Loading...
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 <linux/string_choices.h>
15#include "ctree.h"
16#include "fs.h"
17#include "messages.h"
18#include "misc.h"
19#include "free-space-cache.h"
20#include "transaction.h"
21#include "disk-io.h"
22#include "extent_io.h"
23#include "space-info.h"
24#include "block-group.h"
25#include "discard.h"
26#include "subpage.h"
27#include "inode-item.h"
28#include "accessors.h"
29#include "file-item.h"
30#include "file.h"
31#include "super.h"
32
33#define BITS_PER_BITMAP (PAGE_SIZE * 8UL)
34#define MAX_CACHE_BYTES_PER_GIG SZ_64K
35#define FORCE_EXTENT_THRESHOLD SZ_1M
36
37static struct kmem_cache *btrfs_free_space_cachep;
38static struct kmem_cache *btrfs_free_space_bitmap_cachep;
39
40struct btrfs_trim_range {
41 u64 start;
42 u64 bytes;
43 struct list_head list;
44};
45
46static int link_free_space(struct btrfs_free_space_ctl *ctl,
47 struct btrfs_free_space *info);
48static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
49 struct btrfs_free_space *info, bool update_stat);
50static int search_bitmap(struct btrfs_free_space_ctl *ctl,
51 struct btrfs_free_space *bitmap_info, u64 *offset,
52 u64 *bytes, bool for_alloc);
53static void free_bitmap(struct btrfs_free_space_ctl *ctl,
54 struct btrfs_free_space *bitmap_info);
55static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
56 struct btrfs_free_space *info, u64 offset,
57 u64 bytes, bool update_stats);
58
59static void btrfs_crc32c_final(u32 crc, u8 *result)
60{
61 put_unaligned_le32(~crc, result);
62}
63
64static void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
65{
66 struct btrfs_free_space *info;
67 struct rb_node *node;
68
69 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
70 info = rb_entry(node, struct btrfs_free_space, offset_index);
71 if (!info->bitmap) {
72 unlink_free_space(ctl, info, true);
73 kmem_cache_free(btrfs_free_space_cachep, info);
74 } else {
75 free_bitmap(ctl, info);
76 }
77
78 cond_resched_lock(&ctl->tree_lock);
79 }
80}
81
82static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
83 struct btrfs_path *path,
84 u64 offset)
85{
86 struct btrfs_key key;
87 struct btrfs_key location;
88 struct btrfs_disk_key disk_key;
89 struct btrfs_free_space_header *header;
90 struct extent_buffer *leaf;
91 struct inode *inode = NULL;
92 unsigned nofs_flag;
93 int ret;
94
95 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
96 key.offset = offset;
97 key.type = 0;
98
99 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
100 if (ret < 0)
101 return ERR_PTR(ret);
102 if (ret > 0) {
103 btrfs_release_path(path);
104 return ERR_PTR(-ENOENT);
105 }
106
107 leaf = path->nodes[0];
108 header = btrfs_item_ptr(leaf, path->slots[0],
109 struct btrfs_free_space_header);
110 btrfs_free_space_key(leaf, header, &disk_key);
111 btrfs_disk_key_to_cpu(&location, &disk_key);
112 btrfs_release_path(path);
113
114 /*
115 * We are often under a trans handle at this point, so we need to make
116 * sure NOFS is set to keep us from deadlocking.
117 */
118 nofs_flag = memalloc_nofs_save();
119 inode = btrfs_iget_path(location.objectid, root, path);
120 btrfs_release_path(path);
121 memalloc_nofs_restore(nofs_flag);
122 if (IS_ERR(inode))
123 return inode;
124
125 mapping_set_gfp_mask(inode->i_mapping,
126 mapping_gfp_constraint(inode->i_mapping,
127 ~(__GFP_FS | __GFP_HIGHMEM)));
128
129 return inode;
130}
131
132struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
133 struct btrfs_path *path)
134{
135 struct btrfs_fs_info *fs_info = block_group->fs_info;
136 struct inode *inode = NULL;
137 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
138
139 spin_lock(&block_group->lock);
140 if (block_group->inode)
141 inode = igrab(&block_group->inode->vfs_inode);
142 spin_unlock(&block_group->lock);
143 if (inode)
144 return inode;
145
146 inode = __lookup_free_space_inode(fs_info->tree_root, path,
147 block_group->start);
148 if (IS_ERR(inode))
149 return inode;
150
151 spin_lock(&block_group->lock);
152 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
153 btrfs_info(fs_info, "Old style space inode found, converting.");
154 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
155 BTRFS_INODE_NODATACOW;
156 block_group->disk_cache_state = BTRFS_DC_CLEAR;
157 }
158
159 if (!test_and_set_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags))
160 block_group->inode = BTRFS_I(igrab(inode));
161 spin_unlock(&block_group->lock);
162
163 return inode;
164}
165
166static int __create_free_space_inode(struct btrfs_root *root,
167 struct btrfs_trans_handle *trans,
168 struct btrfs_path *path,
169 u64 ino, u64 offset)
170{
171 struct btrfs_key key;
172 struct btrfs_disk_key disk_key;
173 struct btrfs_free_space_header *header;
174 struct btrfs_inode_item *inode_item;
175 struct extent_buffer *leaf;
176 /* We inline CRCs for the free disk space cache */
177 const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
178 BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
179 int ret;
180
181 ret = btrfs_insert_empty_inode(trans, root, path, ino);
182 if (ret)
183 return ret;
184
185 leaf = path->nodes[0];
186 inode_item = btrfs_item_ptr(leaf, path->slots[0],
187 struct btrfs_inode_item);
188 btrfs_item_key(leaf, &disk_key, path->slots[0]);
189 memzero_extent_buffer(leaf, (unsigned long)inode_item,
190 sizeof(*inode_item));
191 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
192 btrfs_set_inode_size(leaf, inode_item, 0);
193 btrfs_set_inode_nbytes(leaf, inode_item, 0);
194 btrfs_set_inode_uid(leaf, inode_item, 0);
195 btrfs_set_inode_gid(leaf, inode_item, 0);
196 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
197 btrfs_set_inode_flags(leaf, inode_item, flags);
198 btrfs_set_inode_nlink(leaf, inode_item, 1);
199 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
200 btrfs_set_inode_block_group(leaf, inode_item, offset);
201 btrfs_mark_buffer_dirty(trans, leaf);
202 btrfs_release_path(path);
203
204 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
205 key.offset = offset;
206 key.type = 0;
207 ret = btrfs_insert_empty_item(trans, root, path, &key,
208 sizeof(struct btrfs_free_space_header));
209 if (ret < 0) {
210 btrfs_release_path(path);
211 return ret;
212 }
213
214 leaf = path->nodes[0];
215 header = btrfs_item_ptr(leaf, path->slots[0],
216 struct btrfs_free_space_header);
217 memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
218 btrfs_set_free_space_key(leaf, header, &disk_key);
219 btrfs_mark_buffer_dirty(trans, leaf);
220 btrfs_release_path(path);
221
222 return 0;
223}
224
225int create_free_space_inode(struct btrfs_trans_handle *trans,
226 struct btrfs_block_group *block_group,
227 struct btrfs_path *path)
228{
229 int ret;
230 u64 ino;
231
232 ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
233 if (ret < 0)
234 return ret;
235
236 return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
237 ino, block_group->start);
238}
239
240/*
241 * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
242 * handles lookup, otherwise it takes ownership and iputs the inode.
243 * Don't reuse an inode pointer after passing it into this function.
244 */
245int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
246 struct inode *inode,
247 struct btrfs_block_group *block_group)
248{
249 struct btrfs_path *path;
250 struct btrfs_key key;
251 int ret = 0;
252
253 path = btrfs_alloc_path();
254 if (!path)
255 return -ENOMEM;
256
257 if (!inode)
258 inode = lookup_free_space_inode(block_group, path);
259 if (IS_ERR(inode)) {
260 if (PTR_ERR(inode) != -ENOENT)
261 ret = PTR_ERR(inode);
262 goto out;
263 }
264 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
265 if (ret) {
266 btrfs_add_delayed_iput(BTRFS_I(inode));
267 goto out;
268 }
269 clear_nlink(inode);
270 /* One for the block groups ref */
271 spin_lock(&block_group->lock);
272 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags)) {
273 block_group->inode = NULL;
274 spin_unlock(&block_group->lock);
275 iput(inode);
276 } else {
277 spin_unlock(&block_group->lock);
278 }
279 /* One for the lookup ref */
280 btrfs_add_delayed_iput(BTRFS_I(inode));
281
282 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
283 key.type = 0;
284 key.offset = block_group->start;
285 ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
286 -1, 1);
287 if (ret) {
288 if (ret > 0)
289 ret = 0;
290 goto out;
291 }
292 ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
293out:
294 btrfs_free_path(path);
295 return ret;
296}
297
298int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
299 struct btrfs_block_group *block_group,
300 struct inode *vfs_inode)
301{
302 struct btrfs_truncate_control control = {
303 .inode = BTRFS_I(vfs_inode),
304 .new_size = 0,
305 .ino = btrfs_ino(BTRFS_I(vfs_inode)),
306 .min_type = BTRFS_EXTENT_DATA_KEY,
307 .clear_extent_range = true,
308 };
309 struct btrfs_inode *inode = BTRFS_I(vfs_inode);
310 struct btrfs_root *root = inode->root;
311 struct extent_state *cached_state = NULL;
312 int ret = 0;
313 bool locked = false;
314
315 if (block_group) {
316 struct btrfs_path *path = btrfs_alloc_path();
317
318 if (!path) {
319 ret = -ENOMEM;
320 goto fail;
321 }
322 locked = true;
323 mutex_lock(&trans->transaction->cache_write_mutex);
324 if (!list_empty(&block_group->io_list)) {
325 list_del_init(&block_group->io_list);
326
327 btrfs_wait_cache_io(trans, block_group, path);
328 btrfs_put_block_group(block_group);
329 }
330
331 /*
332 * now that we've truncated the cache away, its no longer
333 * setup or written
334 */
335 spin_lock(&block_group->lock);
336 block_group->disk_cache_state = BTRFS_DC_CLEAR;
337 spin_unlock(&block_group->lock);
338 btrfs_free_path(path);
339 }
340
341 btrfs_i_size_write(inode, 0);
342 truncate_pagecache(vfs_inode, 0);
343
344 lock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
345 btrfs_drop_extent_map_range(inode, 0, (u64)-1, false);
346
347 /*
348 * We skip the throttling logic for free space cache inodes, so we don't
349 * need to check for -EAGAIN.
350 */
351 ret = btrfs_truncate_inode_items(trans, root, &control);
352
353 inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
354 btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
355
356 unlock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
357 if (ret)
358 goto fail;
359
360 ret = btrfs_update_inode(trans, inode);
361
362fail:
363 if (locked)
364 mutex_unlock(&trans->transaction->cache_write_mutex);
365 if (ret)
366 btrfs_abort_transaction(trans, ret);
367
368 return ret;
369}
370
371static void readahead_cache(struct inode *inode)
372{
373 struct file_ra_state ra;
374 unsigned long last_index;
375
376 file_ra_state_init(&ra, inode->i_mapping);
377 last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
378
379 page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
380}
381
382static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
383 int write)
384{
385 int num_pages;
386
387 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
388
389 /* Make sure we can fit our crcs and generation into the first page */
390 if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
391 return -ENOSPC;
392
393 memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
394
395 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
396 if (!io_ctl->pages)
397 return -ENOMEM;
398
399 io_ctl->num_pages = num_pages;
400 io_ctl->fs_info = inode_to_fs_info(inode);
401 io_ctl->inode = inode;
402
403 return 0;
404}
405ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
406
407static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
408{
409 kfree(io_ctl->pages);
410 io_ctl->pages = NULL;
411}
412
413static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
414{
415 if (io_ctl->cur) {
416 io_ctl->cur = NULL;
417 io_ctl->orig = NULL;
418 }
419}
420
421static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
422{
423 ASSERT(io_ctl->index < io_ctl->num_pages);
424 io_ctl->page = io_ctl->pages[io_ctl->index++];
425 io_ctl->cur = page_address(io_ctl->page);
426 io_ctl->orig = io_ctl->cur;
427 io_ctl->size = PAGE_SIZE;
428 if (clear)
429 clear_page(io_ctl->cur);
430}
431
432static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
433{
434 int i;
435
436 io_ctl_unmap_page(io_ctl);
437
438 for (i = 0; i < io_ctl->num_pages; i++) {
439 if (io_ctl->pages[i]) {
440 btrfs_folio_clear_checked(io_ctl->fs_info,
441 page_folio(io_ctl->pages[i]),
442 page_offset(io_ctl->pages[i]),
443 PAGE_SIZE);
444 unlock_page(io_ctl->pages[i]);
445 put_page(io_ctl->pages[i]);
446 }
447 }
448}
449
450static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
451{
452 struct page *page;
453 struct inode *inode = io_ctl->inode;
454 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
455 int i;
456
457 for (i = 0; i < io_ctl->num_pages; i++) {
458 int ret;
459
460 page = find_or_create_page(inode->i_mapping, i, mask);
461 if (!page) {
462 io_ctl_drop_pages(io_ctl);
463 return -ENOMEM;
464 }
465
466 ret = set_page_extent_mapped(page);
467 if (ret < 0) {
468 unlock_page(page);
469 put_page(page);
470 io_ctl_drop_pages(io_ctl);
471 return ret;
472 }
473
474 io_ctl->pages[i] = page;
475 if (uptodate && !PageUptodate(page)) {
476 btrfs_read_folio(NULL, page_folio(page));
477 lock_page(page);
478 if (page->mapping != inode->i_mapping) {
479 btrfs_err(BTRFS_I(inode)->root->fs_info,
480 "free space cache page truncated");
481 io_ctl_drop_pages(io_ctl);
482 return -EIO;
483 }
484 if (!PageUptodate(page)) {
485 btrfs_err(BTRFS_I(inode)->root->fs_info,
486 "error reading free space cache");
487 io_ctl_drop_pages(io_ctl);
488 return -EIO;
489 }
490 }
491 }
492
493 for (i = 0; i < io_ctl->num_pages; i++)
494 clear_page_dirty_for_io(io_ctl->pages[i]);
495
496 return 0;
497}
498
499static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
500{
501 io_ctl_map_page(io_ctl, 1);
502
503 /*
504 * Skip the csum areas. If we don't check crcs then we just have a
505 * 64bit chunk at the front of the first page.
506 */
507 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
508 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
509
510 put_unaligned_le64(generation, io_ctl->cur);
511 io_ctl->cur += sizeof(u64);
512}
513
514static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
515{
516 u64 cache_gen;
517
518 /*
519 * Skip the crc area. If we don't check crcs then we just have a 64bit
520 * chunk at the front of the first page.
521 */
522 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
523 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
524
525 cache_gen = get_unaligned_le64(io_ctl->cur);
526 if (cache_gen != generation) {
527 btrfs_err_rl(io_ctl->fs_info,
528 "space cache generation (%llu) does not match inode (%llu)",
529 cache_gen, generation);
530 io_ctl_unmap_page(io_ctl);
531 return -EIO;
532 }
533 io_ctl->cur += sizeof(u64);
534 return 0;
535}
536
537static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
538{
539 u32 *tmp;
540 u32 crc = ~(u32)0;
541 unsigned offset = 0;
542
543 if (index == 0)
544 offset = sizeof(u32) * io_ctl->num_pages;
545
546 crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
547 btrfs_crc32c_final(crc, (u8 *)&crc);
548 io_ctl_unmap_page(io_ctl);
549 tmp = page_address(io_ctl->pages[0]);
550 tmp += index;
551 *tmp = crc;
552}
553
554static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
555{
556 u32 *tmp, val;
557 u32 crc = ~(u32)0;
558 unsigned offset = 0;
559
560 if (index == 0)
561 offset = sizeof(u32) * io_ctl->num_pages;
562
563 tmp = page_address(io_ctl->pages[0]);
564 tmp += index;
565 val = *tmp;
566
567 io_ctl_map_page(io_ctl, 0);
568 crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
569 btrfs_crc32c_final(crc, (u8 *)&crc);
570 if (val != crc) {
571 btrfs_err_rl(io_ctl->fs_info,
572 "csum mismatch on free space cache");
573 io_ctl_unmap_page(io_ctl);
574 return -EIO;
575 }
576
577 return 0;
578}
579
580static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
581 void *bitmap)
582{
583 struct btrfs_free_space_entry *entry;
584
585 if (!io_ctl->cur)
586 return -ENOSPC;
587
588 entry = io_ctl->cur;
589 put_unaligned_le64(offset, &entry->offset);
590 put_unaligned_le64(bytes, &entry->bytes);
591 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
592 BTRFS_FREE_SPACE_EXTENT;
593 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
594 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
595
596 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
597 return 0;
598
599 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
600
601 /* No more pages to map */
602 if (io_ctl->index >= io_ctl->num_pages)
603 return 0;
604
605 /* map the next page */
606 io_ctl_map_page(io_ctl, 1);
607 return 0;
608}
609
610static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
611{
612 if (!io_ctl->cur)
613 return -ENOSPC;
614
615 /*
616 * If we aren't at the start of the current page, unmap this one and
617 * map the next one if there is any left.
618 */
619 if (io_ctl->cur != io_ctl->orig) {
620 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
621 if (io_ctl->index >= io_ctl->num_pages)
622 return -ENOSPC;
623 io_ctl_map_page(io_ctl, 0);
624 }
625
626 copy_page(io_ctl->cur, bitmap);
627 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
628 if (io_ctl->index < io_ctl->num_pages)
629 io_ctl_map_page(io_ctl, 0);
630 return 0;
631}
632
633static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
634{
635 /*
636 * If we're not on the boundary we know we've modified the page and we
637 * need to crc the page.
638 */
639 if (io_ctl->cur != io_ctl->orig)
640 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
641 else
642 io_ctl_unmap_page(io_ctl);
643
644 while (io_ctl->index < io_ctl->num_pages) {
645 io_ctl_map_page(io_ctl, 1);
646 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
647 }
648}
649
650static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
651 struct btrfs_free_space *entry, u8 *type)
652{
653 struct btrfs_free_space_entry *e;
654 int ret;
655
656 if (!io_ctl->cur) {
657 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
658 if (ret)
659 return ret;
660 }
661
662 e = io_ctl->cur;
663 entry->offset = get_unaligned_le64(&e->offset);
664 entry->bytes = get_unaligned_le64(&e->bytes);
665 *type = e->type;
666 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
667 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
668
669 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
670 return 0;
671
672 io_ctl_unmap_page(io_ctl);
673
674 return 0;
675}
676
677static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
678 struct btrfs_free_space *entry)
679{
680 int ret;
681
682 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
683 if (ret)
684 return ret;
685
686 copy_page(entry->bitmap, io_ctl->cur);
687 io_ctl_unmap_page(io_ctl);
688
689 return 0;
690}
691
692static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
693{
694 struct btrfs_block_group *block_group = ctl->block_group;
695 u64 max_bytes;
696 u64 bitmap_bytes;
697 u64 extent_bytes;
698 u64 size = block_group->length;
699 u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
700 u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
701
702 max_bitmaps = max_t(u64, max_bitmaps, 1);
703
704 if (ctl->total_bitmaps > max_bitmaps)
705 btrfs_err(block_group->fs_info,
706"invalid free space control: bg start=%llu len=%llu total_bitmaps=%u unit=%u max_bitmaps=%llu bytes_per_bg=%llu",
707 block_group->start, block_group->length,
708 ctl->total_bitmaps, ctl->unit, max_bitmaps,
709 bytes_per_bg);
710 ASSERT(ctl->total_bitmaps <= max_bitmaps);
711
712 /*
713 * We are trying to keep the total amount of memory used per 1GiB of
714 * space to be MAX_CACHE_BYTES_PER_GIG. However, with a reclamation
715 * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
716 * bitmaps, we may end up using more memory than this.
717 */
718 if (size < SZ_1G)
719 max_bytes = MAX_CACHE_BYTES_PER_GIG;
720 else
721 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
722
723 bitmap_bytes = ctl->total_bitmaps * ctl->unit;
724
725 /*
726 * we want the extent entry threshold to always be at most 1/2 the max
727 * bytes we can have, or whatever is less than that.
728 */
729 extent_bytes = max_bytes - bitmap_bytes;
730 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
731
732 ctl->extents_thresh =
733 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
734}
735
736static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
737 struct btrfs_free_space_ctl *ctl,
738 struct btrfs_path *path, u64 offset)
739{
740 struct btrfs_fs_info *fs_info = root->fs_info;
741 struct btrfs_free_space_header *header;
742 struct extent_buffer *leaf;
743 struct btrfs_io_ctl io_ctl;
744 struct btrfs_key key;
745 struct btrfs_free_space *e, *n;
746 LIST_HEAD(bitmaps);
747 u64 num_entries;
748 u64 num_bitmaps;
749 u64 generation;
750 u8 type;
751 int ret = 0;
752
753 /* Nothing in the space cache, goodbye */
754 if (!i_size_read(inode))
755 return 0;
756
757 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
758 key.offset = offset;
759 key.type = 0;
760
761 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
762 if (ret < 0)
763 return 0;
764 else if (ret > 0) {
765 btrfs_release_path(path);
766 return 0;
767 }
768
769 ret = -1;
770
771 leaf = path->nodes[0];
772 header = btrfs_item_ptr(leaf, path->slots[0],
773 struct btrfs_free_space_header);
774 num_entries = btrfs_free_space_entries(leaf, header);
775 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
776 generation = btrfs_free_space_generation(leaf, header);
777 btrfs_release_path(path);
778
779 if (!BTRFS_I(inode)->generation) {
780 btrfs_info(fs_info,
781 "the free space cache file (%llu) is invalid, skip it",
782 offset);
783 return 0;
784 }
785
786 if (BTRFS_I(inode)->generation != generation) {
787 btrfs_err(fs_info,
788 "free space inode generation (%llu) did not match free space cache generation (%llu)",
789 BTRFS_I(inode)->generation, generation);
790 return 0;
791 }
792
793 if (!num_entries)
794 return 0;
795
796 ret = io_ctl_init(&io_ctl, inode, 0);
797 if (ret)
798 return ret;
799
800 readahead_cache(inode);
801
802 ret = io_ctl_prepare_pages(&io_ctl, true);
803 if (ret)
804 goto out;
805
806 ret = io_ctl_check_crc(&io_ctl, 0);
807 if (ret)
808 goto free_cache;
809
810 ret = io_ctl_check_generation(&io_ctl, generation);
811 if (ret)
812 goto free_cache;
813
814 while (num_entries) {
815 e = kmem_cache_zalloc(btrfs_free_space_cachep,
816 GFP_NOFS);
817 if (!e) {
818 ret = -ENOMEM;
819 goto free_cache;
820 }
821
822 ret = io_ctl_read_entry(&io_ctl, e, &type);
823 if (ret) {
824 kmem_cache_free(btrfs_free_space_cachep, e);
825 goto free_cache;
826 }
827
828 if (!e->bytes) {
829 ret = -1;
830 kmem_cache_free(btrfs_free_space_cachep, e);
831 goto free_cache;
832 }
833
834 if (type == BTRFS_FREE_SPACE_EXTENT) {
835 spin_lock(&ctl->tree_lock);
836 ret = link_free_space(ctl, e);
837 spin_unlock(&ctl->tree_lock);
838 if (ret) {
839 btrfs_err(fs_info,
840 "Duplicate entries in free space cache, dumping");
841 kmem_cache_free(btrfs_free_space_cachep, e);
842 goto free_cache;
843 }
844 } else {
845 ASSERT(num_bitmaps);
846 num_bitmaps--;
847 e->bitmap = kmem_cache_zalloc(
848 btrfs_free_space_bitmap_cachep, GFP_NOFS);
849 if (!e->bitmap) {
850 ret = -ENOMEM;
851 kmem_cache_free(
852 btrfs_free_space_cachep, e);
853 goto free_cache;
854 }
855 spin_lock(&ctl->tree_lock);
856 ret = link_free_space(ctl, e);
857 if (ret) {
858 spin_unlock(&ctl->tree_lock);
859 btrfs_err(fs_info,
860 "Duplicate entries in free space cache, dumping");
861 kmem_cache_free(btrfs_free_space_bitmap_cachep, e->bitmap);
862 kmem_cache_free(btrfs_free_space_cachep, e);
863 goto free_cache;
864 }
865 ctl->total_bitmaps++;
866 recalculate_thresholds(ctl);
867 spin_unlock(&ctl->tree_lock);
868 list_add_tail(&e->list, &bitmaps);
869 }
870
871 num_entries--;
872 }
873
874 io_ctl_unmap_page(&io_ctl);
875
876 /*
877 * We add the bitmaps at the end of the entries in order that
878 * the bitmap entries are added to the cache.
879 */
880 list_for_each_entry_safe(e, n, &bitmaps, list) {
881 list_del_init(&e->list);
882 ret = io_ctl_read_bitmap(&io_ctl, e);
883 if (ret)
884 goto free_cache;
885 }
886
887 io_ctl_drop_pages(&io_ctl);
888 ret = 1;
889out:
890 io_ctl_free(&io_ctl);
891 return ret;
892free_cache:
893 io_ctl_drop_pages(&io_ctl);
894
895 spin_lock(&ctl->tree_lock);
896 __btrfs_remove_free_space_cache(ctl);
897 spin_unlock(&ctl->tree_lock);
898 goto out;
899}
900
901static int copy_free_space_cache(struct btrfs_block_group *block_group,
902 struct btrfs_free_space_ctl *ctl)
903{
904 struct btrfs_free_space *info;
905 struct rb_node *n;
906 int ret = 0;
907
908 while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
909 info = rb_entry(n, struct btrfs_free_space, offset_index);
910 if (!info->bitmap) {
911 const u64 offset = info->offset;
912 const u64 bytes = info->bytes;
913
914 unlink_free_space(ctl, info, true);
915 spin_unlock(&ctl->tree_lock);
916 kmem_cache_free(btrfs_free_space_cachep, info);
917 ret = btrfs_add_free_space(block_group, offset, bytes);
918 spin_lock(&ctl->tree_lock);
919 } else {
920 u64 offset = info->offset;
921 u64 bytes = ctl->unit;
922
923 ret = search_bitmap(ctl, info, &offset, &bytes, false);
924 if (ret == 0) {
925 bitmap_clear_bits(ctl, info, offset, bytes, true);
926 spin_unlock(&ctl->tree_lock);
927 ret = btrfs_add_free_space(block_group, offset,
928 bytes);
929 spin_lock(&ctl->tree_lock);
930 } else {
931 free_bitmap(ctl, info);
932 ret = 0;
933 }
934 }
935 cond_resched_lock(&ctl->tree_lock);
936 }
937 return ret;
938}
939
940static struct lock_class_key btrfs_free_space_inode_key;
941
942int load_free_space_cache(struct btrfs_block_group *block_group)
943{
944 struct btrfs_fs_info *fs_info = block_group->fs_info;
945 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
946 struct btrfs_free_space_ctl tmp_ctl = {};
947 struct inode *inode;
948 struct btrfs_path *path;
949 int ret = 0;
950 bool matched;
951 u64 used = block_group->used;
952
953 /*
954 * Because we could potentially discard our loaded free space, we want
955 * to load everything into a temporary structure first, and then if it's
956 * valid copy it all into the actual free space ctl.
957 */
958 btrfs_init_free_space_ctl(block_group, &tmp_ctl);
959
960 /*
961 * If this block group has been marked to be cleared for one reason or
962 * another then we can't trust the on disk cache, so just return.
963 */
964 spin_lock(&block_group->lock);
965 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
966 spin_unlock(&block_group->lock);
967 return 0;
968 }
969 spin_unlock(&block_group->lock);
970
971 path = btrfs_alloc_path();
972 if (!path)
973 return 0;
974 path->search_commit_root = 1;
975 path->skip_locking = 1;
976
977 /*
978 * We must pass a path with search_commit_root set to btrfs_iget in
979 * order to avoid a deadlock when allocating extents for the tree root.
980 *
981 * When we are COWing an extent buffer from the tree root, when looking
982 * for a free extent, at extent-tree.c:find_free_extent(), we can find
983 * block group without its free space cache loaded. When we find one
984 * we must load its space cache which requires reading its free space
985 * cache's inode item from the root tree. If this inode item is located
986 * in the same leaf that we started COWing before, then we end up in
987 * deadlock on the extent buffer (trying to read lock it when we
988 * previously write locked it).
989 *
990 * It's safe to read the inode item using the commit root because
991 * block groups, once loaded, stay in memory forever (until they are
992 * removed) as well as their space caches once loaded. New block groups
993 * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
994 * we will never try to read their inode item while the fs is mounted.
995 */
996 inode = lookup_free_space_inode(block_group, path);
997 if (IS_ERR(inode)) {
998 btrfs_free_path(path);
999 return 0;
1000 }
1001
1002 /* We may have converted the inode and made the cache invalid. */
1003 spin_lock(&block_group->lock);
1004 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
1005 spin_unlock(&block_group->lock);
1006 btrfs_free_path(path);
1007 goto out;
1008 }
1009 spin_unlock(&block_group->lock);
1010
1011 /*
1012 * Reinitialize the class of struct inode's mapping->invalidate_lock for
1013 * free space inodes to prevent false positives related to locks for normal
1014 * inodes.
1015 */
1016 lockdep_set_class(&(&inode->i_data)->invalidate_lock,
1017 &btrfs_free_space_inode_key);
1018
1019 ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
1020 path, block_group->start);
1021 btrfs_free_path(path);
1022 if (ret <= 0)
1023 goto out;
1024
1025 matched = (tmp_ctl.free_space == (block_group->length - used -
1026 block_group->bytes_super));
1027
1028 if (matched) {
1029 spin_lock(&tmp_ctl.tree_lock);
1030 ret = copy_free_space_cache(block_group, &tmp_ctl);
1031 spin_unlock(&tmp_ctl.tree_lock);
1032 /*
1033 * ret == 1 means we successfully loaded the free space cache,
1034 * so we need to re-set it here.
1035 */
1036 if (ret == 0)
1037 ret = 1;
1038 } else {
1039 /*
1040 * We need to call the _locked variant so we don't try to update
1041 * the discard counters.
1042 */
1043 spin_lock(&tmp_ctl.tree_lock);
1044 __btrfs_remove_free_space_cache(&tmp_ctl);
1045 spin_unlock(&tmp_ctl.tree_lock);
1046 btrfs_warn(fs_info,
1047 "block group %llu has wrong amount of free space",
1048 block_group->start);
1049 ret = -1;
1050 }
1051out:
1052 if (ret < 0) {
1053 /* This cache is bogus, make sure it gets cleared */
1054 spin_lock(&block_group->lock);
1055 block_group->disk_cache_state = BTRFS_DC_CLEAR;
1056 spin_unlock(&block_group->lock);
1057 ret = 0;
1058
1059 btrfs_warn(fs_info,
1060 "failed to load free space cache for block group %llu, rebuilding it now",
1061 block_group->start);
1062 }
1063
1064 spin_lock(&ctl->tree_lock);
1065 btrfs_discard_update_discardable(block_group);
1066 spin_unlock(&ctl->tree_lock);
1067 iput(inode);
1068 return ret;
1069}
1070
1071static noinline_for_stack
1072int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1073 struct btrfs_free_space_ctl *ctl,
1074 struct btrfs_block_group *block_group,
1075 int *entries, int *bitmaps,
1076 struct list_head *bitmap_list)
1077{
1078 int ret;
1079 struct btrfs_free_cluster *cluster = NULL;
1080 struct btrfs_free_cluster *cluster_locked = NULL;
1081 struct rb_node *node = rb_first(&ctl->free_space_offset);
1082 struct btrfs_trim_range *trim_entry;
1083
1084 /* Get the cluster for this block_group if it exists */
1085 if (block_group && !list_empty(&block_group->cluster_list)) {
1086 cluster = list_entry(block_group->cluster_list.next,
1087 struct btrfs_free_cluster,
1088 block_group_list);
1089 }
1090
1091 if (!node && cluster) {
1092 cluster_locked = cluster;
1093 spin_lock(&cluster_locked->lock);
1094 node = rb_first(&cluster->root);
1095 cluster = NULL;
1096 }
1097
1098 /* Write out the extent entries */
1099 while (node) {
1100 struct btrfs_free_space *e;
1101
1102 e = rb_entry(node, struct btrfs_free_space, offset_index);
1103 *entries += 1;
1104
1105 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1106 e->bitmap);
1107 if (ret)
1108 goto fail;
1109
1110 if (e->bitmap) {
1111 list_add_tail(&e->list, bitmap_list);
1112 *bitmaps += 1;
1113 }
1114 node = rb_next(node);
1115 if (!node && cluster) {
1116 node = rb_first(&cluster->root);
1117 cluster_locked = cluster;
1118 spin_lock(&cluster_locked->lock);
1119 cluster = NULL;
1120 }
1121 }
1122 if (cluster_locked) {
1123 spin_unlock(&cluster_locked->lock);
1124 cluster_locked = NULL;
1125 }
1126
1127 /*
1128 * Make sure we don't miss any range that was removed from our rbtree
1129 * because trimming is running. Otherwise after a umount+mount (or crash
1130 * after committing the transaction) we would leak free space and get
1131 * an inconsistent free space cache report from fsck.
1132 */
1133 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1134 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1135 trim_entry->bytes, NULL);
1136 if (ret)
1137 goto fail;
1138 *entries += 1;
1139 }
1140
1141 return 0;
1142fail:
1143 if (cluster_locked)
1144 spin_unlock(&cluster_locked->lock);
1145 return -ENOSPC;
1146}
1147
1148static noinline_for_stack int
1149update_cache_item(struct btrfs_trans_handle *trans,
1150 struct btrfs_root *root,
1151 struct inode *inode,
1152 struct btrfs_path *path, u64 offset,
1153 int entries, int bitmaps)
1154{
1155 struct btrfs_key key;
1156 struct btrfs_free_space_header *header;
1157 struct extent_buffer *leaf;
1158 int ret;
1159
1160 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1161 key.offset = offset;
1162 key.type = 0;
1163
1164 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1165 if (ret < 0) {
1166 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1167 EXTENT_DELALLOC, NULL);
1168 goto fail;
1169 }
1170 leaf = path->nodes[0];
1171 if (ret > 0) {
1172 struct btrfs_key found_key;
1173 ASSERT(path->slots[0]);
1174 path->slots[0]--;
1175 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1176 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1177 found_key.offset != offset) {
1178 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1179 inode->i_size - 1, EXTENT_DELALLOC,
1180 NULL);
1181 btrfs_release_path(path);
1182 goto fail;
1183 }
1184 }
1185
1186 BTRFS_I(inode)->generation = trans->transid;
1187 header = btrfs_item_ptr(leaf, path->slots[0],
1188 struct btrfs_free_space_header);
1189 btrfs_set_free_space_entries(leaf, header, entries);
1190 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1191 btrfs_set_free_space_generation(leaf, header, trans->transid);
1192 btrfs_mark_buffer_dirty(trans, leaf);
1193 btrfs_release_path(path);
1194
1195 return 0;
1196
1197fail:
1198 return -1;
1199}
1200
1201static noinline_for_stack int write_pinned_extent_entries(
1202 struct btrfs_trans_handle *trans,
1203 struct btrfs_block_group *block_group,
1204 struct btrfs_io_ctl *io_ctl,
1205 int *entries)
1206{
1207 u64 start, extent_start, extent_end, len;
1208 struct extent_io_tree *unpin = NULL;
1209 int ret;
1210
1211 if (!block_group)
1212 return 0;
1213
1214 /*
1215 * We want to add any pinned extents to our free space cache
1216 * so we don't leak the space
1217 *
1218 * We shouldn't have switched the pinned extents yet so this is the
1219 * right one
1220 */
1221 unpin = &trans->transaction->pinned_extents;
1222
1223 start = block_group->start;
1224
1225 while (start < block_group->start + block_group->length) {
1226 if (!find_first_extent_bit(unpin, start,
1227 &extent_start, &extent_end,
1228 EXTENT_DIRTY, NULL))
1229 return 0;
1230
1231 /* This pinned extent is out of our range */
1232 if (extent_start >= block_group->start + block_group->length)
1233 return 0;
1234
1235 extent_start = max(extent_start, start);
1236 extent_end = min(block_group->start + block_group->length,
1237 extent_end + 1);
1238 len = extent_end - extent_start;
1239
1240 *entries += 1;
1241 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1242 if (ret)
1243 return -ENOSPC;
1244
1245 start = extent_end;
1246 }
1247
1248 return 0;
1249}
1250
1251static noinline_for_stack int
1252write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1253{
1254 struct btrfs_free_space *entry, *next;
1255 int ret;
1256
1257 /* Write out the bitmaps */
1258 list_for_each_entry_safe(entry, next, bitmap_list, list) {
1259 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1260 if (ret)
1261 return -ENOSPC;
1262 list_del_init(&entry->list);
1263 }
1264
1265 return 0;
1266}
1267
1268static int flush_dirty_cache(struct inode *inode)
1269{
1270 int ret;
1271
1272 ret = btrfs_wait_ordered_range(BTRFS_I(inode), 0, (u64)-1);
1273 if (ret)
1274 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1275 EXTENT_DELALLOC, NULL);
1276
1277 return ret;
1278}
1279
1280static void noinline_for_stack
1281cleanup_bitmap_list(struct list_head *bitmap_list)
1282{
1283 struct btrfs_free_space *entry, *next;
1284
1285 list_for_each_entry_safe(entry, next, bitmap_list, list)
1286 list_del_init(&entry->list);
1287}
1288
1289static void noinline_for_stack
1290cleanup_write_cache_enospc(struct inode *inode,
1291 struct btrfs_io_ctl *io_ctl,
1292 struct extent_state **cached_state)
1293{
1294 io_ctl_drop_pages(io_ctl);
1295 unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1296 cached_state);
1297}
1298
1299static int __btrfs_wait_cache_io(struct btrfs_root *root,
1300 struct btrfs_trans_handle *trans,
1301 struct btrfs_block_group *block_group,
1302 struct btrfs_io_ctl *io_ctl,
1303 struct btrfs_path *path, u64 offset)
1304{
1305 int ret;
1306 struct inode *inode = io_ctl->inode;
1307
1308 if (!inode)
1309 return 0;
1310
1311 /* Flush the dirty pages in the cache file. */
1312 ret = flush_dirty_cache(inode);
1313 if (ret)
1314 goto out;
1315
1316 /* Update the cache item to tell everyone this cache file is valid. */
1317 ret = update_cache_item(trans, root, inode, path, offset,
1318 io_ctl->entries, io_ctl->bitmaps);
1319out:
1320 if (ret) {
1321 invalidate_inode_pages2(inode->i_mapping);
1322 BTRFS_I(inode)->generation = 0;
1323 if (block_group)
1324 btrfs_debug(root->fs_info,
1325 "failed to write free space cache for block group %llu error %d",
1326 block_group->start, ret);
1327 }
1328 btrfs_update_inode(trans, BTRFS_I(inode));
1329
1330 if (block_group) {
1331 /* the dirty list is protected by the dirty_bgs_lock */
1332 spin_lock(&trans->transaction->dirty_bgs_lock);
1333
1334 /* the disk_cache_state is protected by the block group lock */
1335 spin_lock(&block_group->lock);
1336
1337 /*
1338 * only mark this as written if we didn't get put back on
1339 * the dirty list while waiting for IO. Otherwise our
1340 * cache state won't be right, and we won't get written again
1341 */
1342 if (!ret && list_empty(&block_group->dirty_list))
1343 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1344 else if (ret)
1345 block_group->disk_cache_state = BTRFS_DC_ERROR;
1346
1347 spin_unlock(&block_group->lock);
1348 spin_unlock(&trans->transaction->dirty_bgs_lock);
1349 io_ctl->inode = NULL;
1350 iput(inode);
1351 }
1352
1353 return ret;
1354
1355}
1356
1357int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1358 struct btrfs_block_group *block_group,
1359 struct btrfs_path *path)
1360{
1361 return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1362 block_group, &block_group->io_ctl,
1363 path, block_group->start);
1364}
1365
1366/*
1367 * Write out cached info to an inode.
1368 *
1369 * @inode: freespace inode we are writing out
1370 * @ctl: free space cache we are going to write out
1371 * @block_group: block_group for this cache if it belongs to a block_group
1372 * @io_ctl: holds context for the io
1373 * @trans: the trans handle
1374 *
1375 * This function writes out a free space cache struct to disk for quick recovery
1376 * on mount. This will return 0 if it was successful in writing the cache out,
1377 * or an errno if it was not.
1378 */
1379static int __btrfs_write_out_cache(struct inode *inode,
1380 struct btrfs_free_space_ctl *ctl,
1381 struct btrfs_block_group *block_group,
1382 struct btrfs_io_ctl *io_ctl,
1383 struct btrfs_trans_handle *trans)
1384{
1385 struct extent_state *cached_state = NULL;
1386 LIST_HEAD(bitmap_list);
1387 int entries = 0;
1388 int bitmaps = 0;
1389 int ret;
1390 int must_iput = 0;
1391 int i_size;
1392
1393 if (!i_size_read(inode))
1394 return -EIO;
1395
1396 WARN_ON(io_ctl->pages);
1397 ret = io_ctl_init(io_ctl, inode, 1);
1398 if (ret)
1399 return ret;
1400
1401 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1402 down_write(&block_group->data_rwsem);
1403 spin_lock(&block_group->lock);
1404 if (block_group->delalloc_bytes) {
1405 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1406 spin_unlock(&block_group->lock);
1407 up_write(&block_group->data_rwsem);
1408 BTRFS_I(inode)->generation = 0;
1409 ret = 0;
1410 must_iput = 1;
1411 goto out;
1412 }
1413 spin_unlock(&block_group->lock);
1414 }
1415
1416 /* Lock all pages first so we can lock the extent safely. */
1417 ret = io_ctl_prepare_pages(io_ctl, false);
1418 if (ret)
1419 goto out_unlock;
1420
1421 lock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1422 &cached_state);
1423
1424 io_ctl_set_generation(io_ctl, trans->transid);
1425
1426 mutex_lock(&ctl->cache_writeout_mutex);
1427 /* Write out the extent entries in the free space cache */
1428 spin_lock(&ctl->tree_lock);
1429 ret = write_cache_extent_entries(io_ctl, ctl,
1430 block_group, &entries, &bitmaps,
1431 &bitmap_list);
1432 if (ret)
1433 goto out_nospc_locked;
1434
1435 /*
1436 * Some spaces that are freed in the current transaction are pinned,
1437 * they will be added into free space cache after the transaction is
1438 * committed, we shouldn't lose them.
1439 *
1440 * If this changes while we are working we'll get added back to
1441 * the dirty list and redo it. No locking needed
1442 */
1443 ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1444 if (ret)
1445 goto out_nospc_locked;
1446
1447 /*
1448 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1449 * locked while doing it because a concurrent trim can be manipulating
1450 * or freeing the bitmap.
1451 */
1452 ret = write_bitmap_entries(io_ctl, &bitmap_list);
1453 spin_unlock(&ctl->tree_lock);
1454 mutex_unlock(&ctl->cache_writeout_mutex);
1455 if (ret)
1456 goto out_nospc;
1457
1458 /* Zero out the rest of the pages just to make sure */
1459 io_ctl_zero_remaining_pages(io_ctl);
1460
1461 /* Everything is written out, now we dirty the pages in the file. */
1462 i_size = i_size_read(inode);
1463 for (int i = 0; i < round_up(i_size, PAGE_SIZE) / PAGE_SIZE; i++) {
1464 u64 dirty_start = i * PAGE_SIZE;
1465 u64 dirty_len = min_t(u64, dirty_start + PAGE_SIZE, i_size) - dirty_start;
1466
1467 ret = btrfs_dirty_folio(BTRFS_I(inode), page_folio(io_ctl->pages[i]),
1468 dirty_start, dirty_len, &cached_state, false);
1469 if (ret < 0)
1470 goto out_nospc;
1471 }
1472
1473 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1474 up_write(&block_group->data_rwsem);
1475 /*
1476 * Release the pages and unlock the extent, we will flush
1477 * them out later
1478 */
1479 io_ctl_drop_pages(io_ctl);
1480 io_ctl_free(io_ctl);
1481
1482 unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1483 &cached_state);
1484
1485 /*
1486 * at this point the pages are under IO and we're happy,
1487 * The caller is responsible for waiting on them and updating
1488 * the cache and the inode
1489 */
1490 io_ctl->entries = entries;
1491 io_ctl->bitmaps = bitmaps;
1492
1493 ret = btrfs_fdatawrite_range(BTRFS_I(inode), 0, (u64)-1);
1494 if (ret)
1495 goto out;
1496
1497 return 0;
1498
1499out_nospc_locked:
1500 cleanup_bitmap_list(&bitmap_list);
1501 spin_unlock(&ctl->tree_lock);
1502 mutex_unlock(&ctl->cache_writeout_mutex);
1503
1504out_nospc:
1505 cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1506
1507out_unlock:
1508 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1509 up_write(&block_group->data_rwsem);
1510
1511out:
1512 io_ctl->inode = NULL;
1513 io_ctl_free(io_ctl);
1514 if (ret) {
1515 invalidate_inode_pages2(inode->i_mapping);
1516 BTRFS_I(inode)->generation = 0;
1517 }
1518 btrfs_update_inode(trans, BTRFS_I(inode));
1519 if (must_iput)
1520 iput(inode);
1521 return ret;
1522}
1523
1524int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1525 struct btrfs_block_group *block_group,
1526 struct btrfs_path *path)
1527{
1528 struct btrfs_fs_info *fs_info = trans->fs_info;
1529 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1530 struct inode *inode;
1531 int ret = 0;
1532
1533 spin_lock(&block_group->lock);
1534 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1535 spin_unlock(&block_group->lock);
1536 return 0;
1537 }
1538 spin_unlock(&block_group->lock);
1539
1540 inode = lookup_free_space_inode(block_group, path);
1541 if (IS_ERR(inode))
1542 return 0;
1543
1544 ret = __btrfs_write_out_cache(inode, ctl, block_group,
1545 &block_group->io_ctl, trans);
1546 if (ret) {
1547 btrfs_debug(fs_info,
1548 "failed to write free space cache for block group %llu error %d",
1549 block_group->start, ret);
1550 spin_lock(&block_group->lock);
1551 block_group->disk_cache_state = BTRFS_DC_ERROR;
1552 spin_unlock(&block_group->lock);
1553
1554 block_group->io_ctl.inode = NULL;
1555 iput(inode);
1556 }
1557
1558 /*
1559 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1560 * to wait for IO and put the inode
1561 */
1562
1563 return ret;
1564}
1565
1566static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1567 u64 offset)
1568{
1569 ASSERT(offset >= bitmap_start);
1570 offset -= bitmap_start;
1571 return (unsigned long)(div_u64(offset, unit));
1572}
1573
1574static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1575{
1576 return (unsigned long)(div_u64(bytes, unit));
1577}
1578
1579static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1580 u64 offset)
1581{
1582 u64 bitmap_start;
1583 u64 bytes_per_bitmap;
1584
1585 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1586 bitmap_start = offset - ctl->start;
1587 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1588 bitmap_start *= bytes_per_bitmap;
1589 bitmap_start += ctl->start;
1590
1591 return bitmap_start;
1592}
1593
1594static int tree_insert_offset(struct btrfs_free_space_ctl *ctl,
1595 struct btrfs_free_cluster *cluster,
1596 struct btrfs_free_space *new_entry)
1597{
1598 struct rb_root *root;
1599 struct rb_node **p;
1600 struct rb_node *parent = NULL;
1601
1602 lockdep_assert_held(&ctl->tree_lock);
1603
1604 if (cluster) {
1605 lockdep_assert_held(&cluster->lock);
1606 root = &cluster->root;
1607 } else {
1608 root = &ctl->free_space_offset;
1609 }
1610
1611 p = &root->rb_node;
1612
1613 while (*p) {
1614 struct btrfs_free_space *info;
1615
1616 parent = *p;
1617 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1618
1619 if (new_entry->offset < info->offset) {
1620 p = &(*p)->rb_left;
1621 } else if (new_entry->offset > info->offset) {
1622 p = &(*p)->rb_right;
1623 } else {
1624 /*
1625 * we could have a bitmap entry and an extent entry
1626 * share the same offset. If this is the case, we want
1627 * the extent entry to always be found first if we do a
1628 * linear search through the tree, since we want to have
1629 * the quickest allocation time, and allocating from an
1630 * extent is faster than allocating from a bitmap. So
1631 * if we're inserting a bitmap and we find an entry at
1632 * this offset, we want to go right, or after this entry
1633 * logically. If we are inserting an extent and we've
1634 * found a bitmap, we want to go left, or before
1635 * logically.
1636 */
1637 if (new_entry->bitmap) {
1638 if (info->bitmap) {
1639 WARN_ON_ONCE(1);
1640 return -EEXIST;
1641 }
1642 p = &(*p)->rb_right;
1643 } else {
1644 if (!info->bitmap) {
1645 WARN_ON_ONCE(1);
1646 return -EEXIST;
1647 }
1648 p = &(*p)->rb_left;
1649 }
1650 }
1651 }
1652
1653 rb_link_node(&new_entry->offset_index, parent, p);
1654 rb_insert_color(&new_entry->offset_index, root);
1655
1656 return 0;
1657}
1658
1659/*
1660 * This is a little subtle. We *only* have ->max_extent_size set if we actually
1661 * searched through the bitmap and figured out the largest ->max_extent_size,
1662 * otherwise it's 0. In the case that it's 0 we don't want to tell the
1663 * allocator the wrong thing, we want to use the actual real max_extent_size
1664 * we've found already if it's larger, or we want to use ->bytes.
1665 *
1666 * This matters because find_free_space() will skip entries who's ->bytes is
1667 * less than the required bytes. So if we didn't search down this bitmap, we
1668 * may pick some previous entry that has a smaller ->max_extent_size than we
1669 * have. For example, assume we have two entries, one that has
1670 * ->max_extent_size set to 4K and ->bytes set to 1M. A second entry hasn't set
1671 * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous. We will
1672 * call into find_free_space(), and return with max_extent_size == 4K, because
1673 * that first bitmap entry had ->max_extent_size set, but the second one did
1674 * not. If instead we returned 8K we'd come in searching for 8K, and find the
1675 * 8K contiguous range.
1676 *
1677 * Consider the other case, we have 2 8K chunks in that second entry and still
1678 * don't have ->max_extent_size set. We'll return 16K, and the next time the
1679 * allocator comes in it'll fully search our second bitmap, and this time it'll
1680 * get an uptodate value of 8K as the maximum chunk size. Then we'll get the
1681 * right allocation the next loop through.
1682 */
1683static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1684{
1685 if (entry->bitmap && entry->max_extent_size)
1686 return entry->max_extent_size;
1687 return entry->bytes;
1688}
1689
1690/*
1691 * We want the largest entry to be leftmost, so this is inverted from what you'd
1692 * normally expect.
1693 */
1694static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1695{
1696 const struct btrfs_free_space *entry, *exist;
1697
1698 entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1699 exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1700 return get_max_extent_size(exist) < get_max_extent_size(entry);
1701}
1702
1703/*
1704 * searches the tree for the given offset.
1705 *
1706 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1707 * want a section that has at least bytes size and comes at or after the given
1708 * offset.
1709 */
1710static struct btrfs_free_space *
1711tree_search_offset(struct btrfs_free_space_ctl *ctl,
1712 u64 offset, int bitmap_only, int fuzzy)
1713{
1714 struct rb_node *n = ctl->free_space_offset.rb_node;
1715 struct btrfs_free_space *entry = NULL, *prev = NULL;
1716
1717 lockdep_assert_held(&ctl->tree_lock);
1718
1719 /* find entry that is closest to the 'offset' */
1720 while (n) {
1721 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1722 prev = entry;
1723
1724 if (offset < entry->offset)
1725 n = n->rb_left;
1726 else if (offset > entry->offset)
1727 n = n->rb_right;
1728 else
1729 break;
1730
1731 entry = NULL;
1732 }
1733
1734 if (bitmap_only) {
1735 if (!entry)
1736 return NULL;
1737 if (entry->bitmap)
1738 return entry;
1739
1740 /*
1741 * bitmap entry and extent entry may share same offset,
1742 * in that case, bitmap entry comes after extent entry.
1743 */
1744 n = rb_next(n);
1745 if (!n)
1746 return NULL;
1747 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1748 if (entry->offset != offset)
1749 return NULL;
1750
1751 WARN_ON(!entry->bitmap);
1752 return entry;
1753 } else if (entry) {
1754 if (entry->bitmap) {
1755 /*
1756 * if previous extent entry covers the offset,
1757 * we should return it instead of the bitmap entry
1758 */
1759 n = rb_prev(&entry->offset_index);
1760 if (n) {
1761 prev = rb_entry(n, struct btrfs_free_space,
1762 offset_index);
1763 if (!prev->bitmap &&
1764 prev->offset + prev->bytes > offset)
1765 entry = prev;
1766 }
1767 }
1768 return entry;
1769 }
1770
1771 if (!prev)
1772 return NULL;
1773
1774 /* find last entry before the 'offset' */
1775 entry = prev;
1776 if (entry->offset > offset) {
1777 n = rb_prev(&entry->offset_index);
1778 if (n) {
1779 entry = rb_entry(n, struct btrfs_free_space,
1780 offset_index);
1781 ASSERT(entry->offset <= offset);
1782 } else {
1783 if (fuzzy)
1784 return entry;
1785 else
1786 return NULL;
1787 }
1788 }
1789
1790 if (entry->bitmap) {
1791 n = rb_prev(&entry->offset_index);
1792 if (n) {
1793 prev = rb_entry(n, struct btrfs_free_space,
1794 offset_index);
1795 if (!prev->bitmap &&
1796 prev->offset + prev->bytes > offset)
1797 return prev;
1798 }
1799 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1800 return entry;
1801 } else if (entry->offset + entry->bytes > offset)
1802 return entry;
1803
1804 if (!fuzzy)
1805 return NULL;
1806
1807 while (1) {
1808 n = rb_next(&entry->offset_index);
1809 if (!n)
1810 return NULL;
1811 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1812 if (entry->bitmap) {
1813 if (entry->offset + BITS_PER_BITMAP *
1814 ctl->unit > offset)
1815 break;
1816 } else {
1817 if (entry->offset + entry->bytes > offset)
1818 break;
1819 }
1820 }
1821 return entry;
1822}
1823
1824static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1825 struct btrfs_free_space *info,
1826 bool update_stat)
1827{
1828 lockdep_assert_held(&ctl->tree_lock);
1829
1830 rb_erase(&info->offset_index, &ctl->free_space_offset);
1831 rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1832 ctl->free_extents--;
1833
1834 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1835 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1836 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1837 }
1838
1839 if (update_stat)
1840 ctl->free_space -= info->bytes;
1841}
1842
1843static int link_free_space(struct btrfs_free_space_ctl *ctl,
1844 struct btrfs_free_space *info)
1845{
1846 int ret = 0;
1847
1848 lockdep_assert_held(&ctl->tree_lock);
1849
1850 ASSERT(info->bytes || info->bitmap);
1851 ret = tree_insert_offset(ctl, NULL, info);
1852 if (ret)
1853 return ret;
1854
1855 rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1856
1857 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1858 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1859 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1860 }
1861
1862 ctl->free_space += info->bytes;
1863 ctl->free_extents++;
1864 return ret;
1865}
1866
1867static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1868 struct btrfs_free_space *info)
1869{
1870 ASSERT(info->bitmap);
1871
1872 /*
1873 * If our entry is empty it's because we're on a cluster and we don't
1874 * want to re-link it into our ctl bytes index.
1875 */
1876 if (RB_EMPTY_NODE(&info->bytes_index))
1877 return;
1878
1879 lockdep_assert_held(&ctl->tree_lock);
1880
1881 rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1882 rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1883}
1884
1885static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1886 struct btrfs_free_space *info,
1887 u64 offset, u64 bytes, bool update_stat)
1888{
1889 unsigned long start, count, end;
1890 int extent_delta = -1;
1891
1892 start = offset_to_bit(info->offset, ctl->unit, offset);
1893 count = bytes_to_bits(bytes, ctl->unit);
1894 end = start + count;
1895 ASSERT(end <= BITS_PER_BITMAP);
1896
1897 bitmap_clear(info->bitmap, start, count);
1898
1899 info->bytes -= bytes;
1900 if (info->max_extent_size > ctl->unit)
1901 info->max_extent_size = 0;
1902
1903 relink_bitmap_entry(ctl, info);
1904
1905 if (start && test_bit(start - 1, info->bitmap))
1906 extent_delta++;
1907
1908 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1909 extent_delta++;
1910
1911 info->bitmap_extents += extent_delta;
1912 if (!btrfs_free_space_trimmed(info)) {
1913 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1914 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1915 }
1916
1917 if (update_stat)
1918 ctl->free_space -= bytes;
1919}
1920
1921static void btrfs_bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1922 struct btrfs_free_space *info, u64 offset,
1923 u64 bytes)
1924{
1925 unsigned long start, count, end;
1926 int extent_delta = 1;
1927
1928 start = offset_to_bit(info->offset, ctl->unit, offset);
1929 count = bytes_to_bits(bytes, ctl->unit);
1930 end = start + count;
1931 ASSERT(end <= BITS_PER_BITMAP);
1932
1933 bitmap_set(info->bitmap, start, count);
1934
1935 /*
1936 * We set some bytes, we have no idea what the max extent size is
1937 * anymore.
1938 */
1939 info->max_extent_size = 0;
1940 info->bytes += bytes;
1941 ctl->free_space += bytes;
1942
1943 relink_bitmap_entry(ctl, info);
1944
1945 if (start && test_bit(start - 1, info->bitmap))
1946 extent_delta--;
1947
1948 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1949 extent_delta--;
1950
1951 info->bitmap_extents += extent_delta;
1952 if (!btrfs_free_space_trimmed(info)) {
1953 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1954 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1955 }
1956}
1957
1958/*
1959 * If we can not find suitable extent, we will use bytes to record
1960 * the size of the max extent.
1961 */
1962static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1963 struct btrfs_free_space *bitmap_info, u64 *offset,
1964 u64 *bytes, bool for_alloc)
1965{
1966 unsigned long found_bits = 0;
1967 unsigned long max_bits = 0;
1968 unsigned long bits, i;
1969 unsigned long next_zero;
1970 unsigned long extent_bits;
1971
1972 /*
1973 * Skip searching the bitmap if we don't have a contiguous section that
1974 * is large enough for this allocation.
1975 */
1976 if (for_alloc &&
1977 bitmap_info->max_extent_size &&
1978 bitmap_info->max_extent_size < *bytes) {
1979 *bytes = bitmap_info->max_extent_size;
1980 return -1;
1981 }
1982
1983 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1984 max_t(u64, *offset, bitmap_info->offset));
1985 bits = bytes_to_bits(*bytes, ctl->unit);
1986
1987 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1988 if (for_alloc && bits == 1) {
1989 found_bits = 1;
1990 break;
1991 }
1992 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1993 BITS_PER_BITMAP, i);
1994 extent_bits = next_zero - i;
1995 if (extent_bits >= bits) {
1996 found_bits = extent_bits;
1997 break;
1998 } else if (extent_bits > max_bits) {
1999 max_bits = extent_bits;
2000 }
2001 i = next_zero;
2002 }
2003
2004 if (found_bits) {
2005 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
2006 *bytes = (u64)(found_bits) * ctl->unit;
2007 return 0;
2008 }
2009
2010 *bytes = (u64)(max_bits) * ctl->unit;
2011 bitmap_info->max_extent_size = *bytes;
2012 relink_bitmap_entry(ctl, bitmap_info);
2013 return -1;
2014}
2015
2016/* Cache the size of the max extent in bytes */
2017static struct btrfs_free_space *
2018find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
2019 unsigned long align, u64 *max_extent_size, bool use_bytes_index)
2020{
2021 struct btrfs_free_space *entry;
2022 struct rb_node *node;
2023 u64 tmp;
2024 u64 align_off;
2025 int ret;
2026
2027 if (!ctl->free_space_offset.rb_node)
2028 goto out;
2029again:
2030 if (use_bytes_index) {
2031 node = rb_first_cached(&ctl->free_space_bytes);
2032 } else {
2033 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
2034 0, 1);
2035 if (!entry)
2036 goto out;
2037 node = &entry->offset_index;
2038 }
2039
2040 for (; node; node = rb_next(node)) {
2041 if (use_bytes_index)
2042 entry = rb_entry(node, struct btrfs_free_space,
2043 bytes_index);
2044 else
2045 entry = rb_entry(node, struct btrfs_free_space,
2046 offset_index);
2047
2048 /*
2049 * If we are using the bytes index then all subsequent entries
2050 * in this tree are going to be < bytes, so simply set the max
2051 * extent size and exit the loop.
2052 *
2053 * If we're using the offset index then we need to keep going
2054 * through the rest of the tree.
2055 */
2056 if (entry->bytes < *bytes) {
2057 *max_extent_size = max(get_max_extent_size(entry),
2058 *max_extent_size);
2059 if (use_bytes_index)
2060 break;
2061 continue;
2062 }
2063
2064 /* make sure the space returned is big enough
2065 * to match our requested alignment
2066 */
2067 if (*bytes >= align) {
2068 tmp = entry->offset - ctl->start + align - 1;
2069 tmp = div64_u64(tmp, align);
2070 tmp = tmp * align + ctl->start;
2071 align_off = tmp - entry->offset;
2072 } else {
2073 align_off = 0;
2074 tmp = entry->offset;
2075 }
2076
2077 /*
2078 * We don't break here if we're using the bytes index because we
2079 * may have another entry that has the correct alignment that is
2080 * the right size, so we don't want to miss that possibility.
2081 * At worst this adds another loop through the logic, but if we
2082 * broke here we could prematurely ENOSPC.
2083 */
2084 if (entry->bytes < *bytes + align_off) {
2085 *max_extent_size = max(get_max_extent_size(entry),
2086 *max_extent_size);
2087 continue;
2088 }
2089
2090 if (entry->bitmap) {
2091 struct rb_node *old_next = rb_next(node);
2092 u64 size = *bytes;
2093
2094 ret = search_bitmap(ctl, entry, &tmp, &size, true);
2095 if (!ret) {
2096 *offset = tmp;
2097 *bytes = size;
2098 return entry;
2099 } else {
2100 *max_extent_size =
2101 max(get_max_extent_size(entry),
2102 *max_extent_size);
2103 }
2104
2105 /*
2106 * The bitmap may have gotten re-arranged in the space
2107 * index here because the max_extent_size may have been
2108 * updated. Start from the beginning again if this
2109 * happened.
2110 */
2111 if (use_bytes_index && old_next != rb_next(node))
2112 goto again;
2113 continue;
2114 }
2115
2116 *offset = tmp;
2117 *bytes = entry->bytes - align_off;
2118 return entry;
2119 }
2120out:
2121 return NULL;
2122}
2123
2124static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2125 struct btrfs_free_space *info, u64 offset)
2126{
2127 info->offset = offset_to_bitmap(ctl, offset);
2128 info->bytes = 0;
2129 info->bitmap_extents = 0;
2130 INIT_LIST_HEAD(&info->list);
2131 link_free_space(ctl, info);
2132 ctl->total_bitmaps++;
2133 recalculate_thresholds(ctl);
2134}
2135
2136static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2137 struct btrfs_free_space *bitmap_info)
2138{
2139 /*
2140 * Normally when this is called, the bitmap is completely empty. However,
2141 * if we are blowing up the free space cache for one reason or another
2142 * via __btrfs_remove_free_space_cache(), then it may not be freed and
2143 * we may leave stats on the table.
2144 */
2145 if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2146 ctl->discardable_extents[BTRFS_STAT_CURR] -=
2147 bitmap_info->bitmap_extents;
2148 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2149
2150 }
2151 unlink_free_space(ctl, bitmap_info, true);
2152 kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2153 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2154 ctl->total_bitmaps--;
2155 recalculate_thresholds(ctl);
2156}
2157
2158static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2159 struct btrfs_free_space *bitmap_info,
2160 u64 *offset, u64 *bytes)
2161{
2162 u64 end;
2163 u64 search_start, search_bytes;
2164 int ret;
2165
2166again:
2167 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2168
2169 /*
2170 * We need to search for bits in this bitmap. We could only cover some
2171 * of the extent in this bitmap thanks to how we add space, so we need
2172 * to search for as much as it as we can and clear that amount, and then
2173 * go searching for the next bit.
2174 */
2175 search_start = *offset;
2176 search_bytes = ctl->unit;
2177 search_bytes = min(search_bytes, end - search_start + 1);
2178 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2179 false);
2180 if (ret < 0 || search_start != *offset)
2181 return -EINVAL;
2182
2183 /* We may have found more bits than what we need */
2184 search_bytes = min(search_bytes, *bytes);
2185
2186 /* Cannot clear past the end of the bitmap */
2187 search_bytes = min(search_bytes, end - search_start + 1);
2188
2189 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2190 *offset += search_bytes;
2191 *bytes -= search_bytes;
2192
2193 if (*bytes) {
2194 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2195 if (!bitmap_info->bytes)
2196 free_bitmap(ctl, bitmap_info);
2197
2198 /*
2199 * no entry after this bitmap, but we still have bytes to
2200 * remove, so something has gone wrong.
2201 */
2202 if (!next)
2203 return -EINVAL;
2204
2205 bitmap_info = rb_entry(next, struct btrfs_free_space,
2206 offset_index);
2207
2208 /*
2209 * if the next entry isn't a bitmap we need to return to let the
2210 * extent stuff do its work.
2211 */
2212 if (!bitmap_info->bitmap)
2213 return -EAGAIN;
2214
2215 /*
2216 * Ok the next item is a bitmap, but it may not actually hold
2217 * the information for the rest of this free space stuff, so
2218 * look for it, and if we don't find it return so we can try
2219 * everything over again.
2220 */
2221 search_start = *offset;
2222 search_bytes = ctl->unit;
2223 ret = search_bitmap(ctl, bitmap_info, &search_start,
2224 &search_bytes, false);
2225 if (ret < 0 || search_start != *offset)
2226 return -EAGAIN;
2227
2228 goto again;
2229 } else if (!bitmap_info->bytes)
2230 free_bitmap(ctl, bitmap_info);
2231
2232 return 0;
2233}
2234
2235static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2236 struct btrfs_free_space *info, u64 offset,
2237 u64 bytes, enum btrfs_trim_state trim_state)
2238{
2239 u64 bytes_to_set = 0;
2240 u64 end;
2241
2242 /*
2243 * This is a tradeoff to make bitmap trim state minimal. We mark the
2244 * whole bitmap untrimmed if at any point we add untrimmed regions.
2245 */
2246 if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2247 if (btrfs_free_space_trimmed(info)) {
2248 ctl->discardable_extents[BTRFS_STAT_CURR] +=
2249 info->bitmap_extents;
2250 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2251 }
2252 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2253 }
2254
2255 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2256
2257 bytes_to_set = min(end - offset, bytes);
2258
2259 btrfs_bitmap_set_bits(ctl, info, offset, bytes_to_set);
2260
2261 return bytes_to_set;
2262
2263}
2264
2265static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2266 struct btrfs_free_space *info)
2267{
2268 struct btrfs_block_group *block_group = ctl->block_group;
2269 struct btrfs_fs_info *fs_info = block_group->fs_info;
2270 bool forced = false;
2271
2272#ifdef CONFIG_BTRFS_DEBUG
2273 if (btrfs_should_fragment_free_space(block_group))
2274 forced = true;
2275#endif
2276
2277 /* This is a way to reclaim large regions from the bitmaps. */
2278 if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2279 return false;
2280
2281 /*
2282 * If we are below the extents threshold then we can add this as an
2283 * extent, and don't have to deal with the bitmap
2284 */
2285 if (!forced && ctl->free_extents < ctl->extents_thresh) {
2286 /*
2287 * If this block group has some small extents we don't want to
2288 * use up all of our free slots in the cache with them, we want
2289 * to reserve them to larger extents, however if we have plenty
2290 * of cache left then go ahead an dadd them, no sense in adding
2291 * the overhead of a bitmap if we don't have to.
2292 */
2293 if (info->bytes <= fs_info->sectorsize * 8) {
2294 if (ctl->free_extents * 3 <= ctl->extents_thresh)
2295 return false;
2296 } else {
2297 return false;
2298 }
2299 }
2300
2301 /*
2302 * The original block groups from mkfs can be really small, like 8
2303 * megabytes, so don't bother with a bitmap for those entries. However
2304 * some block groups can be smaller than what a bitmap would cover but
2305 * are still large enough that they could overflow the 32k memory limit,
2306 * so allow those block groups to still be allowed to have a bitmap
2307 * entry.
2308 */
2309 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2310 return false;
2311
2312 return true;
2313}
2314
2315static const struct btrfs_free_space_op free_space_op = {
2316 .use_bitmap = use_bitmap,
2317};
2318
2319static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2320 struct btrfs_free_space *info)
2321{
2322 struct btrfs_free_space *bitmap_info;
2323 struct btrfs_block_group *block_group = NULL;
2324 int added = 0;
2325 u64 bytes, offset, bytes_added;
2326 enum btrfs_trim_state trim_state;
2327 int ret;
2328
2329 bytes = info->bytes;
2330 offset = info->offset;
2331 trim_state = info->trim_state;
2332
2333 if (!ctl->op->use_bitmap(ctl, info))
2334 return 0;
2335
2336 if (ctl->op == &free_space_op)
2337 block_group = ctl->block_group;
2338again:
2339 /*
2340 * Since we link bitmaps right into the cluster we need to see if we
2341 * have a cluster here, and if so and it has our bitmap we need to add
2342 * the free space to that bitmap.
2343 */
2344 if (block_group && !list_empty(&block_group->cluster_list)) {
2345 struct btrfs_free_cluster *cluster;
2346 struct rb_node *node;
2347 struct btrfs_free_space *entry;
2348
2349 cluster = list_entry(block_group->cluster_list.next,
2350 struct btrfs_free_cluster,
2351 block_group_list);
2352 spin_lock(&cluster->lock);
2353 node = rb_first(&cluster->root);
2354 if (!node) {
2355 spin_unlock(&cluster->lock);
2356 goto no_cluster_bitmap;
2357 }
2358
2359 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2360 if (!entry->bitmap) {
2361 spin_unlock(&cluster->lock);
2362 goto no_cluster_bitmap;
2363 }
2364
2365 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2366 bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2367 bytes, trim_state);
2368 bytes -= bytes_added;
2369 offset += bytes_added;
2370 }
2371 spin_unlock(&cluster->lock);
2372 if (!bytes) {
2373 ret = 1;
2374 goto out;
2375 }
2376 }
2377
2378no_cluster_bitmap:
2379 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2380 1, 0);
2381 if (!bitmap_info) {
2382 ASSERT(added == 0);
2383 goto new_bitmap;
2384 }
2385
2386 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2387 trim_state);
2388 bytes -= bytes_added;
2389 offset += bytes_added;
2390 added = 0;
2391
2392 if (!bytes) {
2393 ret = 1;
2394 goto out;
2395 } else
2396 goto again;
2397
2398new_bitmap:
2399 if (info && info->bitmap) {
2400 add_new_bitmap(ctl, info, offset);
2401 added = 1;
2402 info = NULL;
2403 goto again;
2404 } else {
2405 spin_unlock(&ctl->tree_lock);
2406
2407 /* no pre-allocated info, allocate a new one */
2408 if (!info) {
2409 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2410 GFP_NOFS);
2411 if (!info) {
2412 spin_lock(&ctl->tree_lock);
2413 ret = -ENOMEM;
2414 goto out;
2415 }
2416 }
2417
2418 /* allocate the bitmap */
2419 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2420 GFP_NOFS);
2421 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2422 spin_lock(&ctl->tree_lock);
2423 if (!info->bitmap) {
2424 ret = -ENOMEM;
2425 goto out;
2426 }
2427 goto again;
2428 }
2429
2430out:
2431 if (info) {
2432 if (info->bitmap)
2433 kmem_cache_free(btrfs_free_space_bitmap_cachep,
2434 info->bitmap);
2435 kmem_cache_free(btrfs_free_space_cachep, info);
2436 }
2437
2438 return ret;
2439}
2440
2441/*
2442 * Free space merging rules:
2443 * 1) Merge trimmed areas together
2444 * 2) Let untrimmed areas coalesce with trimmed areas
2445 * 3) Always pull neighboring regions from bitmaps
2446 *
2447 * The above rules are for when we merge free space based on btrfs_trim_state.
2448 * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2449 * same reason: to promote larger extent regions which makes life easier for
2450 * find_free_extent(). Rule 2 enables coalescing based on the common path
2451 * being returning free space from btrfs_finish_extent_commit(). So when free
2452 * space is trimmed, it will prevent aggregating trimmed new region and
2453 * untrimmed regions in the rb_tree. Rule 3 is purely to obtain larger extents
2454 * and provide find_free_extent() with the largest extents possible hoping for
2455 * the reuse path.
2456 */
2457static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2458 struct btrfs_free_space *info, bool update_stat)
2459{
2460 struct btrfs_free_space *left_info = NULL;
2461 struct btrfs_free_space *right_info;
2462 bool merged = false;
2463 u64 offset = info->offset;
2464 u64 bytes = info->bytes;
2465 const bool is_trimmed = btrfs_free_space_trimmed(info);
2466 struct rb_node *right_prev = NULL;
2467
2468 /*
2469 * first we want to see if there is free space adjacent to the range we
2470 * are adding, if there is remove that struct and add a new one to
2471 * cover the entire range
2472 */
2473 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2474 if (right_info)
2475 right_prev = rb_prev(&right_info->offset_index);
2476
2477 if (right_prev)
2478 left_info = rb_entry(right_prev, struct btrfs_free_space, offset_index);
2479 else if (!right_info)
2480 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2481
2482 /* See try_merge_free_space() comment. */
2483 if (right_info && !right_info->bitmap &&
2484 (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2485 unlink_free_space(ctl, right_info, update_stat);
2486 info->bytes += right_info->bytes;
2487 kmem_cache_free(btrfs_free_space_cachep, right_info);
2488 merged = true;
2489 }
2490
2491 /* See try_merge_free_space() comment. */
2492 if (left_info && !left_info->bitmap &&
2493 left_info->offset + left_info->bytes == offset &&
2494 (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2495 unlink_free_space(ctl, left_info, update_stat);
2496 info->offset = left_info->offset;
2497 info->bytes += left_info->bytes;
2498 kmem_cache_free(btrfs_free_space_cachep, left_info);
2499 merged = true;
2500 }
2501
2502 return merged;
2503}
2504
2505static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2506 struct btrfs_free_space *info,
2507 bool update_stat)
2508{
2509 struct btrfs_free_space *bitmap;
2510 unsigned long i;
2511 unsigned long j;
2512 const u64 end = info->offset + info->bytes;
2513 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2514 u64 bytes;
2515
2516 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2517 if (!bitmap)
2518 return false;
2519
2520 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2521 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2522 if (j == i)
2523 return false;
2524 bytes = (j - i) * ctl->unit;
2525 info->bytes += bytes;
2526
2527 /* See try_merge_free_space() comment. */
2528 if (!btrfs_free_space_trimmed(bitmap))
2529 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2530
2531 bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2532
2533 if (!bitmap->bytes)
2534 free_bitmap(ctl, bitmap);
2535
2536 return true;
2537}
2538
2539static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2540 struct btrfs_free_space *info,
2541 bool update_stat)
2542{
2543 struct btrfs_free_space *bitmap;
2544 u64 bitmap_offset;
2545 unsigned long i;
2546 unsigned long j;
2547 unsigned long prev_j;
2548 u64 bytes;
2549
2550 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2551 /* If we're on a boundary, try the previous logical bitmap. */
2552 if (bitmap_offset == info->offset) {
2553 if (info->offset == 0)
2554 return false;
2555 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2556 }
2557
2558 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2559 if (!bitmap)
2560 return false;
2561
2562 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2563 j = 0;
2564 prev_j = (unsigned long)-1;
2565 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2566 if (j > i)
2567 break;
2568 prev_j = j;
2569 }
2570 if (prev_j == i)
2571 return false;
2572
2573 if (prev_j == (unsigned long)-1)
2574 bytes = (i + 1) * ctl->unit;
2575 else
2576 bytes = (i - prev_j) * ctl->unit;
2577
2578 info->offset -= bytes;
2579 info->bytes += bytes;
2580
2581 /* See try_merge_free_space() comment. */
2582 if (!btrfs_free_space_trimmed(bitmap))
2583 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2584
2585 bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2586
2587 if (!bitmap->bytes)
2588 free_bitmap(ctl, bitmap);
2589
2590 return true;
2591}
2592
2593/*
2594 * We prefer always to allocate from extent entries, both for clustered and
2595 * non-clustered allocation requests. So when attempting to add a new extent
2596 * entry, try to see if there's adjacent free space in bitmap entries, and if
2597 * there is, migrate that space from the bitmaps to the extent.
2598 * Like this we get better chances of satisfying space allocation requests
2599 * because we attempt to satisfy them based on a single cache entry, and never
2600 * on 2 or more entries - even if the entries represent a contiguous free space
2601 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2602 * ends).
2603 */
2604static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2605 struct btrfs_free_space *info,
2606 bool update_stat)
2607{
2608 /*
2609 * Only work with disconnected entries, as we can change their offset,
2610 * and must be extent entries.
2611 */
2612 ASSERT(!info->bitmap);
2613 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2614
2615 if (ctl->total_bitmaps > 0) {
2616 bool stole_end;
2617 bool stole_front = false;
2618
2619 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2620 if (ctl->total_bitmaps > 0)
2621 stole_front = steal_from_bitmap_to_front(ctl, info,
2622 update_stat);
2623
2624 if (stole_end || stole_front)
2625 try_merge_free_space(ctl, info, update_stat);
2626 }
2627}
2628
2629static int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2630 u64 offset, u64 bytes,
2631 enum btrfs_trim_state trim_state)
2632{
2633 struct btrfs_fs_info *fs_info = block_group->fs_info;
2634 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2635 struct btrfs_free_space *info;
2636 int ret = 0;
2637 u64 filter_bytes = bytes;
2638
2639 ASSERT(!btrfs_is_zoned(fs_info));
2640
2641 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2642 if (!info)
2643 return -ENOMEM;
2644
2645 info->offset = offset;
2646 info->bytes = bytes;
2647 info->trim_state = trim_state;
2648 RB_CLEAR_NODE(&info->offset_index);
2649 RB_CLEAR_NODE(&info->bytes_index);
2650
2651 spin_lock(&ctl->tree_lock);
2652
2653 if (try_merge_free_space(ctl, info, true))
2654 goto link;
2655
2656 /*
2657 * There was no extent directly to the left or right of this new
2658 * extent then we know we're going to have to allocate a new extent, so
2659 * before we do that see if we need to drop this into a bitmap
2660 */
2661 ret = insert_into_bitmap(ctl, info);
2662 if (ret < 0) {
2663 goto out;
2664 } else if (ret) {
2665 ret = 0;
2666 goto out;
2667 }
2668link:
2669 /*
2670 * Only steal free space from adjacent bitmaps if we're sure we're not
2671 * going to add the new free space to existing bitmap entries - because
2672 * that would mean unnecessary work that would be reverted. Therefore
2673 * attempt to steal space from bitmaps if we're adding an extent entry.
2674 */
2675 steal_from_bitmap(ctl, info, true);
2676
2677 filter_bytes = max(filter_bytes, info->bytes);
2678
2679 ret = link_free_space(ctl, info);
2680 if (ret)
2681 kmem_cache_free(btrfs_free_space_cachep, info);
2682out:
2683 btrfs_discard_update_discardable(block_group);
2684 spin_unlock(&ctl->tree_lock);
2685
2686 if (ret) {
2687 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2688 ASSERT(ret != -EEXIST);
2689 }
2690
2691 if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2692 btrfs_discard_check_filter(block_group, filter_bytes);
2693 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2694 }
2695
2696 return ret;
2697}
2698
2699static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2700 u64 bytenr, u64 size, bool used)
2701{
2702 struct btrfs_space_info *sinfo = block_group->space_info;
2703 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2704 u64 offset = bytenr - block_group->start;
2705 u64 to_free, to_unusable;
2706 int bg_reclaim_threshold = 0;
2707 bool initial;
2708 u64 reclaimable_unusable;
2709
2710 spin_lock(&block_group->lock);
2711
2712 initial = ((size == block_group->length) && (block_group->alloc_offset == 0));
2713 WARN_ON(!initial && offset + size > block_group->zone_capacity);
2714 if (!initial)
2715 bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2716
2717 if (!used)
2718 to_free = size;
2719 else if (initial)
2720 to_free = block_group->zone_capacity;
2721 else if (offset >= block_group->alloc_offset)
2722 to_free = size;
2723 else if (offset + size <= block_group->alloc_offset)
2724 to_free = 0;
2725 else
2726 to_free = offset + size - block_group->alloc_offset;
2727 to_unusable = size - to_free;
2728
2729 spin_lock(&ctl->tree_lock);
2730 ctl->free_space += to_free;
2731 spin_unlock(&ctl->tree_lock);
2732 /*
2733 * If the block group is read-only, we should account freed space into
2734 * bytes_readonly.
2735 */
2736 if (!block_group->ro) {
2737 block_group->zone_unusable += to_unusable;
2738 WARN_ON(block_group->zone_unusable > block_group->length);
2739 }
2740 if (!used) {
2741 block_group->alloc_offset -= size;
2742 }
2743
2744 reclaimable_unusable = block_group->zone_unusable -
2745 (block_group->length - block_group->zone_capacity);
2746 /* All the region is now unusable. Mark it as unused and reclaim */
2747 if (block_group->zone_unusable == block_group->length) {
2748 btrfs_mark_bg_unused(block_group);
2749 } else if (bg_reclaim_threshold &&
2750 reclaimable_unusable >=
2751 mult_perc(block_group->zone_capacity, bg_reclaim_threshold)) {
2752 btrfs_mark_bg_to_reclaim(block_group);
2753 }
2754
2755 spin_unlock(&block_group->lock);
2756
2757 return 0;
2758}
2759
2760int btrfs_add_free_space(struct btrfs_block_group *block_group,
2761 u64 bytenr, u64 size)
2762{
2763 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2764
2765 if (btrfs_is_zoned(block_group->fs_info))
2766 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2767 true);
2768
2769 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2770 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2771
2772 return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2773}
2774
2775int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2776 u64 bytenr, u64 size)
2777{
2778 if (btrfs_is_zoned(block_group->fs_info))
2779 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2780 false);
2781
2782 return btrfs_add_free_space(block_group, bytenr, size);
2783}
2784
2785/*
2786 * This is a subtle distinction because when adding free space back in general,
2787 * we want it to be added as untrimmed for async. But in the case where we add
2788 * it on loading of a block group, we want to consider it trimmed.
2789 */
2790int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2791 u64 bytenr, u64 size)
2792{
2793 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2794
2795 if (btrfs_is_zoned(block_group->fs_info))
2796 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2797 true);
2798
2799 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2800 btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2801 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2802
2803 return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2804}
2805
2806int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2807 u64 offset, u64 bytes)
2808{
2809 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2810 struct btrfs_free_space *info;
2811 int ret;
2812 bool re_search = false;
2813
2814 if (btrfs_is_zoned(block_group->fs_info)) {
2815 /*
2816 * This can happen with conventional zones when replaying log.
2817 * Since the allocation info of tree-log nodes are not recorded
2818 * to the extent-tree, calculate_alloc_pointer() failed to
2819 * advance the allocation pointer after last allocated tree log
2820 * node blocks.
2821 *
2822 * This function is called from
2823 * btrfs_pin_extent_for_log_replay() when replaying the log.
2824 * Advance the pointer not to overwrite the tree-log nodes.
2825 */
2826 if (block_group->start + block_group->alloc_offset <
2827 offset + bytes) {
2828 block_group->alloc_offset =
2829 offset + bytes - block_group->start;
2830 }
2831 return 0;
2832 }
2833
2834 spin_lock(&ctl->tree_lock);
2835
2836again:
2837 ret = 0;
2838 if (!bytes)
2839 goto out_lock;
2840
2841 info = tree_search_offset(ctl, offset, 0, 0);
2842 if (!info) {
2843 /*
2844 * oops didn't find an extent that matched the space we wanted
2845 * to remove, look for a bitmap instead
2846 */
2847 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2848 1, 0);
2849 if (!info) {
2850 /*
2851 * If we found a partial bit of our free space in a
2852 * bitmap but then couldn't find the other part this may
2853 * be a problem, so WARN about it.
2854 */
2855 WARN_ON(re_search);
2856 goto out_lock;
2857 }
2858 }
2859
2860 re_search = false;
2861 if (!info->bitmap) {
2862 unlink_free_space(ctl, info, true);
2863 if (offset == info->offset) {
2864 u64 to_free = min(bytes, info->bytes);
2865
2866 info->bytes -= to_free;
2867 info->offset += to_free;
2868 if (info->bytes) {
2869 ret = link_free_space(ctl, info);
2870 WARN_ON(ret);
2871 } else {
2872 kmem_cache_free(btrfs_free_space_cachep, info);
2873 }
2874
2875 offset += to_free;
2876 bytes -= to_free;
2877 goto again;
2878 } else {
2879 u64 old_end = info->bytes + info->offset;
2880
2881 info->bytes = offset - info->offset;
2882 ret = link_free_space(ctl, info);
2883 WARN_ON(ret);
2884 if (ret)
2885 goto out_lock;
2886
2887 /* Not enough bytes in this entry to satisfy us */
2888 if (old_end < offset + bytes) {
2889 bytes -= old_end - offset;
2890 offset = old_end;
2891 goto again;
2892 } else if (old_end == offset + bytes) {
2893 /* all done */
2894 goto out_lock;
2895 }
2896 spin_unlock(&ctl->tree_lock);
2897
2898 ret = __btrfs_add_free_space(block_group,
2899 offset + bytes,
2900 old_end - (offset + bytes),
2901 info->trim_state);
2902 WARN_ON(ret);
2903 goto out;
2904 }
2905 }
2906
2907 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2908 if (ret == -EAGAIN) {
2909 re_search = true;
2910 goto again;
2911 }
2912out_lock:
2913 btrfs_discard_update_discardable(block_group);
2914 spin_unlock(&ctl->tree_lock);
2915out:
2916 return ret;
2917}
2918
2919void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2920 u64 bytes)
2921{
2922 struct btrfs_fs_info *fs_info = block_group->fs_info;
2923 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2924 struct btrfs_free_space *info;
2925 struct rb_node *n;
2926 int count = 0;
2927
2928 /*
2929 * Zoned btrfs does not use free space tree and cluster. Just print
2930 * out the free space after the allocation offset.
2931 */
2932 if (btrfs_is_zoned(fs_info)) {
2933 btrfs_info(fs_info, "free space %llu active %d",
2934 block_group->zone_capacity - block_group->alloc_offset,
2935 test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2936 &block_group->runtime_flags));
2937 return;
2938 }
2939
2940 spin_lock(&ctl->tree_lock);
2941 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2942 info = rb_entry(n, struct btrfs_free_space, offset_index);
2943 if (info->bytes >= bytes && !block_group->ro)
2944 count++;
2945 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2946 info->offset, info->bytes, str_yes_no(info->bitmap));
2947 }
2948 spin_unlock(&ctl->tree_lock);
2949 btrfs_info(fs_info, "block group has cluster?: %s",
2950 str_no_yes(list_empty(&block_group->cluster_list)));
2951 btrfs_info(fs_info,
2952 "%d free space entries at or bigger than %llu bytes",
2953 count, bytes);
2954}
2955
2956void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2957 struct btrfs_free_space_ctl *ctl)
2958{
2959 struct btrfs_fs_info *fs_info = block_group->fs_info;
2960
2961 spin_lock_init(&ctl->tree_lock);
2962 ctl->unit = fs_info->sectorsize;
2963 ctl->start = block_group->start;
2964 ctl->block_group = block_group;
2965 ctl->op = &free_space_op;
2966 ctl->free_space_bytes = RB_ROOT_CACHED;
2967 INIT_LIST_HEAD(&ctl->trimming_ranges);
2968 mutex_init(&ctl->cache_writeout_mutex);
2969
2970 /*
2971 * we only want to have 32k of ram per block group for keeping
2972 * track of free space, and if we pass 1/2 of that we want to
2973 * start converting things over to using bitmaps
2974 */
2975 ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2976}
2977
2978/*
2979 * for a given cluster, put all of its extents back into the free
2980 * space cache. If the block group passed doesn't match the block group
2981 * pointed to by the cluster, someone else raced in and freed the
2982 * cluster already. In that case, we just return without changing anything
2983 */
2984static void __btrfs_return_cluster_to_free_space(
2985 struct btrfs_block_group *block_group,
2986 struct btrfs_free_cluster *cluster)
2987{
2988 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2989 struct rb_node *node;
2990
2991 lockdep_assert_held(&ctl->tree_lock);
2992
2993 spin_lock(&cluster->lock);
2994 if (cluster->block_group != block_group) {
2995 spin_unlock(&cluster->lock);
2996 return;
2997 }
2998
2999 cluster->block_group = NULL;
3000 cluster->window_start = 0;
3001 list_del_init(&cluster->block_group_list);
3002
3003 node = rb_first(&cluster->root);
3004 while (node) {
3005 struct btrfs_free_space *entry;
3006
3007 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3008 node = rb_next(&entry->offset_index);
3009 rb_erase(&entry->offset_index, &cluster->root);
3010 RB_CLEAR_NODE(&entry->offset_index);
3011
3012 if (!entry->bitmap) {
3013 /* Merging treats extents as if they were new */
3014 if (!btrfs_free_space_trimmed(entry)) {
3015 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3016 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
3017 entry->bytes;
3018 }
3019
3020 try_merge_free_space(ctl, entry, false);
3021 steal_from_bitmap(ctl, entry, false);
3022
3023 /* As we insert directly, update these statistics */
3024 if (!btrfs_free_space_trimmed(entry)) {
3025 ctl->discardable_extents[BTRFS_STAT_CURR]++;
3026 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
3027 entry->bytes;
3028 }
3029 }
3030 tree_insert_offset(ctl, NULL, entry);
3031 rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
3032 entry_less);
3033 }
3034 cluster->root = RB_ROOT;
3035 spin_unlock(&cluster->lock);
3036 btrfs_put_block_group(block_group);
3037}
3038
3039void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
3040{
3041 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3042 struct btrfs_free_cluster *cluster;
3043 struct list_head *head;
3044
3045 spin_lock(&ctl->tree_lock);
3046 while ((head = block_group->cluster_list.next) !=
3047 &block_group->cluster_list) {
3048 cluster = list_entry(head, struct btrfs_free_cluster,
3049 block_group_list);
3050
3051 WARN_ON(cluster->block_group != block_group);
3052 __btrfs_return_cluster_to_free_space(block_group, cluster);
3053
3054 cond_resched_lock(&ctl->tree_lock);
3055 }
3056 __btrfs_remove_free_space_cache(ctl);
3057 btrfs_discard_update_discardable(block_group);
3058 spin_unlock(&ctl->tree_lock);
3059
3060}
3061
3062/*
3063 * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3064 */
3065bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3066{
3067 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3068 struct btrfs_free_space *info;
3069 struct rb_node *node;
3070 bool ret = true;
3071
3072 spin_lock(&ctl->tree_lock);
3073 node = rb_first(&ctl->free_space_offset);
3074
3075 while (node) {
3076 info = rb_entry(node, struct btrfs_free_space, offset_index);
3077
3078 if (!btrfs_free_space_trimmed(info)) {
3079 ret = false;
3080 break;
3081 }
3082
3083 node = rb_next(node);
3084 }
3085
3086 spin_unlock(&ctl->tree_lock);
3087 return ret;
3088}
3089
3090u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3091 u64 offset, u64 bytes, u64 empty_size,
3092 u64 *max_extent_size)
3093{
3094 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3095 struct btrfs_discard_ctl *discard_ctl =
3096 &block_group->fs_info->discard_ctl;
3097 struct btrfs_free_space *entry = NULL;
3098 u64 bytes_search = bytes + empty_size;
3099 u64 ret = 0;
3100 u64 align_gap = 0;
3101 u64 align_gap_len = 0;
3102 enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3103 bool use_bytes_index = (offset == block_group->start);
3104
3105 ASSERT(!btrfs_is_zoned(block_group->fs_info));
3106
3107 spin_lock(&ctl->tree_lock);
3108 entry = find_free_space(ctl, &offset, &bytes_search,
3109 block_group->full_stripe_len, max_extent_size,
3110 use_bytes_index);
3111 if (!entry)
3112 goto out;
3113
3114 ret = offset;
3115 if (entry->bitmap) {
3116 bitmap_clear_bits(ctl, entry, offset, bytes, true);
3117
3118 if (!btrfs_free_space_trimmed(entry))
3119 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3120
3121 if (!entry->bytes)
3122 free_bitmap(ctl, entry);
3123 } else {
3124 unlink_free_space(ctl, entry, true);
3125 align_gap_len = offset - entry->offset;
3126 align_gap = entry->offset;
3127 align_gap_trim_state = entry->trim_state;
3128
3129 if (!btrfs_free_space_trimmed(entry))
3130 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3131
3132 entry->offset = offset + bytes;
3133 WARN_ON(entry->bytes < bytes + align_gap_len);
3134
3135 entry->bytes -= bytes + align_gap_len;
3136 if (!entry->bytes)
3137 kmem_cache_free(btrfs_free_space_cachep, entry);
3138 else
3139 link_free_space(ctl, entry);
3140 }
3141out:
3142 btrfs_discard_update_discardable(block_group);
3143 spin_unlock(&ctl->tree_lock);
3144
3145 if (align_gap_len)
3146 __btrfs_add_free_space(block_group, align_gap, align_gap_len,
3147 align_gap_trim_state);
3148 return ret;
3149}
3150
3151/*
3152 * given a cluster, put all of its extents back into the free space
3153 * cache. If a block group is passed, this function will only free
3154 * a cluster that belongs to the passed block group.
3155 *
3156 * Otherwise, it'll get a reference on the block group pointed to by the
3157 * cluster and remove the cluster from it.
3158 */
3159void btrfs_return_cluster_to_free_space(
3160 struct btrfs_block_group *block_group,
3161 struct btrfs_free_cluster *cluster)
3162{
3163 struct btrfs_free_space_ctl *ctl;
3164
3165 /* first, get a safe pointer to the block group */
3166 spin_lock(&cluster->lock);
3167 if (!block_group) {
3168 block_group = cluster->block_group;
3169 if (!block_group) {
3170 spin_unlock(&cluster->lock);
3171 return;
3172 }
3173 } else if (cluster->block_group != block_group) {
3174 /* someone else has already freed it don't redo their work */
3175 spin_unlock(&cluster->lock);
3176 return;
3177 }
3178 btrfs_get_block_group(block_group);
3179 spin_unlock(&cluster->lock);
3180
3181 ctl = block_group->free_space_ctl;
3182
3183 /* now return any extents the cluster had on it */
3184 spin_lock(&ctl->tree_lock);
3185 __btrfs_return_cluster_to_free_space(block_group, cluster);
3186 spin_unlock(&ctl->tree_lock);
3187
3188 btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3189
3190 /* finally drop our ref */
3191 btrfs_put_block_group(block_group);
3192}
3193
3194static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3195 struct btrfs_free_cluster *cluster,
3196 struct btrfs_free_space *entry,
3197 u64 bytes, u64 min_start,
3198 u64 *max_extent_size)
3199{
3200 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3201 int err;
3202 u64 search_start = cluster->window_start;
3203 u64 search_bytes = bytes;
3204 u64 ret = 0;
3205
3206 search_start = min_start;
3207 search_bytes = bytes;
3208
3209 err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3210 if (err) {
3211 *max_extent_size = max(get_max_extent_size(entry),
3212 *max_extent_size);
3213 return 0;
3214 }
3215
3216 ret = search_start;
3217 bitmap_clear_bits(ctl, entry, ret, bytes, false);
3218
3219 return ret;
3220}
3221
3222/*
3223 * given a cluster, try to allocate 'bytes' from it, returns 0
3224 * if it couldn't find anything suitably large, or a logical disk offset
3225 * if things worked out
3226 */
3227u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3228 struct btrfs_free_cluster *cluster, u64 bytes,
3229 u64 min_start, u64 *max_extent_size)
3230{
3231 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3232 struct btrfs_discard_ctl *discard_ctl =
3233 &block_group->fs_info->discard_ctl;
3234 struct btrfs_free_space *entry = NULL;
3235 struct rb_node *node;
3236 u64 ret = 0;
3237
3238 ASSERT(!btrfs_is_zoned(block_group->fs_info));
3239
3240 spin_lock(&cluster->lock);
3241 if (bytes > cluster->max_size)
3242 goto out;
3243
3244 if (cluster->block_group != block_group)
3245 goto out;
3246
3247 node = rb_first(&cluster->root);
3248 if (!node)
3249 goto out;
3250
3251 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3252 while (1) {
3253 if (entry->bytes < bytes)
3254 *max_extent_size = max(get_max_extent_size(entry),
3255 *max_extent_size);
3256
3257 if (entry->bytes < bytes ||
3258 (!entry->bitmap && entry->offset < min_start)) {
3259 node = rb_next(&entry->offset_index);
3260 if (!node)
3261 break;
3262 entry = rb_entry(node, struct btrfs_free_space,
3263 offset_index);
3264 continue;
3265 }
3266
3267 if (entry->bitmap) {
3268 ret = btrfs_alloc_from_bitmap(block_group,
3269 cluster, entry, bytes,
3270 cluster->window_start,
3271 max_extent_size);
3272 if (ret == 0) {
3273 node = rb_next(&entry->offset_index);
3274 if (!node)
3275 break;
3276 entry = rb_entry(node, struct btrfs_free_space,
3277 offset_index);
3278 continue;
3279 }
3280 cluster->window_start += bytes;
3281 } else {
3282 ret = entry->offset;
3283
3284 entry->offset += bytes;
3285 entry->bytes -= bytes;
3286 }
3287
3288 break;
3289 }
3290out:
3291 spin_unlock(&cluster->lock);
3292
3293 if (!ret)
3294 return 0;
3295
3296 spin_lock(&ctl->tree_lock);
3297
3298 if (!btrfs_free_space_trimmed(entry))
3299 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3300
3301 ctl->free_space -= bytes;
3302 if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3303 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3304
3305 spin_lock(&cluster->lock);
3306 if (entry->bytes == 0) {
3307 rb_erase(&entry->offset_index, &cluster->root);
3308 ctl->free_extents--;
3309 if (entry->bitmap) {
3310 kmem_cache_free(btrfs_free_space_bitmap_cachep,
3311 entry->bitmap);
3312 ctl->total_bitmaps--;
3313 recalculate_thresholds(ctl);
3314 } else if (!btrfs_free_space_trimmed(entry)) {
3315 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3316 }
3317 kmem_cache_free(btrfs_free_space_cachep, entry);
3318 }
3319
3320 spin_unlock(&cluster->lock);
3321 spin_unlock(&ctl->tree_lock);
3322
3323 return ret;
3324}
3325
3326static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3327 struct btrfs_free_space *entry,
3328 struct btrfs_free_cluster *cluster,
3329 u64 offset, u64 bytes,
3330 u64 cont1_bytes, u64 min_bytes)
3331{
3332 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3333 unsigned long next_zero;
3334 unsigned long i;
3335 unsigned long want_bits;
3336 unsigned long min_bits;
3337 unsigned long found_bits;
3338 unsigned long max_bits = 0;
3339 unsigned long start = 0;
3340 unsigned long total_found = 0;
3341 int ret;
3342
3343 lockdep_assert_held(&ctl->tree_lock);
3344
3345 i = offset_to_bit(entry->offset, ctl->unit,
3346 max_t(u64, offset, entry->offset));
3347 want_bits = bytes_to_bits(bytes, ctl->unit);
3348 min_bits = bytes_to_bits(min_bytes, ctl->unit);
3349
3350 /*
3351 * Don't bother looking for a cluster in this bitmap if it's heavily
3352 * fragmented.
3353 */
3354 if (entry->max_extent_size &&
3355 entry->max_extent_size < cont1_bytes)
3356 return -ENOSPC;
3357again:
3358 found_bits = 0;
3359 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3360 next_zero = find_next_zero_bit(entry->bitmap,
3361 BITS_PER_BITMAP, i);
3362 if (next_zero - i >= min_bits) {
3363 found_bits = next_zero - i;
3364 if (found_bits > max_bits)
3365 max_bits = found_bits;
3366 break;
3367 }
3368 if (next_zero - i > max_bits)
3369 max_bits = next_zero - i;
3370 i = next_zero;
3371 }
3372
3373 if (!found_bits) {
3374 entry->max_extent_size = (u64)max_bits * ctl->unit;
3375 return -ENOSPC;
3376 }
3377
3378 if (!total_found) {
3379 start = i;
3380 cluster->max_size = 0;
3381 }
3382
3383 total_found += found_bits;
3384
3385 if (cluster->max_size < found_bits * ctl->unit)
3386 cluster->max_size = found_bits * ctl->unit;
3387
3388 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3389 i = next_zero + 1;
3390 goto again;
3391 }
3392
3393 cluster->window_start = start * ctl->unit + entry->offset;
3394 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3395 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3396
3397 /*
3398 * We need to know if we're currently on the normal space index when we
3399 * manipulate the bitmap so that we know we need to remove and re-insert
3400 * it into the space_index tree. Clear the bytes_index node here so the
3401 * bitmap manipulation helpers know not to mess with the space_index
3402 * until this bitmap entry is added back into the normal cache.
3403 */
3404 RB_CLEAR_NODE(&entry->bytes_index);
3405
3406 ret = tree_insert_offset(ctl, cluster, entry);
3407 ASSERT(!ret); /* -EEXIST; Logic error */
3408
3409 trace_btrfs_setup_cluster(block_group, cluster,
3410 total_found * ctl->unit, 1);
3411 return 0;
3412}
3413
3414/*
3415 * This searches the block group for just extents to fill the cluster with.
3416 * Try to find a cluster with at least bytes total bytes, at least one
3417 * extent of cont1_bytes, and other clusters of at least min_bytes.
3418 */
3419static noinline int
3420setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3421 struct btrfs_free_cluster *cluster,
3422 struct list_head *bitmaps, u64 offset, u64 bytes,
3423 u64 cont1_bytes, u64 min_bytes)
3424{
3425 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3426 struct btrfs_free_space *first = NULL;
3427 struct btrfs_free_space *entry = NULL;
3428 struct btrfs_free_space *last;
3429 struct rb_node *node;
3430 u64 window_free;
3431 u64 max_extent;
3432 u64 total_size = 0;
3433
3434 lockdep_assert_held(&ctl->tree_lock);
3435
3436 entry = tree_search_offset(ctl, offset, 0, 1);
3437 if (!entry)
3438 return -ENOSPC;
3439
3440 /*
3441 * We don't want bitmaps, so just move along until we find a normal
3442 * extent entry.
3443 */
3444 while (entry->bitmap || entry->bytes < min_bytes) {
3445 if (entry->bitmap && list_empty(&entry->list))
3446 list_add_tail(&entry->list, bitmaps);
3447 node = rb_next(&entry->offset_index);
3448 if (!node)
3449 return -ENOSPC;
3450 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3451 }
3452
3453 window_free = entry->bytes;
3454 max_extent = entry->bytes;
3455 first = entry;
3456 last = entry;
3457
3458 for (node = rb_next(&entry->offset_index); node;
3459 node = rb_next(&entry->offset_index)) {
3460 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3461
3462 if (entry->bitmap) {
3463 if (list_empty(&entry->list))
3464 list_add_tail(&entry->list, bitmaps);
3465 continue;
3466 }
3467
3468 if (entry->bytes < min_bytes)
3469 continue;
3470
3471 last = entry;
3472 window_free += entry->bytes;
3473 if (entry->bytes > max_extent)
3474 max_extent = entry->bytes;
3475 }
3476
3477 if (window_free < bytes || max_extent < cont1_bytes)
3478 return -ENOSPC;
3479
3480 cluster->window_start = first->offset;
3481
3482 node = &first->offset_index;
3483
3484 /*
3485 * now we've found our entries, pull them out of the free space
3486 * cache and put them into the cluster rbtree
3487 */
3488 do {
3489 int ret;
3490
3491 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3492 node = rb_next(&entry->offset_index);
3493 if (entry->bitmap || entry->bytes < min_bytes)
3494 continue;
3495
3496 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3497 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3498 ret = tree_insert_offset(ctl, cluster, entry);
3499 total_size += entry->bytes;
3500 ASSERT(!ret); /* -EEXIST; Logic error */
3501 } while (node && entry != last);
3502
3503 cluster->max_size = max_extent;
3504 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3505 return 0;
3506}
3507
3508/*
3509 * This specifically looks for bitmaps that may work in the cluster, we assume
3510 * that we have already failed to find extents that will work.
3511 */
3512static noinline int
3513setup_cluster_bitmap(struct btrfs_block_group *block_group,
3514 struct btrfs_free_cluster *cluster,
3515 struct list_head *bitmaps, u64 offset, u64 bytes,
3516 u64 cont1_bytes, u64 min_bytes)
3517{
3518 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3519 struct btrfs_free_space *entry = NULL;
3520 int ret = -ENOSPC;
3521 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3522
3523 if (ctl->total_bitmaps == 0)
3524 return -ENOSPC;
3525
3526 /*
3527 * The bitmap that covers offset won't be in the list unless offset
3528 * is just its start offset.
3529 */
3530 if (!list_empty(bitmaps))
3531 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3532
3533 if (!entry || entry->offset != bitmap_offset) {
3534 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3535 if (entry && list_empty(&entry->list))
3536 list_add(&entry->list, bitmaps);
3537 }
3538
3539 list_for_each_entry(entry, bitmaps, list) {
3540 if (entry->bytes < bytes)
3541 continue;
3542 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3543 bytes, cont1_bytes, min_bytes);
3544 if (!ret)
3545 return 0;
3546 }
3547
3548 /*
3549 * The bitmaps list has all the bitmaps that record free space
3550 * starting after offset, so no more search is required.
3551 */
3552 return -ENOSPC;
3553}
3554
3555/*
3556 * here we try to find a cluster of blocks in a block group. The goal
3557 * is to find at least bytes+empty_size.
3558 * We might not find them all in one contiguous area.
3559 *
3560 * returns zero and sets up cluster if things worked out, otherwise
3561 * it returns -enospc
3562 */
3563int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3564 struct btrfs_free_cluster *cluster,
3565 u64 offset, u64 bytes, u64 empty_size)
3566{
3567 struct btrfs_fs_info *fs_info = block_group->fs_info;
3568 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3569 struct btrfs_free_space *entry, *tmp;
3570 LIST_HEAD(bitmaps);
3571 u64 min_bytes;
3572 u64 cont1_bytes;
3573 int ret;
3574
3575 /*
3576 * Choose the minimum extent size we'll require for this
3577 * cluster. For SSD_SPREAD, don't allow any fragmentation.
3578 * For metadata, allow allocates with smaller extents. For
3579 * data, keep it dense.
3580 */
3581 if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3582 cont1_bytes = bytes + empty_size;
3583 min_bytes = cont1_bytes;
3584 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3585 cont1_bytes = bytes;
3586 min_bytes = fs_info->sectorsize;
3587 } else {
3588 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3589 min_bytes = fs_info->sectorsize;
3590 }
3591
3592 spin_lock(&ctl->tree_lock);
3593
3594 /*
3595 * If we know we don't have enough space to make a cluster don't even
3596 * bother doing all the work to try and find one.
3597 */
3598 if (ctl->free_space < bytes) {
3599 spin_unlock(&ctl->tree_lock);
3600 return -ENOSPC;
3601 }
3602
3603 spin_lock(&cluster->lock);
3604
3605 /* someone already found a cluster, hooray */
3606 if (cluster->block_group) {
3607 ret = 0;
3608 goto out;
3609 }
3610
3611 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3612 min_bytes);
3613
3614 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3615 bytes + empty_size,
3616 cont1_bytes, min_bytes);
3617 if (ret)
3618 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3619 offset, bytes + empty_size,
3620 cont1_bytes, min_bytes);
3621
3622 /* Clear our temporary list */
3623 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3624 list_del_init(&entry->list);
3625
3626 if (!ret) {
3627 btrfs_get_block_group(block_group);
3628 list_add_tail(&cluster->block_group_list,
3629 &block_group->cluster_list);
3630 cluster->block_group = block_group;
3631 } else {
3632 trace_btrfs_failed_cluster_setup(block_group);
3633 }
3634out:
3635 spin_unlock(&cluster->lock);
3636 spin_unlock(&ctl->tree_lock);
3637
3638 return ret;
3639}
3640
3641/*
3642 * simple code to zero out a cluster
3643 */
3644void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3645{
3646 spin_lock_init(&cluster->lock);
3647 spin_lock_init(&cluster->refill_lock);
3648 cluster->root = RB_ROOT;
3649 cluster->max_size = 0;
3650 cluster->fragmented = false;
3651 INIT_LIST_HEAD(&cluster->block_group_list);
3652 cluster->block_group = NULL;
3653}
3654
3655static int do_trimming(struct btrfs_block_group *block_group,
3656 u64 *total_trimmed, u64 start, u64 bytes,
3657 u64 reserved_start, u64 reserved_bytes,
3658 enum btrfs_trim_state reserved_trim_state,
3659 struct btrfs_trim_range *trim_entry)
3660{
3661 struct btrfs_space_info *space_info = block_group->space_info;
3662 struct btrfs_fs_info *fs_info = block_group->fs_info;
3663 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3664 int ret;
3665 int update = 0;
3666 const u64 end = start + bytes;
3667 const u64 reserved_end = reserved_start + reserved_bytes;
3668 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3669 u64 trimmed = 0;
3670
3671 spin_lock(&space_info->lock);
3672 spin_lock(&block_group->lock);
3673 if (!block_group->ro) {
3674 block_group->reserved += reserved_bytes;
3675 space_info->bytes_reserved += reserved_bytes;
3676 update = 1;
3677 }
3678 spin_unlock(&block_group->lock);
3679 spin_unlock(&space_info->lock);
3680
3681 ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3682 if (!ret) {
3683 *total_trimmed += trimmed;
3684 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3685 }
3686
3687 mutex_lock(&ctl->cache_writeout_mutex);
3688 if (reserved_start < start)
3689 __btrfs_add_free_space(block_group, reserved_start,
3690 start - reserved_start,
3691 reserved_trim_state);
3692 if (end < reserved_end)
3693 __btrfs_add_free_space(block_group, end, reserved_end - end,
3694 reserved_trim_state);
3695 __btrfs_add_free_space(block_group, start, bytes, trim_state);
3696 list_del(&trim_entry->list);
3697 mutex_unlock(&ctl->cache_writeout_mutex);
3698
3699 if (update) {
3700 spin_lock(&space_info->lock);
3701 spin_lock(&block_group->lock);
3702 if (block_group->ro)
3703 space_info->bytes_readonly += reserved_bytes;
3704 block_group->reserved -= reserved_bytes;
3705 space_info->bytes_reserved -= reserved_bytes;
3706 spin_unlock(&block_group->lock);
3707 spin_unlock(&space_info->lock);
3708 }
3709
3710 return ret;
3711}
3712
3713/*
3714 * If @async is set, then we will trim 1 region and return.
3715 */
3716static int trim_no_bitmap(struct btrfs_block_group *block_group,
3717 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3718 bool async)
3719{
3720 struct btrfs_discard_ctl *discard_ctl =
3721 &block_group->fs_info->discard_ctl;
3722 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3723 struct btrfs_free_space *entry;
3724 struct rb_node *node;
3725 int ret = 0;
3726 u64 extent_start;
3727 u64 extent_bytes;
3728 enum btrfs_trim_state extent_trim_state;
3729 u64 bytes;
3730 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3731
3732 while (start < end) {
3733 struct btrfs_trim_range trim_entry;
3734
3735 mutex_lock(&ctl->cache_writeout_mutex);
3736 spin_lock(&ctl->tree_lock);
3737
3738 if (ctl->free_space < minlen)
3739 goto out_unlock;
3740
3741 entry = tree_search_offset(ctl, start, 0, 1);
3742 if (!entry)
3743 goto out_unlock;
3744
3745 /* Skip bitmaps and if async, already trimmed entries */
3746 while (entry->bitmap ||
3747 (async && btrfs_free_space_trimmed(entry))) {
3748 node = rb_next(&entry->offset_index);
3749 if (!node)
3750 goto out_unlock;
3751 entry = rb_entry(node, struct btrfs_free_space,
3752 offset_index);
3753 }
3754
3755 if (entry->offset >= end)
3756 goto out_unlock;
3757
3758 extent_start = entry->offset;
3759 extent_bytes = entry->bytes;
3760 extent_trim_state = entry->trim_state;
3761 if (async) {
3762 start = entry->offset;
3763 bytes = entry->bytes;
3764 if (bytes < minlen) {
3765 spin_unlock(&ctl->tree_lock);
3766 mutex_unlock(&ctl->cache_writeout_mutex);
3767 goto next;
3768 }
3769 unlink_free_space(ctl, entry, true);
3770 /*
3771 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3772 * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3773 * X when we come back around. So trim it now.
3774 */
3775 if (max_discard_size &&
3776 bytes >= (max_discard_size +
3777 BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3778 bytes = max_discard_size;
3779 extent_bytes = max_discard_size;
3780 entry->offset += max_discard_size;
3781 entry->bytes -= max_discard_size;
3782 link_free_space(ctl, entry);
3783 } else {
3784 kmem_cache_free(btrfs_free_space_cachep, entry);
3785 }
3786 } else {
3787 start = max(start, extent_start);
3788 bytes = min(extent_start + extent_bytes, end) - start;
3789 if (bytes < minlen) {
3790 spin_unlock(&ctl->tree_lock);
3791 mutex_unlock(&ctl->cache_writeout_mutex);
3792 goto next;
3793 }
3794
3795 unlink_free_space(ctl, entry, true);
3796 kmem_cache_free(btrfs_free_space_cachep, entry);
3797 }
3798
3799 spin_unlock(&ctl->tree_lock);
3800 trim_entry.start = extent_start;
3801 trim_entry.bytes = extent_bytes;
3802 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3803 mutex_unlock(&ctl->cache_writeout_mutex);
3804
3805 ret = do_trimming(block_group, total_trimmed, start, bytes,
3806 extent_start, extent_bytes, extent_trim_state,
3807 &trim_entry);
3808 if (ret) {
3809 block_group->discard_cursor = start + bytes;
3810 break;
3811 }
3812next:
3813 start += bytes;
3814 block_group->discard_cursor = start;
3815 if (async && *total_trimmed)
3816 break;
3817
3818 if (btrfs_trim_interrupted()) {
3819 ret = -ERESTARTSYS;
3820 break;
3821 }
3822
3823 cond_resched();
3824 }
3825
3826 return ret;
3827
3828out_unlock:
3829 block_group->discard_cursor = btrfs_block_group_end(block_group);
3830 spin_unlock(&ctl->tree_lock);
3831 mutex_unlock(&ctl->cache_writeout_mutex);
3832
3833 return ret;
3834}
3835
3836/*
3837 * If we break out of trimming a bitmap prematurely, we should reset the
3838 * trimming bit. In a rather contrieved case, it's possible to race here so
3839 * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3840 *
3841 * start = start of bitmap
3842 * end = near end of bitmap
3843 *
3844 * Thread 1: Thread 2:
3845 * trim_bitmaps(start)
3846 * trim_bitmaps(end)
3847 * end_trimming_bitmap()
3848 * reset_trimming_bitmap()
3849 */
3850static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3851{
3852 struct btrfs_free_space *entry;
3853
3854 spin_lock(&ctl->tree_lock);
3855 entry = tree_search_offset(ctl, offset, 1, 0);
3856 if (entry) {
3857 if (btrfs_free_space_trimmed(entry)) {
3858 ctl->discardable_extents[BTRFS_STAT_CURR] +=
3859 entry->bitmap_extents;
3860 ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3861 }
3862 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3863 }
3864
3865 spin_unlock(&ctl->tree_lock);
3866}
3867
3868static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3869 struct btrfs_free_space *entry)
3870{
3871 if (btrfs_free_space_trimming_bitmap(entry)) {
3872 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3873 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3874 entry->bitmap_extents;
3875 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3876 }
3877}
3878
3879/*
3880 * If @async is set, then we will trim 1 region and return.
3881 */
3882static int trim_bitmaps(struct btrfs_block_group *block_group,
3883 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3884 u64 maxlen, bool async)
3885{
3886 struct btrfs_discard_ctl *discard_ctl =
3887 &block_group->fs_info->discard_ctl;
3888 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3889 struct btrfs_free_space *entry;
3890 int ret = 0;
3891 int ret2;
3892 u64 bytes;
3893 u64 offset = offset_to_bitmap(ctl, start);
3894 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3895
3896 while (offset < end) {
3897 bool next_bitmap = false;
3898 struct btrfs_trim_range trim_entry;
3899
3900 mutex_lock(&ctl->cache_writeout_mutex);
3901 spin_lock(&ctl->tree_lock);
3902
3903 if (ctl->free_space < minlen) {
3904 block_group->discard_cursor =
3905 btrfs_block_group_end(block_group);
3906 spin_unlock(&ctl->tree_lock);
3907 mutex_unlock(&ctl->cache_writeout_mutex);
3908 break;
3909 }
3910
3911 entry = tree_search_offset(ctl, offset, 1, 0);
3912 /*
3913 * Bitmaps are marked trimmed lossily now to prevent constant
3914 * discarding of the same bitmap (the reason why we are bound
3915 * by the filters). So, retrim the block group bitmaps when we
3916 * are preparing to punt to the unused_bgs list. This uses
3917 * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3918 * which is the only discard index which sets minlen to 0.
3919 */
3920 if (!entry || (async && minlen && start == offset &&
3921 btrfs_free_space_trimmed(entry))) {
3922 spin_unlock(&ctl->tree_lock);
3923 mutex_unlock(&ctl->cache_writeout_mutex);
3924 next_bitmap = true;
3925 goto next;
3926 }
3927
3928 /*
3929 * Async discard bitmap trimming begins at by setting the start
3930 * to be key.objectid and the offset_to_bitmap() aligns to the
3931 * start of the bitmap. This lets us know we are fully
3932 * scanning the bitmap rather than only some portion of it.
3933 */
3934 if (start == offset)
3935 entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3936
3937 bytes = minlen;
3938 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3939 if (ret2 || start >= end) {
3940 /*
3941 * We lossily consider a bitmap trimmed if we only skip
3942 * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3943 */
3944 if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3945 end_trimming_bitmap(ctl, entry);
3946 else
3947 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3948 spin_unlock(&ctl->tree_lock);
3949 mutex_unlock(&ctl->cache_writeout_mutex);
3950 next_bitmap = true;
3951 goto next;
3952 }
3953
3954 /*
3955 * We already trimmed a region, but are using the locking above
3956 * to reset the trim_state.
3957 */
3958 if (async && *total_trimmed) {
3959 spin_unlock(&ctl->tree_lock);
3960 mutex_unlock(&ctl->cache_writeout_mutex);
3961 goto out;
3962 }
3963
3964 bytes = min(bytes, end - start);
3965 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3966 spin_unlock(&ctl->tree_lock);
3967 mutex_unlock(&ctl->cache_writeout_mutex);
3968 goto next;
3969 }
3970
3971 /*
3972 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3973 * If X < @minlen, we won't trim X when we come back around.
3974 * So trim it now. We differ here from trimming extents as we
3975 * don't keep individual state per bit.
3976 */
3977 if (async &&
3978 max_discard_size &&
3979 bytes > (max_discard_size + minlen))
3980 bytes = max_discard_size;
3981
3982 bitmap_clear_bits(ctl, entry, start, bytes, true);
3983 if (entry->bytes == 0)
3984 free_bitmap(ctl, entry);
3985
3986 spin_unlock(&ctl->tree_lock);
3987 trim_entry.start = start;
3988 trim_entry.bytes = bytes;
3989 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3990 mutex_unlock(&ctl->cache_writeout_mutex);
3991
3992 ret = do_trimming(block_group, total_trimmed, start, bytes,
3993 start, bytes, 0, &trim_entry);
3994 if (ret) {
3995 reset_trimming_bitmap(ctl, offset);
3996 block_group->discard_cursor =
3997 btrfs_block_group_end(block_group);
3998 break;
3999 }
4000next:
4001 if (next_bitmap) {
4002 offset += BITS_PER_BITMAP * ctl->unit;
4003 start = offset;
4004 } else {
4005 start += bytes;
4006 }
4007 block_group->discard_cursor = start;
4008
4009 if (btrfs_trim_interrupted()) {
4010 if (start != offset)
4011 reset_trimming_bitmap(ctl, offset);
4012 ret = -ERESTARTSYS;
4013 break;
4014 }
4015
4016 cond_resched();
4017 }
4018
4019 if (offset >= end)
4020 block_group->discard_cursor = end;
4021
4022out:
4023 return ret;
4024}
4025
4026int btrfs_trim_block_group(struct btrfs_block_group *block_group,
4027 u64 *trimmed, u64 start, u64 end, u64 minlen)
4028{
4029 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
4030 int ret;
4031 u64 rem = 0;
4032
4033 ASSERT(!btrfs_is_zoned(block_group->fs_info));
4034
4035 *trimmed = 0;
4036
4037 spin_lock(&block_group->lock);
4038 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4039 spin_unlock(&block_group->lock);
4040 return 0;
4041 }
4042 btrfs_freeze_block_group(block_group);
4043 spin_unlock(&block_group->lock);
4044
4045 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4046 if (ret)
4047 goto out;
4048
4049 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4050 div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4051 /* If we ended in the middle of a bitmap, reset the trimming flag */
4052 if (rem)
4053 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4054out:
4055 btrfs_unfreeze_block_group(block_group);
4056 return ret;
4057}
4058
4059int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4060 u64 *trimmed, u64 start, u64 end, u64 minlen,
4061 bool async)
4062{
4063 int ret;
4064
4065 *trimmed = 0;
4066
4067 spin_lock(&block_group->lock);
4068 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4069 spin_unlock(&block_group->lock);
4070 return 0;
4071 }
4072 btrfs_freeze_block_group(block_group);
4073 spin_unlock(&block_group->lock);
4074
4075 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4076 btrfs_unfreeze_block_group(block_group);
4077
4078 return ret;
4079}
4080
4081int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4082 u64 *trimmed, u64 start, u64 end, u64 minlen,
4083 u64 maxlen, bool async)
4084{
4085 int ret;
4086
4087 *trimmed = 0;
4088
4089 spin_lock(&block_group->lock);
4090 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4091 spin_unlock(&block_group->lock);
4092 return 0;
4093 }
4094 btrfs_freeze_block_group(block_group);
4095 spin_unlock(&block_group->lock);
4096
4097 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4098 async);
4099
4100 btrfs_unfreeze_block_group(block_group);
4101
4102 return ret;
4103}
4104
4105bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4106{
4107 return btrfs_super_cache_generation(fs_info->super_copy);
4108}
4109
4110static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4111 struct btrfs_trans_handle *trans)
4112{
4113 struct btrfs_block_group *block_group;
4114 struct rb_node *node;
4115 int ret = 0;
4116
4117 btrfs_info(fs_info, "cleaning free space cache v1");
4118
4119 node = rb_first_cached(&fs_info->block_group_cache_tree);
4120 while (node) {
4121 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4122 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4123 if (ret)
4124 goto out;
4125 node = rb_next(node);
4126 }
4127out:
4128 return ret;
4129}
4130
4131int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4132{
4133 struct btrfs_trans_handle *trans;
4134 int ret;
4135
4136 /*
4137 * update_super_roots will appropriately set or unset
4138 * super_copy->cache_generation based on SPACE_CACHE and
4139 * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4140 * transaction commit whether we are enabling space cache v1 and don't
4141 * have any other work to do, or are disabling it and removing free
4142 * space inodes.
4143 */
4144 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4145 if (IS_ERR(trans))
4146 return PTR_ERR(trans);
4147
4148 if (!active) {
4149 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4150 ret = cleanup_free_space_cache_v1(fs_info, trans);
4151 if (ret) {
4152 btrfs_abort_transaction(trans, ret);
4153 btrfs_end_transaction(trans);
4154 goto out;
4155 }
4156 }
4157
4158 ret = btrfs_commit_transaction(trans);
4159out:
4160 clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4161
4162 return ret;
4163}
4164
4165int __init btrfs_free_space_init(void)
4166{
4167 btrfs_free_space_cachep = KMEM_CACHE(btrfs_free_space, 0);
4168 if (!btrfs_free_space_cachep)
4169 return -ENOMEM;
4170
4171 btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
4172 PAGE_SIZE, PAGE_SIZE,
4173 0, NULL);
4174 if (!btrfs_free_space_bitmap_cachep) {
4175 kmem_cache_destroy(btrfs_free_space_cachep);
4176 return -ENOMEM;
4177 }
4178
4179 return 0;
4180}
4181
4182void __cold btrfs_free_space_exit(void)
4183{
4184 kmem_cache_destroy(btrfs_free_space_cachep);
4185 kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
4186}
4187
4188#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4189/*
4190 * Use this if you need to make a bitmap or extent entry specifically, it
4191 * doesn't do any of the merging that add_free_space does, this acts a lot like
4192 * how the free space cache loading stuff works, so you can get really weird
4193 * configurations.
4194 */
4195int test_add_free_space_entry(struct btrfs_block_group *cache,
4196 u64 offset, u64 bytes, bool bitmap)
4197{
4198 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4199 struct btrfs_free_space *info = NULL, *bitmap_info;
4200 void *map = NULL;
4201 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4202 u64 bytes_added;
4203 int ret;
4204
4205again:
4206 if (!info) {
4207 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4208 if (!info)
4209 return -ENOMEM;
4210 }
4211
4212 if (!bitmap) {
4213 spin_lock(&ctl->tree_lock);
4214 info->offset = offset;
4215 info->bytes = bytes;
4216 info->max_extent_size = 0;
4217 ret = link_free_space(ctl, info);
4218 spin_unlock(&ctl->tree_lock);
4219 if (ret)
4220 kmem_cache_free(btrfs_free_space_cachep, info);
4221 return ret;
4222 }
4223
4224 if (!map) {
4225 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4226 if (!map) {
4227 kmem_cache_free(btrfs_free_space_cachep, info);
4228 return -ENOMEM;
4229 }
4230 }
4231
4232 spin_lock(&ctl->tree_lock);
4233 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4234 1, 0);
4235 if (!bitmap_info) {
4236 info->bitmap = map;
4237 map = NULL;
4238 add_new_bitmap(ctl, info, offset);
4239 bitmap_info = info;
4240 info = NULL;
4241 }
4242
4243 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4244 trim_state);
4245
4246 bytes -= bytes_added;
4247 offset += bytes_added;
4248 spin_unlock(&ctl->tree_lock);
4249
4250 if (bytes)
4251 goto again;
4252
4253 if (info)
4254 kmem_cache_free(btrfs_free_space_cachep, info);
4255 if (map)
4256 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4257 return 0;
4258}
4259
4260/*
4261 * Checks to see if the given range is in the free space cache. This is really
4262 * just used to check the absence of space, so if there is free space in the
4263 * range at all we will return 1.
4264 */
4265int test_check_exists(struct btrfs_block_group *cache,
4266 u64 offset, u64 bytes)
4267{
4268 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4269 struct btrfs_free_space *info;
4270 int ret = 0;
4271
4272 spin_lock(&ctl->tree_lock);
4273 info = tree_search_offset(ctl, offset, 0, 0);
4274 if (!info) {
4275 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4276 1, 0);
4277 if (!info)
4278 goto out;
4279 }
4280
4281have_info:
4282 if (info->bitmap) {
4283 u64 bit_off, bit_bytes;
4284 struct rb_node *n;
4285 struct btrfs_free_space *tmp;
4286
4287 bit_off = offset;
4288 bit_bytes = ctl->unit;
4289 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4290 if (!ret) {
4291 if (bit_off == offset) {
4292 ret = 1;
4293 goto out;
4294 } else if (bit_off > offset &&
4295 offset + bytes > bit_off) {
4296 ret = 1;
4297 goto out;
4298 }
4299 }
4300
4301 n = rb_prev(&info->offset_index);
4302 while (n) {
4303 tmp = rb_entry(n, struct btrfs_free_space,
4304 offset_index);
4305 if (tmp->offset + tmp->bytes < offset)
4306 break;
4307 if (offset + bytes < tmp->offset) {
4308 n = rb_prev(&tmp->offset_index);
4309 continue;
4310 }
4311 info = tmp;
4312 goto have_info;
4313 }
4314
4315 n = rb_next(&info->offset_index);
4316 while (n) {
4317 tmp = rb_entry(n, struct btrfs_free_space,
4318 offset_index);
4319 if (offset + bytes < tmp->offset)
4320 break;
4321 if (tmp->offset + tmp->bytes < offset) {
4322 n = rb_next(&tmp->offset_index);
4323 continue;
4324 }
4325 info = tmp;
4326 goto have_info;
4327 }
4328
4329 ret = 0;
4330 goto out;
4331 }
4332
4333 if (info->offset == offset) {
4334 ret = 1;
4335 goto out;
4336 }
4337
4338 if (offset > info->offset && offset < info->offset + info->bytes)
4339 ret = 1;
4340out:
4341 spin_unlock(&ctl->tree_lock);
4342 return ret;
4343}
4344#endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */
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#include "space-info.h"
22#include "delalloc-space.h"
23#include "block-group.h"
24#include "discard.h"
25
26#define BITS_PER_BITMAP (PAGE_SIZE * 8UL)
27#define MAX_CACHE_BYTES_PER_GIG SZ_64K
28#define FORCE_EXTENT_THRESHOLD SZ_1M
29
30struct btrfs_trim_range {
31 u64 start;
32 u64 bytes;
33 struct list_head list;
34};
35
36static int count_bitmap_extents(struct btrfs_free_space_ctl *ctl,
37 struct btrfs_free_space *bitmap_info);
38static int link_free_space(struct btrfs_free_space_ctl *ctl,
39 struct btrfs_free_space *info);
40static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
41 struct btrfs_free_space *info);
42static int btrfs_wait_cache_io_root(struct btrfs_root *root,
43 struct btrfs_trans_handle *trans,
44 struct btrfs_io_ctl *io_ctl,
45 struct btrfs_path *path);
46
47static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
48 struct btrfs_path *path,
49 u64 offset)
50{
51 struct btrfs_fs_info *fs_info = root->fs_info;
52 struct btrfs_key key;
53 struct btrfs_key location;
54 struct btrfs_disk_key disk_key;
55 struct btrfs_free_space_header *header;
56 struct extent_buffer *leaf;
57 struct inode *inode = NULL;
58 unsigned nofs_flag;
59 int ret;
60
61 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
62 key.offset = offset;
63 key.type = 0;
64
65 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
66 if (ret < 0)
67 return ERR_PTR(ret);
68 if (ret > 0) {
69 btrfs_release_path(path);
70 return ERR_PTR(-ENOENT);
71 }
72
73 leaf = path->nodes[0];
74 header = btrfs_item_ptr(leaf, path->slots[0],
75 struct btrfs_free_space_header);
76 btrfs_free_space_key(leaf, header, &disk_key);
77 btrfs_disk_key_to_cpu(&location, &disk_key);
78 btrfs_release_path(path);
79
80 /*
81 * We are often under a trans handle at this point, so we need to make
82 * sure NOFS is set to keep us from deadlocking.
83 */
84 nofs_flag = memalloc_nofs_save();
85 inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
86 btrfs_release_path(path);
87 memalloc_nofs_restore(nofs_flag);
88 if (IS_ERR(inode))
89 return inode;
90
91 mapping_set_gfp_mask(inode->i_mapping,
92 mapping_gfp_constraint(inode->i_mapping,
93 ~(__GFP_FS | __GFP_HIGHMEM)));
94
95 return inode;
96}
97
98struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
99 struct btrfs_path *path)
100{
101 struct btrfs_fs_info *fs_info = block_group->fs_info;
102 struct inode *inode = NULL;
103 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
104
105 spin_lock(&block_group->lock);
106 if (block_group->inode)
107 inode = igrab(block_group->inode);
108 spin_unlock(&block_group->lock);
109 if (inode)
110 return inode;
111
112 inode = __lookup_free_space_inode(fs_info->tree_root, path,
113 block_group->start);
114 if (IS_ERR(inode))
115 return inode;
116
117 spin_lock(&block_group->lock);
118 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
119 btrfs_info(fs_info, "Old style space inode found, converting.");
120 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
121 BTRFS_INODE_NODATACOW;
122 block_group->disk_cache_state = BTRFS_DC_CLEAR;
123 }
124
125 if (!block_group->iref) {
126 block_group->inode = igrab(inode);
127 block_group->iref = 1;
128 }
129 spin_unlock(&block_group->lock);
130
131 return inode;
132}
133
134static int __create_free_space_inode(struct btrfs_root *root,
135 struct btrfs_trans_handle *trans,
136 struct btrfs_path *path,
137 u64 ino, u64 offset)
138{
139 struct btrfs_key key;
140 struct btrfs_disk_key disk_key;
141 struct btrfs_free_space_header *header;
142 struct btrfs_inode_item *inode_item;
143 struct extent_buffer *leaf;
144 u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
145 int ret;
146
147 ret = btrfs_insert_empty_inode(trans, root, path, ino);
148 if (ret)
149 return ret;
150
151 /* We inline crc's for the free disk space cache */
152 if (ino != BTRFS_FREE_INO_OBJECTID)
153 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
154
155 leaf = path->nodes[0];
156 inode_item = btrfs_item_ptr(leaf, path->slots[0],
157 struct btrfs_inode_item);
158 btrfs_item_key(leaf, &disk_key, path->slots[0]);
159 memzero_extent_buffer(leaf, (unsigned long)inode_item,
160 sizeof(*inode_item));
161 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
162 btrfs_set_inode_size(leaf, inode_item, 0);
163 btrfs_set_inode_nbytes(leaf, inode_item, 0);
164 btrfs_set_inode_uid(leaf, inode_item, 0);
165 btrfs_set_inode_gid(leaf, inode_item, 0);
166 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
167 btrfs_set_inode_flags(leaf, inode_item, flags);
168 btrfs_set_inode_nlink(leaf, inode_item, 1);
169 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
170 btrfs_set_inode_block_group(leaf, inode_item, offset);
171 btrfs_mark_buffer_dirty(leaf);
172 btrfs_release_path(path);
173
174 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
175 key.offset = offset;
176 key.type = 0;
177 ret = btrfs_insert_empty_item(trans, root, path, &key,
178 sizeof(struct btrfs_free_space_header));
179 if (ret < 0) {
180 btrfs_release_path(path);
181 return ret;
182 }
183
184 leaf = path->nodes[0];
185 header = btrfs_item_ptr(leaf, path->slots[0],
186 struct btrfs_free_space_header);
187 memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
188 btrfs_set_free_space_key(leaf, header, &disk_key);
189 btrfs_mark_buffer_dirty(leaf);
190 btrfs_release_path(path);
191
192 return 0;
193}
194
195int create_free_space_inode(struct btrfs_trans_handle *trans,
196 struct btrfs_block_group *block_group,
197 struct btrfs_path *path)
198{
199 int ret;
200 u64 ino;
201
202 ret = btrfs_find_free_objectid(trans->fs_info->tree_root, &ino);
203 if (ret < 0)
204 return ret;
205
206 return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
207 ino, block_group->start);
208}
209
210int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
211 struct btrfs_block_rsv *rsv)
212{
213 u64 needed_bytes;
214 int ret;
215
216 /* 1 for slack space, 1 for updating the inode */
217 needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
218 btrfs_calc_metadata_size(fs_info, 1);
219
220 spin_lock(&rsv->lock);
221 if (rsv->reserved < needed_bytes)
222 ret = -ENOSPC;
223 else
224 ret = 0;
225 spin_unlock(&rsv->lock);
226 return ret;
227}
228
229int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
230 struct btrfs_block_group *block_group,
231 struct inode *inode)
232{
233 struct btrfs_root *root = BTRFS_I(inode)->root;
234 int ret = 0;
235 bool locked = false;
236
237 if (block_group) {
238 struct btrfs_path *path = btrfs_alloc_path();
239
240 if (!path) {
241 ret = -ENOMEM;
242 goto fail;
243 }
244 locked = true;
245 mutex_lock(&trans->transaction->cache_write_mutex);
246 if (!list_empty(&block_group->io_list)) {
247 list_del_init(&block_group->io_list);
248
249 btrfs_wait_cache_io(trans, block_group, path);
250 btrfs_put_block_group(block_group);
251 }
252
253 /*
254 * now that we've truncated the cache away, its no longer
255 * setup or written
256 */
257 spin_lock(&block_group->lock);
258 block_group->disk_cache_state = BTRFS_DC_CLEAR;
259 spin_unlock(&block_group->lock);
260 btrfs_free_path(path);
261 }
262
263 btrfs_i_size_write(BTRFS_I(inode), 0);
264 truncate_pagecache(inode, 0);
265
266 /*
267 * We skip the throttling logic for free space cache inodes, so we don't
268 * need to check for -EAGAIN.
269 */
270 ret = btrfs_truncate_inode_items(trans, root, inode,
271 0, BTRFS_EXTENT_DATA_KEY);
272 if (ret)
273 goto fail;
274
275 ret = btrfs_update_inode(trans, root, inode);
276
277fail:
278 if (locked)
279 mutex_unlock(&trans->transaction->cache_write_mutex);
280 if (ret)
281 btrfs_abort_transaction(trans, ret);
282
283 return ret;
284}
285
286static void readahead_cache(struct inode *inode)
287{
288 struct file_ra_state *ra;
289 unsigned long last_index;
290
291 ra = kzalloc(sizeof(*ra), GFP_NOFS);
292 if (!ra)
293 return;
294
295 file_ra_state_init(ra, inode->i_mapping);
296 last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
297
298 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
299
300 kfree(ra);
301}
302
303static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
304 int write)
305{
306 int num_pages;
307 int check_crcs = 0;
308
309 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
310
311 if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
312 check_crcs = 1;
313
314 /* Make sure we can fit our crcs and generation into the first page */
315 if (write && check_crcs &&
316 (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
317 return -ENOSPC;
318
319 memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
320
321 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
322 if (!io_ctl->pages)
323 return -ENOMEM;
324
325 io_ctl->num_pages = num_pages;
326 io_ctl->fs_info = btrfs_sb(inode->i_sb);
327 io_ctl->check_crcs = check_crcs;
328 io_ctl->inode = inode;
329
330 return 0;
331}
332ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
333
334static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
335{
336 kfree(io_ctl->pages);
337 io_ctl->pages = NULL;
338}
339
340static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
341{
342 if (io_ctl->cur) {
343 io_ctl->cur = NULL;
344 io_ctl->orig = NULL;
345 }
346}
347
348static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
349{
350 ASSERT(io_ctl->index < io_ctl->num_pages);
351 io_ctl->page = io_ctl->pages[io_ctl->index++];
352 io_ctl->cur = page_address(io_ctl->page);
353 io_ctl->orig = io_ctl->cur;
354 io_ctl->size = PAGE_SIZE;
355 if (clear)
356 clear_page(io_ctl->cur);
357}
358
359static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
360{
361 int i;
362
363 io_ctl_unmap_page(io_ctl);
364
365 for (i = 0; i < io_ctl->num_pages; i++) {
366 if (io_ctl->pages[i]) {
367 ClearPageChecked(io_ctl->pages[i]);
368 unlock_page(io_ctl->pages[i]);
369 put_page(io_ctl->pages[i]);
370 }
371 }
372}
373
374static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
375{
376 struct page *page;
377 struct inode *inode = io_ctl->inode;
378 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
379 int i;
380
381 for (i = 0; i < io_ctl->num_pages; i++) {
382 page = find_or_create_page(inode->i_mapping, i, mask);
383 if (!page) {
384 io_ctl_drop_pages(io_ctl);
385 return -ENOMEM;
386 }
387 io_ctl->pages[i] = page;
388 if (uptodate && !PageUptodate(page)) {
389 btrfs_readpage(NULL, page);
390 lock_page(page);
391 if (page->mapping != inode->i_mapping) {
392 btrfs_err(BTRFS_I(inode)->root->fs_info,
393 "free space cache page truncated");
394 io_ctl_drop_pages(io_ctl);
395 return -EIO;
396 }
397 if (!PageUptodate(page)) {
398 btrfs_err(BTRFS_I(inode)->root->fs_info,
399 "error reading free space cache");
400 io_ctl_drop_pages(io_ctl);
401 return -EIO;
402 }
403 }
404 }
405
406 for (i = 0; i < io_ctl->num_pages; i++) {
407 clear_page_dirty_for_io(io_ctl->pages[i]);
408 set_page_extent_mapped(io_ctl->pages[i]);
409 }
410
411 return 0;
412}
413
414static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
415{
416 __le64 *val;
417
418 io_ctl_map_page(io_ctl, 1);
419
420 /*
421 * Skip the csum areas. If we don't check crcs then we just have a
422 * 64bit chunk at the front of the first page.
423 */
424 if (io_ctl->check_crcs) {
425 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
426 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
427 } else {
428 io_ctl->cur += sizeof(u64);
429 io_ctl->size -= sizeof(u64) * 2;
430 }
431
432 val = io_ctl->cur;
433 *val = cpu_to_le64(generation);
434 io_ctl->cur += sizeof(u64);
435}
436
437static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
438{
439 __le64 *gen;
440
441 /*
442 * Skip the crc area. If we don't check crcs then we just have a 64bit
443 * chunk at the front of the first page.
444 */
445 if (io_ctl->check_crcs) {
446 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
447 io_ctl->size -= sizeof(u64) +
448 (sizeof(u32) * io_ctl->num_pages);
449 } else {
450 io_ctl->cur += sizeof(u64);
451 io_ctl->size -= sizeof(u64) * 2;
452 }
453
454 gen = io_ctl->cur;
455 if (le64_to_cpu(*gen) != generation) {
456 btrfs_err_rl(io_ctl->fs_info,
457 "space cache generation (%llu) does not match inode (%llu)",
458 *gen, generation);
459 io_ctl_unmap_page(io_ctl);
460 return -EIO;
461 }
462 io_ctl->cur += sizeof(u64);
463 return 0;
464}
465
466static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
467{
468 u32 *tmp;
469 u32 crc = ~(u32)0;
470 unsigned offset = 0;
471
472 if (!io_ctl->check_crcs) {
473 io_ctl_unmap_page(io_ctl);
474 return;
475 }
476
477 if (index == 0)
478 offset = sizeof(u32) * io_ctl->num_pages;
479
480 crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
481 btrfs_crc32c_final(crc, (u8 *)&crc);
482 io_ctl_unmap_page(io_ctl);
483 tmp = page_address(io_ctl->pages[0]);
484 tmp += index;
485 *tmp = crc;
486}
487
488static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
489{
490 u32 *tmp, val;
491 u32 crc = ~(u32)0;
492 unsigned offset = 0;
493
494 if (!io_ctl->check_crcs) {
495 io_ctl_map_page(io_ctl, 0);
496 return 0;
497 }
498
499 if (index == 0)
500 offset = sizeof(u32) * io_ctl->num_pages;
501
502 tmp = page_address(io_ctl->pages[0]);
503 tmp += index;
504 val = *tmp;
505
506 io_ctl_map_page(io_ctl, 0);
507 crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
508 btrfs_crc32c_final(crc, (u8 *)&crc);
509 if (val != crc) {
510 btrfs_err_rl(io_ctl->fs_info,
511 "csum mismatch on free space cache");
512 io_ctl_unmap_page(io_ctl);
513 return -EIO;
514 }
515
516 return 0;
517}
518
519static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
520 void *bitmap)
521{
522 struct btrfs_free_space_entry *entry;
523
524 if (!io_ctl->cur)
525 return -ENOSPC;
526
527 entry = io_ctl->cur;
528 entry->offset = cpu_to_le64(offset);
529 entry->bytes = cpu_to_le64(bytes);
530 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
531 BTRFS_FREE_SPACE_EXTENT;
532 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
533 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
534
535 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
536 return 0;
537
538 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
539
540 /* No more pages to map */
541 if (io_ctl->index >= io_ctl->num_pages)
542 return 0;
543
544 /* map the next page */
545 io_ctl_map_page(io_ctl, 1);
546 return 0;
547}
548
549static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
550{
551 if (!io_ctl->cur)
552 return -ENOSPC;
553
554 /*
555 * If we aren't at the start of the current page, unmap this one and
556 * map the next one if there is any left.
557 */
558 if (io_ctl->cur != io_ctl->orig) {
559 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
560 if (io_ctl->index >= io_ctl->num_pages)
561 return -ENOSPC;
562 io_ctl_map_page(io_ctl, 0);
563 }
564
565 copy_page(io_ctl->cur, bitmap);
566 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
567 if (io_ctl->index < io_ctl->num_pages)
568 io_ctl_map_page(io_ctl, 0);
569 return 0;
570}
571
572static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
573{
574 /*
575 * If we're not on the boundary we know we've modified the page and we
576 * need to crc the page.
577 */
578 if (io_ctl->cur != io_ctl->orig)
579 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
580 else
581 io_ctl_unmap_page(io_ctl);
582
583 while (io_ctl->index < io_ctl->num_pages) {
584 io_ctl_map_page(io_ctl, 1);
585 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
586 }
587}
588
589static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
590 struct btrfs_free_space *entry, u8 *type)
591{
592 struct btrfs_free_space_entry *e;
593 int ret;
594
595 if (!io_ctl->cur) {
596 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
597 if (ret)
598 return ret;
599 }
600
601 e = io_ctl->cur;
602 entry->offset = le64_to_cpu(e->offset);
603 entry->bytes = le64_to_cpu(e->bytes);
604 *type = e->type;
605 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
606 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
607
608 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
609 return 0;
610
611 io_ctl_unmap_page(io_ctl);
612
613 return 0;
614}
615
616static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
617 struct btrfs_free_space *entry)
618{
619 int ret;
620
621 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
622 if (ret)
623 return ret;
624
625 copy_page(entry->bitmap, io_ctl->cur);
626 io_ctl_unmap_page(io_ctl);
627
628 return 0;
629}
630
631/*
632 * Since we attach pinned extents after the fact we can have contiguous sections
633 * of free space that are split up in entries. This poses a problem with the
634 * tree logging stuff since it could have allocated across what appears to be 2
635 * entries since we would have merged the entries when adding the pinned extents
636 * back to the free space cache. So run through the space cache that we just
637 * loaded and merge contiguous entries. This will make the log replay stuff not
638 * blow up and it will make for nicer allocator behavior.
639 */
640static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
641{
642 struct btrfs_free_space *e, *prev = NULL;
643 struct rb_node *n;
644
645again:
646 spin_lock(&ctl->tree_lock);
647 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
648 e = rb_entry(n, struct btrfs_free_space, offset_index);
649 if (!prev)
650 goto next;
651 if (e->bitmap || prev->bitmap)
652 goto next;
653 if (prev->offset + prev->bytes == e->offset) {
654 unlink_free_space(ctl, prev);
655 unlink_free_space(ctl, e);
656 prev->bytes += e->bytes;
657 kmem_cache_free(btrfs_free_space_cachep, e);
658 link_free_space(ctl, prev);
659 prev = NULL;
660 spin_unlock(&ctl->tree_lock);
661 goto again;
662 }
663next:
664 prev = e;
665 }
666 spin_unlock(&ctl->tree_lock);
667}
668
669static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
670 struct btrfs_free_space_ctl *ctl,
671 struct btrfs_path *path, u64 offset)
672{
673 struct btrfs_fs_info *fs_info = root->fs_info;
674 struct btrfs_free_space_header *header;
675 struct extent_buffer *leaf;
676 struct btrfs_io_ctl io_ctl;
677 struct btrfs_key key;
678 struct btrfs_free_space *e, *n;
679 LIST_HEAD(bitmaps);
680 u64 num_entries;
681 u64 num_bitmaps;
682 u64 generation;
683 u8 type;
684 int ret = 0;
685
686 /* Nothing in the space cache, goodbye */
687 if (!i_size_read(inode))
688 return 0;
689
690 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
691 key.offset = offset;
692 key.type = 0;
693
694 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
695 if (ret < 0)
696 return 0;
697 else if (ret > 0) {
698 btrfs_release_path(path);
699 return 0;
700 }
701
702 ret = -1;
703
704 leaf = path->nodes[0];
705 header = btrfs_item_ptr(leaf, path->slots[0],
706 struct btrfs_free_space_header);
707 num_entries = btrfs_free_space_entries(leaf, header);
708 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
709 generation = btrfs_free_space_generation(leaf, header);
710 btrfs_release_path(path);
711
712 if (!BTRFS_I(inode)->generation) {
713 btrfs_info(fs_info,
714 "the free space cache file (%llu) is invalid, skip it",
715 offset);
716 return 0;
717 }
718
719 if (BTRFS_I(inode)->generation != generation) {
720 btrfs_err(fs_info,
721 "free space inode generation (%llu) did not match free space cache generation (%llu)",
722 BTRFS_I(inode)->generation, generation);
723 return 0;
724 }
725
726 if (!num_entries)
727 return 0;
728
729 ret = io_ctl_init(&io_ctl, inode, 0);
730 if (ret)
731 return ret;
732
733 readahead_cache(inode);
734
735 ret = io_ctl_prepare_pages(&io_ctl, true);
736 if (ret)
737 goto out;
738
739 ret = io_ctl_check_crc(&io_ctl, 0);
740 if (ret)
741 goto free_cache;
742
743 ret = io_ctl_check_generation(&io_ctl, generation);
744 if (ret)
745 goto free_cache;
746
747 while (num_entries) {
748 e = kmem_cache_zalloc(btrfs_free_space_cachep,
749 GFP_NOFS);
750 if (!e)
751 goto free_cache;
752
753 ret = io_ctl_read_entry(&io_ctl, e, &type);
754 if (ret) {
755 kmem_cache_free(btrfs_free_space_cachep, e);
756 goto free_cache;
757 }
758
759 /*
760 * Sync discard ensures that the free space cache is always
761 * trimmed. So when reading this in, the state should reflect
762 * that. We also do this for async as a stop gap for lack of
763 * persistence.
764 */
765 if (btrfs_test_opt(fs_info, DISCARD_SYNC) ||
766 btrfs_test_opt(fs_info, DISCARD_ASYNC))
767 e->trim_state = BTRFS_TRIM_STATE_TRIMMED;
768
769 if (!e->bytes) {
770 kmem_cache_free(btrfs_free_space_cachep, e);
771 goto free_cache;
772 }
773
774 if (type == BTRFS_FREE_SPACE_EXTENT) {
775 spin_lock(&ctl->tree_lock);
776 ret = link_free_space(ctl, e);
777 spin_unlock(&ctl->tree_lock);
778 if (ret) {
779 btrfs_err(fs_info,
780 "Duplicate entries in free space cache, dumping");
781 kmem_cache_free(btrfs_free_space_cachep, e);
782 goto free_cache;
783 }
784 } else {
785 ASSERT(num_bitmaps);
786 num_bitmaps--;
787 e->bitmap = kmem_cache_zalloc(
788 btrfs_free_space_bitmap_cachep, GFP_NOFS);
789 if (!e->bitmap) {
790 kmem_cache_free(
791 btrfs_free_space_cachep, e);
792 goto free_cache;
793 }
794 spin_lock(&ctl->tree_lock);
795 ret = link_free_space(ctl, e);
796 ctl->total_bitmaps++;
797 ctl->op->recalc_thresholds(ctl);
798 spin_unlock(&ctl->tree_lock);
799 if (ret) {
800 btrfs_err(fs_info,
801 "Duplicate entries in free space cache, dumping");
802 kmem_cache_free(btrfs_free_space_cachep, e);
803 goto free_cache;
804 }
805 list_add_tail(&e->list, &bitmaps);
806 }
807
808 num_entries--;
809 }
810
811 io_ctl_unmap_page(&io_ctl);
812
813 /*
814 * We add the bitmaps at the end of the entries in order that
815 * the bitmap entries are added to the cache.
816 */
817 list_for_each_entry_safe(e, n, &bitmaps, list) {
818 list_del_init(&e->list);
819 ret = io_ctl_read_bitmap(&io_ctl, e);
820 if (ret)
821 goto free_cache;
822 e->bitmap_extents = count_bitmap_extents(ctl, e);
823 if (!btrfs_free_space_trimmed(e)) {
824 ctl->discardable_extents[BTRFS_STAT_CURR] +=
825 e->bitmap_extents;
826 ctl->discardable_bytes[BTRFS_STAT_CURR] += e->bytes;
827 }
828 }
829
830 io_ctl_drop_pages(&io_ctl);
831 merge_space_tree(ctl);
832 ret = 1;
833out:
834 btrfs_discard_update_discardable(ctl->private, ctl);
835 io_ctl_free(&io_ctl);
836 return ret;
837free_cache:
838 io_ctl_drop_pages(&io_ctl);
839 __btrfs_remove_free_space_cache(ctl);
840 goto out;
841}
842
843int load_free_space_cache(struct btrfs_block_group *block_group)
844{
845 struct btrfs_fs_info *fs_info = block_group->fs_info;
846 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
847 struct inode *inode;
848 struct btrfs_path *path;
849 int ret = 0;
850 bool matched;
851 u64 used = block_group->used;
852
853 /*
854 * If this block group has been marked to be cleared for one reason or
855 * another then we can't trust the on disk cache, so just return.
856 */
857 spin_lock(&block_group->lock);
858 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
859 spin_unlock(&block_group->lock);
860 return 0;
861 }
862 spin_unlock(&block_group->lock);
863
864 path = btrfs_alloc_path();
865 if (!path)
866 return 0;
867 path->search_commit_root = 1;
868 path->skip_locking = 1;
869
870 /*
871 * We must pass a path with search_commit_root set to btrfs_iget in
872 * order to avoid a deadlock when allocating extents for the tree root.
873 *
874 * When we are COWing an extent buffer from the tree root, when looking
875 * for a free extent, at extent-tree.c:find_free_extent(), we can find
876 * block group without its free space cache loaded. When we find one
877 * we must load its space cache which requires reading its free space
878 * cache's inode item from the root tree. If this inode item is located
879 * in the same leaf that we started COWing before, then we end up in
880 * deadlock on the extent buffer (trying to read lock it when we
881 * previously write locked it).
882 *
883 * It's safe to read the inode item using the commit root because
884 * block groups, once loaded, stay in memory forever (until they are
885 * removed) as well as their space caches once loaded. New block groups
886 * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
887 * we will never try to read their inode item while the fs is mounted.
888 */
889 inode = lookup_free_space_inode(block_group, path);
890 if (IS_ERR(inode)) {
891 btrfs_free_path(path);
892 return 0;
893 }
894
895 /* We may have converted the inode and made the cache invalid. */
896 spin_lock(&block_group->lock);
897 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
898 spin_unlock(&block_group->lock);
899 btrfs_free_path(path);
900 goto out;
901 }
902 spin_unlock(&block_group->lock);
903
904 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
905 path, block_group->start);
906 btrfs_free_path(path);
907 if (ret <= 0)
908 goto out;
909
910 spin_lock(&ctl->tree_lock);
911 matched = (ctl->free_space == (block_group->length - used -
912 block_group->bytes_super));
913 spin_unlock(&ctl->tree_lock);
914
915 if (!matched) {
916 __btrfs_remove_free_space_cache(ctl);
917 btrfs_warn(fs_info,
918 "block group %llu has wrong amount of free space",
919 block_group->start);
920 ret = -1;
921 }
922out:
923 if (ret < 0) {
924 /* This cache is bogus, make sure it gets cleared */
925 spin_lock(&block_group->lock);
926 block_group->disk_cache_state = BTRFS_DC_CLEAR;
927 spin_unlock(&block_group->lock);
928 ret = 0;
929
930 btrfs_warn(fs_info,
931 "failed to load free space cache for block group %llu, rebuilding it now",
932 block_group->start);
933 }
934
935 iput(inode);
936 return ret;
937}
938
939static noinline_for_stack
940int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
941 struct btrfs_free_space_ctl *ctl,
942 struct btrfs_block_group *block_group,
943 int *entries, int *bitmaps,
944 struct list_head *bitmap_list)
945{
946 int ret;
947 struct btrfs_free_cluster *cluster = NULL;
948 struct btrfs_free_cluster *cluster_locked = NULL;
949 struct rb_node *node = rb_first(&ctl->free_space_offset);
950 struct btrfs_trim_range *trim_entry;
951
952 /* Get the cluster for this block_group if it exists */
953 if (block_group && !list_empty(&block_group->cluster_list)) {
954 cluster = list_entry(block_group->cluster_list.next,
955 struct btrfs_free_cluster,
956 block_group_list);
957 }
958
959 if (!node && cluster) {
960 cluster_locked = cluster;
961 spin_lock(&cluster_locked->lock);
962 node = rb_first(&cluster->root);
963 cluster = NULL;
964 }
965
966 /* Write out the extent entries */
967 while (node) {
968 struct btrfs_free_space *e;
969
970 e = rb_entry(node, struct btrfs_free_space, offset_index);
971 *entries += 1;
972
973 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
974 e->bitmap);
975 if (ret)
976 goto fail;
977
978 if (e->bitmap) {
979 list_add_tail(&e->list, bitmap_list);
980 *bitmaps += 1;
981 }
982 node = rb_next(node);
983 if (!node && cluster) {
984 node = rb_first(&cluster->root);
985 cluster_locked = cluster;
986 spin_lock(&cluster_locked->lock);
987 cluster = NULL;
988 }
989 }
990 if (cluster_locked) {
991 spin_unlock(&cluster_locked->lock);
992 cluster_locked = NULL;
993 }
994
995 /*
996 * Make sure we don't miss any range that was removed from our rbtree
997 * because trimming is running. Otherwise after a umount+mount (or crash
998 * after committing the transaction) we would leak free space and get
999 * an inconsistent free space cache report from fsck.
1000 */
1001 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1002 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1003 trim_entry->bytes, NULL);
1004 if (ret)
1005 goto fail;
1006 *entries += 1;
1007 }
1008
1009 return 0;
1010fail:
1011 if (cluster_locked)
1012 spin_unlock(&cluster_locked->lock);
1013 return -ENOSPC;
1014}
1015
1016static noinline_for_stack int
1017update_cache_item(struct btrfs_trans_handle *trans,
1018 struct btrfs_root *root,
1019 struct inode *inode,
1020 struct btrfs_path *path, u64 offset,
1021 int entries, int bitmaps)
1022{
1023 struct btrfs_key key;
1024 struct btrfs_free_space_header *header;
1025 struct extent_buffer *leaf;
1026 int ret;
1027
1028 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1029 key.offset = offset;
1030 key.type = 0;
1031
1032 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1033 if (ret < 0) {
1034 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1035 EXTENT_DELALLOC, 0, 0, NULL);
1036 goto fail;
1037 }
1038 leaf = path->nodes[0];
1039 if (ret > 0) {
1040 struct btrfs_key found_key;
1041 ASSERT(path->slots[0]);
1042 path->slots[0]--;
1043 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1044 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1045 found_key.offset != offset) {
1046 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1047 inode->i_size - 1, EXTENT_DELALLOC, 0,
1048 0, NULL);
1049 btrfs_release_path(path);
1050 goto fail;
1051 }
1052 }
1053
1054 BTRFS_I(inode)->generation = trans->transid;
1055 header = btrfs_item_ptr(leaf, path->slots[0],
1056 struct btrfs_free_space_header);
1057 btrfs_set_free_space_entries(leaf, header, entries);
1058 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1059 btrfs_set_free_space_generation(leaf, header, trans->transid);
1060 btrfs_mark_buffer_dirty(leaf);
1061 btrfs_release_path(path);
1062
1063 return 0;
1064
1065fail:
1066 return -1;
1067}
1068
1069static noinline_for_stack int write_pinned_extent_entries(
1070 struct btrfs_trans_handle *trans,
1071 struct btrfs_block_group *block_group,
1072 struct btrfs_io_ctl *io_ctl,
1073 int *entries)
1074{
1075 u64 start, extent_start, extent_end, len;
1076 struct extent_io_tree *unpin = NULL;
1077 int ret;
1078
1079 if (!block_group)
1080 return 0;
1081
1082 /*
1083 * We want to add any pinned extents to our free space cache
1084 * so we don't leak the space
1085 *
1086 * We shouldn't have switched the pinned extents yet so this is the
1087 * right one
1088 */
1089 unpin = &trans->transaction->pinned_extents;
1090
1091 start = block_group->start;
1092
1093 while (start < block_group->start + block_group->length) {
1094 ret = find_first_extent_bit(unpin, start,
1095 &extent_start, &extent_end,
1096 EXTENT_DIRTY, NULL);
1097 if (ret)
1098 return 0;
1099
1100 /* This pinned extent is out of our range */
1101 if (extent_start >= block_group->start + block_group->length)
1102 return 0;
1103
1104 extent_start = max(extent_start, start);
1105 extent_end = min(block_group->start + block_group->length,
1106 extent_end + 1);
1107 len = extent_end - extent_start;
1108
1109 *entries += 1;
1110 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1111 if (ret)
1112 return -ENOSPC;
1113
1114 start = extent_end;
1115 }
1116
1117 return 0;
1118}
1119
1120static noinline_for_stack int
1121write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1122{
1123 struct btrfs_free_space *entry, *next;
1124 int ret;
1125
1126 /* Write out the bitmaps */
1127 list_for_each_entry_safe(entry, next, bitmap_list, list) {
1128 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1129 if (ret)
1130 return -ENOSPC;
1131 list_del_init(&entry->list);
1132 }
1133
1134 return 0;
1135}
1136
1137static int flush_dirty_cache(struct inode *inode)
1138{
1139 int ret;
1140
1141 ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1142 if (ret)
1143 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1144 EXTENT_DELALLOC, 0, 0, NULL);
1145
1146 return ret;
1147}
1148
1149static void noinline_for_stack
1150cleanup_bitmap_list(struct list_head *bitmap_list)
1151{
1152 struct btrfs_free_space *entry, *next;
1153
1154 list_for_each_entry_safe(entry, next, bitmap_list, list)
1155 list_del_init(&entry->list);
1156}
1157
1158static void noinline_for_stack
1159cleanup_write_cache_enospc(struct inode *inode,
1160 struct btrfs_io_ctl *io_ctl,
1161 struct extent_state **cached_state)
1162{
1163 io_ctl_drop_pages(io_ctl);
1164 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1165 i_size_read(inode) - 1, cached_state);
1166}
1167
1168static int __btrfs_wait_cache_io(struct btrfs_root *root,
1169 struct btrfs_trans_handle *trans,
1170 struct btrfs_block_group *block_group,
1171 struct btrfs_io_ctl *io_ctl,
1172 struct btrfs_path *path, u64 offset)
1173{
1174 int ret;
1175 struct inode *inode = io_ctl->inode;
1176
1177 if (!inode)
1178 return 0;
1179
1180 /* Flush the dirty pages in the cache file. */
1181 ret = flush_dirty_cache(inode);
1182 if (ret)
1183 goto out;
1184
1185 /* Update the cache item to tell everyone this cache file is valid. */
1186 ret = update_cache_item(trans, root, inode, path, offset,
1187 io_ctl->entries, io_ctl->bitmaps);
1188out:
1189 if (ret) {
1190 invalidate_inode_pages2(inode->i_mapping);
1191 BTRFS_I(inode)->generation = 0;
1192 if (block_group)
1193 btrfs_debug(root->fs_info,
1194 "failed to write free space cache for block group %llu error %d",
1195 block_group->start, ret);
1196 }
1197 btrfs_update_inode(trans, root, inode);
1198
1199 if (block_group) {
1200 /* the dirty list is protected by the dirty_bgs_lock */
1201 spin_lock(&trans->transaction->dirty_bgs_lock);
1202
1203 /* the disk_cache_state is protected by the block group lock */
1204 spin_lock(&block_group->lock);
1205
1206 /*
1207 * only mark this as written if we didn't get put back on
1208 * the dirty list while waiting for IO. Otherwise our
1209 * cache state won't be right, and we won't get written again
1210 */
1211 if (!ret && list_empty(&block_group->dirty_list))
1212 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1213 else if (ret)
1214 block_group->disk_cache_state = BTRFS_DC_ERROR;
1215
1216 spin_unlock(&block_group->lock);
1217 spin_unlock(&trans->transaction->dirty_bgs_lock);
1218 io_ctl->inode = NULL;
1219 iput(inode);
1220 }
1221
1222 return ret;
1223
1224}
1225
1226static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1227 struct btrfs_trans_handle *trans,
1228 struct btrfs_io_ctl *io_ctl,
1229 struct btrfs_path *path)
1230{
1231 return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1232}
1233
1234int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1235 struct btrfs_block_group *block_group,
1236 struct btrfs_path *path)
1237{
1238 return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1239 block_group, &block_group->io_ctl,
1240 path, block_group->start);
1241}
1242
1243/**
1244 * __btrfs_write_out_cache - write out cached info to an inode
1245 * @root - the root the inode belongs to
1246 * @ctl - the free space cache we are going to write out
1247 * @block_group - the block_group for this cache if it belongs to a block_group
1248 * @trans - the trans handle
1249 *
1250 * This function writes out a free space cache struct to disk for quick recovery
1251 * on mount. This will return 0 if it was successful in writing the cache out,
1252 * or an errno if it was not.
1253 */
1254static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1255 struct btrfs_free_space_ctl *ctl,
1256 struct btrfs_block_group *block_group,
1257 struct btrfs_io_ctl *io_ctl,
1258 struct btrfs_trans_handle *trans)
1259{
1260 struct extent_state *cached_state = NULL;
1261 LIST_HEAD(bitmap_list);
1262 int entries = 0;
1263 int bitmaps = 0;
1264 int ret;
1265 int must_iput = 0;
1266
1267 if (!i_size_read(inode))
1268 return -EIO;
1269
1270 WARN_ON(io_ctl->pages);
1271 ret = io_ctl_init(io_ctl, inode, 1);
1272 if (ret)
1273 return ret;
1274
1275 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1276 down_write(&block_group->data_rwsem);
1277 spin_lock(&block_group->lock);
1278 if (block_group->delalloc_bytes) {
1279 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1280 spin_unlock(&block_group->lock);
1281 up_write(&block_group->data_rwsem);
1282 BTRFS_I(inode)->generation = 0;
1283 ret = 0;
1284 must_iput = 1;
1285 goto out;
1286 }
1287 spin_unlock(&block_group->lock);
1288 }
1289
1290 /* Lock all pages first so we can lock the extent safely. */
1291 ret = io_ctl_prepare_pages(io_ctl, false);
1292 if (ret)
1293 goto out_unlock;
1294
1295 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1296 &cached_state);
1297
1298 io_ctl_set_generation(io_ctl, trans->transid);
1299
1300 mutex_lock(&ctl->cache_writeout_mutex);
1301 /* Write out the extent entries in the free space cache */
1302 spin_lock(&ctl->tree_lock);
1303 ret = write_cache_extent_entries(io_ctl, ctl,
1304 block_group, &entries, &bitmaps,
1305 &bitmap_list);
1306 if (ret)
1307 goto out_nospc_locked;
1308
1309 /*
1310 * Some spaces that are freed in the current transaction are pinned,
1311 * they will be added into free space cache after the transaction is
1312 * committed, we shouldn't lose them.
1313 *
1314 * If this changes while we are working we'll get added back to
1315 * the dirty list and redo it. No locking needed
1316 */
1317 ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1318 if (ret)
1319 goto out_nospc_locked;
1320
1321 /*
1322 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1323 * locked while doing it because a concurrent trim can be manipulating
1324 * or freeing the bitmap.
1325 */
1326 ret = write_bitmap_entries(io_ctl, &bitmap_list);
1327 spin_unlock(&ctl->tree_lock);
1328 mutex_unlock(&ctl->cache_writeout_mutex);
1329 if (ret)
1330 goto out_nospc;
1331
1332 /* Zero out the rest of the pages just to make sure */
1333 io_ctl_zero_remaining_pages(io_ctl);
1334
1335 /* Everything is written out, now we dirty the pages in the file. */
1336 ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1337 io_ctl->num_pages, 0, i_size_read(inode),
1338 &cached_state);
1339 if (ret)
1340 goto out_nospc;
1341
1342 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1343 up_write(&block_group->data_rwsem);
1344 /*
1345 * Release the pages and unlock the extent, we will flush
1346 * them out later
1347 */
1348 io_ctl_drop_pages(io_ctl);
1349 io_ctl_free(io_ctl);
1350
1351 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1352 i_size_read(inode) - 1, &cached_state);
1353
1354 /*
1355 * at this point the pages are under IO and we're happy,
1356 * The caller is responsible for waiting on them and updating the
1357 * the cache and the inode
1358 */
1359 io_ctl->entries = entries;
1360 io_ctl->bitmaps = bitmaps;
1361
1362 ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1363 if (ret)
1364 goto out;
1365
1366 return 0;
1367
1368out_nospc_locked:
1369 cleanup_bitmap_list(&bitmap_list);
1370 spin_unlock(&ctl->tree_lock);
1371 mutex_unlock(&ctl->cache_writeout_mutex);
1372
1373out_nospc:
1374 cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1375
1376out_unlock:
1377 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1378 up_write(&block_group->data_rwsem);
1379
1380out:
1381 io_ctl->inode = NULL;
1382 io_ctl_free(io_ctl);
1383 if (ret) {
1384 invalidate_inode_pages2(inode->i_mapping);
1385 BTRFS_I(inode)->generation = 0;
1386 }
1387 btrfs_update_inode(trans, root, inode);
1388 if (must_iput)
1389 iput(inode);
1390 return ret;
1391}
1392
1393int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1394 struct btrfs_block_group *block_group,
1395 struct btrfs_path *path)
1396{
1397 struct btrfs_fs_info *fs_info = trans->fs_info;
1398 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1399 struct inode *inode;
1400 int ret = 0;
1401
1402 spin_lock(&block_group->lock);
1403 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1404 spin_unlock(&block_group->lock);
1405 return 0;
1406 }
1407 spin_unlock(&block_group->lock);
1408
1409 inode = lookup_free_space_inode(block_group, path);
1410 if (IS_ERR(inode))
1411 return 0;
1412
1413 ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1414 block_group, &block_group->io_ctl, trans);
1415 if (ret) {
1416 btrfs_debug(fs_info,
1417 "failed to write free space cache for block group %llu error %d",
1418 block_group->start, ret);
1419 spin_lock(&block_group->lock);
1420 block_group->disk_cache_state = BTRFS_DC_ERROR;
1421 spin_unlock(&block_group->lock);
1422
1423 block_group->io_ctl.inode = NULL;
1424 iput(inode);
1425 }
1426
1427 /*
1428 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1429 * to wait for IO and put the inode
1430 */
1431
1432 return ret;
1433}
1434
1435static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1436 u64 offset)
1437{
1438 ASSERT(offset >= bitmap_start);
1439 offset -= bitmap_start;
1440 return (unsigned long)(div_u64(offset, unit));
1441}
1442
1443static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1444{
1445 return (unsigned long)(div_u64(bytes, unit));
1446}
1447
1448static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1449 u64 offset)
1450{
1451 u64 bitmap_start;
1452 u64 bytes_per_bitmap;
1453
1454 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1455 bitmap_start = offset - ctl->start;
1456 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1457 bitmap_start *= bytes_per_bitmap;
1458 bitmap_start += ctl->start;
1459
1460 return bitmap_start;
1461}
1462
1463static int tree_insert_offset(struct rb_root *root, u64 offset,
1464 struct rb_node *node, int bitmap)
1465{
1466 struct rb_node **p = &root->rb_node;
1467 struct rb_node *parent = NULL;
1468 struct btrfs_free_space *info;
1469
1470 while (*p) {
1471 parent = *p;
1472 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1473
1474 if (offset < info->offset) {
1475 p = &(*p)->rb_left;
1476 } else if (offset > info->offset) {
1477 p = &(*p)->rb_right;
1478 } else {
1479 /*
1480 * we could have a bitmap entry and an extent entry
1481 * share the same offset. If this is the case, we want
1482 * the extent entry to always be found first if we do a
1483 * linear search through the tree, since we want to have
1484 * the quickest allocation time, and allocating from an
1485 * extent is faster than allocating from a bitmap. So
1486 * if we're inserting a bitmap and we find an entry at
1487 * this offset, we want to go right, or after this entry
1488 * logically. If we are inserting an extent and we've
1489 * found a bitmap, we want to go left, or before
1490 * logically.
1491 */
1492 if (bitmap) {
1493 if (info->bitmap) {
1494 WARN_ON_ONCE(1);
1495 return -EEXIST;
1496 }
1497 p = &(*p)->rb_right;
1498 } else {
1499 if (!info->bitmap) {
1500 WARN_ON_ONCE(1);
1501 return -EEXIST;
1502 }
1503 p = &(*p)->rb_left;
1504 }
1505 }
1506 }
1507
1508 rb_link_node(node, parent, p);
1509 rb_insert_color(node, root);
1510
1511 return 0;
1512}
1513
1514/*
1515 * searches the tree for the given offset.
1516 *
1517 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1518 * want a section that has at least bytes size and comes at or after the given
1519 * offset.
1520 */
1521static struct btrfs_free_space *
1522tree_search_offset(struct btrfs_free_space_ctl *ctl,
1523 u64 offset, int bitmap_only, int fuzzy)
1524{
1525 struct rb_node *n = ctl->free_space_offset.rb_node;
1526 struct btrfs_free_space *entry, *prev = NULL;
1527
1528 /* find entry that is closest to the 'offset' */
1529 while (1) {
1530 if (!n) {
1531 entry = NULL;
1532 break;
1533 }
1534
1535 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1536 prev = entry;
1537
1538 if (offset < entry->offset)
1539 n = n->rb_left;
1540 else if (offset > entry->offset)
1541 n = n->rb_right;
1542 else
1543 break;
1544 }
1545
1546 if (bitmap_only) {
1547 if (!entry)
1548 return NULL;
1549 if (entry->bitmap)
1550 return entry;
1551
1552 /*
1553 * bitmap entry and extent entry may share same offset,
1554 * in that case, bitmap entry comes after extent entry.
1555 */
1556 n = rb_next(n);
1557 if (!n)
1558 return NULL;
1559 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1560 if (entry->offset != offset)
1561 return NULL;
1562
1563 WARN_ON(!entry->bitmap);
1564 return entry;
1565 } else if (entry) {
1566 if (entry->bitmap) {
1567 /*
1568 * if previous extent entry covers the offset,
1569 * we should return it instead of the bitmap entry
1570 */
1571 n = rb_prev(&entry->offset_index);
1572 if (n) {
1573 prev = rb_entry(n, struct btrfs_free_space,
1574 offset_index);
1575 if (!prev->bitmap &&
1576 prev->offset + prev->bytes > offset)
1577 entry = prev;
1578 }
1579 }
1580 return entry;
1581 }
1582
1583 if (!prev)
1584 return NULL;
1585
1586 /* find last entry before the 'offset' */
1587 entry = prev;
1588 if (entry->offset > offset) {
1589 n = rb_prev(&entry->offset_index);
1590 if (n) {
1591 entry = rb_entry(n, struct btrfs_free_space,
1592 offset_index);
1593 ASSERT(entry->offset <= offset);
1594 } else {
1595 if (fuzzy)
1596 return entry;
1597 else
1598 return NULL;
1599 }
1600 }
1601
1602 if (entry->bitmap) {
1603 n = rb_prev(&entry->offset_index);
1604 if (n) {
1605 prev = rb_entry(n, struct btrfs_free_space,
1606 offset_index);
1607 if (!prev->bitmap &&
1608 prev->offset + prev->bytes > offset)
1609 return prev;
1610 }
1611 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1612 return entry;
1613 } else if (entry->offset + entry->bytes > offset)
1614 return entry;
1615
1616 if (!fuzzy)
1617 return NULL;
1618
1619 while (1) {
1620 if (entry->bitmap) {
1621 if (entry->offset + BITS_PER_BITMAP *
1622 ctl->unit > offset)
1623 break;
1624 } else {
1625 if (entry->offset + entry->bytes > offset)
1626 break;
1627 }
1628
1629 n = rb_next(&entry->offset_index);
1630 if (!n)
1631 return NULL;
1632 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1633 }
1634 return entry;
1635}
1636
1637static inline void
1638__unlink_free_space(struct btrfs_free_space_ctl *ctl,
1639 struct btrfs_free_space *info)
1640{
1641 rb_erase(&info->offset_index, &ctl->free_space_offset);
1642 ctl->free_extents--;
1643
1644 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1645 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1646 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1647 }
1648}
1649
1650static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1651 struct btrfs_free_space *info)
1652{
1653 __unlink_free_space(ctl, info);
1654 ctl->free_space -= info->bytes;
1655}
1656
1657static int link_free_space(struct btrfs_free_space_ctl *ctl,
1658 struct btrfs_free_space *info)
1659{
1660 int ret = 0;
1661
1662 ASSERT(info->bytes || info->bitmap);
1663 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1664 &info->offset_index, (info->bitmap != NULL));
1665 if (ret)
1666 return ret;
1667
1668 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1669 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1670 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1671 }
1672
1673 ctl->free_space += info->bytes;
1674 ctl->free_extents++;
1675 return ret;
1676}
1677
1678static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1679{
1680 struct btrfs_block_group *block_group = ctl->private;
1681 u64 max_bytes;
1682 u64 bitmap_bytes;
1683 u64 extent_bytes;
1684 u64 size = block_group->length;
1685 u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1686 u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1687
1688 max_bitmaps = max_t(u64, max_bitmaps, 1);
1689
1690 ASSERT(ctl->total_bitmaps <= max_bitmaps);
1691
1692 /*
1693 * We are trying to keep the total amount of memory used per 1GiB of
1694 * space to be MAX_CACHE_BYTES_PER_GIG. However, with a reclamation
1695 * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
1696 * bitmaps, we may end up using more memory than this.
1697 */
1698 if (size < SZ_1G)
1699 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1700 else
1701 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1702
1703 bitmap_bytes = ctl->total_bitmaps * ctl->unit;
1704
1705 /*
1706 * we want the extent entry threshold to always be at most 1/2 the max
1707 * bytes we can have, or whatever is less than that.
1708 */
1709 extent_bytes = max_bytes - bitmap_bytes;
1710 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1711
1712 ctl->extents_thresh =
1713 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1714}
1715
1716static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1717 struct btrfs_free_space *info,
1718 u64 offset, u64 bytes)
1719{
1720 unsigned long start, count, end;
1721 int extent_delta = -1;
1722
1723 start = offset_to_bit(info->offset, ctl->unit, offset);
1724 count = bytes_to_bits(bytes, ctl->unit);
1725 end = start + count;
1726 ASSERT(end <= BITS_PER_BITMAP);
1727
1728 bitmap_clear(info->bitmap, start, count);
1729
1730 info->bytes -= bytes;
1731 if (info->max_extent_size > ctl->unit)
1732 info->max_extent_size = 0;
1733
1734 if (start && test_bit(start - 1, info->bitmap))
1735 extent_delta++;
1736
1737 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1738 extent_delta++;
1739
1740 info->bitmap_extents += extent_delta;
1741 if (!btrfs_free_space_trimmed(info)) {
1742 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1743 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1744 }
1745}
1746
1747static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1748 struct btrfs_free_space *info, u64 offset,
1749 u64 bytes)
1750{
1751 __bitmap_clear_bits(ctl, info, offset, bytes);
1752 ctl->free_space -= bytes;
1753}
1754
1755static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1756 struct btrfs_free_space *info, u64 offset,
1757 u64 bytes)
1758{
1759 unsigned long start, count, end;
1760 int extent_delta = 1;
1761
1762 start = offset_to_bit(info->offset, ctl->unit, offset);
1763 count = bytes_to_bits(bytes, ctl->unit);
1764 end = start + count;
1765 ASSERT(end <= BITS_PER_BITMAP);
1766
1767 bitmap_set(info->bitmap, start, count);
1768
1769 info->bytes += bytes;
1770 ctl->free_space += bytes;
1771
1772 if (start && test_bit(start - 1, info->bitmap))
1773 extent_delta--;
1774
1775 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1776 extent_delta--;
1777
1778 info->bitmap_extents += extent_delta;
1779 if (!btrfs_free_space_trimmed(info)) {
1780 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1781 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1782 }
1783}
1784
1785/*
1786 * If we can not find suitable extent, we will use bytes to record
1787 * the size of the max extent.
1788 */
1789static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1790 struct btrfs_free_space *bitmap_info, u64 *offset,
1791 u64 *bytes, bool for_alloc)
1792{
1793 unsigned long found_bits = 0;
1794 unsigned long max_bits = 0;
1795 unsigned long bits, i;
1796 unsigned long next_zero;
1797 unsigned long extent_bits;
1798
1799 /*
1800 * Skip searching the bitmap if we don't have a contiguous section that
1801 * is large enough for this allocation.
1802 */
1803 if (for_alloc &&
1804 bitmap_info->max_extent_size &&
1805 bitmap_info->max_extent_size < *bytes) {
1806 *bytes = bitmap_info->max_extent_size;
1807 return -1;
1808 }
1809
1810 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1811 max_t(u64, *offset, bitmap_info->offset));
1812 bits = bytes_to_bits(*bytes, ctl->unit);
1813
1814 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1815 if (for_alloc && bits == 1) {
1816 found_bits = 1;
1817 break;
1818 }
1819 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1820 BITS_PER_BITMAP, i);
1821 extent_bits = next_zero - i;
1822 if (extent_bits >= bits) {
1823 found_bits = extent_bits;
1824 break;
1825 } else if (extent_bits > max_bits) {
1826 max_bits = extent_bits;
1827 }
1828 i = next_zero;
1829 }
1830
1831 if (found_bits) {
1832 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1833 *bytes = (u64)(found_bits) * ctl->unit;
1834 return 0;
1835 }
1836
1837 *bytes = (u64)(max_bits) * ctl->unit;
1838 bitmap_info->max_extent_size = *bytes;
1839 return -1;
1840}
1841
1842static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1843{
1844 if (entry->bitmap)
1845 return entry->max_extent_size;
1846 return entry->bytes;
1847}
1848
1849/* Cache the size of the max extent in bytes */
1850static struct btrfs_free_space *
1851find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1852 unsigned long align, u64 *max_extent_size)
1853{
1854 struct btrfs_free_space *entry;
1855 struct rb_node *node;
1856 u64 tmp;
1857 u64 align_off;
1858 int ret;
1859
1860 if (!ctl->free_space_offset.rb_node)
1861 goto out;
1862
1863 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1864 if (!entry)
1865 goto out;
1866
1867 for (node = &entry->offset_index; node; node = rb_next(node)) {
1868 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1869 if (entry->bytes < *bytes) {
1870 *max_extent_size = max(get_max_extent_size(entry),
1871 *max_extent_size);
1872 continue;
1873 }
1874
1875 /* make sure the space returned is big enough
1876 * to match our requested alignment
1877 */
1878 if (*bytes >= align) {
1879 tmp = entry->offset - ctl->start + align - 1;
1880 tmp = div64_u64(tmp, align);
1881 tmp = tmp * align + ctl->start;
1882 align_off = tmp - entry->offset;
1883 } else {
1884 align_off = 0;
1885 tmp = entry->offset;
1886 }
1887
1888 if (entry->bytes < *bytes + align_off) {
1889 *max_extent_size = max(get_max_extent_size(entry),
1890 *max_extent_size);
1891 continue;
1892 }
1893
1894 if (entry->bitmap) {
1895 u64 size = *bytes;
1896
1897 ret = search_bitmap(ctl, entry, &tmp, &size, true);
1898 if (!ret) {
1899 *offset = tmp;
1900 *bytes = size;
1901 return entry;
1902 } else {
1903 *max_extent_size =
1904 max(get_max_extent_size(entry),
1905 *max_extent_size);
1906 }
1907 continue;
1908 }
1909
1910 *offset = tmp;
1911 *bytes = entry->bytes - align_off;
1912 return entry;
1913 }
1914out:
1915 return NULL;
1916}
1917
1918static int count_bitmap_extents(struct btrfs_free_space_ctl *ctl,
1919 struct btrfs_free_space *bitmap_info)
1920{
1921 struct btrfs_block_group *block_group = ctl->private;
1922 u64 bytes = bitmap_info->bytes;
1923 unsigned int rs, re;
1924 int count = 0;
1925
1926 if (!block_group || !bytes)
1927 return count;
1928
1929 bitmap_for_each_set_region(bitmap_info->bitmap, rs, re, 0,
1930 BITS_PER_BITMAP) {
1931 bytes -= (rs - re) * ctl->unit;
1932 count++;
1933
1934 if (!bytes)
1935 break;
1936 }
1937
1938 return count;
1939}
1940
1941static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1942 struct btrfs_free_space *info, u64 offset)
1943{
1944 info->offset = offset_to_bitmap(ctl, offset);
1945 info->bytes = 0;
1946 info->bitmap_extents = 0;
1947 INIT_LIST_HEAD(&info->list);
1948 link_free_space(ctl, info);
1949 ctl->total_bitmaps++;
1950
1951 ctl->op->recalc_thresholds(ctl);
1952}
1953
1954static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1955 struct btrfs_free_space *bitmap_info)
1956{
1957 /*
1958 * Normally when this is called, the bitmap is completely empty. However,
1959 * if we are blowing up the free space cache for one reason or another
1960 * via __btrfs_remove_free_space_cache(), then it may not be freed and
1961 * we may leave stats on the table.
1962 */
1963 if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
1964 ctl->discardable_extents[BTRFS_STAT_CURR] -=
1965 bitmap_info->bitmap_extents;
1966 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
1967
1968 }
1969 unlink_free_space(ctl, bitmap_info);
1970 kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1971 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1972 ctl->total_bitmaps--;
1973 ctl->op->recalc_thresholds(ctl);
1974}
1975
1976static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1977 struct btrfs_free_space *bitmap_info,
1978 u64 *offset, u64 *bytes)
1979{
1980 u64 end;
1981 u64 search_start, search_bytes;
1982 int ret;
1983
1984again:
1985 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1986
1987 /*
1988 * We need to search for bits in this bitmap. We could only cover some
1989 * of the extent in this bitmap thanks to how we add space, so we need
1990 * to search for as much as it as we can and clear that amount, and then
1991 * go searching for the next bit.
1992 */
1993 search_start = *offset;
1994 search_bytes = ctl->unit;
1995 search_bytes = min(search_bytes, end - search_start + 1);
1996 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1997 false);
1998 if (ret < 0 || search_start != *offset)
1999 return -EINVAL;
2000
2001 /* We may have found more bits than what we need */
2002 search_bytes = min(search_bytes, *bytes);
2003
2004 /* Cannot clear past the end of the bitmap */
2005 search_bytes = min(search_bytes, end - search_start + 1);
2006
2007 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
2008 *offset += search_bytes;
2009 *bytes -= search_bytes;
2010
2011 if (*bytes) {
2012 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2013 if (!bitmap_info->bytes)
2014 free_bitmap(ctl, bitmap_info);
2015
2016 /*
2017 * no entry after this bitmap, but we still have bytes to
2018 * remove, so something has gone wrong.
2019 */
2020 if (!next)
2021 return -EINVAL;
2022
2023 bitmap_info = rb_entry(next, struct btrfs_free_space,
2024 offset_index);
2025
2026 /*
2027 * if the next entry isn't a bitmap we need to return to let the
2028 * extent stuff do its work.
2029 */
2030 if (!bitmap_info->bitmap)
2031 return -EAGAIN;
2032
2033 /*
2034 * Ok the next item is a bitmap, but it may not actually hold
2035 * the information for the rest of this free space stuff, so
2036 * look for it, and if we don't find it return so we can try
2037 * everything over again.
2038 */
2039 search_start = *offset;
2040 search_bytes = ctl->unit;
2041 ret = search_bitmap(ctl, bitmap_info, &search_start,
2042 &search_bytes, false);
2043 if (ret < 0 || search_start != *offset)
2044 return -EAGAIN;
2045
2046 goto again;
2047 } else if (!bitmap_info->bytes)
2048 free_bitmap(ctl, bitmap_info);
2049
2050 return 0;
2051}
2052
2053static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2054 struct btrfs_free_space *info, u64 offset,
2055 u64 bytes, enum btrfs_trim_state trim_state)
2056{
2057 u64 bytes_to_set = 0;
2058 u64 end;
2059
2060 /*
2061 * This is a tradeoff to make bitmap trim state minimal. We mark the
2062 * whole bitmap untrimmed if at any point we add untrimmed regions.
2063 */
2064 if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2065 if (btrfs_free_space_trimmed(info)) {
2066 ctl->discardable_extents[BTRFS_STAT_CURR] +=
2067 info->bitmap_extents;
2068 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2069 }
2070 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2071 }
2072
2073 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2074
2075 bytes_to_set = min(end - offset, bytes);
2076
2077 bitmap_set_bits(ctl, info, offset, bytes_to_set);
2078
2079 /*
2080 * We set some bytes, we have no idea what the max extent size is
2081 * anymore.
2082 */
2083 info->max_extent_size = 0;
2084
2085 return bytes_to_set;
2086
2087}
2088
2089static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2090 struct btrfs_free_space *info)
2091{
2092 struct btrfs_block_group *block_group = ctl->private;
2093 struct btrfs_fs_info *fs_info = block_group->fs_info;
2094 bool forced = false;
2095
2096#ifdef CONFIG_BTRFS_DEBUG
2097 if (btrfs_should_fragment_free_space(block_group))
2098 forced = true;
2099#endif
2100
2101 /* This is a way to reclaim large regions from the bitmaps. */
2102 if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2103 return false;
2104
2105 /*
2106 * If we are below the extents threshold then we can add this as an
2107 * extent, and don't have to deal with the bitmap
2108 */
2109 if (!forced && ctl->free_extents < ctl->extents_thresh) {
2110 /*
2111 * If this block group has some small extents we don't want to
2112 * use up all of our free slots in the cache with them, we want
2113 * to reserve them to larger extents, however if we have plenty
2114 * of cache left then go ahead an dadd them, no sense in adding
2115 * the overhead of a bitmap if we don't have to.
2116 */
2117 if (info->bytes <= fs_info->sectorsize * 8) {
2118 if (ctl->free_extents * 3 <= ctl->extents_thresh)
2119 return false;
2120 } else {
2121 return false;
2122 }
2123 }
2124
2125 /*
2126 * The original block groups from mkfs can be really small, like 8
2127 * megabytes, so don't bother with a bitmap for those entries. However
2128 * some block groups can be smaller than what a bitmap would cover but
2129 * are still large enough that they could overflow the 32k memory limit,
2130 * so allow those block groups to still be allowed to have a bitmap
2131 * entry.
2132 */
2133 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2134 return false;
2135
2136 return true;
2137}
2138
2139static const struct btrfs_free_space_op free_space_op = {
2140 .recalc_thresholds = recalculate_thresholds,
2141 .use_bitmap = use_bitmap,
2142};
2143
2144static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2145 struct btrfs_free_space *info)
2146{
2147 struct btrfs_free_space *bitmap_info;
2148 struct btrfs_block_group *block_group = NULL;
2149 int added = 0;
2150 u64 bytes, offset, bytes_added;
2151 enum btrfs_trim_state trim_state;
2152 int ret;
2153
2154 bytes = info->bytes;
2155 offset = info->offset;
2156 trim_state = info->trim_state;
2157
2158 if (!ctl->op->use_bitmap(ctl, info))
2159 return 0;
2160
2161 if (ctl->op == &free_space_op)
2162 block_group = ctl->private;
2163again:
2164 /*
2165 * Since we link bitmaps right into the cluster we need to see if we
2166 * have a cluster here, and if so and it has our bitmap we need to add
2167 * the free space to that bitmap.
2168 */
2169 if (block_group && !list_empty(&block_group->cluster_list)) {
2170 struct btrfs_free_cluster *cluster;
2171 struct rb_node *node;
2172 struct btrfs_free_space *entry;
2173
2174 cluster = list_entry(block_group->cluster_list.next,
2175 struct btrfs_free_cluster,
2176 block_group_list);
2177 spin_lock(&cluster->lock);
2178 node = rb_first(&cluster->root);
2179 if (!node) {
2180 spin_unlock(&cluster->lock);
2181 goto no_cluster_bitmap;
2182 }
2183
2184 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2185 if (!entry->bitmap) {
2186 spin_unlock(&cluster->lock);
2187 goto no_cluster_bitmap;
2188 }
2189
2190 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2191 bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2192 bytes, trim_state);
2193 bytes -= bytes_added;
2194 offset += bytes_added;
2195 }
2196 spin_unlock(&cluster->lock);
2197 if (!bytes) {
2198 ret = 1;
2199 goto out;
2200 }
2201 }
2202
2203no_cluster_bitmap:
2204 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2205 1, 0);
2206 if (!bitmap_info) {
2207 ASSERT(added == 0);
2208 goto new_bitmap;
2209 }
2210
2211 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2212 trim_state);
2213 bytes -= bytes_added;
2214 offset += bytes_added;
2215 added = 0;
2216
2217 if (!bytes) {
2218 ret = 1;
2219 goto out;
2220 } else
2221 goto again;
2222
2223new_bitmap:
2224 if (info && info->bitmap) {
2225 add_new_bitmap(ctl, info, offset);
2226 added = 1;
2227 info = NULL;
2228 goto again;
2229 } else {
2230 spin_unlock(&ctl->tree_lock);
2231
2232 /* no pre-allocated info, allocate a new one */
2233 if (!info) {
2234 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2235 GFP_NOFS);
2236 if (!info) {
2237 spin_lock(&ctl->tree_lock);
2238 ret = -ENOMEM;
2239 goto out;
2240 }
2241 }
2242
2243 /* allocate the bitmap */
2244 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2245 GFP_NOFS);
2246 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2247 spin_lock(&ctl->tree_lock);
2248 if (!info->bitmap) {
2249 ret = -ENOMEM;
2250 goto out;
2251 }
2252 goto again;
2253 }
2254
2255out:
2256 if (info) {
2257 if (info->bitmap)
2258 kmem_cache_free(btrfs_free_space_bitmap_cachep,
2259 info->bitmap);
2260 kmem_cache_free(btrfs_free_space_cachep, info);
2261 }
2262
2263 return ret;
2264}
2265
2266/*
2267 * Free space merging rules:
2268 * 1) Merge trimmed areas together
2269 * 2) Let untrimmed areas coalesce with trimmed areas
2270 * 3) Always pull neighboring regions from bitmaps
2271 *
2272 * The above rules are for when we merge free space based on btrfs_trim_state.
2273 * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2274 * same reason: to promote larger extent regions which makes life easier for
2275 * find_free_extent(). Rule 2 enables coalescing based on the common path
2276 * being returning free space from btrfs_finish_extent_commit(). So when free
2277 * space is trimmed, it will prevent aggregating trimmed new region and
2278 * untrimmed regions in the rb_tree. Rule 3 is purely to obtain larger extents
2279 * and provide find_free_extent() with the largest extents possible hoping for
2280 * the reuse path.
2281 */
2282static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2283 struct btrfs_free_space *info, bool update_stat)
2284{
2285 struct btrfs_free_space *left_info = NULL;
2286 struct btrfs_free_space *right_info;
2287 bool merged = false;
2288 u64 offset = info->offset;
2289 u64 bytes = info->bytes;
2290 const bool is_trimmed = btrfs_free_space_trimmed(info);
2291
2292 /*
2293 * first we want to see if there is free space adjacent to the range we
2294 * are adding, if there is remove that struct and add a new one to
2295 * cover the entire range
2296 */
2297 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2298 if (right_info && rb_prev(&right_info->offset_index))
2299 left_info = rb_entry(rb_prev(&right_info->offset_index),
2300 struct btrfs_free_space, offset_index);
2301 else if (!right_info)
2302 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2303
2304 /* See try_merge_free_space() comment. */
2305 if (right_info && !right_info->bitmap &&
2306 (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2307 if (update_stat)
2308 unlink_free_space(ctl, right_info);
2309 else
2310 __unlink_free_space(ctl, right_info);
2311 info->bytes += right_info->bytes;
2312 kmem_cache_free(btrfs_free_space_cachep, right_info);
2313 merged = true;
2314 }
2315
2316 /* See try_merge_free_space() comment. */
2317 if (left_info && !left_info->bitmap &&
2318 left_info->offset + left_info->bytes == offset &&
2319 (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2320 if (update_stat)
2321 unlink_free_space(ctl, left_info);
2322 else
2323 __unlink_free_space(ctl, left_info);
2324 info->offset = left_info->offset;
2325 info->bytes += left_info->bytes;
2326 kmem_cache_free(btrfs_free_space_cachep, left_info);
2327 merged = true;
2328 }
2329
2330 return merged;
2331}
2332
2333static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2334 struct btrfs_free_space *info,
2335 bool update_stat)
2336{
2337 struct btrfs_free_space *bitmap;
2338 unsigned long i;
2339 unsigned long j;
2340 const u64 end = info->offset + info->bytes;
2341 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2342 u64 bytes;
2343
2344 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2345 if (!bitmap)
2346 return false;
2347
2348 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2349 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2350 if (j == i)
2351 return false;
2352 bytes = (j - i) * ctl->unit;
2353 info->bytes += bytes;
2354
2355 /* See try_merge_free_space() comment. */
2356 if (!btrfs_free_space_trimmed(bitmap))
2357 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2358
2359 if (update_stat)
2360 bitmap_clear_bits(ctl, bitmap, end, bytes);
2361 else
2362 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2363
2364 if (!bitmap->bytes)
2365 free_bitmap(ctl, bitmap);
2366
2367 return true;
2368}
2369
2370static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2371 struct btrfs_free_space *info,
2372 bool update_stat)
2373{
2374 struct btrfs_free_space *bitmap;
2375 u64 bitmap_offset;
2376 unsigned long i;
2377 unsigned long j;
2378 unsigned long prev_j;
2379 u64 bytes;
2380
2381 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2382 /* If we're on a boundary, try the previous logical bitmap. */
2383 if (bitmap_offset == info->offset) {
2384 if (info->offset == 0)
2385 return false;
2386 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2387 }
2388
2389 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2390 if (!bitmap)
2391 return false;
2392
2393 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2394 j = 0;
2395 prev_j = (unsigned long)-1;
2396 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2397 if (j > i)
2398 break;
2399 prev_j = j;
2400 }
2401 if (prev_j == i)
2402 return false;
2403
2404 if (prev_j == (unsigned long)-1)
2405 bytes = (i + 1) * ctl->unit;
2406 else
2407 bytes = (i - prev_j) * ctl->unit;
2408
2409 info->offset -= bytes;
2410 info->bytes += bytes;
2411
2412 /* See try_merge_free_space() comment. */
2413 if (!btrfs_free_space_trimmed(bitmap))
2414 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2415
2416 if (update_stat)
2417 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2418 else
2419 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2420
2421 if (!bitmap->bytes)
2422 free_bitmap(ctl, bitmap);
2423
2424 return true;
2425}
2426
2427/*
2428 * We prefer always to allocate from extent entries, both for clustered and
2429 * non-clustered allocation requests. So when attempting to add a new extent
2430 * entry, try to see if there's adjacent free space in bitmap entries, and if
2431 * there is, migrate that space from the bitmaps to the extent.
2432 * Like this we get better chances of satisfying space allocation requests
2433 * because we attempt to satisfy them based on a single cache entry, and never
2434 * on 2 or more entries - even if the entries represent a contiguous free space
2435 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2436 * ends).
2437 */
2438static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2439 struct btrfs_free_space *info,
2440 bool update_stat)
2441{
2442 /*
2443 * Only work with disconnected entries, as we can change their offset,
2444 * and must be extent entries.
2445 */
2446 ASSERT(!info->bitmap);
2447 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2448
2449 if (ctl->total_bitmaps > 0) {
2450 bool stole_end;
2451 bool stole_front = false;
2452
2453 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2454 if (ctl->total_bitmaps > 0)
2455 stole_front = steal_from_bitmap_to_front(ctl, info,
2456 update_stat);
2457
2458 if (stole_end || stole_front)
2459 try_merge_free_space(ctl, info, update_stat);
2460 }
2461}
2462
2463int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2464 struct btrfs_free_space_ctl *ctl,
2465 u64 offset, u64 bytes,
2466 enum btrfs_trim_state trim_state)
2467{
2468 struct btrfs_block_group *block_group = ctl->private;
2469 struct btrfs_free_space *info;
2470 int ret = 0;
2471 u64 filter_bytes = bytes;
2472
2473 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2474 if (!info)
2475 return -ENOMEM;
2476
2477 info->offset = offset;
2478 info->bytes = bytes;
2479 info->trim_state = trim_state;
2480 RB_CLEAR_NODE(&info->offset_index);
2481
2482 spin_lock(&ctl->tree_lock);
2483
2484 if (try_merge_free_space(ctl, info, true))
2485 goto link;
2486
2487 /*
2488 * There was no extent directly to the left or right of this new
2489 * extent then we know we're going to have to allocate a new extent, so
2490 * before we do that see if we need to drop this into a bitmap
2491 */
2492 ret = insert_into_bitmap(ctl, info);
2493 if (ret < 0) {
2494 goto out;
2495 } else if (ret) {
2496 ret = 0;
2497 goto out;
2498 }
2499link:
2500 /*
2501 * Only steal free space from adjacent bitmaps if we're sure we're not
2502 * going to add the new free space to existing bitmap entries - because
2503 * that would mean unnecessary work that would be reverted. Therefore
2504 * attempt to steal space from bitmaps if we're adding an extent entry.
2505 */
2506 steal_from_bitmap(ctl, info, true);
2507
2508 filter_bytes = max(filter_bytes, info->bytes);
2509
2510 ret = link_free_space(ctl, info);
2511 if (ret)
2512 kmem_cache_free(btrfs_free_space_cachep, info);
2513out:
2514 btrfs_discard_update_discardable(block_group, ctl);
2515 spin_unlock(&ctl->tree_lock);
2516
2517 if (ret) {
2518 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2519 ASSERT(ret != -EEXIST);
2520 }
2521
2522 if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2523 btrfs_discard_check_filter(block_group, filter_bytes);
2524 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2525 }
2526
2527 return ret;
2528}
2529
2530int btrfs_add_free_space(struct btrfs_block_group *block_group,
2531 u64 bytenr, u64 size)
2532{
2533 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2534
2535 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2536 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2537
2538 return __btrfs_add_free_space(block_group->fs_info,
2539 block_group->free_space_ctl,
2540 bytenr, size, trim_state);
2541}
2542
2543/*
2544 * This is a subtle distinction because when adding free space back in general,
2545 * we want it to be added as untrimmed for async. But in the case where we add
2546 * it on loading of a block group, we want to consider it trimmed.
2547 */
2548int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2549 u64 bytenr, u64 size)
2550{
2551 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2552
2553 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2554 btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2555 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2556
2557 return __btrfs_add_free_space(block_group->fs_info,
2558 block_group->free_space_ctl,
2559 bytenr, size, trim_state);
2560}
2561
2562int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2563 u64 offset, u64 bytes)
2564{
2565 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2566 struct btrfs_free_space *info;
2567 int ret;
2568 bool re_search = false;
2569
2570 spin_lock(&ctl->tree_lock);
2571
2572again:
2573 ret = 0;
2574 if (!bytes)
2575 goto out_lock;
2576
2577 info = tree_search_offset(ctl, offset, 0, 0);
2578 if (!info) {
2579 /*
2580 * oops didn't find an extent that matched the space we wanted
2581 * to remove, look for a bitmap instead
2582 */
2583 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2584 1, 0);
2585 if (!info) {
2586 /*
2587 * If we found a partial bit of our free space in a
2588 * bitmap but then couldn't find the other part this may
2589 * be a problem, so WARN about it.
2590 */
2591 WARN_ON(re_search);
2592 goto out_lock;
2593 }
2594 }
2595
2596 re_search = false;
2597 if (!info->bitmap) {
2598 unlink_free_space(ctl, info);
2599 if (offset == info->offset) {
2600 u64 to_free = min(bytes, info->bytes);
2601
2602 info->bytes -= to_free;
2603 info->offset += to_free;
2604 if (info->bytes) {
2605 ret = link_free_space(ctl, info);
2606 WARN_ON(ret);
2607 } else {
2608 kmem_cache_free(btrfs_free_space_cachep, info);
2609 }
2610
2611 offset += to_free;
2612 bytes -= to_free;
2613 goto again;
2614 } else {
2615 u64 old_end = info->bytes + info->offset;
2616
2617 info->bytes = offset - info->offset;
2618 ret = link_free_space(ctl, info);
2619 WARN_ON(ret);
2620 if (ret)
2621 goto out_lock;
2622
2623 /* Not enough bytes in this entry to satisfy us */
2624 if (old_end < offset + bytes) {
2625 bytes -= old_end - offset;
2626 offset = old_end;
2627 goto again;
2628 } else if (old_end == offset + bytes) {
2629 /* all done */
2630 goto out_lock;
2631 }
2632 spin_unlock(&ctl->tree_lock);
2633
2634 ret = __btrfs_add_free_space(block_group->fs_info, ctl,
2635 offset + bytes,
2636 old_end - (offset + bytes),
2637 info->trim_state);
2638 WARN_ON(ret);
2639 goto out;
2640 }
2641 }
2642
2643 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2644 if (ret == -EAGAIN) {
2645 re_search = true;
2646 goto again;
2647 }
2648out_lock:
2649 btrfs_discard_update_discardable(block_group, ctl);
2650 spin_unlock(&ctl->tree_lock);
2651out:
2652 return ret;
2653}
2654
2655void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2656 u64 bytes)
2657{
2658 struct btrfs_fs_info *fs_info = block_group->fs_info;
2659 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2660 struct btrfs_free_space *info;
2661 struct rb_node *n;
2662 int count = 0;
2663
2664 spin_lock(&ctl->tree_lock);
2665 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2666 info = rb_entry(n, struct btrfs_free_space, offset_index);
2667 if (info->bytes >= bytes && !block_group->ro)
2668 count++;
2669 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2670 info->offset, info->bytes,
2671 (info->bitmap) ? "yes" : "no");
2672 }
2673 spin_unlock(&ctl->tree_lock);
2674 btrfs_info(fs_info, "block group has cluster?: %s",
2675 list_empty(&block_group->cluster_list) ? "no" : "yes");
2676 btrfs_info(fs_info,
2677 "%d blocks of free space at or bigger than bytes is", count);
2678}
2679
2680void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group)
2681{
2682 struct btrfs_fs_info *fs_info = block_group->fs_info;
2683 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2684
2685 spin_lock_init(&ctl->tree_lock);
2686 ctl->unit = fs_info->sectorsize;
2687 ctl->start = block_group->start;
2688 ctl->private = block_group;
2689 ctl->op = &free_space_op;
2690 INIT_LIST_HEAD(&ctl->trimming_ranges);
2691 mutex_init(&ctl->cache_writeout_mutex);
2692
2693 /*
2694 * we only want to have 32k of ram per block group for keeping
2695 * track of free space, and if we pass 1/2 of that we want to
2696 * start converting things over to using bitmaps
2697 */
2698 ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2699}
2700
2701/*
2702 * for a given cluster, put all of its extents back into the free
2703 * space cache. If the block group passed doesn't match the block group
2704 * pointed to by the cluster, someone else raced in and freed the
2705 * cluster already. In that case, we just return without changing anything
2706 */
2707static void __btrfs_return_cluster_to_free_space(
2708 struct btrfs_block_group *block_group,
2709 struct btrfs_free_cluster *cluster)
2710{
2711 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2712 struct btrfs_free_space *entry;
2713 struct rb_node *node;
2714
2715 spin_lock(&cluster->lock);
2716 if (cluster->block_group != block_group)
2717 goto out;
2718
2719 cluster->block_group = NULL;
2720 cluster->window_start = 0;
2721 list_del_init(&cluster->block_group_list);
2722
2723 node = rb_first(&cluster->root);
2724 while (node) {
2725 bool bitmap;
2726
2727 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2728 node = rb_next(&entry->offset_index);
2729 rb_erase(&entry->offset_index, &cluster->root);
2730 RB_CLEAR_NODE(&entry->offset_index);
2731
2732 bitmap = (entry->bitmap != NULL);
2733 if (!bitmap) {
2734 /* Merging treats extents as if they were new */
2735 if (!btrfs_free_space_trimmed(entry)) {
2736 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2737 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2738 entry->bytes;
2739 }
2740
2741 try_merge_free_space(ctl, entry, false);
2742 steal_from_bitmap(ctl, entry, false);
2743
2744 /* As we insert directly, update these statistics */
2745 if (!btrfs_free_space_trimmed(entry)) {
2746 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2747 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2748 entry->bytes;
2749 }
2750 }
2751 tree_insert_offset(&ctl->free_space_offset,
2752 entry->offset, &entry->offset_index, bitmap);
2753 }
2754 cluster->root = RB_ROOT;
2755
2756out:
2757 spin_unlock(&cluster->lock);
2758 btrfs_put_block_group(block_group);
2759}
2760
2761static void __btrfs_remove_free_space_cache_locked(
2762 struct btrfs_free_space_ctl *ctl)
2763{
2764 struct btrfs_free_space *info;
2765 struct rb_node *node;
2766
2767 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2768 info = rb_entry(node, struct btrfs_free_space, offset_index);
2769 if (!info->bitmap) {
2770 unlink_free_space(ctl, info);
2771 kmem_cache_free(btrfs_free_space_cachep, info);
2772 } else {
2773 free_bitmap(ctl, info);
2774 }
2775
2776 cond_resched_lock(&ctl->tree_lock);
2777 }
2778}
2779
2780void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2781{
2782 spin_lock(&ctl->tree_lock);
2783 __btrfs_remove_free_space_cache_locked(ctl);
2784 if (ctl->private)
2785 btrfs_discard_update_discardable(ctl->private, ctl);
2786 spin_unlock(&ctl->tree_lock);
2787}
2788
2789void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2790{
2791 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2792 struct btrfs_free_cluster *cluster;
2793 struct list_head *head;
2794
2795 spin_lock(&ctl->tree_lock);
2796 while ((head = block_group->cluster_list.next) !=
2797 &block_group->cluster_list) {
2798 cluster = list_entry(head, struct btrfs_free_cluster,
2799 block_group_list);
2800
2801 WARN_ON(cluster->block_group != block_group);
2802 __btrfs_return_cluster_to_free_space(block_group, cluster);
2803
2804 cond_resched_lock(&ctl->tree_lock);
2805 }
2806 __btrfs_remove_free_space_cache_locked(ctl);
2807 btrfs_discard_update_discardable(block_group, ctl);
2808 spin_unlock(&ctl->tree_lock);
2809
2810}
2811
2812/**
2813 * btrfs_is_free_space_trimmed - see if everything is trimmed
2814 * @block_group: block_group of interest
2815 *
2816 * Walk @block_group's free space rb_tree to determine if everything is trimmed.
2817 */
2818bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
2819{
2820 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2821 struct btrfs_free_space *info;
2822 struct rb_node *node;
2823 bool ret = true;
2824
2825 spin_lock(&ctl->tree_lock);
2826 node = rb_first(&ctl->free_space_offset);
2827
2828 while (node) {
2829 info = rb_entry(node, struct btrfs_free_space, offset_index);
2830
2831 if (!btrfs_free_space_trimmed(info)) {
2832 ret = false;
2833 break;
2834 }
2835
2836 node = rb_next(node);
2837 }
2838
2839 spin_unlock(&ctl->tree_lock);
2840 return ret;
2841}
2842
2843u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2844 u64 offset, u64 bytes, u64 empty_size,
2845 u64 *max_extent_size)
2846{
2847 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2848 struct btrfs_discard_ctl *discard_ctl =
2849 &block_group->fs_info->discard_ctl;
2850 struct btrfs_free_space *entry = NULL;
2851 u64 bytes_search = bytes + empty_size;
2852 u64 ret = 0;
2853 u64 align_gap = 0;
2854 u64 align_gap_len = 0;
2855 enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2856
2857 spin_lock(&ctl->tree_lock);
2858 entry = find_free_space(ctl, &offset, &bytes_search,
2859 block_group->full_stripe_len, max_extent_size);
2860 if (!entry)
2861 goto out;
2862
2863 ret = offset;
2864 if (entry->bitmap) {
2865 bitmap_clear_bits(ctl, entry, offset, bytes);
2866
2867 if (!btrfs_free_space_trimmed(entry))
2868 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2869
2870 if (!entry->bytes)
2871 free_bitmap(ctl, entry);
2872 } else {
2873 unlink_free_space(ctl, entry);
2874 align_gap_len = offset - entry->offset;
2875 align_gap = entry->offset;
2876 align_gap_trim_state = entry->trim_state;
2877
2878 if (!btrfs_free_space_trimmed(entry))
2879 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2880
2881 entry->offset = offset + bytes;
2882 WARN_ON(entry->bytes < bytes + align_gap_len);
2883
2884 entry->bytes -= bytes + align_gap_len;
2885 if (!entry->bytes)
2886 kmem_cache_free(btrfs_free_space_cachep, entry);
2887 else
2888 link_free_space(ctl, entry);
2889 }
2890out:
2891 btrfs_discard_update_discardable(block_group, ctl);
2892 spin_unlock(&ctl->tree_lock);
2893
2894 if (align_gap_len)
2895 __btrfs_add_free_space(block_group->fs_info, ctl,
2896 align_gap, align_gap_len,
2897 align_gap_trim_state);
2898 return ret;
2899}
2900
2901/*
2902 * given a cluster, put all of its extents back into the free space
2903 * cache. If a block group is passed, this function will only free
2904 * a cluster that belongs to the passed block group.
2905 *
2906 * Otherwise, it'll get a reference on the block group pointed to by the
2907 * cluster and remove the cluster from it.
2908 */
2909void btrfs_return_cluster_to_free_space(
2910 struct btrfs_block_group *block_group,
2911 struct btrfs_free_cluster *cluster)
2912{
2913 struct btrfs_free_space_ctl *ctl;
2914
2915 /* first, get a safe pointer to the block group */
2916 spin_lock(&cluster->lock);
2917 if (!block_group) {
2918 block_group = cluster->block_group;
2919 if (!block_group) {
2920 spin_unlock(&cluster->lock);
2921 return;
2922 }
2923 } else if (cluster->block_group != block_group) {
2924 /* someone else has already freed it don't redo their work */
2925 spin_unlock(&cluster->lock);
2926 return;
2927 }
2928 btrfs_get_block_group(block_group);
2929 spin_unlock(&cluster->lock);
2930
2931 ctl = block_group->free_space_ctl;
2932
2933 /* now return any extents the cluster had on it */
2934 spin_lock(&ctl->tree_lock);
2935 __btrfs_return_cluster_to_free_space(block_group, cluster);
2936 spin_unlock(&ctl->tree_lock);
2937
2938 btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
2939
2940 /* finally drop our ref */
2941 btrfs_put_block_group(block_group);
2942}
2943
2944static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
2945 struct btrfs_free_cluster *cluster,
2946 struct btrfs_free_space *entry,
2947 u64 bytes, u64 min_start,
2948 u64 *max_extent_size)
2949{
2950 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2951 int err;
2952 u64 search_start = cluster->window_start;
2953 u64 search_bytes = bytes;
2954 u64 ret = 0;
2955
2956 search_start = min_start;
2957 search_bytes = bytes;
2958
2959 err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2960 if (err) {
2961 *max_extent_size = max(get_max_extent_size(entry),
2962 *max_extent_size);
2963 return 0;
2964 }
2965
2966 ret = search_start;
2967 __bitmap_clear_bits(ctl, entry, ret, bytes);
2968
2969 return ret;
2970}
2971
2972/*
2973 * given a cluster, try to allocate 'bytes' from it, returns 0
2974 * if it couldn't find anything suitably large, or a logical disk offset
2975 * if things worked out
2976 */
2977u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
2978 struct btrfs_free_cluster *cluster, u64 bytes,
2979 u64 min_start, u64 *max_extent_size)
2980{
2981 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2982 struct btrfs_discard_ctl *discard_ctl =
2983 &block_group->fs_info->discard_ctl;
2984 struct btrfs_free_space *entry = NULL;
2985 struct rb_node *node;
2986 u64 ret = 0;
2987
2988 spin_lock(&cluster->lock);
2989 if (bytes > cluster->max_size)
2990 goto out;
2991
2992 if (cluster->block_group != block_group)
2993 goto out;
2994
2995 node = rb_first(&cluster->root);
2996 if (!node)
2997 goto out;
2998
2999 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3000 while (1) {
3001 if (entry->bytes < bytes)
3002 *max_extent_size = max(get_max_extent_size(entry),
3003 *max_extent_size);
3004
3005 if (entry->bytes < bytes ||
3006 (!entry->bitmap && entry->offset < min_start)) {
3007 node = rb_next(&entry->offset_index);
3008 if (!node)
3009 break;
3010 entry = rb_entry(node, struct btrfs_free_space,
3011 offset_index);
3012 continue;
3013 }
3014
3015 if (entry->bitmap) {
3016 ret = btrfs_alloc_from_bitmap(block_group,
3017 cluster, entry, bytes,
3018 cluster->window_start,
3019 max_extent_size);
3020 if (ret == 0) {
3021 node = rb_next(&entry->offset_index);
3022 if (!node)
3023 break;
3024 entry = rb_entry(node, struct btrfs_free_space,
3025 offset_index);
3026 continue;
3027 }
3028 cluster->window_start += bytes;
3029 } else {
3030 ret = entry->offset;
3031
3032 entry->offset += bytes;
3033 entry->bytes -= bytes;
3034 }
3035
3036 if (entry->bytes == 0)
3037 rb_erase(&entry->offset_index, &cluster->root);
3038 break;
3039 }
3040out:
3041 spin_unlock(&cluster->lock);
3042
3043 if (!ret)
3044 return 0;
3045
3046 spin_lock(&ctl->tree_lock);
3047
3048 if (!btrfs_free_space_trimmed(entry))
3049 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3050
3051 ctl->free_space -= bytes;
3052 if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3053 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3054 if (entry->bytes == 0) {
3055 ctl->free_extents--;
3056 if (entry->bitmap) {
3057 kmem_cache_free(btrfs_free_space_bitmap_cachep,
3058 entry->bitmap);
3059 ctl->total_bitmaps--;
3060 ctl->op->recalc_thresholds(ctl);
3061 } else if (!btrfs_free_space_trimmed(entry)) {
3062 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3063 }
3064 kmem_cache_free(btrfs_free_space_cachep, entry);
3065 }
3066
3067 spin_unlock(&ctl->tree_lock);
3068
3069 return ret;
3070}
3071
3072static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3073 struct btrfs_free_space *entry,
3074 struct btrfs_free_cluster *cluster,
3075 u64 offset, u64 bytes,
3076 u64 cont1_bytes, u64 min_bytes)
3077{
3078 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3079 unsigned long next_zero;
3080 unsigned long i;
3081 unsigned long want_bits;
3082 unsigned long min_bits;
3083 unsigned long found_bits;
3084 unsigned long max_bits = 0;
3085 unsigned long start = 0;
3086 unsigned long total_found = 0;
3087 int ret;
3088
3089 i = offset_to_bit(entry->offset, ctl->unit,
3090 max_t(u64, offset, entry->offset));
3091 want_bits = bytes_to_bits(bytes, ctl->unit);
3092 min_bits = bytes_to_bits(min_bytes, ctl->unit);
3093
3094 /*
3095 * Don't bother looking for a cluster in this bitmap if it's heavily
3096 * fragmented.
3097 */
3098 if (entry->max_extent_size &&
3099 entry->max_extent_size < cont1_bytes)
3100 return -ENOSPC;
3101again:
3102 found_bits = 0;
3103 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3104 next_zero = find_next_zero_bit(entry->bitmap,
3105 BITS_PER_BITMAP, i);
3106 if (next_zero - i >= min_bits) {
3107 found_bits = next_zero - i;
3108 if (found_bits > max_bits)
3109 max_bits = found_bits;
3110 break;
3111 }
3112 if (next_zero - i > max_bits)
3113 max_bits = next_zero - i;
3114 i = next_zero;
3115 }
3116
3117 if (!found_bits) {
3118 entry->max_extent_size = (u64)max_bits * ctl->unit;
3119 return -ENOSPC;
3120 }
3121
3122 if (!total_found) {
3123 start = i;
3124 cluster->max_size = 0;
3125 }
3126
3127 total_found += found_bits;
3128
3129 if (cluster->max_size < found_bits * ctl->unit)
3130 cluster->max_size = found_bits * ctl->unit;
3131
3132 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3133 i = next_zero + 1;
3134 goto again;
3135 }
3136
3137 cluster->window_start = start * ctl->unit + entry->offset;
3138 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3139 ret = tree_insert_offset(&cluster->root, entry->offset,
3140 &entry->offset_index, 1);
3141 ASSERT(!ret); /* -EEXIST; Logic error */
3142
3143 trace_btrfs_setup_cluster(block_group, cluster,
3144 total_found * ctl->unit, 1);
3145 return 0;
3146}
3147
3148/*
3149 * This searches the block group for just extents to fill the cluster with.
3150 * Try to find a cluster with at least bytes total bytes, at least one
3151 * extent of cont1_bytes, and other clusters of at least min_bytes.
3152 */
3153static noinline int
3154setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3155 struct btrfs_free_cluster *cluster,
3156 struct list_head *bitmaps, u64 offset, u64 bytes,
3157 u64 cont1_bytes, u64 min_bytes)
3158{
3159 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3160 struct btrfs_free_space *first = NULL;
3161 struct btrfs_free_space *entry = NULL;
3162 struct btrfs_free_space *last;
3163 struct rb_node *node;
3164 u64 window_free;
3165 u64 max_extent;
3166 u64 total_size = 0;
3167
3168 entry = tree_search_offset(ctl, offset, 0, 1);
3169 if (!entry)
3170 return -ENOSPC;
3171
3172 /*
3173 * We don't want bitmaps, so just move along until we find a normal
3174 * extent entry.
3175 */
3176 while (entry->bitmap || entry->bytes < min_bytes) {
3177 if (entry->bitmap && list_empty(&entry->list))
3178 list_add_tail(&entry->list, bitmaps);
3179 node = rb_next(&entry->offset_index);
3180 if (!node)
3181 return -ENOSPC;
3182 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3183 }
3184
3185 window_free = entry->bytes;
3186 max_extent = entry->bytes;
3187 first = entry;
3188 last = entry;
3189
3190 for (node = rb_next(&entry->offset_index); node;
3191 node = rb_next(&entry->offset_index)) {
3192 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3193
3194 if (entry->bitmap) {
3195 if (list_empty(&entry->list))
3196 list_add_tail(&entry->list, bitmaps);
3197 continue;
3198 }
3199
3200 if (entry->bytes < min_bytes)
3201 continue;
3202
3203 last = entry;
3204 window_free += entry->bytes;
3205 if (entry->bytes > max_extent)
3206 max_extent = entry->bytes;
3207 }
3208
3209 if (window_free < bytes || max_extent < cont1_bytes)
3210 return -ENOSPC;
3211
3212 cluster->window_start = first->offset;
3213
3214 node = &first->offset_index;
3215
3216 /*
3217 * now we've found our entries, pull them out of the free space
3218 * cache and put them into the cluster rbtree
3219 */
3220 do {
3221 int ret;
3222
3223 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3224 node = rb_next(&entry->offset_index);
3225 if (entry->bitmap || entry->bytes < min_bytes)
3226 continue;
3227
3228 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3229 ret = tree_insert_offset(&cluster->root, entry->offset,
3230 &entry->offset_index, 0);
3231 total_size += entry->bytes;
3232 ASSERT(!ret); /* -EEXIST; Logic error */
3233 } while (node && entry != last);
3234
3235 cluster->max_size = max_extent;
3236 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3237 return 0;
3238}
3239
3240/*
3241 * This specifically looks for bitmaps that may work in the cluster, we assume
3242 * that we have already failed to find extents that will work.
3243 */
3244static noinline int
3245setup_cluster_bitmap(struct btrfs_block_group *block_group,
3246 struct btrfs_free_cluster *cluster,
3247 struct list_head *bitmaps, u64 offset, u64 bytes,
3248 u64 cont1_bytes, u64 min_bytes)
3249{
3250 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3251 struct btrfs_free_space *entry = NULL;
3252 int ret = -ENOSPC;
3253 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3254
3255 if (ctl->total_bitmaps == 0)
3256 return -ENOSPC;
3257
3258 /*
3259 * The bitmap that covers offset won't be in the list unless offset
3260 * is just its start offset.
3261 */
3262 if (!list_empty(bitmaps))
3263 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3264
3265 if (!entry || entry->offset != bitmap_offset) {
3266 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3267 if (entry && list_empty(&entry->list))
3268 list_add(&entry->list, bitmaps);
3269 }
3270
3271 list_for_each_entry(entry, bitmaps, list) {
3272 if (entry->bytes < bytes)
3273 continue;
3274 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3275 bytes, cont1_bytes, min_bytes);
3276 if (!ret)
3277 return 0;
3278 }
3279
3280 /*
3281 * The bitmaps list has all the bitmaps that record free space
3282 * starting after offset, so no more search is required.
3283 */
3284 return -ENOSPC;
3285}
3286
3287/*
3288 * here we try to find a cluster of blocks in a block group. The goal
3289 * is to find at least bytes+empty_size.
3290 * We might not find them all in one contiguous area.
3291 *
3292 * returns zero and sets up cluster if things worked out, otherwise
3293 * it returns -enospc
3294 */
3295int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3296 struct btrfs_free_cluster *cluster,
3297 u64 offset, u64 bytes, u64 empty_size)
3298{
3299 struct btrfs_fs_info *fs_info = block_group->fs_info;
3300 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3301 struct btrfs_free_space *entry, *tmp;
3302 LIST_HEAD(bitmaps);
3303 u64 min_bytes;
3304 u64 cont1_bytes;
3305 int ret;
3306
3307 /*
3308 * Choose the minimum extent size we'll require for this
3309 * cluster. For SSD_SPREAD, don't allow any fragmentation.
3310 * For metadata, allow allocates with smaller extents. For
3311 * data, keep it dense.
3312 */
3313 if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3314 cont1_bytes = min_bytes = bytes + empty_size;
3315 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3316 cont1_bytes = bytes;
3317 min_bytes = fs_info->sectorsize;
3318 } else {
3319 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3320 min_bytes = fs_info->sectorsize;
3321 }
3322
3323 spin_lock(&ctl->tree_lock);
3324
3325 /*
3326 * If we know we don't have enough space to make a cluster don't even
3327 * bother doing all the work to try and find one.
3328 */
3329 if (ctl->free_space < bytes) {
3330 spin_unlock(&ctl->tree_lock);
3331 return -ENOSPC;
3332 }
3333
3334 spin_lock(&cluster->lock);
3335
3336 /* someone already found a cluster, hooray */
3337 if (cluster->block_group) {
3338 ret = 0;
3339 goto out;
3340 }
3341
3342 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3343 min_bytes);
3344
3345 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3346 bytes + empty_size,
3347 cont1_bytes, min_bytes);
3348 if (ret)
3349 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3350 offset, bytes + empty_size,
3351 cont1_bytes, min_bytes);
3352
3353 /* Clear our temporary list */
3354 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3355 list_del_init(&entry->list);
3356
3357 if (!ret) {
3358 btrfs_get_block_group(block_group);
3359 list_add_tail(&cluster->block_group_list,
3360 &block_group->cluster_list);
3361 cluster->block_group = block_group;
3362 } else {
3363 trace_btrfs_failed_cluster_setup(block_group);
3364 }
3365out:
3366 spin_unlock(&cluster->lock);
3367 spin_unlock(&ctl->tree_lock);
3368
3369 return ret;
3370}
3371
3372/*
3373 * simple code to zero out a cluster
3374 */
3375void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3376{
3377 spin_lock_init(&cluster->lock);
3378 spin_lock_init(&cluster->refill_lock);
3379 cluster->root = RB_ROOT;
3380 cluster->max_size = 0;
3381 cluster->fragmented = false;
3382 INIT_LIST_HEAD(&cluster->block_group_list);
3383 cluster->block_group = NULL;
3384}
3385
3386static int do_trimming(struct btrfs_block_group *block_group,
3387 u64 *total_trimmed, u64 start, u64 bytes,
3388 u64 reserved_start, u64 reserved_bytes,
3389 enum btrfs_trim_state reserved_trim_state,
3390 struct btrfs_trim_range *trim_entry)
3391{
3392 struct btrfs_space_info *space_info = block_group->space_info;
3393 struct btrfs_fs_info *fs_info = block_group->fs_info;
3394 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3395 int ret;
3396 int update = 0;
3397 const u64 end = start + bytes;
3398 const u64 reserved_end = reserved_start + reserved_bytes;
3399 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3400 u64 trimmed = 0;
3401
3402 spin_lock(&space_info->lock);
3403 spin_lock(&block_group->lock);
3404 if (!block_group->ro) {
3405 block_group->reserved += reserved_bytes;
3406 space_info->bytes_reserved += reserved_bytes;
3407 update = 1;
3408 }
3409 spin_unlock(&block_group->lock);
3410 spin_unlock(&space_info->lock);
3411
3412 ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3413 if (!ret) {
3414 *total_trimmed += trimmed;
3415 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3416 }
3417
3418 mutex_lock(&ctl->cache_writeout_mutex);
3419 if (reserved_start < start)
3420 __btrfs_add_free_space(fs_info, ctl, reserved_start,
3421 start - reserved_start,
3422 reserved_trim_state);
3423 if (start + bytes < reserved_start + reserved_bytes)
3424 __btrfs_add_free_space(fs_info, ctl, end, reserved_end - end,
3425 reserved_trim_state);
3426 __btrfs_add_free_space(fs_info, ctl, start, bytes, trim_state);
3427 list_del(&trim_entry->list);
3428 mutex_unlock(&ctl->cache_writeout_mutex);
3429
3430 if (update) {
3431 spin_lock(&space_info->lock);
3432 spin_lock(&block_group->lock);
3433 if (block_group->ro)
3434 space_info->bytes_readonly += reserved_bytes;
3435 block_group->reserved -= reserved_bytes;
3436 space_info->bytes_reserved -= reserved_bytes;
3437 spin_unlock(&block_group->lock);
3438 spin_unlock(&space_info->lock);
3439 }
3440
3441 return ret;
3442}
3443
3444/*
3445 * If @async is set, then we will trim 1 region and return.
3446 */
3447static int trim_no_bitmap(struct btrfs_block_group *block_group,
3448 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3449 bool async)
3450{
3451 struct btrfs_discard_ctl *discard_ctl =
3452 &block_group->fs_info->discard_ctl;
3453 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3454 struct btrfs_free_space *entry;
3455 struct rb_node *node;
3456 int ret = 0;
3457 u64 extent_start;
3458 u64 extent_bytes;
3459 enum btrfs_trim_state extent_trim_state;
3460 u64 bytes;
3461 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3462
3463 while (start < end) {
3464 struct btrfs_trim_range trim_entry;
3465
3466 mutex_lock(&ctl->cache_writeout_mutex);
3467 spin_lock(&ctl->tree_lock);
3468
3469 if (ctl->free_space < minlen)
3470 goto out_unlock;
3471
3472 entry = tree_search_offset(ctl, start, 0, 1);
3473 if (!entry)
3474 goto out_unlock;
3475
3476 /* Skip bitmaps and if async, already trimmed entries */
3477 while (entry->bitmap ||
3478 (async && btrfs_free_space_trimmed(entry))) {
3479 node = rb_next(&entry->offset_index);
3480 if (!node)
3481 goto out_unlock;
3482 entry = rb_entry(node, struct btrfs_free_space,
3483 offset_index);
3484 }
3485
3486 if (entry->offset >= end)
3487 goto out_unlock;
3488
3489 extent_start = entry->offset;
3490 extent_bytes = entry->bytes;
3491 extent_trim_state = entry->trim_state;
3492 if (async) {
3493 start = entry->offset;
3494 bytes = entry->bytes;
3495 if (bytes < minlen) {
3496 spin_unlock(&ctl->tree_lock);
3497 mutex_unlock(&ctl->cache_writeout_mutex);
3498 goto next;
3499 }
3500 unlink_free_space(ctl, entry);
3501 /*
3502 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3503 * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3504 * X when we come back around. So trim it now.
3505 */
3506 if (max_discard_size &&
3507 bytes >= (max_discard_size +
3508 BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3509 bytes = max_discard_size;
3510 extent_bytes = max_discard_size;
3511 entry->offset += max_discard_size;
3512 entry->bytes -= max_discard_size;
3513 link_free_space(ctl, entry);
3514 } else {
3515 kmem_cache_free(btrfs_free_space_cachep, entry);
3516 }
3517 } else {
3518 start = max(start, extent_start);
3519 bytes = min(extent_start + extent_bytes, end) - start;
3520 if (bytes < minlen) {
3521 spin_unlock(&ctl->tree_lock);
3522 mutex_unlock(&ctl->cache_writeout_mutex);
3523 goto next;
3524 }
3525
3526 unlink_free_space(ctl, entry);
3527 kmem_cache_free(btrfs_free_space_cachep, entry);
3528 }
3529
3530 spin_unlock(&ctl->tree_lock);
3531 trim_entry.start = extent_start;
3532 trim_entry.bytes = extent_bytes;
3533 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3534 mutex_unlock(&ctl->cache_writeout_mutex);
3535
3536 ret = do_trimming(block_group, total_trimmed, start, bytes,
3537 extent_start, extent_bytes, extent_trim_state,
3538 &trim_entry);
3539 if (ret) {
3540 block_group->discard_cursor = start + bytes;
3541 break;
3542 }
3543next:
3544 start += bytes;
3545 block_group->discard_cursor = start;
3546 if (async && *total_trimmed)
3547 break;
3548
3549 if (fatal_signal_pending(current)) {
3550 ret = -ERESTARTSYS;
3551 break;
3552 }
3553
3554 cond_resched();
3555 }
3556
3557 return ret;
3558
3559out_unlock:
3560 block_group->discard_cursor = btrfs_block_group_end(block_group);
3561 spin_unlock(&ctl->tree_lock);
3562 mutex_unlock(&ctl->cache_writeout_mutex);
3563
3564 return ret;
3565}
3566
3567/*
3568 * If we break out of trimming a bitmap prematurely, we should reset the
3569 * trimming bit. In a rather contrieved case, it's possible to race here so
3570 * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3571 *
3572 * start = start of bitmap
3573 * end = near end of bitmap
3574 *
3575 * Thread 1: Thread 2:
3576 * trim_bitmaps(start)
3577 * trim_bitmaps(end)
3578 * end_trimming_bitmap()
3579 * reset_trimming_bitmap()
3580 */
3581static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3582{
3583 struct btrfs_free_space *entry;
3584
3585 spin_lock(&ctl->tree_lock);
3586 entry = tree_search_offset(ctl, offset, 1, 0);
3587 if (entry) {
3588 if (btrfs_free_space_trimmed(entry)) {
3589 ctl->discardable_extents[BTRFS_STAT_CURR] +=
3590 entry->bitmap_extents;
3591 ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3592 }
3593 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3594 }
3595
3596 spin_unlock(&ctl->tree_lock);
3597}
3598
3599static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3600 struct btrfs_free_space *entry)
3601{
3602 if (btrfs_free_space_trimming_bitmap(entry)) {
3603 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3604 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3605 entry->bitmap_extents;
3606 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3607 }
3608}
3609
3610/*
3611 * If @async is set, then we will trim 1 region and return.
3612 */
3613static int trim_bitmaps(struct btrfs_block_group *block_group,
3614 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3615 u64 maxlen, bool async)
3616{
3617 struct btrfs_discard_ctl *discard_ctl =
3618 &block_group->fs_info->discard_ctl;
3619 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3620 struct btrfs_free_space *entry;
3621 int ret = 0;
3622 int ret2;
3623 u64 bytes;
3624 u64 offset = offset_to_bitmap(ctl, start);
3625 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3626
3627 while (offset < end) {
3628 bool next_bitmap = false;
3629 struct btrfs_trim_range trim_entry;
3630
3631 mutex_lock(&ctl->cache_writeout_mutex);
3632 spin_lock(&ctl->tree_lock);
3633
3634 if (ctl->free_space < minlen) {
3635 block_group->discard_cursor =
3636 btrfs_block_group_end(block_group);
3637 spin_unlock(&ctl->tree_lock);
3638 mutex_unlock(&ctl->cache_writeout_mutex);
3639 break;
3640 }
3641
3642 entry = tree_search_offset(ctl, offset, 1, 0);
3643 /*
3644 * Bitmaps are marked trimmed lossily now to prevent constant
3645 * discarding of the same bitmap (the reason why we are bound
3646 * by the filters). So, retrim the block group bitmaps when we
3647 * are preparing to punt to the unused_bgs list. This uses
3648 * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3649 * which is the only discard index which sets minlen to 0.
3650 */
3651 if (!entry || (async && minlen && start == offset &&
3652 btrfs_free_space_trimmed(entry))) {
3653 spin_unlock(&ctl->tree_lock);
3654 mutex_unlock(&ctl->cache_writeout_mutex);
3655 next_bitmap = true;
3656 goto next;
3657 }
3658
3659 /*
3660 * Async discard bitmap trimming begins at by setting the start
3661 * to be key.objectid and the offset_to_bitmap() aligns to the
3662 * start of the bitmap. This lets us know we are fully
3663 * scanning the bitmap rather than only some portion of it.
3664 */
3665 if (start == offset)
3666 entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3667
3668 bytes = minlen;
3669 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3670 if (ret2 || start >= end) {
3671 /*
3672 * We lossily consider a bitmap trimmed if we only skip
3673 * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3674 */
3675 if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3676 end_trimming_bitmap(ctl, entry);
3677 else
3678 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3679 spin_unlock(&ctl->tree_lock);
3680 mutex_unlock(&ctl->cache_writeout_mutex);
3681 next_bitmap = true;
3682 goto next;
3683 }
3684
3685 /*
3686 * We already trimmed a region, but are using the locking above
3687 * to reset the trim_state.
3688 */
3689 if (async && *total_trimmed) {
3690 spin_unlock(&ctl->tree_lock);
3691 mutex_unlock(&ctl->cache_writeout_mutex);
3692 goto out;
3693 }
3694
3695 bytes = min(bytes, end - start);
3696 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3697 spin_unlock(&ctl->tree_lock);
3698 mutex_unlock(&ctl->cache_writeout_mutex);
3699 goto next;
3700 }
3701
3702 /*
3703 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3704 * If X < @minlen, we won't trim X when we come back around.
3705 * So trim it now. We differ here from trimming extents as we
3706 * don't keep individual state per bit.
3707 */
3708 if (async &&
3709 max_discard_size &&
3710 bytes > (max_discard_size + minlen))
3711 bytes = max_discard_size;
3712
3713 bitmap_clear_bits(ctl, entry, start, bytes);
3714 if (entry->bytes == 0)
3715 free_bitmap(ctl, entry);
3716
3717 spin_unlock(&ctl->tree_lock);
3718 trim_entry.start = start;
3719 trim_entry.bytes = bytes;
3720 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3721 mutex_unlock(&ctl->cache_writeout_mutex);
3722
3723 ret = do_trimming(block_group, total_trimmed, start, bytes,
3724 start, bytes, 0, &trim_entry);
3725 if (ret) {
3726 reset_trimming_bitmap(ctl, offset);
3727 block_group->discard_cursor =
3728 btrfs_block_group_end(block_group);
3729 break;
3730 }
3731next:
3732 if (next_bitmap) {
3733 offset += BITS_PER_BITMAP * ctl->unit;
3734 start = offset;
3735 } else {
3736 start += bytes;
3737 }
3738 block_group->discard_cursor = start;
3739
3740 if (fatal_signal_pending(current)) {
3741 if (start != offset)
3742 reset_trimming_bitmap(ctl, offset);
3743 ret = -ERESTARTSYS;
3744 break;
3745 }
3746
3747 cond_resched();
3748 }
3749
3750 if (offset >= end)
3751 block_group->discard_cursor = end;
3752
3753out:
3754 return ret;
3755}
3756
3757int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3758 u64 *trimmed, u64 start, u64 end, u64 minlen)
3759{
3760 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3761 int ret;
3762 u64 rem = 0;
3763
3764 *trimmed = 0;
3765
3766 spin_lock(&block_group->lock);
3767 if (block_group->removed) {
3768 spin_unlock(&block_group->lock);
3769 return 0;
3770 }
3771 btrfs_freeze_block_group(block_group);
3772 spin_unlock(&block_group->lock);
3773
3774 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
3775 if (ret)
3776 goto out;
3777
3778 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
3779 div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
3780 /* If we ended in the middle of a bitmap, reset the trimming flag */
3781 if (rem)
3782 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
3783out:
3784 btrfs_unfreeze_block_group(block_group);
3785 return ret;
3786}
3787
3788int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
3789 u64 *trimmed, u64 start, u64 end, u64 minlen,
3790 bool async)
3791{
3792 int ret;
3793
3794 *trimmed = 0;
3795
3796 spin_lock(&block_group->lock);
3797 if (block_group->removed) {
3798 spin_unlock(&block_group->lock);
3799 return 0;
3800 }
3801 btrfs_freeze_block_group(block_group);
3802 spin_unlock(&block_group->lock);
3803
3804 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
3805 btrfs_unfreeze_block_group(block_group);
3806
3807 return ret;
3808}
3809
3810int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
3811 u64 *trimmed, u64 start, u64 end, u64 minlen,
3812 u64 maxlen, bool async)
3813{
3814 int ret;
3815
3816 *trimmed = 0;
3817
3818 spin_lock(&block_group->lock);
3819 if (block_group->removed) {
3820 spin_unlock(&block_group->lock);
3821 return 0;
3822 }
3823 btrfs_freeze_block_group(block_group);
3824 spin_unlock(&block_group->lock);
3825
3826 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
3827 async);
3828
3829 btrfs_unfreeze_block_group(block_group);
3830
3831 return ret;
3832}
3833
3834/*
3835 * Find the left-most item in the cache tree, and then return the
3836 * smallest inode number in the item.
3837 *
3838 * Note: the returned inode number may not be the smallest one in
3839 * the tree, if the left-most item is a bitmap.
3840 */
3841u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3842{
3843 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3844 struct btrfs_free_space *entry = NULL;
3845 u64 ino = 0;
3846
3847 spin_lock(&ctl->tree_lock);
3848
3849 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3850 goto out;
3851
3852 entry = rb_entry(rb_first(&ctl->free_space_offset),
3853 struct btrfs_free_space, offset_index);
3854
3855 if (!entry->bitmap) {
3856 ino = entry->offset;
3857
3858 unlink_free_space(ctl, entry);
3859 entry->offset++;
3860 entry->bytes--;
3861 if (!entry->bytes)
3862 kmem_cache_free(btrfs_free_space_cachep, entry);
3863 else
3864 link_free_space(ctl, entry);
3865 } else {
3866 u64 offset = 0;
3867 u64 count = 1;
3868 int ret;
3869
3870 ret = search_bitmap(ctl, entry, &offset, &count, true);
3871 /* Logic error; Should be empty if it can't find anything */
3872 ASSERT(!ret);
3873
3874 ino = offset;
3875 bitmap_clear_bits(ctl, entry, offset, 1);
3876 if (entry->bytes == 0)
3877 free_bitmap(ctl, entry);
3878 }
3879out:
3880 spin_unlock(&ctl->tree_lock);
3881
3882 return ino;
3883}
3884
3885struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3886 struct btrfs_path *path)
3887{
3888 struct inode *inode = NULL;
3889
3890 spin_lock(&root->ino_cache_lock);
3891 if (root->ino_cache_inode)
3892 inode = igrab(root->ino_cache_inode);
3893 spin_unlock(&root->ino_cache_lock);
3894 if (inode)
3895 return inode;
3896
3897 inode = __lookup_free_space_inode(root, path, 0);
3898 if (IS_ERR(inode))
3899 return inode;
3900
3901 spin_lock(&root->ino_cache_lock);
3902 if (!btrfs_fs_closing(root->fs_info))
3903 root->ino_cache_inode = igrab(inode);
3904 spin_unlock(&root->ino_cache_lock);
3905
3906 return inode;
3907}
3908
3909int create_free_ino_inode(struct btrfs_root *root,
3910 struct btrfs_trans_handle *trans,
3911 struct btrfs_path *path)
3912{
3913 return __create_free_space_inode(root, trans, path,
3914 BTRFS_FREE_INO_OBJECTID, 0);
3915}
3916
3917int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3918{
3919 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3920 struct btrfs_path *path;
3921 struct inode *inode;
3922 int ret = 0;
3923 u64 root_gen = btrfs_root_generation(&root->root_item);
3924
3925 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3926 return 0;
3927
3928 /*
3929 * If we're unmounting then just return, since this does a search on the
3930 * normal root and not the commit root and we could deadlock.
3931 */
3932 if (btrfs_fs_closing(fs_info))
3933 return 0;
3934
3935 path = btrfs_alloc_path();
3936 if (!path)
3937 return 0;
3938
3939 inode = lookup_free_ino_inode(root, path);
3940 if (IS_ERR(inode))
3941 goto out;
3942
3943 if (root_gen != BTRFS_I(inode)->generation)
3944 goto out_put;
3945
3946 ret = __load_free_space_cache(root, inode, ctl, path, 0);
3947
3948 if (ret < 0)
3949 btrfs_err(fs_info,
3950 "failed to load free ino cache for root %llu",
3951 root->root_key.objectid);
3952out_put:
3953 iput(inode);
3954out:
3955 btrfs_free_path(path);
3956 return ret;
3957}
3958
3959int btrfs_write_out_ino_cache(struct btrfs_root *root,
3960 struct btrfs_trans_handle *trans,
3961 struct btrfs_path *path,
3962 struct inode *inode)
3963{
3964 struct btrfs_fs_info *fs_info = root->fs_info;
3965 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3966 int ret;
3967 struct btrfs_io_ctl io_ctl;
3968 bool release_metadata = true;
3969
3970 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3971 return 0;
3972
3973 memset(&io_ctl, 0, sizeof(io_ctl));
3974 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3975 if (!ret) {
3976 /*
3977 * At this point writepages() didn't error out, so our metadata
3978 * reservation is released when the writeback finishes, at
3979 * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3980 * with or without an error.
3981 */
3982 release_metadata = false;
3983 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3984 }
3985
3986 if (ret) {
3987 if (release_metadata)
3988 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3989 inode->i_size, true);
3990 btrfs_debug(fs_info,
3991 "failed to write free ino cache for root %llu error %d",
3992 root->root_key.objectid, ret);
3993 }
3994
3995 return ret;
3996}
3997
3998#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3999/*
4000 * Use this if you need to make a bitmap or extent entry specifically, it
4001 * doesn't do any of the merging that add_free_space does, this acts a lot like
4002 * how the free space cache loading stuff works, so you can get really weird
4003 * configurations.
4004 */
4005int test_add_free_space_entry(struct btrfs_block_group *cache,
4006 u64 offset, u64 bytes, bool bitmap)
4007{
4008 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4009 struct btrfs_free_space *info = NULL, *bitmap_info;
4010 void *map = NULL;
4011 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4012 u64 bytes_added;
4013 int ret;
4014
4015again:
4016 if (!info) {
4017 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4018 if (!info)
4019 return -ENOMEM;
4020 }
4021
4022 if (!bitmap) {
4023 spin_lock(&ctl->tree_lock);
4024 info->offset = offset;
4025 info->bytes = bytes;
4026 info->max_extent_size = 0;
4027 ret = link_free_space(ctl, info);
4028 spin_unlock(&ctl->tree_lock);
4029 if (ret)
4030 kmem_cache_free(btrfs_free_space_cachep, info);
4031 return ret;
4032 }
4033
4034 if (!map) {
4035 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4036 if (!map) {
4037 kmem_cache_free(btrfs_free_space_cachep, info);
4038 return -ENOMEM;
4039 }
4040 }
4041
4042 spin_lock(&ctl->tree_lock);
4043 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4044 1, 0);
4045 if (!bitmap_info) {
4046 info->bitmap = map;
4047 map = NULL;
4048 add_new_bitmap(ctl, info, offset);
4049 bitmap_info = info;
4050 info = NULL;
4051 }
4052
4053 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4054 trim_state);
4055
4056 bytes -= bytes_added;
4057 offset += bytes_added;
4058 spin_unlock(&ctl->tree_lock);
4059
4060 if (bytes)
4061 goto again;
4062
4063 if (info)
4064 kmem_cache_free(btrfs_free_space_cachep, info);
4065 if (map)
4066 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4067 return 0;
4068}
4069
4070/*
4071 * Checks to see if the given range is in the free space cache. This is really
4072 * just used to check the absence of space, so if there is free space in the
4073 * range at all we will return 1.
4074 */
4075int test_check_exists(struct btrfs_block_group *cache,
4076 u64 offset, u64 bytes)
4077{
4078 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4079 struct btrfs_free_space *info;
4080 int ret = 0;
4081
4082 spin_lock(&ctl->tree_lock);
4083 info = tree_search_offset(ctl, offset, 0, 0);
4084 if (!info) {
4085 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4086 1, 0);
4087 if (!info)
4088 goto out;
4089 }
4090
4091have_info:
4092 if (info->bitmap) {
4093 u64 bit_off, bit_bytes;
4094 struct rb_node *n;
4095 struct btrfs_free_space *tmp;
4096
4097 bit_off = offset;
4098 bit_bytes = ctl->unit;
4099 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4100 if (!ret) {
4101 if (bit_off == offset) {
4102 ret = 1;
4103 goto out;
4104 } else if (bit_off > offset &&
4105 offset + bytes > bit_off) {
4106 ret = 1;
4107 goto out;
4108 }
4109 }
4110
4111 n = rb_prev(&info->offset_index);
4112 while (n) {
4113 tmp = rb_entry(n, struct btrfs_free_space,
4114 offset_index);
4115 if (tmp->offset + tmp->bytes < offset)
4116 break;
4117 if (offset + bytes < tmp->offset) {
4118 n = rb_prev(&tmp->offset_index);
4119 continue;
4120 }
4121 info = tmp;
4122 goto have_info;
4123 }
4124
4125 n = rb_next(&info->offset_index);
4126 while (n) {
4127 tmp = rb_entry(n, struct btrfs_free_space,
4128 offset_index);
4129 if (offset + bytes < tmp->offset)
4130 break;
4131 if (tmp->offset + tmp->bytes < offset) {
4132 n = rb_next(&tmp->offset_index);
4133 continue;
4134 }
4135 info = tmp;
4136 goto have_info;
4137 }
4138
4139 ret = 0;
4140 goto out;
4141 }
4142
4143 if (info->offset == offset) {
4144 ret = 1;
4145 goto out;
4146 }
4147
4148 if (offset > info->offset && offset < info->offset + info->bytes)
4149 ret = 1;
4150out:
4151 spin_unlock(&ctl->tree_lock);
4152 return ret;
4153}
4154#endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */