?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.19.133
[linux/fpc-iii.git] / fs / btrfs / free-space-cache.c
blobc9965e89097fdc2d593723c00f03fa6f5bce35f2
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_path(fs_info->sb, &location, root, NULL, path);
79 btrfs_release_path(path);
80 memalloc_nofs_restore(nofs_flag);
81 if (IS_ERR(inode))
82 return inode;
83
84 mapping_set_gfp_mask(inode->i_mapping,
85 mapping_gfp_constraint(inode->i_mapping,
86 ~(__GFP_FS | __GFP_HIGHMEM)));
87
88 return inode;
89 }
90
91 struct inode *lookup_free_space_inode(struct btrfs_fs_info *fs_info,
92 struct btrfs_block_group_cache
93 *block_group, struct btrfs_path *path)
94 {
95 struct inode *inode = NULL;
96 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
97
98 spin_lock(&block_group->lock);
99 if (block_group->inode)
100 inode = igrab(block_group->inode);
101 spin_unlock(&block_group->lock);
102 if (inode)
103 return inode;
104
105 inode = __lookup_free_space_inode(fs_info->tree_root, path,
106 block_group->key.objectid);
107 if (IS_ERR(inode))
108 return inode;
109
110 spin_lock(&block_group->lock);
111 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
112 btrfs_info(fs_info, "Old style space inode found, converting.");
113 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
114 BTRFS_INODE_NODATACOW;
115 block_group->disk_cache_state = BTRFS_DC_CLEAR;
116 }
117
118 if (!block_group->iref) {
119 block_group->inode = igrab(inode);
120 block_group->iref = 1;
121 }
122 spin_unlock(&block_group->lock);
123
124 return inode;
125 }
126
127 static int __create_free_space_inode(struct btrfs_root *root,
128 struct btrfs_trans_handle *trans,
129 struct btrfs_path *path,
130 u64 ino, u64 offset)
131 {
132 struct btrfs_key key;
133 struct btrfs_disk_key disk_key;
134 struct btrfs_free_space_header *header;
135 struct btrfs_inode_item *inode_item;
136 struct extent_buffer *leaf;
137 u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
138 int ret;
139
140 ret = btrfs_insert_empty_inode(trans, root, path, ino);
141 if (ret)
142 return ret;
143
144 /* We inline crc's for the free disk space cache */
145 if (ino != BTRFS_FREE_INO_OBJECTID)
146 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
147
148 leaf = path->nodes[0];
149 inode_item = btrfs_item_ptr(leaf, path->slots[0],
150 struct btrfs_inode_item);
151 btrfs_item_key(leaf, &disk_key, path->slots[0]);
152 memzero_extent_buffer(leaf, (unsigned long)inode_item,
153 sizeof(*inode_item));
154 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
155 btrfs_set_inode_size(leaf, inode_item, 0);
156 btrfs_set_inode_nbytes(leaf, inode_item, 0);
157 btrfs_set_inode_uid(leaf, inode_item, 0);
158 btrfs_set_inode_gid(leaf, inode_item, 0);
159 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
160 btrfs_set_inode_flags(leaf, inode_item, flags);
161 btrfs_set_inode_nlink(leaf, inode_item, 1);
162 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
163 btrfs_set_inode_block_group(leaf, inode_item, offset);
164 btrfs_mark_buffer_dirty(leaf);
165 btrfs_release_path(path);
166
167 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
168 key.offset = offset;
169 key.type = 0;
170 ret = btrfs_insert_empty_item(trans, root, path, &key,
171 sizeof(struct btrfs_free_space_header));
172 if (ret < 0) {
173 btrfs_release_path(path);
174 return ret;
175 }
176
177 leaf = path->nodes[0];
178 header = btrfs_item_ptr(leaf, path->slots[0],
179 struct btrfs_free_space_header);
180 memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
181 btrfs_set_free_space_key(leaf, header, &disk_key);
182 btrfs_mark_buffer_dirty(leaf);
183 btrfs_release_path(path);
184
185 return 0;
186 }
187
188 int create_free_space_inode(struct btrfs_fs_info *fs_info,
189 struct btrfs_trans_handle *trans,
190 struct btrfs_block_group_cache *block_group,
191 struct btrfs_path *path)
192 {
193 int ret;
194 u64 ino;
195
196 ret = btrfs_find_free_objectid(fs_info->tree_root, &ino);
197 if (ret < 0)
198 return ret;
199
200 return __create_free_space_inode(fs_info->tree_root, trans, path, ino,
201 block_group->key.objectid);
202 }
203
204 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
205 struct btrfs_block_rsv *rsv)
206 {
207 u64 needed_bytes;
208 int ret;
209
210 /* 1 for slack space, 1 for updating the inode */
211 needed_bytes = btrfs_calc_trunc_metadata_size(fs_info, 1) +
212 btrfs_calc_trans_metadata_size(fs_info, 1);
213
214 spin_lock(&rsv->lock);
215 if (rsv->reserved < needed_bytes)
216 ret = -ENOSPC;
217 else
218 ret = 0;
219 spin_unlock(&rsv->lock);
220 return ret;
221 }
222
223 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
224 struct btrfs_block_group_cache *block_group,
225 struct inode *inode)
226 {
227 struct btrfs_root *root = BTRFS_I(inode)->root;
228 int ret = 0;
229 bool locked = false;
230
231 if (block_group) {
232 struct btrfs_path *path = btrfs_alloc_path();
233
234 if (!path) {
235 ret = -ENOMEM;
236 goto fail;
237 }
238 locked = true;
239 mutex_lock(&trans->transaction->cache_write_mutex);
240 if (!list_empty(&block_group->io_list)) {
241 list_del_init(&block_group->io_list);
242
243 btrfs_wait_cache_io(trans, block_group, path);
244 btrfs_put_block_group(block_group);
245 }
246
247 /*
248 * now that we've truncated the cache away, its no longer
249 * setup or written
250 */
251 spin_lock(&block_group->lock);
252 block_group->disk_cache_state = BTRFS_DC_CLEAR;
253 spin_unlock(&block_group->lock);
254 btrfs_free_path(path);
255 }
256
257 btrfs_i_size_write(BTRFS_I(inode), 0);
258 truncate_pagecache(inode, 0);
259
260 /*
261 * We skip the throttling logic for free space cache inodes, so we don't
262 * need to check for -EAGAIN.
263 */
264 ret = btrfs_truncate_inode_items(trans, root, inode,
265 0, BTRFS_EXTENT_DATA_KEY);
266 if (ret)
267 goto fail;
268
269 ret = btrfs_update_inode(trans, root, inode);
270
271 fail:
272 if (locked)
273 mutex_unlock(&trans->transaction->cache_write_mutex);
274 if (ret)
275 btrfs_abort_transaction(trans, ret);
276
277 return ret;
278 }
279
280 static void readahead_cache(struct inode *inode)
281 {
282 struct file_ra_state *ra;
283 unsigned long last_index;
284
285 ra = kzalloc(sizeof(*ra), GFP_NOFS);
286 if (!ra)
287 return;
288
289 file_ra_state_init(ra, inode->i_mapping);
290 last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
291
292 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
293
294 kfree(ra);
295 }
296
297 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
298 int write)
299 {
300 int num_pages;
301 int check_crcs = 0;
302
303 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
304
305 if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
306 check_crcs = 1;
307
308 /* Make sure we can fit our crcs and generation into the first page */
309 if (write && check_crcs &&
310 (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
311 return -ENOSPC;
312
313 memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
314
315 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
316 if (!io_ctl->pages)
317 return -ENOMEM;
318
319 io_ctl->num_pages = num_pages;
320 io_ctl->fs_info = btrfs_sb(inode->i_sb);
321 io_ctl->check_crcs = check_crcs;
322 io_ctl->inode = inode;
323
324 return 0;
325 }
326 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
327
328 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
329 {
330 kfree(io_ctl->pages);
331 io_ctl->pages = NULL;
332 }
333
334 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
335 {
336 if (io_ctl->cur) {
337 io_ctl->cur = NULL;
338 io_ctl->orig = NULL;
339 }
340 }
341
342 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
343 {
344 ASSERT(io_ctl->index < io_ctl->num_pages);
345 io_ctl->page = io_ctl->pages[io_ctl->index++];
346 io_ctl->cur = page_address(io_ctl->page);
347 io_ctl->orig = io_ctl->cur;
348 io_ctl->size = PAGE_SIZE;
349 if (clear)
350 clear_page(io_ctl->cur);
351 }
352
353 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
354 {
355 int i;
356
357 io_ctl_unmap_page(io_ctl);
358
359 for (i = 0; i < io_ctl->num_pages; i++) {
360 if (io_ctl->pages[i]) {
361 ClearPageChecked(io_ctl->pages[i]);
362 unlock_page(io_ctl->pages[i]);
363 put_page(io_ctl->pages[i]);
364 }
365 }
366 }
367
368 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
369 int uptodate)
370 {
371 struct page *page;
372 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
373 int i;
374
375 for (i = 0; i < io_ctl->num_pages; i++) {
376 page = find_or_create_page(inode->i_mapping, i, mask);
377 if (!page) {
378 io_ctl_drop_pages(io_ctl);
379 return -ENOMEM;
380 }
381 io_ctl->pages[i] = page;
382 if (uptodate && !PageUptodate(page)) {
383 btrfs_readpage(NULL, page);
384 lock_page(page);
385 if (page->mapping != inode->i_mapping) {
386 btrfs_err(BTRFS_I(inode)->root->fs_info,
387 "free space cache page truncated");
388 io_ctl_drop_pages(io_ctl);
389 return -EIO;
390 }
391 if (!PageUptodate(page)) {
392 btrfs_err(BTRFS_I(inode)->root->fs_info,
393 "error reading free space cache");
394 io_ctl_drop_pages(io_ctl);
395 return -EIO;
396 }
397 }
398 }
399
400 for (i = 0; i < io_ctl->num_pages; i++) {
401 clear_page_dirty_for_io(io_ctl->pages[i]);
402 set_page_extent_mapped(io_ctl->pages[i]);
403 }
404
405 return 0;
406 }
407
408 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
409 {
410 __le64 *val;
411
412 io_ctl_map_page(io_ctl, 1);
413
414 /*
415 * Skip the csum areas. If we don't check crcs then we just have a
416 * 64bit chunk at the front of the first page.
417 */
418 if (io_ctl->check_crcs) {
419 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
420 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
421 } else {
422 io_ctl->cur += sizeof(u64);
423 io_ctl->size -= sizeof(u64) * 2;
424 }
425
426 val = io_ctl->cur;
427 *val = cpu_to_le64(generation);
428 io_ctl->cur += sizeof(u64);
429 }
430
431 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
432 {
433 __le64 *gen;
434
435 /*
436 * Skip the crc area. If we don't check crcs then we just have a 64bit
437 * chunk at the front of the first page.
438 */
439 if (io_ctl->check_crcs) {
440 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
441 io_ctl->size -= sizeof(u64) +
442 (sizeof(u32) * io_ctl->num_pages);
443 } else {
444 io_ctl->cur += sizeof(u64);
445 io_ctl->size -= sizeof(u64) * 2;
446 }
447
448 gen = io_ctl->cur;
449 if (le64_to_cpu(*gen) != generation) {
450 btrfs_err_rl(io_ctl->fs_info,
451 "space cache generation (%llu) does not match inode (%llu)",
452 *gen, generation);
453 io_ctl_unmap_page(io_ctl);
454 return -EIO;
455 }
456 io_ctl->cur += sizeof(u64);
457 return 0;
458 }
459
460 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
461 {
462 u32 *tmp;
463 u32 crc = ~(u32)0;
464 unsigned offset = 0;
465
466 if (!io_ctl->check_crcs) {
467 io_ctl_unmap_page(io_ctl);
468 return;
469 }
470
471 if (index == 0)
472 offset = sizeof(u32) * io_ctl->num_pages;
473
474 crc = btrfs_csum_data(io_ctl->orig + offset, crc,
475 PAGE_SIZE - offset);
476 btrfs_csum_final(crc, (u8 *)&crc);
477 io_ctl_unmap_page(io_ctl);
478 tmp = page_address(io_ctl->pages[0]);
479 tmp += index;
480 *tmp = crc;
481 }
482
483 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
484 {
485 u32 *tmp, val;
486 u32 crc = ~(u32)0;
487 unsigned offset = 0;
488
489 if (!io_ctl->check_crcs) {
490 io_ctl_map_page(io_ctl, 0);
491 return 0;
492 }
493
494 if (index == 0)
495 offset = sizeof(u32) * io_ctl->num_pages;
496
497 tmp = page_address(io_ctl->pages[0]);
498 tmp += index;
499 val = *tmp;
500
501 io_ctl_map_page(io_ctl, 0);
502 crc = btrfs_csum_data(io_ctl->orig + offset, crc,
503 PAGE_SIZE - offset);
504 btrfs_csum_final(crc, (u8 *)&crc);
505 if (val != crc) {
506 btrfs_err_rl(io_ctl->fs_info,
507 "csum mismatch on free space cache");
508 io_ctl_unmap_page(io_ctl);
509 return -EIO;
510 }
511
512 return 0;
513 }
514
515 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
516 void *bitmap)
517 {
518 struct btrfs_free_space_entry *entry;
519
520 if (!io_ctl->cur)
521 return -ENOSPC;
522
523 entry = io_ctl->cur;
524 entry->offset = cpu_to_le64(offset);
525 entry->bytes = cpu_to_le64(bytes);
526 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
527 BTRFS_FREE_SPACE_EXTENT;
528 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
529 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
530
531 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
532 return 0;
533
534 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
535
536 /* No more pages to map */
537 if (io_ctl->index >= io_ctl->num_pages)
538 return 0;
539
540 /* map the next page */
541 io_ctl_map_page(io_ctl, 1);
542 return 0;
543 }
544
545 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
546 {
547 if (!io_ctl->cur)
548 return -ENOSPC;
549
550 /*
551 * If we aren't at the start of the current page, unmap this one and
552 * map the next one if there is any left.
553 */
554 if (io_ctl->cur != io_ctl->orig) {
555 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
556 if (io_ctl->index >= io_ctl->num_pages)
557 return -ENOSPC;
558 io_ctl_map_page(io_ctl, 0);
559 }
560
561 copy_page(io_ctl->cur, bitmap);
562 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
563 if (io_ctl->index < io_ctl->num_pages)
564 io_ctl_map_page(io_ctl, 0);
565 return 0;
566 }
567
568 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
569 {
570 /*
571 * If we're not on the boundary we know we've modified the page and we
572 * need to crc the page.
573 */
574 if (io_ctl->cur != io_ctl->orig)
575 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
576 else
577 io_ctl_unmap_page(io_ctl);
578
579 while (io_ctl->index < io_ctl->num_pages) {
580 io_ctl_map_page(io_ctl, 1);
581 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
582 }
583 }
584
585 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
586 struct btrfs_free_space *entry, u8 *type)
587 {
588 struct btrfs_free_space_entry *e;
589 int ret;
590
591 if (!io_ctl->cur) {
592 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
593 if (ret)
594 return ret;
595 }
596
597 e = io_ctl->cur;
598 entry->offset = le64_to_cpu(e->offset);
599 entry->bytes = le64_to_cpu(e->bytes);
600 *type = e->type;
601 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
602 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
603
604 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
605 return 0;
606
607 io_ctl_unmap_page(io_ctl);
608
609 return 0;
610 }
611
612 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
613 struct btrfs_free_space *entry)
614 {
615 int ret;
616
617 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
618 if (ret)
619 return ret;
620
621 copy_page(entry->bitmap, io_ctl->cur);
622 io_ctl_unmap_page(io_ctl);
623
624 return 0;
625 }
626
627 /*
628 * Since we attach pinned extents after the fact we can have contiguous sections
629 * of free space that are split up in entries. This poses a problem with the
630 * tree logging stuff since it could have allocated across what appears to be 2
631 * entries since we would have merged the entries when adding the pinned extents
632 * back to the free space cache. So run through the space cache that we just
633 * loaded and merge contiguous entries. This will make the log replay stuff not
634 * blow up and it will make for nicer allocator behavior.
635 */
636 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
637 {
638 struct btrfs_free_space *e, *prev = NULL;
639 struct rb_node *n;
640
641 again:
642 spin_lock(&ctl->tree_lock);
643 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
644 e = rb_entry(n, struct btrfs_free_space, offset_index);
645 if (!prev)
646 goto next;
647 if (e->bitmap || prev->bitmap)
648 goto next;
649 if (prev->offset + prev->bytes == e->offset) {
650 unlink_free_space(ctl, prev);
651 unlink_free_space(ctl, e);
652 prev->bytes += e->bytes;
653 kmem_cache_free(btrfs_free_space_cachep, e);
654 link_free_space(ctl, prev);
655 prev = NULL;
656 spin_unlock(&ctl->tree_lock);
657 goto again;
658 }
659 next:
660 prev = e;
661 }
662 spin_unlock(&ctl->tree_lock);
663 }
664
665 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
666 struct btrfs_free_space_ctl *ctl,
667 struct btrfs_path *path, u64 offset)
668 {
669 struct btrfs_fs_info *fs_info = root->fs_info;
670 struct btrfs_free_space_header *header;
671 struct extent_buffer *leaf;
672 struct btrfs_io_ctl io_ctl;
673 struct btrfs_key key;
674 struct btrfs_free_space *e, *n;
675 LIST_HEAD(bitmaps);
676 u64 num_entries;
677 u64 num_bitmaps;
678 u64 generation;
679 u8 type;
680 int ret = 0;
681
682 /* Nothing in the space cache, goodbye */
683 if (!i_size_read(inode))
684 return 0;
685
686 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
687 key.offset = offset;
688 key.type = 0;
689
690 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
691 if (ret < 0)
692 return 0;
693 else if (ret > 0) {
694 btrfs_release_path(path);
695 return 0;
696 }
697
698 ret = -1;
699
700 leaf = path->nodes[0];
701 header = btrfs_item_ptr(leaf, path->slots[0],
702 struct btrfs_free_space_header);
703 num_entries = btrfs_free_space_entries(leaf, header);
704 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
705 generation = btrfs_free_space_generation(leaf, header);
706 btrfs_release_path(path);
707
708 if (!BTRFS_I(inode)->generation) {
709 btrfs_info(fs_info,
710 "the free space cache file (%llu) is invalid, skip it",
711 offset);
712 return 0;
713 }
714
715 if (BTRFS_I(inode)->generation != generation) {
716 btrfs_err(fs_info,
717 "free space inode generation (%llu) did not match free space cache generation (%llu)",
718 BTRFS_I(inode)->generation, generation);
719 return 0;
720 }
721
722 if (!num_entries)
723 return 0;
724
725 ret = io_ctl_init(&io_ctl, inode, 0);
726 if (ret)
727 return ret;
728
729 readahead_cache(inode);
730
731 ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
732 if (ret)
733 goto out;
734
735 ret = io_ctl_check_crc(&io_ctl, 0);
736 if (ret)
737 goto free_cache;
738
739 ret = io_ctl_check_generation(&io_ctl, generation);
740 if (ret)
741 goto free_cache;
742
743 while (num_entries) {
744 e = kmem_cache_zalloc(btrfs_free_space_cachep,
745 GFP_NOFS);
746 if (!e)
747 goto free_cache;
748
749 ret = io_ctl_read_entry(&io_ctl, e, &type);
750 if (ret) {
751 kmem_cache_free(btrfs_free_space_cachep, e);
752 goto free_cache;
753 }
754
755 if (!e->bytes) {
756 kmem_cache_free(btrfs_free_space_cachep, e);
757 goto free_cache;
758 }
759
760 if (type == BTRFS_FREE_SPACE_EXTENT) {
761 spin_lock(&ctl->tree_lock);
762 ret = link_free_space(ctl, e);
763 spin_unlock(&ctl->tree_lock);
764 if (ret) {
765 btrfs_err(fs_info,
766 "Duplicate entries in free space cache, dumping");
767 kmem_cache_free(btrfs_free_space_cachep, e);
768 goto free_cache;
769 }
770 } else {
771 ASSERT(num_bitmaps);
772 num_bitmaps--;
773 e->bitmap = kmem_cache_zalloc(
774 btrfs_free_space_bitmap_cachep, GFP_NOFS);
775 if (!e->bitmap) {
776 kmem_cache_free(
777 btrfs_free_space_cachep, e);
778 goto free_cache;
779 }
780 spin_lock(&ctl->tree_lock);
781 ret = link_free_space(ctl, e);
782 ctl->total_bitmaps++;
783 ctl->op->recalc_thresholds(ctl);
784 spin_unlock(&ctl->tree_lock);
785 if (ret) {
786 btrfs_err(fs_info,
787 "Duplicate entries in free space cache, dumping");
788 kmem_cache_free(btrfs_free_space_cachep, e);
789 goto free_cache;
790 }
791 list_add_tail(&e->list, &bitmaps);
792 }
793
794 num_entries--;
795 }
796
797 io_ctl_unmap_page(&io_ctl);
798
799 /*
800 * We add the bitmaps at the end of the entries in order that
801 * the bitmap entries are added to the cache.
802 */
803 list_for_each_entry_safe(e, n, &bitmaps, list) {
804 list_del_init(&e->list);
805 ret = io_ctl_read_bitmap(&io_ctl, e);
806 if (ret)
807 goto free_cache;
808 }
809
810 io_ctl_drop_pages(&io_ctl);
811 merge_space_tree(ctl);
812 ret = 1;
813 out:
814 io_ctl_free(&io_ctl);
815 return ret;
816 free_cache:
817 io_ctl_drop_pages(&io_ctl);
818 __btrfs_remove_free_space_cache(ctl);
819 goto out;
820 }
821
822 int load_free_space_cache(struct btrfs_fs_info *fs_info,
823 struct btrfs_block_group_cache *block_group)
824 {
825 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
826 struct inode *inode;
827 struct btrfs_path *path;
828 int ret = 0;
829 bool matched;
830 u64 used = btrfs_block_group_used(&block_group->item);
831
832 /*
833 * If this block group has been marked to be cleared for one reason or
834 * another then we can't trust the on disk cache, so just return.
835 */
836 spin_lock(&block_group->lock);
837 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
838 spin_unlock(&block_group->lock);
839 return 0;
840 }
841 spin_unlock(&block_group->lock);
842
843 path = btrfs_alloc_path();
844 if (!path)
845 return 0;
846 path->search_commit_root = 1;
847 path->skip_locking = 1;
848
849 /*
850 * We must pass a path with search_commit_root set to btrfs_iget in
851 * order to avoid a deadlock when allocating extents for the tree root.
852 *
853 * When we are COWing an extent buffer from the tree root, when looking
854 * for a free extent, at extent-tree.c:find_free_extent(), we can find
855 * block group without its free space cache loaded. When we find one
856 * we must load its space cache which requires reading its free space
857 * cache's inode item from the root tree. If this inode item is located
858 * in the same leaf that we started COWing before, then we end up in
859 * deadlock on the extent buffer (trying to read lock it when we
860 * previously write locked it).
861 *
862 * It's safe to read the inode item using the commit root because
863 * block groups, once loaded, stay in memory forever (until they are
864 * removed) as well as their space caches once loaded. New block groups
865 * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
866 * we will never try to read their inode item while the fs is mounted.
867 */
868 inode = lookup_free_space_inode(fs_info, block_group, path);
869 if (IS_ERR(inode)) {
870 btrfs_free_path(path);
871 return 0;
872 }
873
874 /* We may have converted the inode and made the cache invalid. */
875 spin_lock(&block_group->lock);
876 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
877 spin_unlock(&block_group->lock);
878 btrfs_free_path(path);
879 goto out;
880 }
881 spin_unlock(&block_group->lock);
882
883 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
884 path, block_group->key.objectid);
885 btrfs_free_path(path);
886 if (ret <= 0)
887 goto out;
888
889 spin_lock(&ctl->tree_lock);
890 matched = (ctl->free_space == (block_group->key.offset - used -
891 block_group->bytes_super));
892 spin_unlock(&ctl->tree_lock);
893
894 if (!matched) {
895 __btrfs_remove_free_space_cache(ctl);
896 btrfs_warn(fs_info,
897 "block group %llu has wrong amount of free space",
898 block_group->key.objectid);
899 ret = -1;
900 }
901 out:
902 if (ret < 0) {
903 /* This cache is bogus, make sure it gets cleared */
904 spin_lock(&block_group->lock);
905 block_group->disk_cache_state = BTRFS_DC_CLEAR;
906 spin_unlock(&block_group->lock);
907 ret = 0;
908
909 btrfs_warn(fs_info,
910 "failed to load free space cache for block group %llu, rebuilding it now",
911 block_group->key.objectid);
912 }
913
914 iput(inode);
915 return ret;
916 }
917
918 static noinline_for_stack
919 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
920 struct btrfs_free_space_ctl *ctl,
921 struct btrfs_block_group_cache *block_group,
922 int *entries, int *bitmaps,
923 struct list_head *bitmap_list)
924 {
925 int ret;
926 struct btrfs_free_cluster *cluster = NULL;
927 struct btrfs_free_cluster *cluster_locked = NULL;
928 struct rb_node *node = rb_first(&ctl->free_space_offset);
929 struct btrfs_trim_range *trim_entry;
930
931 /* Get the cluster for this block_group if it exists */
932 if (block_group && !list_empty(&block_group->cluster_list)) {
933 cluster = list_entry(block_group->cluster_list.next,
934 struct btrfs_free_cluster,
935 block_group_list);
936 }
937
938 if (!node && cluster) {
939 cluster_locked = cluster;
940 spin_lock(&cluster_locked->lock);
941 node = rb_first(&cluster->root);
942 cluster = NULL;
943 }
944
945 /* Write out the extent entries */
946 while (node) {
947 struct btrfs_free_space *e;
948
949 e = rb_entry(node, struct btrfs_free_space, offset_index);
950 *entries += 1;
951
952 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
953 e->bitmap);
954 if (ret)
955 goto fail;
956
957 if (e->bitmap) {
958 list_add_tail(&e->list, bitmap_list);
959 *bitmaps += 1;
960 }
961 node = rb_next(node);
962 if (!node && cluster) {
963 node = rb_first(&cluster->root);
964 cluster_locked = cluster;
965 spin_lock(&cluster_locked->lock);
966 cluster = NULL;
967 }
968 }
969 if (cluster_locked) {
970 spin_unlock(&cluster_locked->lock);
971 cluster_locked = NULL;
972 }
973
974 /*
975 * Make sure we don't miss any range that was removed from our rbtree
976 * because trimming is running. Otherwise after a umount+mount (or crash
977 * after committing the transaction) we would leak free space and get
978 * an inconsistent free space cache report from fsck.
979 */
980 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
981 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
982 trim_entry->bytes, NULL);
983 if (ret)
984 goto fail;
985 *entries += 1;
986 }
987
988 return 0;
989 fail:
990 if (cluster_locked)
991 spin_unlock(&cluster_locked->lock);
992 return -ENOSPC;
993 }
994
995 static noinline_for_stack int
996 update_cache_item(struct btrfs_trans_handle *trans,
997 struct btrfs_root *root,
998 struct inode *inode,
999 struct btrfs_path *path, u64 offset,
1000 int entries, int bitmaps)
1001 {
1002 struct btrfs_key key;
1003 struct btrfs_free_space_header *header;
1004 struct extent_buffer *leaf;
1005 int ret;
1006
1007 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1008 key.offset = offset;
1009 key.type = 0;
1010
1011 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1012 if (ret < 0) {
1013 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1014 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
1015 goto fail;
1016 }
1017 leaf = path->nodes[0];
1018 if (ret > 0) {
1019 struct btrfs_key found_key;
1020 ASSERT(path->slots[0]);
1021 path->slots[0]--;
1022 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1023 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1024 found_key.offset != offset) {
1025 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1026 inode->i_size - 1,
1027 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1028 NULL);
1029 btrfs_release_path(path);
1030 goto fail;
1031 }
1032 }
1033
1034 BTRFS_I(inode)->generation = trans->transid;
1035 header = btrfs_item_ptr(leaf, path->slots[0],
1036 struct btrfs_free_space_header);
1037 btrfs_set_free_space_entries(leaf, header, entries);
1038 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1039 btrfs_set_free_space_generation(leaf, header, trans->transid);
1040 btrfs_mark_buffer_dirty(leaf);
1041 btrfs_release_path(path);
1042
1043 return 0;
1044
1045 fail:
1046 return -1;
1047 }
1048
1049 static noinline_for_stack int
1050 write_pinned_extent_entries(struct btrfs_fs_info *fs_info,
1051 struct btrfs_block_group_cache *block_group,
1052 struct btrfs_io_ctl *io_ctl,
1053 int *entries)
1054 {
1055 u64 start, extent_start, extent_end, len;
1056 struct extent_io_tree *unpin = NULL;
1057 int ret;
1058
1059 if (!block_group)
1060 return 0;
1061
1062 /*
1063 * We want to add any pinned extents to our free space cache
1064 * so we don't leak the space
1065 *
1066 * We shouldn't have switched the pinned extents yet so this is the
1067 * right one
1068 */
1069 unpin = fs_info->pinned_extents;
1070
1071 start = block_group->key.objectid;
1072
1073 while (start < block_group->key.objectid + block_group->key.offset) {
1074 ret = find_first_extent_bit(unpin, start,
1075 &extent_start, &extent_end,
1076 EXTENT_DIRTY, NULL);
1077 if (ret)
1078 return 0;
1079
1080 /* This pinned extent is out of our range */
1081 if (extent_start >= block_group->key.objectid +
1082 block_group->key.offset)
1083 return 0;
1084
1085 extent_start = max(extent_start, start);
1086 extent_end = min(block_group->key.objectid +
1087 block_group->key.offset, extent_end + 1);
1088 len = extent_end - extent_start;
1089
1090 *entries += 1;
1091 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1092 if (ret)
1093 return -ENOSPC;
1094
1095 start = extent_end;
1096 }
1097
1098 return 0;
1099 }
1100
1101 static noinline_for_stack int
1102 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1103 {
1104 struct btrfs_free_space *entry, *next;
1105 int ret;
1106
1107 /* Write out the bitmaps */
1108 list_for_each_entry_safe(entry, next, bitmap_list, list) {
1109 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1110 if (ret)
1111 return -ENOSPC;
1112 list_del_init(&entry->list);
1113 }
1114
1115 return 0;
1116 }
1117
1118 static int flush_dirty_cache(struct inode *inode)
1119 {
1120 int ret;
1121
1122 ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1123 if (ret)
1124 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1125 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
1126
1127 return ret;
1128 }
1129
1130 static void noinline_for_stack
1131 cleanup_bitmap_list(struct list_head *bitmap_list)
1132 {
1133 struct btrfs_free_space *entry, *next;
1134
1135 list_for_each_entry_safe(entry, next, bitmap_list, list)
1136 list_del_init(&entry->list);
1137 }
1138
1139 static void noinline_for_stack
1140 cleanup_write_cache_enospc(struct inode *inode,
1141 struct btrfs_io_ctl *io_ctl,
1142 struct extent_state **cached_state)
1143 {
1144 io_ctl_drop_pages(io_ctl);
1145 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1146 i_size_read(inode) - 1, cached_state);
1147 }
1148
1149 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1150 struct btrfs_trans_handle *trans,
1151 struct btrfs_block_group_cache *block_group,
1152 struct btrfs_io_ctl *io_ctl,
1153 struct btrfs_path *path, u64 offset)
1154 {
1155 int ret;
1156 struct inode *inode = io_ctl->inode;
1157
1158 if (!inode)
1159 return 0;
1160
1161 /* Flush the dirty pages in the cache file. */
1162 ret = flush_dirty_cache(inode);
1163 if (ret)
1164 goto out;
1165
1166 /* Update the cache item to tell everyone this cache file is valid. */
1167 ret = update_cache_item(trans, root, inode, path, offset,
1168 io_ctl->entries, io_ctl->bitmaps);
1169 out:
1170 io_ctl_free(io_ctl);
1171 if (ret) {
1172 invalidate_inode_pages2(inode->i_mapping);
1173 BTRFS_I(inode)->generation = 0;
1174 if (block_group) {
1175 #ifdef DEBUG
1176 btrfs_err(root->fs_info,
1177 "failed to write free space cache for block group %llu",
1178 block_group->key.objectid);
1179 #endif
1180 }
1181 }
1182 btrfs_update_inode(trans, root, inode);
1183
1184 if (block_group) {
1185 /* the dirty list is protected by the dirty_bgs_lock */
1186 spin_lock(&trans->transaction->dirty_bgs_lock);
1187
1188 /* the disk_cache_state is protected by the block group lock */
1189 spin_lock(&block_group->lock);
1190
1191 /*
1192 * only mark this as written if we didn't get put back on
1193 * the dirty list while waiting for IO. Otherwise our
1194 * cache state won't be right, and we won't get written again
1195 */
1196 if (!ret && list_empty(&block_group->dirty_list))
1197 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1198 else if (ret)
1199 block_group->disk_cache_state = BTRFS_DC_ERROR;
1200
1201 spin_unlock(&block_group->lock);
1202 spin_unlock(&trans->transaction->dirty_bgs_lock);
1203 io_ctl->inode = NULL;
1204 iput(inode);
1205 }
1206
1207 return ret;
1208
1209 }
1210
1211 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1212 struct btrfs_trans_handle *trans,
1213 struct btrfs_io_ctl *io_ctl,
1214 struct btrfs_path *path)
1215 {
1216 return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1217 }
1218
1219 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1220 struct btrfs_block_group_cache *block_group,
1221 struct btrfs_path *path)
1222 {
1223 return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1224 block_group, &block_group->io_ctl,
1225 path, block_group->key.objectid);
1226 }
1227
1228 /**
1229 * __btrfs_write_out_cache - write out cached info to an inode
1230 * @root - the root the inode belongs to
1231 * @ctl - the free space cache we are going to write out
1232 * @block_group - the block_group for this cache if it belongs to a block_group
1233 * @trans - the trans handle
1234 *
1235 * This function writes out a free space cache struct to disk for quick recovery
1236 * on mount. This will return 0 if it was successful in writing the cache out,
1237 * or an errno if it was not.
1238 */
1239 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1240 struct btrfs_free_space_ctl *ctl,
1241 struct btrfs_block_group_cache *block_group,
1242 struct btrfs_io_ctl *io_ctl,
1243 struct btrfs_trans_handle *trans)
1244 {
1245 struct btrfs_fs_info *fs_info = root->fs_info;
1246 struct extent_state *cached_state = NULL;
1247 LIST_HEAD(bitmap_list);
1248 int entries = 0;
1249 int bitmaps = 0;
1250 int ret;
1251 int must_iput = 0;
1252
1253 if (!i_size_read(inode))
1254 return -EIO;
1255
1256 WARN_ON(io_ctl->pages);
1257 ret = io_ctl_init(io_ctl, inode, 1);
1258 if (ret)
1259 return ret;
1260
1261 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1262 down_write(&block_group->data_rwsem);
1263 spin_lock(&block_group->lock);
1264 if (block_group->delalloc_bytes) {
1265 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1266 spin_unlock(&block_group->lock);
1267 up_write(&block_group->data_rwsem);
1268 BTRFS_I(inode)->generation = 0;
1269 ret = 0;
1270 must_iput = 1;
1271 goto out;
1272 }
1273 spin_unlock(&block_group->lock);
1274 }
1275
1276 /* Lock all pages first so we can lock the extent safely. */
1277 ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1278 if (ret)
1279 goto out_unlock;
1280
1281 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1282 &cached_state);
1283
1284 io_ctl_set_generation(io_ctl, trans->transid);
1285
1286 mutex_lock(&ctl->cache_writeout_mutex);
1287 /* Write out the extent entries in the free space cache */
1288 spin_lock(&ctl->tree_lock);
1289 ret = write_cache_extent_entries(io_ctl, ctl,
1290 block_group, &entries, &bitmaps,
1291 &bitmap_list);
1292 if (ret)
1293 goto out_nospc_locked;
1294
1295 /*
1296 * Some spaces that are freed in the current transaction are pinned,
1297 * they will be added into free space cache after the transaction is
1298 * committed, we shouldn't lose them.
1299 *
1300 * If this changes while we are working we'll get added back to
1301 * the dirty list and redo it. No locking needed
1302 */
1303 ret = write_pinned_extent_entries(fs_info, block_group,
1304 io_ctl, &entries);
1305 if (ret)
1306 goto out_nospc_locked;
1307
1308 /*
1309 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1310 * locked while doing it because a concurrent trim can be manipulating
1311 * or freeing the bitmap.
1312 */
1313 ret = write_bitmap_entries(io_ctl, &bitmap_list);
1314 spin_unlock(&ctl->tree_lock);
1315 mutex_unlock(&ctl->cache_writeout_mutex);
1316 if (ret)
1317 goto out_nospc;
1318
1319 /* Zero out the rest of the pages just to make sure */
1320 io_ctl_zero_remaining_pages(io_ctl);
1321
1322 /* Everything is written out, now we dirty the pages in the file. */
1323 ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1324 i_size_read(inode), &cached_state);
1325 if (ret)
1326 goto out_nospc;
1327
1328 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1329 up_write(&block_group->data_rwsem);
1330 /*
1331 * Release the pages and unlock the extent, we will flush
1332 * them out later
1333 */
1334 io_ctl_drop_pages(io_ctl);
1335
1336 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1337 i_size_read(inode) - 1, &cached_state);
1338
1339 /*
1340 * at this point the pages are under IO and we're happy,
1341 * The caller is responsible for waiting on them and updating the
1342 * the cache and the inode
1343 */
1344 io_ctl->entries = entries;
1345 io_ctl->bitmaps = bitmaps;
1346
1347 ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1348 if (ret)
1349 goto out;
1350
1351 return 0;
1352
1353 out:
1354 io_ctl->inode = NULL;
1355 io_ctl_free(io_ctl);
1356 if (ret) {
1357 invalidate_inode_pages2(inode->i_mapping);
1358 BTRFS_I(inode)->generation = 0;
1359 }
1360 btrfs_update_inode(trans, root, inode);
1361 if (must_iput)
1362 iput(inode);
1363 return ret;
1364
1365 out_nospc_locked:
1366 cleanup_bitmap_list(&bitmap_list);
1367 spin_unlock(&ctl->tree_lock);
1368 mutex_unlock(&ctl->cache_writeout_mutex);
1369
1370 out_nospc:
1371 cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1372
1373 out_unlock:
1374 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1375 up_write(&block_group->data_rwsem);
1376
1377 goto out;
1378 }
1379
1380 int btrfs_write_out_cache(struct btrfs_fs_info *fs_info,
1381 struct btrfs_trans_handle *trans,
1382 struct btrfs_block_group_cache *block_group,
1383 struct btrfs_path *path)
1384 {
1385 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1386 struct inode *inode;
1387 int ret = 0;
1388
1389 spin_lock(&block_group->lock);
1390 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1391 spin_unlock(&block_group->lock);
1392 return 0;
1393 }
1394 spin_unlock(&block_group->lock);
1395
1396 inode = lookup_free_space_inode(fs_info, block_group, path);
1397 if (IS_ERR(inode))
1398 return 0;
1399
1400 ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1401 block_group, &block_group->io_ctl, trans);
1402 if (ret) {
1403 #ifdef DEBUG
1404 btrfs_err(fs_info,
1405 "failed to write free space cache for block group %llu",
1406 block_group->key.objectid);
1407 #endif
1408 spin_lock(&block_group->lock);
1409 block_group->disk_cache_state = BTRFS_DC_ERROR;
1410 spin_unlock(&block_group->lock);
1411
1412 block_group->io_ctl.inode = NULL;
1413 iput(inode);
1414 }
1415
1416 /*
1417 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1418 * to wait for IO and put the inode
1419 */
1420
1421 return ret;
1422 }
1423
1424 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1425 u64 offset)
1426 {
1427 ASSERT(offset >= bitmap_start);
1428 offset -= bitmap_start;
1429 return (unsigned long)(div_u64(offset, unit));
1430 }
1431
1432 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1433 {
1434 return (unsigned long)(div_u64(bytes, unit));
1435 }
1436
1437 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1438 u64 offset)
1439 {
1440 u64 bitmap_start;
1441 u64 bytes_per_bitmap;
1442
1443 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1444 bitmap_start = offset - ctl->start;
1445 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1446 bitmap_start *= bytes_per_bitmap;
1447 bitmap_start += ctl->start;
1448
1449 return bitmap_start;
1450 }
1451
1452 static int tree_insert_offset(struct rb_root *root, u64 offset,
1453 struct rb_node *node, int bitmap)
1454 {
1455 struct rb_node **p = &root->rb_node;
1456 struct rb_node *parent = NULL;
1457 struct btrfs_free_space *info;
1458
1459 while (*p) {
1460 parent = *p;
1461 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1462
1463 if (offset < info->offset) {
1464 p = &(*p)->rb_left;
1465 } else if (offset > info->offset) {
1466 p = &(*p)->rb_right;
1467 } else {
1468 /*
1469 * we could have a bitmap entry and an extent entry
1470 * share the same offset. If this is the case, we want
1471 * the extent entry to always be found first if we do a
1472 * linear search through the tree, since we want to have
1473 * the quickest allocation time, and allocating from an
1474 * extent is faster than allocating from a bitmap. So
1475 * if we're inserting a bitmap and we find an entry at
1476 * this offset, we want to go right, or after this entry
1477 * logically. If we are inserting an extent and we've
1478 * found a bitmap, we want to go left, or before
1479 * logically.
1480 */
1481 if (bitmap) {
1482 if (info->bitmap) {
1483 WARN_ON_ONCE(1);
1484 return -EEXIST;
1485 }
1486 p = &(*p)->rb_right;
1487 } else {
1488 if (!info->bitmap) {
1489 WARN_ON_ONCE(1);
1490 return -EEXIST;
1491 }
1492 p = &(*p)->rb_left;
1493 }
1494 }
1495 }
1496
1497 rb_link_node(node, parent, p);
1498 rb_insert_color(node, root);
1499
1500 return 0;
1501 }
1502
1503 /*
1504 * searches the tree for the given offset.
1505 *
1506 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1507 * want a section that has at least bytes size and comes at or after the given
1508 * offset.
1509 */
1510 static struct btrfs_free_space *
1511 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1512 u64 offset, int bitmap_only, int fuzzy)
1513 {
1514 struct rb_node *n = ctl->free_space_offset.rb_node;
1515 struct btrfs_free_space *entry, *prev = NULL;
1516
1517 /* find entry that is closest to the 'offset' */
1518 while (1) {
1519 if (!n) {
1520 entry = NULL;
1521 break;
1522 }
1523
1524 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1525 prev = entry;
1526
1527 if (offset < entry->offset)
1528 n = n->rb_left;
1529 else if (offset > entry->offset)
1530 n = n->rb_right;
1531 else
1532 break;
1533 }
1534
1535 if (bitmap_only) {
1536 if (!entry)
1537 return NULL;
1538 if (entry->bitmap)
1539 return entry;
1540
1541 /*
1542 * bitmap entry and extent entry may share same offset,
1543 * in that case, bitmap entry comes after extent entry.
1544 */
1545 n = rb_next(n);
1546 if (!n)
1547 return NULL;
1548 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1549 if (entry->offset != offset)
1550 return NULL;
1551
1552 WARN_ON(!entry->bitmap);
1553 return entry;
1554 } else if (entry) {
1555 if (entry->bitmap) {
1556 /*
1557 * if previous extent entry covers the offset,
1558 * we should return it instead of the bitmap entry
1559 */
1560 n = rb_prev(&entry->offset_index);
1561 if (n) {
1562 prev = rb_entry(n, struct btrfs_free_space,
1563 offset_index);
1564 if (!prev->bitmap &&
1565 prev->offset + prev->bytes > offset)
1566 entry = prev;
1567 }
1568 }
1569 return entry;
1570 }
1571
1572 if (!prev)
1573 return NULL;
1574
1575 /* find last entry before the 'offset' */
1576 entry = prev;
1577 if (entry->offset > offset) {
1578 n = rb_prev(&entry->offset_index);
1579 if (n) {
1580 entry = rb_entry(n, struct btrfs_free_space,
1581 offset_index);
1582 ASSERT(entry->offset <= offset);
1583 } else {
1584 if (fuzzy)
1585 return entry;
1586 else
1587 return NULL;
1588 }
1589 }
1590
1591 if (entry->bitmap) {
1592 n = rb_prev(&entry->offset_index);
1593 if (n) {
1594 prev = rb_entry(n, struct btrfs_free_space,
1595 offset_index);
1596 if (!prev->bitmap &&
1597 prev->offset + prev->bytes > offset)
1598 return prev;
1599 }
1600 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1601 return entry;
1602 } else if (entry->offset + entry->bytes > offset)
1603 return entry;
1604
1605 if (!fuzzy)
1606 return NULL;
1607
1608 while (1) {
1609 if (entry->bitmap) {
1610 if (entry->offset + BITS_PER_BITMAP *
1611 ctl->unit > offset)
1612 break;
1613 } else {
1614 if (entry->offset + entry->bytes > offset)
1615 break;
1616 }
1617
1618 n = rb_next(&entry->offset_index);
1619 if (!n)
1620 return NULL;
1621 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1622 }
1623 return entry;
1624 }
1625
1626 static inline void
1627 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1628 struct btrfs_free_space *info)
1629 {
1630 rb_erase(&info->offset_index, &ctl->free_space_offset);
1631 ctl->free_extents--;
1632 }
1633
1634 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1635 struct btrfs_free_space *info)
1636 {
1637 __unlink_free_space(ctl, info);
1638 ctl->free_space -= info->bytes;
1639 }
1640
1641 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1642 struct btrfs_free_space *info)
1643 {
1644 int ret = 0;
1645
1646 ASSERT(info->bytes || info->bitmap);
1647 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1648 &info->offset_index, (info->bitmap != NULL));
1649 if (ret)
1650 return ret;
1651
1652 ctl->free_space += info->bytes;
1653 ctl->free_extents++;
1654 return ret;
1655 }
1656
1657 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1658 {
1659 struct btrfs_block_group_cache *block_group = ctl->private;
1660 u64 max_bytes;
1661 u64 bitmap_bytes;
1662 u64 extent_bytes;
1663 u64 size = block_group->key.offset;
1664 u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1665 u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1666
1667 max_bitmaps = max_t(u64, max_bitmaps, 1);
1668
1669 ASSERT(ctl->total_bitmaps <= max_bitmaps);
1670
1671 /*
1672 * The goal is to keep the total amount of memory used per 1gb of space
1673 * at or below 32k, so we need to adjust how much memory we allow to be
1674 * used by extent based free space tracking
1675 */
1676 if (size < SZ_1G)
1677 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1678 else
1679 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1680
1681 /*
1682 * we want to account for 1 more bitmap than what we have so we can make
1683 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1684 * we add more bitmaps.
1685 */
1686 bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1687
1688 if (bitmap_bytes >= max_bytes) {
1689 ctl->extents_thresh = 0;
1690 return;
1691 }
1692
1693 /*
1694 * we want the extent entry threshold to always be at most 1/2 the max
1695 * bytes we can have, or whatever is less than that.
1696 */
1697 extent_bytes = max_bytes - bitmap_bytes;
1698 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1699
1700 ctl->extents_thresh =
1701 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1702 }
1703
1704 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1705 struct btrfs_free_space *info,
1706 u64 offset, u64 bytes)
1707 {
1708 unsigned long start, count;
1709
1710 start = offset_to_bit(info->offset, ctl->unit, offset);
1711 count = bytes_to_bits(bytes, ctl->unit);
1712 ASSERT(start + count <= BITS_PER_BITMAP);
1713
1714 bitmap_clear(info->bitmap, start, count);
1715
1716 info->bytes -= bytes;
1717 if (info->max_extent_size > ctl->unit)
1718 info->max_extent_size = 0;
1719 }
1720
1721 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1722 struct btrfs_free_space *info, u64 offset,
1723 u64 bytes)
1724 {
1725 __bitmap_clear_bits(ctl, info, offset, bytes);
1726 ctl->free_space -= bytes;
1727 }
1728
1729 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1730 struct btrfs_free_space *info, u64 offset,
1731 u64 bytes)
1732 {
1733 unsigned long start, count;
1734
1735 start = offset_to_bit(info->offset, ctl->unit, offset);
1736 count = bytes_to_bits(bytes, ctl->unit);
1737 ASSERT(start + count <= BITS_PER_BITMAP);
1738
1739 bitmap_set(info->bitmap, start, count);
1740
1741 info->bytes += bytes;
1742 ctl->free_space += bytes;
1743 }
1744
1745 /*
1746 * If we can not find suitable extent, we will use bytes to record
1747 * the size of the max extent.
1748 */
1749 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1750 struct btrfs_free_space *bitmap_info, u64 *offset,
1751 u64 *bytes, bool for_alloc)
1752 {
1753 unsigned long found_bits = 0;
1754 unsigned long max_bits = 0;
1755 unsigned long bits, i;
1756 unsigned long next_zero;
1757 unsigned long extent_bits;
1758
1759 /*
1760 * Skip searching the bitmap if we don't have a contiguous section that
1761 * is large enough for this allocation.
1762 */
1763 if (for_alloc &&
1764 bitmap_info->max_extent_size &&
1765 bitmap_info->max_extent_size < *bytes) {
1766 *bytes = bitmap_info->max_extent_size;
1767 return -1;
1768 }
1769
1770 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1771 max_t(u64, *offset, bitmap_info->offset));
1772 bits = bytes_to_bits(*bytes, ctl->unit);
1773
1774 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1775 if (for_alloc && bits == 1) {
1776 found_bits = 1;
1777 break;
1778 }
1779 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1780 BITS_PER_BITMAP, i);
1781 extent_bits = next_zero - i;
1782 if (extent_bits >= bits) {
1783 found_bits = extent_bits;
1784 break;
1785 } else if (extent_bits > max_bits) {
1786 max_bits = extent_bits;
1787 }
1788 i = next_zero;
1789 }
1790
1791 if (found_bits) {
1792 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1793 *bytes = (u64)(found_bits) * ctl->unit;
1794 return 0;
1795 }
1796
1797 *bytes = (u64)(max_bits) * ctl->unit;
1798 bitmap_info->max_extent_size = *bytes;
1799 return -1;
1800 }
1801
1802 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1803 {
1804 if (entry->bitmap)
1805 return entry->max_extent_size;
1806 return entry->bytes;
1807 }
1808
1809 /* Cache the size of the max extent in bytes */
1810 static struct btrfs_free_space *
1811 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1812 unsigned long align, u64 *max_extent_size)
1813 {
1814 struct btrfs_free_space *entry;
1815 struct rb_node *node;
1816 u64 tmp;
1817 u64 align_off;
1818 int ret;
1819
1820 if (!ctl->free_space_offset.rb_node)
1821 goto out;
1822
1823 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1824 if (!entry)
1825 goto out;
1826
1827 for (node = &entry->offset_index; node; node = rb_next(node)) {
1828 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1829 if (entry->bytes < *bytes) {
1830 *max_extent_size = max(get_max_extent_size(entry),
1831 *max_extent_size);
1832 continue;
1833 }
1834
1835 /* make sure the space returned is big enough
1836 * to match our requested alignment
1837 */
1838 if (*bytes >= align) {
1839 tmp = entry->offset - ctl->start + align - 1;
1840 tmp = div64_u64(tmp, align);
1841 tmp = tmp * align + ctl->start;
1842 align_off = tmp - entry->offset;
1843 } else {
1844 align_off = 0;
1845 tmp = entry->offset;
1846 }
1847
1848 if (entry->bytes < *bytes + align_off) {
1849 *max_extent_size = max(get_max_extent_size(entry),
1850 *max_extent_size);
1851 continue;
1852 }
1853
1854 if (entry->bitmap) {
1855 u64 size = *bytes;
1856
1857 ret = search_bitmap(ctl, entry, &tmp, &size, true);
1858 if (!ret) {
1859 *offset = tmp;
1860 *bytes = size;
1861 return entry;
1862 } else {
1863 *max_extent_size =
1864 max(get_max_extent_size(entry),
1865 *max_extent_size);
1866 }
1867 continue;
1868 }
1869
1870 *offset = tmp;
1871 *bytes = entry->bytes - align_off;
1872 return entry;
1873 }
1874 out:
1875 return NULL;
1876 }
1877
1878 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1879 struct btrfs_free_space *info, u64 offset)
1880 {
1881 info->offset = offset_to_bitmap(ctl, offset);
1882 info->bytes = 0;
1883 INIT_LIST_HEAD(&info->list);
1884 link_free_space(ctl, info);
1885 ctl->total_bitmaps++;
1886
1887 ctl->op->recalc_thresholds(ctl);
1888 }
1889
1890 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1891 struct btrfs_free_space *bitmap_info)
1892 {
1893 unlink_free_space(ctl, bitmap_info);
1894 kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1895 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1896 ctl->total_bitmaps--;
1897 ctl->op->recalc_thresholds(ctl);
1898 }
1899
1900 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1901 struct btrfs_free_space *bitmap_info,
1902 u64 *offset, u64 *bytes)
1903 {
1904 u64 end;
1905 u64 search_start, search_bytes;
1906 int ret;
1907
1908 again:
1909 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1910
1911 /*
1912 * We need to search for bits in this bitmap. We could only cover some
1913 * of the extent in this bitmap thanks to how we add space, so we need
1914 * to search for as much as it as we can and clear that amount, and then
1915 * go searching for the next bit.
1916 */
1917 search_start = *offset;
1918 search_bytes = ctl->unit;
1919 search_bytes = min(search_bytes, end - search_start + 1);
1920 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1921 false);
1922 if (ret < 0 || search_start != *offset)
1923 return -EINVAL;
1924
1925 /* We may have found more bits than what we need */
1926 search_bytes = min(search_bytes, *bytes);
1927
1928 /* Cannot clear past the end of the bitmap */
1929 search_bytes = min(search_bytes, end - search_start + 1);
1930
1931 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1932 *offset += search_bytes;
1933 *bytes -= search_bytes;
1934
1935 if (*bytes) {
1936 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1937 if (!bitmap_info->bytes)
1938 free_bitmap(ctl, bitmap_info);
1939
1940 /*
1941 * no entry after this bitmap, but we still have bytes to
1942 * remove, so something has gone wrong.
1943 */
1944 if (!next)
1945 return -EINVAL;
1946
1947 bitmap_info = rb_entry(next, struct btrfs_free_space,
1948 offset_index);
1949
1950 /*
1951 * if the next entry isn't a bitmap we need to return to let the
1952 * extent stuff do its work.
1953 */
1954 if (!bitmap_info->bitmap)
1955 return -EAGAIN;
1956
1957 /*
1958 * Ok the next item is a bitmap, but it may not actually hold
1959 * the information for the rest of this free space stuff, so
1960 * look for it, and if we don't find it return so we can try
1961 * everything over again.
1962 */
1963 search_start = *offset;
1964 search_bytes = ctl->unit;
1965 ret = search_bitmap(ctl, bitmap_info, &search_start,
1966 &search_bytes, false);
1967 if (ret < 0 || search_start != *offset)
1968 return -EAGAIN;
1969
1970 goto again;
1971 } else if (!bitmap_info->bytes)
1972 free_bitmap(ctl, bitmap_info);
1973
1974 return 0;
1975 }
1976
1977 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1978 struct btrfs_free_space *info, u64 offset,
1979 u64 bytes)
1980 {
1981 u64 bytes_to_set = 0;
1982 u64 end;
1983
1984 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1985
1986 bytes_to_set = min(end - offset, bytes);
1987
1988 bitmap_set_bits(ctl, info, offset, bytes_to_set);
1989
1990 /*
1991 * We set some bytes, we have no idea what the max extent size is
1992 * anymore.
1993 */
1994 info->max_extent_size = 0;
1995
1996 return bytes_to_set;
1997
1998 }
1999
2000 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2001 struct btrfs_free_space *info)
2002 {
2003 struct btrfs_block_group_cache *block_group = ctl->private;
2004 struct btrfs_fs_info *fs_info = block_group->fs_info;
2005 bool forced = false;
2006
2007 #ifdef CONFIG_BTRFS_DEBUG
2008 if (btrfs_should_fragment_free_space(block_group))
2009 forced = true;
2010 #endif
2011
2012 /*
2013 * If we are below the extents threshold then we can add this as an
2014 * extent, and don't have to deal with the bitmap
2015 */
2016 if (!forced && ctl->free_extents < ctl->extents_thresh) {
2017 /*
2018 * If this block group has some small extents we don't want to
2019 * use up all of our free slots in the cache with them, we want
2020 * to reserve them to larger extents, however if we have plenty
2021 * of cache left then go ahead an dadd them, no sense in adding
2022 * the overhead of a bitmap if we don't have to.
2023 */
2024 if (info->bytes <= fs_info->sectorsize * 4) {
2025 if (ctl->free_extents * 2 <= ctl->extents_thresh)
2026 return false;
2027 } else {
2028 return false;
2029 }
2030 }
2031
2032 /*
2033 * The original block groups from mkfs can be really small, like 8
2034 * megabytes, so don't bother with a bitmap for those entries. However
2035 * some block groups can be smaller than what a bitmap would cover but
2036 * are still large enough that they could overflow the 32k memory limit,
2037 * so allow those block groups to still be allowed to have a bitmap
2038 * entry.
2039 */
2040 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2041 return false;
2042
2043 return true;
2044 }
2045
2046 static const struct btrfs_free_space_op free_space_op = {
2047 .recalc_thresholds = recalculate_thresholds,
2048 .use_bitmap = use_bitmap,
2049 };
2050
2051 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2052 struct btrfs_free_space *info)
2053 {
2054 struct btrfs_free_space *bitmap_info;
2055 struct btrfs_block_group_cache *block_group = NULL;
2056 int added = 0;
2057 u64 bytes, offset, bytes_added;
2058 int ret;
2059
2060 bytes = info->bytes;
2061 offset = info->offset;
2062
2063 if (!ctl->op->use_bitmap(ctl, info))
2064 return 0;
2065
2066 if (ctl->op == &free_space_op)
2067 block_group = ctl->private;
2068 again:
2069 /*
2070 * Since we link bitmaps right into the cluster we need to see if we
2071 * have a cluster here, and if so and it has our bitmap we need to add
2072 * the free space to that bitmap.
2073 */
2074 if (block_group && !list_empty(&block_group->cluster_list)) {
2075 struct btrfs_free_cluster *cluster;
2076 struct rb_node *node;
2077 struct btrfs_free_space *entry;
2078
2079 cluster = list_entry(block_group->cluster_list.next,
2080 struct btrfs_free_cluster,
2081 block_group_list);
2082 spin_lock(&cluster->lock);
2083 node = rb_first(&cluster->root);
2084 if (!node) {
2085 spin_unlock(&cluster->lock);
2086 goto no_cluster_bitmap;
2087 }
2088
2089 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2090 if (!entry->bitmap) {
2091 spin_unlock(&cluster->lock);
2092 goto no_cluster_bitmap;
2093 }
2094
2095 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2096 bytes_added = add_bytes_to_bitmap(ctl, entry,
2097 offset, bytes);
2098 bytes -= bytes_added;
2099 offset += bytes_added;
2100 }
2101 spin_unlock(&cluster->lock);
2102 if (!bytes) {
2103 ret = 1;
2104 goto out;
2105 }
2106 }
2107
2108 no_cluster_bitmap:
2109 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2110 1, 0);
2111 if (!bitmap_info) {
2112 ASSERT(added == 0);
2113 goto new_bitmap;
2114 }
2115
2116 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2117 bytes -= bytes_added;
2118 offset += bytes_added;
2119 added = 0;
2120
2121 if (!bytes) {
2122 ret = 1;
2123 goto out;
2124 } else
2125 goto again;
2126
2127 new_bitmap:
2128 if (info && info->bitmap) {
2129 add_new_bitmap(ctl, info, offset);
2130 added = 1;
2131 info = NULL;
2132 goto again;
2133 } else {
2134 spin_unlock(&ctl->tree_lock);
2135
2136 /* no pre-allocated info, allocate a new one */
2137 if (!info) {
2138 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2139 GFP_NOFS);
2140 if (!info) {
2141 spin_lock(&ctl->tree_lock);
2142 ret = -ENOMEM;
2143 goto out;
2144 }
2145 }
2146
2147 /* allocate the bitmap */
2148 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2149 GFP_NOFS);
2150 spin_lock(&ctl->tree_lock);
2151 if (!info->bitmap) {
2152 ret = -ENOMEM;
2153 goto out;
2154 }
2155 goto again;
2156 }
2157
2158 out:
2159 if (info) {
2160 if (info->bitmap)
2161 kmem_cache_free(btrfs_free_space_bitmap_cachep,
2162 info->bitmap);
2163 kmem_cache_free(btrfs_free_space_cachep, info);
2164 }
2165
2166 return ret;
2167 }
2168
2169 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2170 struct btrfs_free_space *info, bool update_stat)
2171 {
2172 struct btrfs_free_space *left_info;
2173 struct btrfs_free_space *right_info;
2174 bool merged = false;
2175 u64 offset = info->offset;
2176 u64 bytes = info->bytes;
2177
2178 /*
2179 * first we want to see if there is free space adjacent to the range we
2180 * are adding, if there is remove that struct and add a new one to
2181 * cover the entire range
2182 */
2183 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2184 if (right_info && rb_prev(&right_info->offset_index))
2185 left_info = rb_entry(rb_prev(&right_info->offset_index),
2186 struct btrfs_free_space, offset_index);
2187 else
2188 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2189
2190 if (right_info && !right_info->bitmap) {
2191 if (update_stat)
2192 unlink_free_space(ctl, right_info);
2193 else
2194 __unlink_free_space(ctl, right_info);
2195 info->bytes += right_info->bytes;
2196 kmem_cache_free(btrfs_free_space_cachep, right_info);
2197 merged = true;
2198 }
2199
2200 if (left_info && !left_info->bitmap &&
2201 left_info->offset + left_info->bytes == offset) {
2202 if (update_stat)
2203 unlink_free_space(ctl, left_info);
2204 else
2205 __unlink_free_space(ctl, left_info);
2206 info->offset = left_info->offset;
2207 info->bytes += left_info->bytes;
2208 kmem_cache_free(btrfs_free_space_cachep, left_info);
2209 merged = true;
2210 }
2211
2212 return merged;
2213 }
2214
2215 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2216 struct btrfs_free_space *info,
2217 bool update_stat)
2218 {
2219 struct btrfs_free_space *bitmap;
2220 unsigned long i;
2221 unsigned long j;
2222 const u64 end = info->offset + info->bytes;
2223 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2224 u64 bytes;
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, end);
2231 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2232 if (j == i)
2233 return false;
2234 bytes = (j - i) * ctl->unit;
2235 info->bytes += bytes;
2236
2237 if (update_stat)
2238 bitmap_clear_bits(ctl, bitmap, end, bytes);
2239 else
2240 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2241
2242 if (!bitmap->bytes)
2243 free_bitmap(ctl, bitmap);
2244
2245 return true;
2246 }
2247
2248 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2249 struct btrfs_free_space *info,
2250 bool update_stat)
2251 {
2252 struct btrfs_free_space *bitmap;
2253 u64 bitmap_offset;
2254 unsigned long i;
2255 unsigned long j;
2256 unsigned long prev_j;
2257 u64 bytes;
2258
2259 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2260 /* If we're on a boundary, try the previous logical bitmap. */
2261 if (bitmap_offset == info->offset) {
2262 if (info->offset == 0)
2263 return false;
2264 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2265 }
2266
2267 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2268 if (!bitmap)
2269 return false;
2270
2271 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2272 j = 0;
2273 prev_j = (unsigned long)-1;
2274 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2275 if (j > i)
2276 break;
2277 prev_j = j;
2278 }
2279 if (prev_j == i)
2280 return false;
2281
2282 if (prev_j == (unsigned long)-1)
2283 bytes = (i + 1) * ctl->unit;
2284 else
2285 bytes = (i - prev_j) * ctl->unit;
2286
2287 info->offset -= bytes;
2288 info->bytes += bytes;
2289
2290 if (update_stat)
2291 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2292 else
2293 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2294
2295 if (!bitmap->bytes)
2296 free_bitmap(ctl, bitmap);
2297
2298 return true;
2299 }
2300
2301 /*
2302 * We prefer always to allocate from extent entries, both for clustered and
2303 * non-clustered allocation requests. So when attempting to add a new extent
2304 * entry, try to see if there's adjacent free space in bitmap entries, and if
2305 * there is, migrate that space from the bitmaps to the extent.
2306 * Like this we get better chances of satisfying space allocation requests
2307 * because we attempt to satisfy them based on a single cache entry, and never
2308 * on 2 or more entries - even if the entries represent a contiguous free space
2309 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2310 * ends).
2311 */
2312 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2313 struct btrfs_free_space *info,
2314 bool update_stat)
2315 {
2316 /*
2317 * Only work with disconnected entries, as we can change their offset,
2318 * and must be extent entries.
2319 */
2320 ASSERT(!info->bitmap);
2321 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2322
2323 if (ctl->total_bitmaps > 0) {
2324 bool stole_end;
2325 bool stole_front = false;
2326
2327 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2328 if (ctl->total_bitmaps > 0)
2329 stole_front = steal_from_bitmap_to_front(ctl, info,
2330 update_stat);
2331
2332 if (stole_end || stole_front)
2333 try_merge_free_space(ctl, info, update_stat);
2334 }
2335 }
2336
2337 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2338 struct btrfs_free_space_ctl *ctl,
2339 u64 offset, u64 bytes)
2340 {
2341 struct btrfs_free_space *info;
2342 int ret = 0;
2343
2344 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2345 if (!info)
2346 return -ENOMEM;
2347
2348 info->offset = offset;
2349 info->bytes = bytes;
2350 RB_CLEAR_NODE(&info->offset_index);
2351
2352 spin_lock(&ctl->tree_lock);
2353
2354 if (try_merge_free_space(ctl, info, true))
2355 goto link;
2356
2357 /*
2358 * There was no extent directly to the left or right of this new
2359 * extent then we know we're going to have to allocate a new extent, so
2360 * before we do that see if we need to drop this into a bitmap
2361 */
2362 ret = insert_into_bitmap(ctl, info);
2363 if (ret < 0) {
2364 goto out;
2365 } else if (ret) {
2366 ret = 0;
2367 goto out;
2368 }
2369 link:
2370 /*
2371 * Only steal free space from adjacent bitmaps if we're sure we're not
2372 * going to add the new free space to existing bitmap entries - because
2373 * that would mean unnecessary work that would be reverted. Therefore
2374 * attempt to steal space from bitmaps if we're adding an extent entry.
2375 */
2376 steal_from_bitmap(ctl, info, true);
2377
2378 ret = link_free_space(ctl, info);
2379 if (ret)
2380 kmem_cache_free(btrfs_free_space_cachep, info);
2381 out:
2382 spin_unlock(&ctl->tree_lock);
2383
2384 if (ret) {
2385 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2386 ASSERT(ret != -EEXIST);
2387 }
2388
2389 return ret;
2390 }
2391
2392 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2393 u64 offset, u64 bytes)
2394 {
2395 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2396 struct btrfs_free_space *info;
2397 int ret;
2398 bool re_search = false;
2399
2400 spin_lock(&ctl->tree_lock);
2401
2402 again:
2403 ret = 0;
2404 if (!bytes)
2405 goto out_lock;
2406
2407 info = tree_search_offset(ctl, offset, 0, 0);
2408 if (!info) {
2409 /*
2410 * oops didn't find an extent that matched the space we wanted
2411 * to remove, look for a bitmap instead
2412 */
2413 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2414 1, 0);
2415 if (!info) {
2416 /*
2417 * If we found a partial bit of our free space in a
2418 * bitmap but then couldn't find the other part this may
2419 * be a problem, so WARN about it.
2420 */
2421 WARN_ON(re_search);
2422 goto out_lock;
2423 }
2424 }
2425
2426 re_search = false;
2427 if (!info->bitmap) {
2428 unlink_free_space(ctl, info);
2429 if (offset == info->offset) {
2430 u64 to_free = min(bytes, info->bytes);
2431
2432 info->bytes -= to_free;
2433 info->offset += to_free;
2434 if (info->bytes) {
2435 ret = link_free_space(ctl, info);
2436 WARN_ON(ret);
2437 } else {
2438 kmem_cache_free(btrfs_free_space_cachep, info);
2439 }
2440
2441 offset += to_free;
2442 bytes -= to_free;
2443 goto again;
2444 } else {
2445 u64 old_end = info->bytes + info->offset;
2446
2447 info->bytes = offset - info->offset;
2448 ret = link_free_space(ctl, info);
2449 WARN_ON(ret);
2450 if (ret)
2451 goto out_lock;
2452
2453 /* Not enough bytes in this entry to satisfy us */
2454 if (old_end < offset + bytes) {
2455 bytes -= old_end - offset;
2456 offset = old_end;
2457 goto again;
2458 } else if (old_end == offset + bytes) {
2459 /* all done */
2460 goto out_lock;
2461 }
2462 spin_unlock(&ctl->tree_lock);
2463
2464 ret = btrfs_add_free_space(block_group, offset + bytes,
2465 old_end - (offset + bytes));
2466 WARN_ON(ret);
2467 goto out;
2468 }
2469 }
2470
2471 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2472 if (ret == -EAGAIN) {
2473 re_search = true;
2474 goto again;
2475 }
2476 out_lock:
2477 spin_unlock(&ctl->tree_lock);
2478 out:
2479 return ret;
2480 }
2481
2482 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2483 u64 bytes)
2484 {
2485 struct btrfs_fs_info *fs_info = block_group->fs_info;
2486 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2487 struct btrfs_free_space *info;
2488 struct rb_node *n;
2489 int count = 0;
2490
2491 spin_lock(&ctl->tree_lock);
2492 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2493 info = rb_entry(n, struct btrfs_free_space, offset_index);
2494 if (info->bytes >= bytes && !block_group->ro)
2495 count++;
2496 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2497 info->offset, info->bytes,
2498 (info->bitmap) ? "yes" : "no");
2499 }
2500 spin_unlock(&ctl->tree_lock);
2501 btrfs_info(fs_info, "block group has cluster?: %s",
2502 list_empty(&block_group->cluster_list) ? "no" : "yes");
2503 btrfs_info(fs_info,
2504 "%d blocks of free space at or bigger than bytes is", count);
2505 }
2506
2507 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2508 {
2509 struct btrfs_fs_info *fs_info = block_group->fs_info;
2510 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2511
2512 spin_lock_init(&ctl->tree_lock);
2513 ctl->unit = fs_info->sectorsize;
2514 ctl->start = block_group->key.objectid;
2515 ctl->private = block_group;
2516 ctl->op = &free_space_op;
2517 INIT_LIST_HEAD(&ctl->trimming_ranges);
2518 mutex_init(&ctl->cache_writeout_mutex);
2519
2520 /*
2521 * we only want to have 32k of ram per block group for keeping
2522 * track of free space, and if we pass 1/2 of that we want to
2523 * start converting things over to using bitmaps
2524 */
2525 ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2526 }
2527
2528 /*
2529 * for a given cluster, put all of its extents back into the free
2530 * space cache. If the block group passed doesn't match the block group
2531 * pointed to by the cluster, someone else raced in and freed the
2532 * cluster already. In that case, we just return without changing anything
2533 */
2534 static int
2535 __btrfs_return_cluster_to_free_space(
2536 struct btrfs_block_group_cache *block_group,
2537 struct btrfs_free_cluster *cluster)
2538 {
2539 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2540 struct btrfs_free_space *entry;
2541 struct rb_node *node;
2542
2543 spin_lock(&cluster->lock);
2544 if (cluster->block_group != block_group)
2545 goto out;
2546
2547 cluster->block_group = NULL;
2548 cluster->window_start = 0;
2549 list_del_init(&cluster->block_group_list);
2550
2551 node = rb_first(&cluster->root);
2552 while (node) {
2553 bool bitmap;
2554
2555 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2556 node = rb_next(&entry->offset_index);
2557 rb_erase(&entry->offset_index, &cluster->root);
2558 RB_CLEAR_NODE(&entry->offset_index);
2559
2560 bitmap = (entry->bitmap != NULL);
2561 if (!bitmap) {
2562 try_merge_free_space(ctl, entry, false);
2563 steal_from_bitmap(ctl, entry, false);
2564 }
2565 tree_insert_offset(&ctl->free_space_offset,
2566 entry->offset, &entry->offset_index, bitmap);
2567 }
2568 cluster->root = RB_ROOT;
2569
2570 out:
2571 spin_unlock(&cluster->lock);
2572 btrfs_put_block_group(block_group);
2573 return 0;
2574 }
2575
2576 static void __btrfs_remove_free_space_cache_locked(
2577 struct btrfs_free_space_ctl *ctl)
2578 {
2579 struct btrfs_free_space *info;
2580 struct rb_node *node;
2581
2582 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2583 info = rb_entry(node, struct btrfs_free_space, offset_index);
2584 if (!info->bitmap) {
2585 unlink_free_space(ctl, info);
2586 kmem_cache_free(btrfs_free_space_cachep, info);
2587 } else {
2588 free_bitmap(ctl, info);
2589 }
2590
2591 cond_resched_lock(&ctl->tree_lock);
2592 }
2593 }
2594
2595 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2596 {
2597 spin_lock(&ctl->tree_lock);
2598 __btrfs_remove_free_space_cache_locked(ctl);
2599 spin_unlock(&ctl->tree_lock);
2600 }
2601
2602 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2603 {
2604 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2605 struct btrfs_free_cluster *cluster;
2606 struct list_head *head;
2607
2608 spin_lock(&ctl->tree_lock);
2609 while ((head = block_group->cluster_list.next) !=
2610 &block_group->cluster_list) {
2611 cluster = list_entry(head, struct btrfs_free_cluster,
2612 block_group_list);
2613
2614 WARN_ON(cluster->block_group != block_group);
2615 __btrfs_return_cluster_to_free_space(block_group, cluster);
2616
2617 cond_resched_lock(&ctl->tree_lock);
2618 }
2619 __btrfs_remove_free_space_cache_locked(ctl);
2620 spin_unlock(&ctl->tree_lock);
2621
2622 }
2623
2624 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2625 u64 offset, u64 bytes, u64 empty_size,
2626 u64 *max_extent_size)
2627 {
2628 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2629 struct btrfs_free_space *entry = NULL;
2630 u64 bytes_search = bytes + empty_size;
2631 u64 ret = 0;
2632 u64 align_gap = 0;
2633 u64 align_gap_len = 0;
2634
2635 spin_lock(&ctl->tree_lock);
2636 entry = find_free_space(ctl, &offset, &bytes_search,
2637 block_group->full_stripe_len, max_extent_size);
2638 if (!entry)
2639 goto out;
2640
2641 ret = offset;
2642 if (entry->bitmap) {
2643 bitmap_clear_bits(ctl, entry, offset, bytes);
2644 if (!entry->bytes)
2645 free_bitmap(ctl, entry);
2646 } else {
2647 unlink_free_space(ctl, entry);
2648 align_gap_len = offset - entry->offset;
2649 align_gap = entry->offset;
2650
2651 entry->offset = offset + bytes;
2652 WARN_ON(entry->bytes < bytes + align_gap_len);
2653
2654 entry->bytes -= bytes + align_gap_len;
2655 if (!entry->bytes)
2656 kmem_cache_free(btrfs_free_space_cachep, entry);
2657 else
2658 link_free_space(ctl, entry);
2659 }
2660 out:
2661 spin_unlock(&ctl->tree_lock);
2662
2663 if (align_gap_len)
2664 __btrfs_add_free_space(block_group->fs_info, ctl,
2665 align_gap, align_gap_len);
2666 return ret;
2667 }
2668
2669 /*
2670 * given a cluster, put all of its extents back into the free space
2671 * cache. If a block group is passed, this function will only free
2672 * a cluster that belongs to the passed block group.
2673 *
2674 * Otherwise, it'll get a reference on the block group pointed to by the
2675 * cluster and remove the cluster from it.
2676 */
2677 int btrfs_return_cluster_to_free_space(
2678 struct btrfs_block_group_cache *block_group,
2679 struct btrfs_free_cluster *cluster)
2680 {
2681 struct btrfs_free_space_ctl *ctl;
2682 int ret;
2683
2684 /* first, get a safe pointer to the block group */
2685 spin_lock(&cluster->lock);
2686 if (!block_group) {
2687 block_group = cluster->block_group;
2688 if (!block_group) {
2689 spin_unlock(&cluster->lock);
2690 return 0;
2691 }
2692 } else if (cluster->block_group != block_group) {
2693 /* someone else has already freed it don't redo their work */
2694 spin_unlock(&cluster->lock);
2695 return 0;
2696 }
2697 atomic_inc(&block_group->count);
2698 spin_unlock(&cluster->lock);
2699
2700 ctl = block_group->free_space_ctl;
2701
2702 /* now return any extents the cluster had on it */
2703 spin_lock(&ctl->tree_lock);
2704 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2705 spin_unlock(&ctl->tree_lock);
2706
2707 /* finally drop our ref */
2708 btrfs_put_block_group(block_group);
2709 return ret;
2710 }
2711
2712 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2713 struct btrfs_free_cluster *cluster,
2714 struct btrfs_free_space *entry,
2715 u64 bytes, u64 min_start,
2716 u64 *max_extent_size)
2717 {
2718 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2719 int err;
2720 u64 search_start = cluster->window_start;
2721 u64 search_bytes = bytes;
2722 u64 ret = 0;
2723
2724 search_start = min_start;
2725 search_bytes = bytes;
2726
2727 err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2728 if (err) {
2729 *max_extent_size = max(get_max_extent_size(entry),
2730 *max_extent_size);
2731 return 0;
2732 }
2733
2734 ret = search_start;
2735 __bitmap_clear_bits(ctl, entry, ret, bytes);
2736
2737 return ret;
2738 }
2739
2740 /*
2741 * given a cluster, try to allocate 'bytes' from it, returns 0
2742 * if it couldn't find anything suitably large, or a logical disk offset
2743 * if things worked out
2744 */
2745 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2746 struct btrfs_free_cluster *cluster, u64 bytes,
2747 u64 min_start, u64 *max_extent_size)
2748 {
2749 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2750 struct btrfs_free_space *entry = NULL;
2751 struct rb_node *node;
2752 u64 ret = 0;
2753
2754 spin_lock(&cluster->lock);
2755 if (bytes > cluster->max_size)
2756 goto out;
2757
2758 if (cluster->block_group != block_group)
2759 goto out;
2760
2761 node = rb_first(&cluster->root);
2762 if (!node)
2763 goto out;
2764
2765 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2766 while (1) {
2767 if (entry->bytes < bytes)
2768 *max_extent_size = max(get_max_extent_size(entry),
2769 *max_extent_size);
2770
2771 if (entry->bytes < bytes ||
2772 (!entry->bitmap && entry->offset < min_start)) {
2773 node = rb_next(&entry->offset_index);
2774 if (!node)
2775 break;
2776 entry = rb_entry(node, struct btrfs_free_space,
2777 offset_index);
2778 continue;
2779 }
2780
2781 if (entry->bitmap) {
2782 ret = btrfs_alloc_from_bitmap(block_group,
2783 cluster, entry, bytes,
2784 cluster->window_start,
2785 max_extent_size);
2786 if (ret == 0) {
2787 node = rb_next(&entry->offset_index);
2788 if (!node)
2789 break;
2790 entry = rb_entry(node, struct btrfs_free_space,
2791 offset_index);
2792 continue;
2793 }
2794 cluster->window_start += bytes;
2795 } else {
2796 ret = entry->offset;
2797
2798 entry->offset += bytes;
2799 entry->bytes -= bytes;
2800 }
2801
2802 if (entry->bytes == 0)
2803 rb_erase(&entry->offset_index, &cluster->root);
2804 break;
2805 }
2806 out:
2807 spin_unlock(&cluster->lock);
2808
2809 if (!ret)
2810 return 0;
2811
2812 spin_lock(&ctl->tree_lock);
2813
2814 ctl->free_space -= bytes;
2815 if (entry->bytes == 0) {
2816 ctl->free_extents--;
2817 if (entry->bitmap) {
2818 kmem_cache_free(btrfs_free_space_bitmap_cachep,
2819 entry->bitmap);
2820 ctl->total_bitmaps--;
2821 ctl->op->recalc_thresholds(ctl);
2822 }
2823 kmem_cache_free(btrfs_free_space_cachep, entry);
2824 }
2825
2826 spin_unlock(&ctl->tree_lock);
2827
2828 return ret;
2829 }
2830
2831 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2832 struct btrfs_free_space *entry,
2833 struct btrfs_free_cluster *cluster,
2834 u64 offset, u64 bytes,
2835 u64 cont1_bytes, u64 min_bytes)
2836 {
2837 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2838 unsigned long next_zero;
2839 unsigned long i;
2840 unsigned long want_bits;
2841 unsigned long min_bits;
2842 unsigned long found_bits;
2843 unsigned long max_bits = 0;
2844 unsigned long start = 0;
2845 unsigned long total_found = 0;
2846 int ret;
2847
2848 i = offset_to_bit(entry->offset, ctl->unit,
2849 max_t(u64, offset, entry->offset));
2850 want_bits = bytes_to_bits(bytes, ctl->unit);
2851 min_bits = bytes_to_bits(min_bytes, ctl->unit);
2852
2853 /*
2854 * Don't bother looking for a cluster in this bitmap if it's heavily
2855 * fragmented.
2856 */
2857 if (entry->max_extent_size &&
2858 entry->max_extent_size < cont1_bytes)
2859 return -ENOSPC;
2860 again:
2861 found_bits = 0;
2862 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2863 next_zero = find_next_zero_bit(entry->bitmap,
2864 BITS_PER_BITMAP, i);
2865 if (next_zero - i >= min_bits) {
2866 found_bits = next_zero - i;
2867 if (found_bits > max_bits)
2868 max_bits = found_bits;
2869 break;
2870 }
2871 if (next_zero - i > max_bits)
2872 max_bits = next_zero - i;
2873 i = next_zero;
2874 }
2875
2876 if (!found_bits) {
2877 entry->max_extent_size = (u64)max_bits * ctl->unit;
2878 return -ENOSPC;
2879 }
2880
2881 if (!total_found) {
2882 start = i;
2883 cluster->max_size = 0;
2884 }
2885
2886 total_found += found_bits;
2887
2888 if (cluster->max_size < found_bits * ctl->unit)
2889 cluster->max_size = found_bits * ctl->unit;
2890
2891 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2892 i = next_zero + 1;
2893 goto again;
2894 }
2895
2896 cluster->window_start = start * ctl->unit + entry->offset;
2897 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2898 ret = tree_insert_offset(&cluster->root, entry->offset,
2899 &entry->offset_index, 1);
2900 ASSERT(!ret); /* -EEXIST; Logic error */
2901
2902 trace_btrfs_setup_cluster(block_group, cluster,
2903 total_found * ctl->unit, 1);
2904 return 0;
2905 }
2906
2907 /*
2908 * This searches the block group for just extents to fill the cluster with.
2909 * Try to find a cluster with at least bytes total bytes, at least one
2910 * extent of cont1_bytes, and other clusters of at least min_bytes.
2911 */
2912 static noinline int
2913 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2914 struct btrfs_free_cluster *cluster,
2915 struct list_head *bitmaps, u64 offset, u64 bytes,
2916 u64 cont1_bytes, u64 min_bytes)
2917 {
2918 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2919 struct btrfs_free_space *first = NULL;
2920 struct btrfs_free_space *entry = NULL;
2921 struct btrfs_free_space *last;
2922 struct rb_node *node;
2923 u64 window_free;
2924 u64 max_extent;
2925 u64 total_size = 0;
2926
2927 entry = tree_search_offset(ctl, offset, 0, 1);
2928 if (!entry)
2929 return -ENOSPC;
2930
2931 /*
2932 * We don't want bitmaps, so just move along until we find a normal
2933 * extent entry.
2934 */
2935 while (entry->bitmap || entry->bytes < min_bytes) {
2936 if (entry->bitmap && list_empty(&entry->list))
2937 list_add_tail(&entry->list, bitmaps);
2938 node = rb_next(&entry->offset_index);
2939 if (!node)
2940 return -ENOSPC;
2941 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2942 }
2943
2944 window_free = entry->bytes;
2945 max_extent = entry->bytes;
2946 first = entry;
2947 last = entry;
2948
2949 for (node = rb_next(&entry->offset_index); node;
2950 node = rb_next(&entry->offset_index)) {
2951 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2952
2953 if (entry->bitmap) {
2954 if (list_empty(&entry->list))
2955 list_add_tail(&entry->list, bitmaps);
2956 continue;
2957 }
2958
2959 if (entry->bytes < min_bytes)
2960 continue;
2961
2962 last = entry;
2963 window_free += entry->bytes;
2964 if (entry->bytes > max_extent)
2965 max_extent = entry->bytes;
2966 }
2967
2968 if (window_free < bytes || max_extent < cont1_bytes)
2969 return -ENOSPC;
2970
2971 cluster->window_start = first->offset;
2972
2973 node = &first->offset_index;
2974
2975 /*
2976 * now we've found our entries, pull them out of the free space
2977 * cache and put them into the cluster rbtree
2978 */
2979 do {
2980 int ret;
2981
2982 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2983 node = rb_next(&entry->offset_index);
2984 if (entry->bitmap || entry->bytes < min_bytes)
2985 continue;
2986
2987 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2988 ret = tree_insert_offset(&cluster->root, entry->offset,
2989 &entry->offset_index, 0);
2990 total_size += entry->bytes;
2991 ASSERT(!ret); /* -EEXIST; Logic error */
2992 } while (node && entry != last);
2993
2994 cluster->max_size = max_extent;
2995 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2996 return 0;
2997 }
2998
2999 /*
3000 * This specifically looks for bitmaps that may work in the cluster, we assume
3001 * that we have already failed to find extents that will work.
3002 */
3003 static noinline int
3004 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
3005 struct btrfs_free_cluster *cluster,
3006 struct list_head *bitmaps, u64 offset, u64 bytes,
3007 u64 cont1_bytes, u64 min_bytes)
3008 {
3009 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3010 struct btrfs_free_space *entry = NULL;
3011 int ret = -ENOSPC;
3012 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3013
3014 if (ctl->total_bitmaps == 0)
3015 return -ENOSPC;
3016
3017 /*
3018 * The bitmap that covers offset won't be in the list unless offset
3019 * is just its start offset.
3020 */
3021 if (!list_empty(bitmaps))
3022 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3023
3024 if (!entry || entry->offset != bitmap_offset) {
3025 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3026 if (entry && list_empty(&entry->list))
3027 list_add(&entry->list, bitmaps);
3028 }
3029
3030 list_for_each_entry(entry, bitmaps, list) {
3031 if (entry->bytes < bytes)
3032 continue;
3033 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3034 bytes, cont1_bytes, min_bytes);
3035 if (!ret)
3036 return 0;
3037 }
3038
3039 /*
3040 * The bitmaps list has all the bitmaps that record free space
3041 * starting after offset, so no more search is required.
3042 */
3043 return -ENOSPC;
3044 }
3045
3046 /*
3047 * here we try to find a cluster of blocks in a block group. The goal
3048 * is to find at least bytes+empty_size.
3049 * We might not find them all in one contiguous area.
3050 *
3051 * returns zero and sets up cluster if things worked out, otherwise
3052 * it returns -enospc
3053 */
3054 int btrfs_find_space_cluster(struct btrfs_fs_info *fs_info,
3055 struct btrfs_block_group_cache *block_group,
3056 struct btrfs_free_cluster *cluster,
3057 u64 offset, u64 bytes, u64 empty_size)
3058 {
3059 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3060 struct btrfs_free_space *entry, *tmp;
3061 LIST_HEAD(bitmaps);
3062 u64 min_bytes;
3063 u64 cont1_bytes;
3064 int ret;
3065
3066 /*
3067 * Choose the minimum extent size we'll require for this
3068 * cluster. For SSD_SPREAD, don't allow any fragmentation.
3069 * For metadata, allow allocates with smaller extents. For
3070 * data, keep it dense.
3071 */
3072 if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3073 cont1_bytes = min_bytes = bytes + empty_size;
3074 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3075 cont1_bytes = bytes;
3076 min_bytes = fs_info->sectorsize;
3077 } else {
3078 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3079 min_bytes = fs_info->sectorsize;
3080 }
3081
3082 spin_lock(&ctl->tree_lock);
3083
3084 /*
3085 * If we know we don't have enough space to make a cluster don't even
3086 * bother doing all the work to try and find one.
3087 */
3088 if (ctl->free_space < bytes) {
3089 spin_unlock(&ctl->tree_lock);
3090 return -ENOSPC;
3091 }
3092
3093 spin_lock(&cluster->lock);
3094
3095 /* someone already found a cluster, hooray */
3096 if (cluster->block_group) {
3097 ret = 0;
3098 goto out;
3099 }
3100
3101 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3102 min_bytes);
3103
3104 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3105 bytes + empty_size,
3106 cont1_bytes, min_bytes);
3107 if (ret)
3108 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3109 offset, bytes + empty_size,
3110 cont1_bytes, min_bytes);
3111
3112 /* Clear our temporary list */
3113 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3114 list_del_init(&entry->list);
3115
3116 if (!ret) {
3117 atomic_inc(&block_group->count);
3118 list_add_tail(&cluster->block_group_list,
3119 &block_group->cluster_list);
3120 cluster->block_group = block_group;
3121 } else {
3122 trace_btrfs_failed_cluster_setup(block_group);
3123 }
3124 out:
3125 spin_unlock(&cluster->lock);
3126 spin_unlock(&ctl->tree_lock);
3127
3128 return ret;
3129 }
3130
3131 /*
3132 * simple code to zero out a cluster
3133 */
3134 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3135 {
3136 spin_lock_init(&cluster->lock);
3137 spin_lock_init(&cluster->refill_lock);
3138 cluster->root = RB_ROOT;
3139 cluster->max_size = 0;
3140 cluster->fragmented = false;
3141 INIT_LIST_HEAD(&cluster->block_group_list);
3142 cluster->block_group = NULL;
3143 }
3144
3145 static int do_trimming(struct btrfs_block_group_cache *block_group,
3146 u64 *total_trimmed, u64 start, u64 bytes,
3147 u64 reserved_start, u64 reserved_bytes,
3148 struct btrfs_trim_range *trim_entry)
3149 {
3150 struct btrfs_space_info *space_info = block_group->space_info;
3151 struct btrfs_fs_info *fs_info = block_group->fs_info;
3152 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3153 int ret;
3154 int update = 0;
3155 u64 trimmed = 0;
3156
3157 spin_lock(&space_info->lock);
3158 spin_lock(&block_group->lock);
3159 if (!block_group->ro) {
3160 block_group->reserved += reserved_bytes;
3161 space_info->bytes_reserved += reserved_bytes;
3162 update = 1;
3163 }
3164 spin_unlock(&block_group->lock);
3165 spin_unlock(&space_info->lock);
3166
3167 ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3168 if (!ret)
3169 *total_trimmed += trimmed;
3170
3171 mutex_lock(&ctl->cache_writeout_mutex);
3172 btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3173 list_del(&trim_entry->list);
3174 mutex_unlock(&ctl->cache_writeout_mutex);
3175
3176 if (update) {
3177 spin_lock(&space_info->lock);
3178 spin_lock(&block_group->lock);
3179 if (block_group->ro)
3180 space_info->bytes_readonly += reserved_bytes;
3181 block_group->reserved -= reserved_bytes;
3182 space_info->bytes_reserved -= reserved_bytes;
3183 spin_unlock(&space_info->lock);
3184 spin_unlock(&block_group->lock);
3185 }
3186
3187 return ret;
3188 }
3189
3190 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3191 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3192 {
3193 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3194 struct btrfs_free_space *entry;
3195 struct rb_node *node;
3196 int ret = 0;
3197 u64 extent_start;
3198 u64 extent_bytes;
3199 u64 bytes;
3200
3201 while (start < end) {
3202 struct btrfs_trim_range trim_entry;
3203
3204 mutex_lock(&ctl->cache_writeout_mutex);
3205 spin_lock(&ctl->tree_lock);
3206
3207 if (ctl->free_space < minlen) {
3208 spin_unlock(&ctl->tree_lock);
3209 mutex_unlock(&ctl->cache_writeout_mutex);
3210 break;
3211 }
3212
3213 entry = tree_search_offset(ctl, start, 0, 1);
3214 if (!entry) {
3215 spin_unlock(&ctl->tree_lock);
3216 mutex_unlock(&ctl->cache_writeout_mutex);
3217 break;
3218 }
3219
3220 /* skip bitmaps */
3221 while (entry->bitmap) {
3222 node = rb_next(&entry->offset_index);
3223 if (!node) {
3224 spin_unlock(&ctl->tree_lock);
3225 mutex_unlock(&ctl->cache_writeout_mutex);
3226 goto out;
3227 }
3228 entry = rb_entry(node, struct btrfs_free_space,
3229 offset_index);
3230 }
3231
3232 if (entry->offset >= end) {
3233 spin_unlock(&ctl->tree_lock);
3234 mutex_unlock(&ctl->cache_writeout_mutex);
3235 break;
3236 }
3237
3238 extent_start = entry->offset;
3239 extent_bytes = entry->bytes;
3240 start = max(start, extent_start);
3241 bytes = min(extent_start + extent_bytes, end) - start;
3242 if (bytes < minlen) {
3243 spin_unlock(&ctl->tree_lock);
3244 mutex_unlock(&ctl->cache_writeout_mutex);
3245 goto next;
3246 }
3247
3248 unlink_free_space(ctl, entry);
3249 kmem_cache_free(btrfs_free_space_cachep, entry);
3250
3251 spin_unlock(&ctl->tree_lock);
3252 trim_entry.start = extent_start;
3253 trim_entry.bytes = extent_bytes;
3254 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3255 mutex_unlock(&ctl->cache_writeout_mutex);
3256
3257 ret = do_trimming(block_group, total_trimmed, start, bytes,
3258 extent_start, extent_bytes, &trim_entry);
3259 if (ret)
3260 break;
3261 next:
3262 start += bytes;
3263
3264 if (fatal_signal_pending(current)) {
3265 ret = -ERESTARTSYS;
3266 break;
3267 }
3268
3269 cond_resched();
3270 }
3271 out:
3272 return ret;
3273 }
3274
3275 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3276 u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3277 {
3278 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3279 struct btrfs_free_space *entry;
3280 int ret = 0;
3281 int ret2;
3282 u64 bytes;
3283 u64 offset = offset_to_bitmap(ctl, start);
3284
3285 while (offset < end) {
3286 bool next_bitmap = false;
3287 struct btrfs_trim_range trim_entry;
3288
3289 mutex_lock(&ctl->cache_writeout_mutex);
3290 spin_lock(&ctl->tree_lock);
3291
3292 if (ctl->free_space < minlen) {
3293 spin_unlock(&ctl->tree_lock);
3294 mutex_unlock(&ctl->cache_writeout_mutex);
3295 break;
3296 }
3297
3298 entry = tree_search_offset(ctl, offset, 1, 0);
3299 if (!entry) {
3300 spin_unlock(&ctl->tree_lock);
3301 mutex_unlock(&ctl->cache_writeout_mutex);
3302 next_bitmap = true;
3303 goto next;
3304 }
3305
3306 bytes = minlen;
3307 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3308 if (ret2 || start >= end) {
3309 spin_unlock(&ctl->tree_lock);
3310 mutex_unlock(&ctl->cache_writeout_mutex);
3311 next_bitmap = true;
3312 goto next;
3313 }
3314
3315 bytes = min(bytes, end - start);
3316 if (bytes < minlen) {
3317 spin_unlock(&ctl->tree_lock);
3318 mutex_unlock(&ctl->cache_writeout_mutex);
3319 goto next;
3320 }
3321
3322 bitmap_clear_bits(ctl, entry, start, bytes);
3323 if (entry->bytes == 0)
3324 free_bitmap(ctl, entry);
3325
3326 spin_unlock(&ctl->tree_lock);
3327 trim_entry.start = start;
3328 trim_entry.bytes = bytes;
3329 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3330 mutex_unlock(&ctl->cache_writeout_mutex);
3331
3332 ret = do_trimming(block_group, total_trimmed, start, bytes,
3333 start, bytes, &trim_entry);
3334 if (ret)
3335 break;
3336 next:
3337 if (next_bitmap) {
3338 offset += BITS_PER_BITMAP * ctl->unit;
3339 } else {
3340 start += bytes;
3341 if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3342 offset += BITS_PER_BITMAP * ctl->unit;
3343 }
3344
3345 if (fatal_signal_pending(current)) {
3346 ret = -ERESTARTSYS;
3347 break;
3348 }
3349
3350 cond_resched();
3351 }
3352
3353 return ret;
3354 }
3355
3356 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3357 {
3358 atomic_inc(&cache->trimming);
3359 }
3360
3361 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3362 {
3363 struct btrfs_fs_info *fs_info = block_group->fs_info;
3364 struct extent_map_tree *em_tree;
3365 struct extent_map *em;
3366 bool cleanup;
3367
3368 spin_lock(&block_group->lock);
3369 cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3370 block_group->removed);
3371 spin_unlock(&block_group->lock);
3372
3373 if (cleanup) {
3374 mutex_lock(&fs_info->chunk_mutex);
3375 em_tree = &fs_info->mapping_tree.map_tree;
3376 write_lock(&em_tree->lock);
3377 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3378 1);
3379 BUG_ON(!em); /* logic error, can't happen */
3380 /*
3381 * remove_extent_mapping() will delete us from the pinned_chunks
3382 * list, which is protected by the chunk mutex.
3383 */
3384 remove_extent_mapping(em_tree, em);
3385 write_unlock(&em_tree->lock);
3386 mutex_unlock(&fs_info->chunk_mutex);
3387
3388 /* once for us and once for the tree */
3389 free_extent_map(em);
3390 free_extent_map(em);
3391
3392 /*
3393 * We've left one free space entry and other tasks trimming
3394 * this block group have left 1 entry each one. Free them.
3395 */
3396 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3397 }
3398 }
3399
3400 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3401 u64 *trimmed, u64 start, u64 end, u64 minlen)
3402 {
3403 int ret;
3404
3405 *trimmed = 0;
3406
3407 spin_lock(&block_group->lock);
3408 if (block_group->removed) {
3409 spin_unlock(&block_group->lock);
3410 return 0;
3411 }
3412 btrfs_get_block_group_trimming(block_group);
3413 spin_unlock(&block_group->lock);
3414
3415 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3416 if (ret)
3417 goto out;
3418
3419 ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3420 out:
3421 btrfs_put_block_group_trimming(block_group);
3422 return ret;
3423 }
3424
3425 /*
3426 * Find the left-most item in the cache tree, and then return the
3427 * smallest inode number in the item.
3428 *
3429 * Note: the returned inode number may not be the smallest one in
3430 * the tree, if the left-most item is a bitmap.
3431 */
3432 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3433 {
3434 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3435 struct btrfs_free_space *entry = NULL;
3436 u64 ino = 0;
3437
3438 spin_lock(&ctl->tree_lock);
3439
3440 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3441 goto out;
3442
3443 entry = rb_entry(rb_first(&ctl->free_space_offset),
3444 struct btrfs_free_space, offset_index);
3445
3446 if (!entry->bitmap) {
3447 ino = entry->offset;
3448
3449 unlink_free_space(ctl, entry);
3450 entry->offset++;
3451 entry->bytes--;
3452 if (!entry->bytes)
3453 kmem_cache_free(btrfs_free_space_cachep, entry);
3454 else
3455 link_free_space(ctl, entry);
3456 } else {
3457 u64 offset = 0;
3458 u64 count = 1;
3459 int ret;
3460
3461 ret = search_bitmap(ctl, entry, &offset, &count, true);
3462 /* Logic error; Should be empty if it can't find anything */
3463 ASSERT(!ret);
3464
3465 ino = offset;
3466 bitmap_clear_bits(ctl, entry, offset, 1);
3467 if (entry->bytes == 0)
3468 free_bitmap(ctl, entry);
3469 }
3470 out:
3471 spin_unlock(&ctl->tree_lock);
3472
3473 return ino;
3474 }
3475
3476 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3477 struct btrfs_path *path)
3478 {
3479 struct inode *inode = NULL;
3480
3481 spin_lock(&root->ino_cache_lock);
3482 if (root->ino_cache_inode)
3483 inode = igrab(root->ino_cache_inode);
3484 spin_unlock(&root->ino_cache_lock);
3485 if (inode)
3486 return inode;
3487
3488 inode = __lookup_free_space_inode(root, path, 0);
3489 if (IS_ERR(inode))
3490 return inode;
3491
3492 spin_lock(&root->ino_cache_lock);
3493 if (!btrfs_fs_closing(root->fs_info))
3494 root->ino_cache_inode = igrab(inode);
3495 spin_unlock(&root->ino_cache_lock);
3496
3497 return inode;
3498 }
3499
3500 int create_free_ino_inode(struct btrfs_root *root,
3501 struct btrfs_trans_handle *trans,
3502 struct btrfs_path *path)
3503 {
3504 return __create_free_space_inode(root, trans, path,
3505 BTRFS_FREE_INO_OBJECTID, 0);
3506 }
3507
3508 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3509 {
3510 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3511 struct btrfs_path *path;
3512 struct inode *inode;
3513 int ret = 0;
3514 u64 root_gen = btrfs_root_generation(&root->root_item);
3515
3516 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3517 return 0;
3518
3519 /*
3520 * If we're unmounting then just return, since this does a search on the
3521 * normal root and not the commit root and we could deadlock.
3522 */
3523 if (btrfs_fs_closing(fs_info))
3524 return 0;
3525
3526 path = btrfs_alloc_path();
3527 if (!path)
3528 return 0;
3529
3530 inode = lookup_free_ino_inode(root, path);
3531 if (IS_ERR(inode))
3532 goto out;
3533
3534 if (root_gen != BTRFS_I(inode)->generation)
3535 goto out_put;
3536
3537 ret = __load_free_space_cache(root, inode, ctl, path, 0);
3538
3539 if (ret < 0)
3540 btrfs_err(fs_info,
3541 "failed to load free ino cache for root %llu",
3542 root->root_key.objectid);
3543 out_put:
3544 iput(inode);
3545 out:
3546 btrfs_free_path(path);
3547 return ret;
3548 }
3549
3550 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3551 struct btrfs_trans_handle *trans,
3552 struct btrfs_path *path,
3553 struct inode *inode)
3554 {
3555 struct btrfs_fs_info *fs_info = root->fs_info;
3556 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3557 int ret;
3558 struct btrfs_io_ctl io_ctl;
3559 bool release_metadata = true;
3560
3561 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3562 return 0;
3563
3564 memset(&io_ctl, 0, sizeof(io_ctl));
3565 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3566 if (!ret) {
3567 /*
3568 * At this point writepages() didn't error out, so our metadata
3569 * reservation is released when the writeback finishes, at
3570 * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3571 * with or without an error.
3572 */
3573 release_metadata = false;
3574 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3575 }
3576
3577 if (ret) {
3578 if (release_metadata)
3579 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3580 inode->i_size, true);
3581 #ifdef DEBUG
3582 btrfs_err(fs_info,
3583 "failed to write free ino cache for root %llu",
3584 root->root_key.objectid);
3585 #endif
3586 }
3587
3588 return ret;
3589 }
3590
3591 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3592 /*
3593 * Use this if you need to make a bitmap or extent entry specifically, it
3594 * doesn't do any of the merging that add_free_space does, this acts a lot like
3595 * how the free space cache loading stuff works, so you can get really weird
3596 * configurations.
3597 */
3598 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3599 u64 offset, u64 bytes, bool bitmap)
3600 {
3601 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3602 struct btrfs_free_space *info = NULL, *bitmap_info;
3603 void *map = NULL;
3604 u64 bytes_added;
3605 int ret;
3606
3607 again:
3608 if (!info) {
3609 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3610 if (!info)
3611 return -ENOMEM;
3612 }
3613
3614 if (!bitmap) {
3615 spin_lock(&ctl->tree_lock);
3616 info->offset = offset;
3617 info->bytes = bytes;
3618 info->max_extent_size = 0;
3619 ret = link_free_space(ctl, info);
3620 spin_unlock(&ctl->tree_lock);
3621 if (ret)
3622 kmem_cache_free(btrfs_free_space_cachep, info);
3623 return ret;
3624 }
3625
3626 if (!map) {
3627 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
3628 if (!map) {
3629 kmem_cache_free(btrfs_free_space_cachep, info);
3630 return -ENOMEM;
3631 }
3632 }
3633
3634 spin_lock(&ctl->tree_lock);
3635 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3636 1, 0);
3637 if (!bitmap_info) {
3638 info->bitmap = map;
3639 map = NULL;
3640 add_new_bitmap(ctl, info, offset);
3641 bitmap_info = info;
3642 info = NULL;
3643 }
3644
3645 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3646
3647 bytes -= bytes_added;
3648 offset += bytes_added;
3649 spin_unlock(&ctl->tree_lock);
3650
3651 if (bytes)
3652 goto again;
3653
3654 if (info)
3655 kmem_cache_free(btrfs_free_space_cachep, info);
3656 if (map)
3657 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
3658 return 0;
3659 }
3660
3661 /*
3662 * Checks to see if the given range is in the free space cache. This is really
3663 * just used to check the absence of space, so if there is free space in the
3664 * range at all we will return 1.
3665 */
3666 int test_check_exists(struct btrfs_block_group_cache *cache,
3667 u64 offset, u64 bytes)
3668 {
3669 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3670 struct btrfs_free_space *info;
3671 int ret = 0;
3672
3673 spin_lock(&ctl->tree_lock);
3674 info = tree_search_offset(ctl, offset, 0, 0);
3675 if (!info) {
3676 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3677 1, 0);
3678 if (!info)
3679 goto out;
3680 }
3681
3682 have_info:
3683 if (info->bitmap) {
3684 u64 bit_off, bit_bytes;
3685 struct rb_node *n;
3686 struct btrfs_free_space *tmp;
3687
3688 bit_off = offset;
3689 bit_bytes = ctl->unit;
3690 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3691 if (!ret) {
3692 if (bit_off == offset) {
3693 ret = 1;
3694 goto out;
3695 } else if (bit_off > offset &&
3696 offset + bytes > bit_off) {
3697 ret = 1;
3698 goto out;
3699 }
3700 }
3701
3702 n = rb_prev(&info->offset_index);
3703 while (n) {
3704 tmp = rb_entry(n, struct btrfs_free_space,
3705 offset_index);
3706 if (tmp->offset + tmp->bytes < offset)
3707 break;
3708 if (offset + bytes < tmp->offset) {
3709 n = rb_prev(&tmp->offset_index);
3710 continue;
3711 }
3712 info = tmp;
3713 goto have_info;
3714 }
3715
3716 n = rb_next(&info->offset_index);
3717 while (n) {
3718 tmp = rb_entry(n, struct btrfs_free_space,
3719 offset_index);
3720 if (offset + bytes < tmp->offset)
3721 break;
3722 if (tmp->offset + tmp->bytes < offset) {
3723 n = rb_next(&tmp->offset_index);
3724 continue;
3725 }
3726 info = tmp;
3727 goto have_info;
3728 }
3729
3730 ret = 0;
3731 goto out;
3732 }
3733
3734 if (info->offset == offset) {
3735 ret = 1;
3736 goto out;
3737 }
3738
3739 if (offset > info->offset && offset < info->offset + info->bytes)
3740 ret = 1;
3741 out:
3742 spin_unlock(&ctl->tree_lock);
3743 return ret;
3744 }
3745 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */
Linux with added FPC-III support