goto again;
1933         } else if (!bitmap_info->bytes)
1934                 free_bitmap(ctl, bitmap_info);
1935
1936         return 0;
1937 }
1938
1939 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1940                                struct btrfs_free_space *info, u64 offset,
1941                                u64 bytes)
1942 {
1943         u64 bytes_to_set = 0;
1944         u64 end;
1945
1946         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1947
1948         bytes_to_set = min(end - offset, bytes);
1949
1950         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1951
1952         /*
1953          * We set some bytes, we have no idea what the max extent size is
1954          * anymore.
1955          */
1956         info->max_extent_size = 0;
1957
1958         return bytes_to_set;
1959
1960 }
1961
1962 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1963                       struct btrfs_free_space *info)
1964 {
1965         struct btrfs_block_group_cache *block_group = ctl->private;
1966         struct btrfs_fs_info *fs_info = block_group->fs_info;
1967         bool forced = false;
1968
1969 #ifdef CONFIG_BTRFS_DEBUG
1970         if (btrfs_should_fragment_free_space(block_group))
1971                 forced = true;
1972 #endif
1973
1974         /*
1975          * If we are below the extents threshold then we can add this as an
1976          * extent, and don't have to deal with the bitmap
1977          */
1978         if (!forced && ctl->free_extents < ctl->extents_thresh) {
1979                 /*
1980                  * If this block group has some small extents we don't want to
1981                  * use up all of our free slots in the cache with them, we want
1982                  * to reserve them to larger extents, however if we have plenty
1983                  * of cache left then go ahead an dadd them, no sense in adding
1984                  * the overhead of a bitmap if we don't have to.
1985                  */
1986                 if (info->bytes <= fs_info->sectorsize * 4) {
1987                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1988                                 return false;
1989                 } else {
1990                         return false;
1991                 }
1992         }
1993
1994         /*
1995          * The original block groups from mkfs can be really small, like 8
1996          * megabytes, so don't bother with a bitmap for those entries.  However
1997          * some block groups can be smaller than what a bitmap would cover but
1998          * are still large enough that they could overflow the 32k memory limit,
1999          * so allow those block groups to still be allowed to have a bitmap
2000          * entry.
2001          */
2002         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2003                 return false;
2004
2005         return true;
2006 }
2007
2008 static const struct btrfs_free_space_op free_space_op = {
2009         .recalc_thresholds      = recalculate_thresholds,
2010         .use_bitmap             = use_bitmap,
2011 };
2012
2013 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2014                               struct btrfs_free_space *info)
2015 {
2016         struct btrfs_free_space *bitmap_info;
2017         struct btrfs_block_group_cache *block_group = NULL;
2018         int added = 0;
2019         u64 bytes, offset, bytes_added;
2020         int ret;
2021
2022         bytes = info->bytes;
2023         offset = info->offset;
2024
2025         if (!ctl->op->use_bitmap(ctl, info))
2026                 return 0;
2027
2028         if (ctl->op == &free_space_op)
2029                 block_group = ctl->private;
2030 again:
2031         /*
2032          * Since we link bitmaps right into the cluster we need to see if we
2033          * have a cluster here, and if so and it has our bitmap we need to add
2034          * the free space to that bitmap.
2035          */
2036         if (block_group && !list_empty(&block_group->cluster_list)) {
2037                 struct btrfs_free_cluster *cluster;
2038                 struct rb_node *node;
2039                 struct btrfs_free_space *entry;
2040
2041                 cluster = list_entry(block_group->cluster_list.next,
2042                                      struct btrfs_free_cluster,
2043                                      block_group_list);
2044                 spin_lock(&cluster->lock);
2045                 node = rb_first(&cluster->root);
2046                 if (!node) {
2047                         spin_unlock(&cluster->lock);
2048                         goto no_cluster_bitmap;
2049                 }
2050
2051                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2052                 if (!entry->bitmap) {
2053                         spin_unlock(&cluster->lock);
2054                         goto no_cluster_bitmap;
2055                 }
2056
2057                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2058                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2059                                                           offset, bytes);
2060                         bytes -= bytes_added;
2061                         offset += bytes_added;
2062                 }
2063                 spin_unlock(&cluster->lock);
2064                 if (!bytes) {
2065                         ret = 1;
2066                         goto out;
2067                 }
2068         }
2069
2070 no_cluster_bitmap:
2071         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2072                                          1, 0);
2073         if (!bitmap_info) {
2074                 ASSERT(added == 0);
2075                 goto new_bitmap;
2076         }
2077
2078         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2079         bytes -= bytes_added;
2080         offset += bytes_added;
2081         added = 0;
2082
2083         if (!bytes) {
2084                 ret = 1;
2085                 goto out;
2086         } else
2087                 goto again;
2088
2089 new_bitmap:
2090         if (info && info->bitmap) {
2091                 add_new_bitmap(ctl, info, offset);
2092                 added = 1;
2093                 info = NULL;
2094                 goto again;
2095         } else {
2096                 spin_unlock(&ctl->tree_lock);
2097
2098                 /* no pre-allocated info, allocate a new one */
2099                 if (!info) {
2100                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2101                                                  GFP_NOFS);
2102                         if (!info) {
2103                                 spin_lock(&ctl->tree_lock);
2104                                 ret = -ENOMEM;
2105                                 goto out;
2106                         }
2107                 }
2108
2109                 /* allocate the bitmap */
2110                 info->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
2111                 spin_lock(&ctl->tree_lock);
2112                 if (!info->bitmap) {
2113                         ret = -ENOMEM;
2114                         goto out;
2115                 }
2116                 goto again;
2117         }
2118
2119 out:
2120         if (info) {
2121                 kfree(info->bitmap);
2122                 kmem_cache_free(btrfs_free_space_cachep, info);
2123         }
2124
2125         return ret;
2126 }
2127
2128 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2129                           struct btrfs_free_space *info, bool update_stat)
2130 {
2131         struct btrfs_free_space *left_info;
2132         struct btrfs_free_space *right_info;
2133         bool merged = false;
2134         u64 offset = info->offset;
2135         u64 bytes = info->bytes;
2136
2137         /*
2138          * first we want to see if there is free space adjacent to the range we
2139          * are adding, if there is remove that struct and add a new one to
2140          * cover the entire range
2141          */
2142         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2143         if (right_info && rb_prev(&right_info->offset_index))
2144                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2145                                      struct btrfs_free_space, offset_index);
2146         else
2147                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2148
2149         if (right_info && !right_info->bitmap) {
2150                 if (update_stat)
2151                         unlink_free_space(ctl, right_info);
2152                 else
2153                         __unlink_free_space(ctl, right_info);
2154                 info->bytes += right_info->bytes;
2155                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2156                 merged = true;
2157         }
2158
2159         if (left_info && !left_info->bitmap &&
2160             left_info->offset + left_info->bytes == offset) {
2161                 if (update_stat)
2162                         unlink_free_space(ctl, left_info);
2163                 else
2164                         __unlink_free_space(ctl, left_info);
2165                 info->offset = left_info->offset;
2166                 info->bytes += left_info->bytes;
2167                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2168                 merged = true;
2169         }
2170
2171         return merged;
2172 }
2173
2174 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2175                                      struct btrfs_free_space *info,
2176                                      bool update_stat)
2177 {
2178         struct btrfs_free_space *bitmap;
2179         unsigned long i;
2180         unsigned long j;
2181         const u64 end = info->offset + info->bytes;
2182         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2183         u64 bytes;
2184
2185         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2186         if (!bitmap)
2187                 return false;
2188
2189         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2190         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2191         if (j == i)
2192                 return false;
2193         bytes = (j - i) * ctl->unit;
2194         info->bytes += bytes;
2195
2196         if (update_stat)
2197                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2198         else
2199                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2200
2201         if (!bitmap->bytes)
2202                 free_bitmap(ctl, bitmap);
2203
2204         return true;
2205 }
2206
2207 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2208                                        struct btrfs_free_space *info,
2209                                        bool update_stat)
2210 {
2211         struct btrfs_free_space *bitmap;
2212         u64 bitmap_offset;
2213         unsigned long i;
2214         unsigned long j;
2215         unsigned long prev_j;
2216         u64 bytes;
2217
2218         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2219         /* If we're on a boundary, try the previous logical bitmap. */
2220         if (bitmap_offset == info->offset) {
2221                 if (info->offset == 0)
2222                         return false;
2223                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2224         }
2225
2226         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2227         if (!bitmap)
2228                 return false;
2229
2230         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2231         j = 0;
2232         prev_j = (unsigned long)-1;
2233         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2234                 if (j > i)
2235                         break;
2236                 prev_j = j;
2237         }
2238         if (prev_j == i)
2239                 return false;
2240
2241         if (prev_j == (unsigned long)-1)
2242                 bytes = (i + 1) * ctl->unit;
2243         else
2244                 bytes = (i - prev_j) * ctl->unit;
2245
2246         info->offset -= bytes;
2247         info->bytes += bytes;
2248
2249         if (update_stat)
2250                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2251         else
2252                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2253
2254         if (!bitmap->bytes)
2255                 free_bitmap(ctl, bitmap);
2256
2257         return true;
2258 }
2259
2260 /*
2261  * We prefer always to allocate from extent entries, both for clustered and
2262  * non-clustered allocation requests. So when attempting to add a new extent
2263  * entry, try to see if there's adjacent free space in bitmap entries, and if
2264  * there is, migrate that space from the bitmaps to the extent.
2265  * Like this we get better chances of satisfying space allocation requests
2266  * because we attempt to satisfy them based on a single cache entry, and never
2267  * on 2 or more entries - even if the entries represent a contiguous free space
2268  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2269  * ends).
2270  */
2271 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2272                               struct btrfs_free_space *info,
2273                               bool update_stat)
2274 {
2275         /*
2276          * Only work with disconnected entries, as we can change their offset,
2277          * and must be extent entries.
2278          */
2279         ASSERT(!info->bitmap);
2280         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2281
2282         if (ctl->total_bitmaps > 0) {
2283                 bool stole_end;
2284                 bool stole_front = false;
2285
2286                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2287                 if (ctl->total_bitmaps > 0)
2288                         stole_front = steal_from_bitmap_to_front(ctl, info,
2289                                                                  update_stat);
2290
2291                 if (stole_end || stole_front)
2292                         try_merge_free_space(ctl, info, update_stat);
2293         }
2294 }
2295
2296 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2297                            struct btrfs_free_space_ctl *ctl,
2298                            u64 offset, u64 bytes)
2299 {
2300         struct btrfs_free_space *info;
2301         int ret = 0;
2302
2303         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2304         if (!info)
2305                 return -ENOMEM;
2306
2307         info->offset = offset;
2308         info->bytes = bytes;
2309         RB_CLEAR_NODE(&info->offset_index);
2310
2311         spin_lock(&ctl->tree_lock);
2312
2313         if (try_merge_free_space(ctl, info, true))
2314                 goto link;
2315
2316         /*
2317          * There was no extent directly to the left or right of this new
2318          * extent then we know we're going to have to allocate a new extent, so
2319          * before we do that see if we need to drop this into a bitmap
2320          */
2321         ret = insert_into_bitmap(ctl, info);
2322         if (ret < 0) {
2323                 goto out;
2324         } else if (ret) {
2325                 ret = 0;
2326                 goto out;
2327         }
2328 link:
2329         /*
2330          * Only steal free space from adjacent bitmaps if we're sure we're not
2331          * going to add the new free space to existing bitmap entries - because
2332          * that would mean unnecessary work that would be reverted. Therefore
2333          * attempt to steal space from bitmaps if we're adding an extent entry.
2334          */
2335         steal_from_bitmap(ctl, info, true);
2336
2337         ret = link_free_space(ctl, info);
2338         if (ret)
2339                 kmem_cache_free(btrfs_free_space_cachep, info);
2340 out:
2341         spin_unlock(&ctl->tree_lock);
2342
2343         if (ret) {
2344                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2345                 ASSERT(ret != -EEXIST);
2346         }
2347
2348         return ret;
2349 }
2350
2351 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2352                             u64 offset, u64 bytes)
2353 {
2354         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2355         struct btrfs_free_space *info;
2356         int ret;
2357         bool re_search = false;
2358
2359         spin_lock(&ctl->tree_lock);
2360
2361 again:
2362         ret = 0;
2363         if (!bytes)
2364                 goto out_lock;
2365
2366         info = tree_search_offset(ctl, offset, 0, 0);
2367         if (!info) {
2368                 /*
2369                  * oops didn't find an extent that matched the space we wanted
2370                  * to remove, look for a bitmap instead
2371                  */
2372                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2373                                           1, 0);
2374                 if (!info) {
2375                         /*
2376                          * If we found a partial bit of our free space in a
2377                          * bitmap but then couldn't find the other part this may
2378                          * be a problem, so WARN about it.
2379                          */
2380                         WARN_ON(re_search);
2381                         goto out_lock;
2382                 }
2383         }
2384
2385         re_search = false;
2386         if (!info->bitmap) {
2387                 unlink_free_space(ctl, info);
2388                 if (offset == info->offset) {
2389                         u64 to_free = min(bytes, info->bytes);
2390
2391                         info->bytes -= to_free;
2392                         info->offset += to_free;
2393                         if (info->bytes) {
2394                                 ret = link_free_space(ctl, info);
2395                                 WARN_ON(ret);
2396                         } else {
2397                                 kmem_cache_free(btrfs_free_space_cachep, info);
2398                         }
2399
2400                         offset += to_free;
2401                         bytes -= to_free;
2402                         goto again;
2403                 } else {
2404                         u64 old_end = info->bytes + info->offset;
2405
2406                         info->bytes = offset - info->offset;
2407                         ret = link_free_space(ctl, info);
2408                         WARN_ON(ret);
2409                         if (ret)
2410                                 goto out_lock;
2411
2412                         /* Not enough bytes in this entry to satisfy us */
2413                         if (old_end < offset + bytes) {
2414                                 bytes -= old_end - offset;
2415                                 offset = old_end;
2416                                 goto again;
2417                         } else if (old_end == offset + bytes) {
2418                                 /* all done */
2419                                 goto out_lock;
2420                         }
2421                         spin_unlock(&ctl->tree_lock);
2422
2423                         ret = btrfs_add_free_space(block_group, offset + bytes,
2424                                                    old_end - (offset + bytes));
2425                         WARN_ON(ret);
2426                         goto out;
2427                 }
2428         }
2429
2430         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2431         if (ret == -EAGAIN) {
2432                 re_search = true;
2433                 goto again;
2434         }
2435 out_lock:
2436         spin_unlock(&ctl->tree_lock);
2437 out:
2438         return ret;
2439 }
2440
2441 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2442                            u64 bytes)
2443 {
2444         struct btrfs_fs_info *fs_info = block_group->fs_info;
2445         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2446         struct btrfs_free_space *info;
2447         struct rb_node *n;
2448         int count = 0;
2449
2450         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2451                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2452                 if (info->bytes >= bytes && !block_group->ro)
2453                         count++;
2454                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2455                            info->offset, info->bytes,
2456                        (info->bitmap) ? "yes" : "no");
2457         }
2458         btrfs_info(fs_info, "block group has cluster?: %s",
2459                list_empty(&block_group->cluster_list) ? "no" : "yes");
2460         btrfs_info(fs_info,
2461                    "%d blocks of free space at or bigger than bytes is", count);
2462 }
2463
2464 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2465 {
2466         struct btrfs_fs_info *fs_info = block_group->fs_info;
2467         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2468
2469         spin_lock_init(&ctl->tree_lock);
2470         ctl->unit = fs_info->sectorsize;
2471         ctl->start = block_group->key.objectid;
2472         ctl->private = block_group;
2473         ctl->op = &free_space_op;
2474         INIT_LIST_HEAD(&ctl->trimming_ranges);
2475         mutex_init(&ctl->cache_writeout_mutex);
2476
2477         /*
2478          * we only want to have 32k of ram per block group for keeping
2479          * track of free space, and if we pass 1/2 of that we want to
2480          * start converting things over to using bitmaps
2481          */
2482         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2483 }
2484
2485 /*
2486  * for a given cluster, put all of its extents back into the free
2487  * space cache.  If the block group passed doesn't match the block group
2488  * pointed to by the cluster, someone else raced in and freed the
2489  * cluster already.  In that case, we just return without changing anything
2490  */
2491 static int
2492 __btrfs_return_cluster_to_free_space(
2493                              struct btrfs_block_group_cache *block_group,
2494                              struct btrfs_free_cluster *cluster)
2495 {
2496         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2497         struct btrfs_free_space *entry;
2498         struct rb_node *node;
2499
2500         spin_lock(&cluster->lock);
2501         if (cluster->block_group != block_group)
2502                 goto out;
2503
2504         cluster->block_group = NULL;
2505         cluster->window_start = 0;
2506         list_del_init(&cluster->block_group_list);
2507
2508         node = rb_first(&cluster->root);
2509         while (node) {
2510                 bool bitmap;
2511
2512                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2513                 node = rb_next(&entry->offset_index);
2514                 rb_erase(&entry->offset_index, &cluster->root);
2515                 RB_CLEAR_NODE(&entry->offset_index);
2516
2517                 bitmap = (entry->bitmap != NULL);
2518                 if (!bitmap) {
2519                         try_merge_free_space(ctl, entry, false);
2520                         steal_from_bitmap(ctl, entry, false);
2521                 }
2522                 tree_insert_offset(&ctl->free_space_offset,
2523                                    entry->offset, &entry->offset_index, bitmap);
2524         }
2525         cluster->root = RB_ROOT;
2526
2527 out:
2528         spin_unlock(&cluster->lock);
2529         btrfs_put_block_group(block_group);
2530         return 0;
2531 }
2532
2533 static void __btrfs_remove_free_space_cache_locked(
2534                                 struct btrfs_free_space_ctl *ctl)
2535 {
2536         struct btrfs_free_space *info;
2537         struct rb_node *node;
2538
2539         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2540                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2541                 if (!info->bitmap) {
2542                         unlink_free_space(ctl, info);
2543                         kmem_cache_free(btrfs_free_space_cachep, info);
2544                 } else {
2545                         free_bitmap(ctl, info);
2546                 }
2547
2548                 cond_resched_lock(&ctl->tree_lock);
2549         }
2550 }
2551
2552 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2553 {
2554         spin_lock(&ctl->tree_lock);
2555         __btrfs_remove_free_space_cache_locked(ctl);
2556         spin_unlock(&ctl->tree_lock);
2557 }
2558
2559 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2560 {
2561         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2562         struct btrfs_free_cluster *cluster;
2563         struct list_head *head;
2564
2565         spin_lock(&ctl->tree_lock);
2566         while ((head = block_group->cluster_list.next) !=
2567                &block_group->cluster_list) {
2568                 cluster = list_entry(head, struct btrfs_free_cluster,
2569                                      block_group_list);
2570
2571                 WARN_ON(cluster->block_group != block_group);
2572                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2573
2574                 cond_resched_lock(&ctl->tree_lock);
2575         }
2576         __btrfs_remove_free_space_cache_locked(ctl);
2577         spin_unlock(&ctl->tree_lock);
2578
2579 }
2580
2581 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2582                                u64 offset, u64 bytes, u64 empty_size,
2583                                u64 *max_extent_size)
2584 {
2585         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2586         struct btrfs_free_space *entry = NULL;
2587         u64 bytes_search = bytes + empty_size;
2588         u64 ret = 0;
2589         u64 align_gap = 0;
2590         u64 align_gap_len = 0;
2591
2592         spin_lock(&ctl->tree_lock);
2593         entry = find_free_space(ctl, &offset, &bytes_search,
2594                                 block_group->full_stripe_len, max_extent_size);
2595         if (!entry)
2596                 goto out;
2597
2598         ret = offset;
2599         if (entry->bitmap) {
2600                 bitmap_clear_bits(ctl, entry, offset, bytes);
2601                 if (!entry->bytes)
2602                         free_bitmap(ctl, entry);
2603         } else {
2604                 unlink_free_space(ctl, entry);
2605                 align_gap_len = offset - entry->offset;
2606                 align_gap = entry->offset;
2607
2608                 entry->offset = offset + bytes;
2609                 WARN_ON(entry->bytes < bytes + align_gap_len);
2610
2611                 entry->bytes -= bytes + align_gap_len;
2612                 if (!entry->bytes)
2613                         kmem_cache_free(btrfs_free_space_cachep, entry);
2614                 else
2615                         link_free_space(ctl, entry);
2616         }
2617 out:
2618         spin_unlock(&ctl->tree_lock);
2619
2620         if (align_gap_len)
2621                 __btrfs_add_free_space(block_group->fs_info, ctl,
2622                                        align_gap, align_gap_len);
2623         return ret;
2624 }
2625
2626 /*
2627  * given a cluster, put all of its extents back into the free space
2628  * cache.  If a block group is passed, this function will only free
2629  * a cluster that belongs to the passed block group.
2630  *
2631  * Otherwise, it'll get a reference on the block group pointed to by the
2632  * cluster and remove the cluster from it.
2633  */
2634 int btrfs_return_cluster_to_free_space(
2635                                struct btrfs_block_group_cache *block_group,
2636                                struct btrfs_free_cluster *cluster)
2637 {
2638         struct btrfs_free_space_ctl *ctl;
2639         int ret;
2640
2641         /* first, get a safe pointer to the block group */
2642         spin_lock(&cluster->lock);
2643         if (!block_group) {
2644                 block_group = cluster->block_group;
2645                 if (!block_group) {
2646                         spin_unlock(&cluster->lock);
2647                         return 0;
2648                 }
2649         } else if (cluster->block_group != block_group) {
2650                 /* someone else has already freed it don't redo their work */
2651                 spin_unlock(&cluster->lock);
2652                 return 0;
2653         }
2654         atomic_inc(&block_group->count);
2655         spin_unlock(&cluster->lock);
2656
2657         ctl = block_group->free_space_ctl;
2658
2659         /* now return any extents the cluster had on it */
2660         spin_lock(&ctl->tree_lock);
2661         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2662         spin_unlock(&ctl->tree_lock);
2663
2664         /* finally drop our ref */
2665         btrfs_put_block_group(block_group);
2666         return ret;
2667 }
2668
2669 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2670                                    struct btrfs_free_cluster *cluster,
2671                                    struct btrfs_free_space *entry,
2672                                    u64 bytes, u64 min_start,
2673                                    u64 *max_extent_size)
2674 {
2675         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2676         int err;
2677         u64 search_start = cluster->window_start;
2678         u64 search_bytes = bytes;
2679         u64 ret = 0;
2680
2681         search_start = min_start;
2682         search_bytes = bytes;
2683
2684         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2685         if (err) {
2686                 if (search_bytes > *max_extent_size)
2687                         *max_extent_size = search_bytes;
2688                 return 0;
2689         }
2690
2691         ret = search_start;
2692         __bitmap_clear_bits(ctl, entry, ret, bytes);
2693
2694         return ret;
2695 }
2696
2697 /*
2698  * given a cluster, try to allocate 'bytes' from it, returns 0
2699  * if it couldn't find anything suitably large, or a logical disk offset
2700  * if things worked out
2701  */
2702 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2703                              struct btrfs_free_cluster *cluster, u64 bytes,
2704                              u64 min_start, u64 *max_extent_size)
2705 {
2706         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2707         struct btrfs_free_space *entry = NULL;
2708         struct rb_node *node;
2709         u64 ret = 0;
2710
2711         spin_lock(&cluster->lock);
2712         if (bytes > cluster->max_size)
2713                 goto out;
2714
2715         if (cluster->block_group != block_group)
2716                 goto out;
2717
2718         node = rb_first(&cluster->root);
2719         if (!node)
2720                 goto out;
2721
2722         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2723         while (1) {
2724                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2725                         *max_extent_size = entry->bytes;
2726
2727                 if (entry->bytes < bytes ||
2728                     (!entry->bitmap && entry->offset < min_start)) {
2729                         node = rb_next(&entry->offset_index);
2730                         if (!node)
2731                                 break;
2732                         entry = rb_entry(node, struct btrfs_free_space,
2733                                          offset_index);
2734                         continue;
2735                 }
2736
2737                 if (entry->bitmap) {
2738                         ret = btrfs_alloc_from_bitmap(block_group,
2739                                                       cluster, entry, bytes,
2740                                                       cluster->window_start,
2741                                                       max_extent_size);
2742                         if (ret == 0) {
2743                                 node = rb_next(&entry->offset_index);
2744                                 if (!node)
2745                                         break;
2746                                 entry = rb_entry(node, struct btrfs_free_space,
2747                                                  offset_index);
2748                                 continue;
2749                         }
2750                         cluster->window_start += bytes;
2751                 } else {
2752                         ret = entry->offset;
2753
2754                         entry->offset += bytes;
2755                         entry->bytes -= bytes;
2756                 }
2757
2758                 if (entry->bytes == 0)
2759                         rb_erase(&entry->offset_index, &cluster->root);
2760                 break;
2761         }
2762 out:
2763         spin_unlock(&cluster->lock);
2764
2765         if (!ret)
2766                 return 0;
2767
2768         spin_lock(&ctl->tree_lock);
2769
2770         ctl->free_space -= bytes;
2771         if (entry->bytes == 0) {
2772                 ctl->free_extents--;
2773                 if (entry->bitmap) {
2774                         kfree(entry->bitmap);
2775                         ctl->total_bitmaps--;
2776                         ctl->op->recalc_thresholds(ctl);
2777                 }
2778                 kmem_cache_free(btrfs_free_space_cachep, entry);
2779         }
2780
2781         spin_unlock(&ctl->tree_lock);
2782
2783         return ret;
2784 }
2785
2786 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2787                                 struct btrfs_free_space *entry,
2788                                 struct btrfs_free_cluster *cluster,
2789                                 u64 offset, u64 bytes,
2790                                 u64 cont1_bytes, u64 min_bytes)
2791 {
2792         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2793         unsigned long next_zero;
2794         unsigned long i;
2795         unsigned long want_bits;
2796         unsigned long min_bits;
2797         unsigned long found_bits;
2798         unsigned long max_bits = 0;
2799         unsigned long start = 0;
2800         unsigned long total_found = 0;
2801         int ret;
2802
2803         i = offset_to_bit(entry->offset, ctl->unit,
2804                           max_t(u64, offset, entry->offset));
2805         want_bits = bytes_to_bits(bytes, ctl->unit);
2806         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2807
2808         /*
2809          * Don't bother looking for a cluster in this bitmap if it's heavily
2810          * fragmented.
2811          */
2812         if (entry->max_extent_size &&
2813             entry->max_extent_size < cont1_bytes)
2814                 return -ENOSPC;
2815 again:
2816         found_bits = 0;
2817         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2818                 next_zero = find_next_zero_bit(entry->bitmap,
2819                                                BITS_PER_BITMAP, i);
2820                 if (next_zero - i >= min_bits) {
2821                         found_bits = next_zero - i;
2822                         if (found_bits > max_bits)
2823                                 max_bits = found_bits;
2824                         break;
2825                 }
2826                 if (next_zero - i > max_bits)
2827                         max_bits = next_zero - i;
2828                 i = next_zero;
2829         }
2830
2831         if (!found_bits) {
2832                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2833                 return -ENOSPC;
2834         }
2835
2836         if (!total_found) {
2837                 start = i;
2838                 cluster->max_size = 0;
2839         }
2840
2841         total_found += found_bits;
2842
2843         if (cluster->max_size < found_bits * ctl->unit)
2844                 cluster->max_size = found_bits * ctl->unit;
2845
2846         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2847                 i = next_zero + 1;
2848                 goto again;
2849         }
2850
2851         cluster->window_start = start * ctl->unit + entry->offset;
2852         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2853         ret = tree_insert_offset(&cluster->root, entry->offset,
2854                                  &entry->offset_index, 1);
2855         ASSERT(!ret); /* -EEXIST; Logic error */
2856
2857         trace_btrfs_setup_cluster(block_group, cluster,
2858                                   total_found * ctl->unit, 1);
2859         return 0;
2860 }
2861
2862 /*
2863  * This searches the block group for just extents to fill the cluster with.
2864  * Try to find a cluster with at least bytes total bytes, at least one
2865  * extent of cont1_bytes, and other clusters of at least min_bytes.
2866  */
2867 static noinline int
2868 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2869                         struct btrfs_free_cluster *cluster,
2870                         struct list_head *bitmaps, u64 offset, u64 bytes,
2871                         u64 cont1_bytes, u64 min_bytes)
2872 {
2873         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2874         struct btrfs_free_space *first = NULL;
2875         struct btrfs_free_space *entry = NULL;
2876         struct btrfs_free_space *last;
2877         struct rb_node *node;
2878         u64 window_free;
2879         u64 max_extent;
2880         u64 total_size = 0;
2881
2882         entry = tree_search_offset(ctl, offset, 0, 1);
2883         if (!entry)
2884                 return -ENOSPC;
2885
2886         /*
2887          * We don't want bitmaps, so just move along until we find a normal
2888          * extent entry.
2889          */
2890         while (entry->bitmap || entry->bytes < min_bytes) {
2891                 if (entry->bitmap && list_empty(&entry->list))
2892                         list_add_tail(&entry->list, bitmaps);
2893                 node = rb_next(&entry->offset_index);
2894                 if (!node)
2895                         return -ENOSPC;
2896                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2897         }
2898
2899         window_free = entry->bytes;
2900         max_extent = entry->bytes;
2901         first = entry;
2902         last = entry;
2903
2904         for (node = rb_next(&entry->offset_index); node;
2905              node = rb_next(&entry->offset_index)) {
2906                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2907
2908                 if (entry->bitmap) {
2909                         if (list_empty(&entry->list))
2910                                 list_add_tail(&entry->list, bitmaps);
2911                         continue;
2912                 }
2913
2914                 if (entry->bytes < min_bytes)
2915                         continue;
2916
2917                 last = entry;
2918                 window_free += entry->bytes;
2919                 if (entry->bytes > max_extent)
2920                         max_extent = entry->bytes;
2921         }
2922
2923         if (window_free < bytes || max_extent < cont1_bytes)
2924                 return -ENOSPC;
2925
2926         cluster->window_start = first->offset;
2927
2928         node = &first->offset_index;
2929
2930         /*
2931          * now we've found our entries, pull them out of the free space
2932          * cache and put them into the cluster rbtree
2933          */
2934         do {
2935                 int ret;
2936
2937                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2938                 node = rb_next(&entry->offset_index);
2939                 if (entry->bitmap || entry->bytes < min_bytes)
2940                         continue;
2941
2942                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2943                 ret = tree_insert_offset(&cluster->root, entry->offset,
2944                                          &entry->offset_index, 0);
2945                 total_size += entry->bytes;
2946                 ASSERT(!ret); /* -EEXIST; Logic error */
2947         } while (node && entry != last);
2948
2949         cluster->max_size = max_extent;
2950         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2951         return 0;
2952 }
2953
2954 /*
2955  * This specifically looks for bitmaps that may work in the cluster, we assume
2956  * that we have already failed to find extents that will work.
2957  */
2958 static noinline int
2959 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2960                      struct btrfs_free_cluster *cluster,
2961                      struct list_head *bitmaps, u64 offset, u64 bytes,
2962                      u64 cont1_bytes, u64 min_bytes)
2963 {
2964         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2965         struct btrfs_free_space *entry = NULL;
2966         int ret = -ENOSPC;
2967         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2968
2969         if (ctl->total_bitmaps == 0)
2970                 return -ENOSPC;
2971
2972         /*
2973          * The bitmap that covers offset won't be in the list unless offset
2974          * is just its start offset.
2975          */
2976         if (!list_empty(bitmaps))
2977                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2978
2979         if (!entry || entry->offset != bitmap_offset) {
2980                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2981                 if (entry && list_empty(&entry->list))
2982                         list_add(&entry->list, bitmaps);
2983         }
2984
2985         list_for_each_entry(entry, bitmaps, list) {
2986                 if (entry->bytes < bytes)
2987                         continue;
2988                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2989                                            bytes, cont1_bytes, min_bytes);
2990                 if (!ret)
2991                         return 0;
2992         }
2993
2994         /*
2995          * The bitmaps list has all the bitmaps that record free space
2996          * starting after offset, so no more search is required.
2997          */
2998         return -ENOSPC;
2999 }
3000
3001 /*
3002  * here we try to find a cluster of blocks in a block group.  The goal
3003  * is to find at least bytes+empty_size.
3004  * We might not find them all in one contiguous area.
3005  *
3006  * returns zero and sets up cluster if things worked out, otherwise
3007  * it returns -enospc
3008  */
3009 int btrfs_find_space_cluster(struct btrfs_fs_info *fs_info,
3010                              struct btrfs_block_group_cache *block_group,
3011                              struct btrfs_free_cluster *cluster,
3012                              u64 offset, u64 bytes, u64 empty_size)
3013 {
3014         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3015         struct btrfs_free_space *entry, *tmp;
3016         LIST_HEAD(bitmaps);
3017         u64 min_bytes;
3018         u64 cont1_bytes;
3019         int ret;
3020
3021         /*
3022          * Choose the minimum extent size we'll require for this
3023          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3024          * For metadata, allow allocates with smaller extents.  For
3025          * data, keep it dense.
3026          */
3027         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3028                 cont1_bytes = min_bytes = bytes + empty_size;
3029         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3030                 cont1_bytes = bytes;
3031                 min_bytes = fs_info->sectorsize;
3032         } else {
3033                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3034                 min_bytes = fs_info->sectorsize;
3035         }
3036
3037         spin_lock(&ctl->tree_lock);
3038
3039         /*
3040          * If we know we don't have enough space to make a cluster don't even
3041          * bother doing all the work to try and find one.
3042          */
3043         if (ctl->free_space < bytes) {
3044                 spin_unlock(&ctl->tree_lock);
3045                 return -ENOSPC;
3046         }
3047
3048         spin_lock(&cluster->lock);
3049
3050         /* someone already found a cluster, hooray */
3051         if (cluster->block_group) {
3052                 ret = 0;
3053                 goto out;
3054         }
3055
3056         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3057                                  min_bytes);
3058
3059         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3060                                       bytes + empty_size,
3061                                       cont1_bytes, min_bytes);
3062         if (ret)
3063                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3064                                            offset, bytes + empty_size,
3065                                            cont1_bytes, min_bytes);
3066
3067         /* Clear our temporary list */
3068         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3069                 list_del_init(&entry->list);
3070
3071         if (!ret) {
3072                 atomic_inc(&block_group->count);
3073                 list_add_tail(&cluster->block_group_list,
3074                               &block_group->cluster_list);
3075                 cluster->block_group = block_group;
3076         } else {
3077                 trace_btrfs_failed_cluster_setup(block_group);
3078         }
3079 out:
3080         spin_unlock(&cluster->lock);
3081         spin_unlock(&ctl->tree_lock);
3082
3083         return ret;
3084 }
3085
3086 /*
3087  * simple code to zero out a cluster
3088  */
3089 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3090 {
3091         spin_lock_init(&cluster->lock);
3092         spin_lock_init(&cluster->refill_lock);
3093         cluster->root = RB_ROOT;
3094         cluster->max_size = 0;
3095         cluster->fragmented = false;
3096         INIT_LIST_HEAD(&cluster->block_group_list);
3097         cluster->block_group = NULL;
3098 }
3099
3100 static int do_trimming(struct btrfs_block_group_cache *block_group,
3101                        u64 *total_trimmed, u64 start, u64 bytes,
3102                        u64 reserved_start, u64 reserved_bytes,
3103                        struct btrfs_trim_range *trim_entry)
3104 {
3105         struct btrfs_space_info *space_info = block_group->space_info;
3106         struct btrfs_fs_info *fs_info = block_group->fs_info;
3107         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3108         int ret;
3109         int update = 0;
3110         u64 trimmed = 0;
3111
3112         spin_lock(&space_info->lock);
3113         spin_lock(&block_group->lock);
3114         if (!block_group->ro) {
3115                 block_group->reserved += reserved_bytes;
3116                 space_info->bytes_reserved += reserved_bytes;
3117                 update = 1;
3118         }
3119         spin_unlock(&block_group->lock);
3120         spin_unlock(&space_info->lock);
3121
3122         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3123         if (!ret)
3124                 *total_trimmed += trimmed;
3125
3126         mutex_lock(&ctl->cache_writeout_mutex);
3127         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3128         list_del(&trim_entry->list);
3129         mutex_unlock(&ctl->cache_writeout_mutex);
3130
3131         if (update) {
3132                 spin_lock(&space_info->lock);