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 "ctree.h"
15#include "fs.h"
16#include "messages.h"
17#include "misc.h"
18#include "free-space-cache.h"
19#include "transaction.h"
20#include "disk-io.h"
21#include "extent_io.h"
22#include "volumes.h"
23#include "space-info.h"
24#include "delalloc-space.h"
25#include "block-group.h"
26#include "discard.h"
27#include "subpage.h"
28#include "inode-item.h"
29#include "accessors.h"
30#include "file-item.h"
31#include "file.h"
32#include "super.h"
33
34#define BITS_PER_BITMAP (PAGE_SIZE * 8UL)
35#define MAX_CACHE_BYTES_PER_GIG SZ_64K
36#define FORCE_EXTENT_THRESHOLD SZ_1M
37
38static struct kmem_cache *btrfs_free_space_cachep;
39static struct kmem_cache *btrfs_free_space_bitmap_cachep;
40
41struct btrfs_trim_range {
42 u64 start;
43 u64 bytes;
44 struct list_head list;
45};
46
47static int link_free_space(struct btrfs_free_space_ctl *ctl,
48 struct btrfs_free_space *info);
49static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
50 struct btrfs_free_space *info, bool update_stat);
51static int search_bitmap(struct btrfs_free_space_ctl *ctl,
52 struct btrfs_free_space *bitmap_info, u64 *offset,
53 u64 *bytes, bool for_alloc);
54static void free_bitmap(struct btrfs_free_space_ctl *ctl,
55 struct btrfs_free_space *bitmap_info);
56static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
57 struct btrfs_free_space *info, u64 offset,
58 u64 bytes, bool update_stats);
59
60static void btrfs_crc32c_final(u32 crc, u8 *result)
61{
62 put_unaligned_le32(~crc, result);
63}
64
65static void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
66{
67 struct btrfs_free_space *info;
68 struct rb_node *node;
69
70 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
71 info = rb_entry(node, struct btrfs_free_space, offset_index);
72 if (!info->bitmap) {
73 unlink_free_space(ctl, info, true);
74 kmem_cache_free(btrfs_free_space_cachep, info);
75 } else {
76 free_bitmap(ctl, info);
77 }
78
79 cond_resched_lock(&ctl->tree_lock);
80 }
81}
82
83static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
84 struct btrfs_path *path,
85 u64 offset)
86{
87 struct btrfs_fs_info *fs_info = root->fs_info;
88 struct btrfs_key key;
89 struct btrfs_key location;
90 struct btrfs_disk_key disk_key;
91 struct btrfs_free_space_header *header;
92 struct extent_buffer *leaf;
93 struct inode *inode = NULL;
94 unsigned nofs_flag;
95 int ret;
96
97 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
98 key.offset = offset;
99 key.type = 0;
100
101 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
102 if (ret < 0)
103 return ERR_PTR(ret);
104 if (ret > 0) {
105 btrfs_release_path(path);
106 return ERR_PTR(-ENOENT);
107 }
108
109 leaf = path->nodes[0];
110 header = btrfs_item_ptr(leaf, path->slots[0],
111 struct btrfs_free_space_header);
112 btrfs_free_space_key(leaf, header, &disk_key);
113 btrfs_disk_key_to_cpu(&location, &disk_key);
114 btrfs_release_path(path);
115
116 /*
117 * We are often under a trans handle at this point, so we need to make
118 * sure NOFS is set to keep us from deadlocking.
119 */
120 nofs_flag = memalloc_nofs_save();
121 inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
122 btrfs_release_path(path);
123 memalloc_nofs_restore(nofs_flag);
124 if (IS_ERR(inode))
125 return inode;
126
127 mapping_set_gfp_mask(inode->i_mapping,
128 mapping_gfp_constraint(inode->i_mapping,
129 ~(__GFP_FS | __GFP_HIGHMEM)));
130
131 return inode;
132}
133
134struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
135 struct btrfs_path *path)
136{
137 struct btrfs_fs_info *fs_info = block_group->fs_info;
138 struct inode *inode = NULL;
139 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
140
141 spin_lock(&block_group->lock);
142 if (block_group->inode)
143 inode = igrab(block_group->inode);
144 spin_unlock(&block_group->lock);
145 if (inode)
146 return inode;
147
148 inode = __lookup_free_space_inode(fs_info->tree_root, path,
149 block_group->start);
150 if (IS_ERR(inode))
151 return inode;
152
153 spin_lock(&block_group->lock);
154 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
155 btrfs_info(fs_info, "Old style space inode found, converting.");
156 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
157 BTRFS_INODE_NODATACOW;
158 block_group->disk_cache_state = BTRFS_DC_CLEAR;
159 }
160
161 if (!test_and_set_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags))
162 block_group->inode = igrab(inode);
163 spin_unlock(&block_group->lock);
164
165 return inode;
166}
167
168static int __create_free_space_inode(struct btrfs_root *root,
169 struct btrfs_trans_handle *trans,
170 struct btrfs_path *path,
171 u64 ino, u64 offset)
172{
173 struct btrfs_key key;
174 struct btrfs_disk_key disk_key;
175 struct btrfs_free_space_header *header;
176 struct btrfs_inode_item *inode_item;
177 struct extent_buffer *leaf;
178 /* We inline CRCs for the free disk space cache */
179 const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
180 BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
181 int ret;
182
183 ret = btrfs_insert_empty_inode(trans, root, path, ino);
184 if (ret)
185 return ret;
186
187 leaf = path->nodes[0];
188 inode_item = btrfs_item_ptr(leaf, path->slots[0],
189 struct btrfs_inode_item);
190 btrfs_item_key(leaf, &disk_key, path->slots[0]);
191 memzero_extent_buffer(leaf, (unsigned long)inode_item,
192 sizeof(*inode_item));
193 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
194 btrfs_set_inode_size(leaf, inode_item, 0);
195 btrfs_set_inode_nbytes(leaf, inode_item, 0);
196 btrfs_set_inode_uid(leaf, inode_item, 0);
197 btrfs_set_inode_gid(leaf, inode_item, 0);
198 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
199 btrfs_set_inode_flags(leaf, inode_item, flags);
200 btrfs_set_inode_nlink(leaf, inode_item, 1);
201 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
202 btrfs_set_inode_block_group(leaf, inode_item, offset);
203 btrfs_mark_buffer_dirty(trans, leaf);
204 btrfs_release_path(path);
205
206 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
207 key.offset = offset;
208 key.type = 0;
209 ret = btrfs_insert_empty_item(trans, root, path, &key,
210 sizeof(struct btrfs_free_space_header));
211 if (ret < 0) {
212 btrfs_release_path(path);
213 return ret;
214 }
215
216 leaf = path->nodes[0];
217 header = btrfs_item_ptr(leaf, path->slots[0],
218 struct btrfs_free_space_header);
219 memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
220 btrfs_set_free_space_key(leaf, header, &disk_key);
221 btrfs_mark_buffer_dirty(trans, leaf);
222 btrfs_release_path(path);
223
224 return 0;
225}
226
227int create_free_space_inode(struct btrfs_trans_handle *trans,
228 struct btrfs_block_group *block_group,
229 struct btrfs_path *path)
230{
231 int ret;
232 u64 ino;
233
234 ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
235 if (ret < 0)
236 return ret;
237
238 return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
239 ino, block_group->start);
240}
241
242/*
243 * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
244 * handles lookup, otherwise it takes ownership and iputs the inode.
245 * Don't reuse an inode pointer after passing it into this function.
246 */
247int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
248 struct inode *inode,
249 struct btrfs_block_group *block_group)
250{
251 struct btrfs_path *path;
252 struct btrfs_key key;
253 int ret = 0;
254
255 path = btrfs_alloc_path();
256 if (!path)
257 return -ENOMEM;
258
259 if (!inode)
260 inode = lookup_free_space_inode(block_group, path);
261 if (IS_ERR(inode)) {
262 if (PTR_ERR(inode) != -ENOENT)
263 ret = PTR_ERR(inode);
264 goto out;
265 }
266 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
267 if (ret) {
268 btrfs_add_delayed_iput(BTRFS_I(inode));
269 goto out;
270 }
271 clear_nlink(inode);
272 /* One for the block groups ref */
273 spin_lock(&block_group->lock);
274 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags)) {
275 block_group->inode = NULL;
276 spin_unlock(&block_group->lock);
277 iput(inode);
278 } else {
279 spin_unlock(&block_group->lock);
280 }
281 /* One for the lookup ref */
282 btrfs_add_delayed_iput(BTRFS_I(inode));
283
284 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
285 key.type = 0;
286 key.offset = block_group->start;
287 ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
288 -1, 1);
289 if (ret) {
290 if (ret > 0)
291 ret = 0;
292 goto out;
293 }
294 ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
295out:
296 btrfs_free_path(path);
297 return ret;
298}
299
300int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
301 struct btrfs_block_group *block_group,
302 struct inode *vfs_inode)
303{
304 struct btrfs_truncate_control control = {
305 .inode = BTRFS_I(vfs_inode),
306 .new_size = 0,
307 .ino = btrfs_ino(BTRFS_I(vfs_inode)),
308 .min_type = BTRFS_EXTENT_DATA_KEY,
309 .clear_extent_range = true,
310 };
311 struct btrfs_inode *inode = BTRFS_I(vfs_inode);
312 struct btrfs_root *root = inode->root;
313 struct extent_state *cached_state = NULL;
314 int ret = 0;
315 bool locked = false;
316
317 if (block_group) {
318 struct btrfs_path *path = btrfs_alloc_path();
319
320 if (!path) {
321 ret = -ENOMEM;
322 goto fail;
323 }
324 locked = true;
325 mutex_lock(&trans->transaction->cache_write_mutex);
326 if (!list_empty(&block_group->io_list)) {
327 list_del_init(&block_group->io_list);
328
329 btrfs_wait_cache_io(trans, block_group, path);
330 btrfs_put_block_group(block_group);
331 }
332
333 /*
334 * now that we've truncated the cache away, its no longer
335 * setup or written
336 */
337 spin_lock(&block_group->lock);
338 block_group->disk_cache_state = BTRFS_DC_CLEAR;
339 spin_unlock(&block_group->lock);
340 btrfs_free_path(path);
341 }
342
343 btrfs_i_size_write(inode, 0);
344 truncate_pagecache(vfs_inode, 0);
345
346 lock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
347 btrfs_drop_extent_map_range(inode, 0, (u64)-1, false);
348
349 /*
350 * We skip the throttling logic for free space cache inodes, so we don't
351 * need to check for -EAGAIN.
352 */
353 ret = btrfs_truncate_inode_items(trans, root, &control);
354
355 inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
356 btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
357
358 unlock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
359 if (ret)
360 goto fail;
361
362 ret = btrfs_update_inode(trans, inode);
363
364fail:
365 if (locked)
366 mutex_unlock(&trans->transaction->cache_write_mutex);
367 if (ret)
368 btrfs_abort_transaction(trans, ret);
369
370 return ret;
371}
372
373static void readahead_cache(struct inode *inode)
374{
375 struct file_ra_state ra;
376 unsigned long last_index;
377
378 file_ra_state_init(&ra, inode->i_mapping);
379 last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
380
381 page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
382}
383
384static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
385 int write)
386{
387 int num_pages;
388
389 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
390
391 /* Make sure we can fit our crcs and generation into the first page */
392 if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
393 return -ENOSPC;
394
395 memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
396
397 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
398 if (!io_ctl->pages)
399 return -ENOMEM;
400
401 io_ctl->num_pages = num_pages;
402 io_ctl->fs_info = btrfs_sb(inode->i_sb);
403 io_ctl->inode = inode;
404
405 return 0;
406}
407ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
408
409static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
410{
411 kfree(io_ctl->pages);
412 io_ctl->pages = NULL;
413}
414
415static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
416{
417 if (io_ctl->cur) {
418 io_ctl->cur = NULL;
419 io_ctl->orig = NULL;
420 }
421}
422
423static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
424{
425 ASSERT(io_ctl->index < io_ctl->num_pages);
426 io_ctl->page = io_ctl->pages[io_ctl->index++];
427 io_ctl->cur = page_address(io_ctl->page);
428 io_ctl->orig = io_ctl->cur;
429 io_ctl->size = PAGE_SIZE;
430 if (clear)
431 clear_page(io_ctl->cur);
432}
433
434static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
435{
436 int i;
437
438 io_ctl_unmap_page(io_ctl);
439
440 for (i = 0; i < io_ctl->num_pages; i++) {
441 if (io_ctl->pages[i]) {
442 btrfs_folio_clear_checked(io_ctl->fs_info,
443 page_folio(io_ctl->pages[i]),
444 page_offset(io_ctl->pages[i]),
445 PAGE_SIZE);
446 unlock_page(io_ctl->pages[i]);
447 put_page(io_ctl->pages[i]);
448 }
449 }
450}
451
452static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
453{
454 struct page *page;
455 struct inode *inode = io_ctl->inode;
456 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
457 int i;
458
459 for (i = 0; i < io_ctl->num_pages; i++) {
460 int ret;
461
462 page = find_or_create_page(inode->i_mapping, i, mask);
463 if (!page) {
464 io_ctl_drop_pages(io_ctl);
465 return -ENOMEM;
466 }
467
468 ret = set_page_extent_mapped(page);
469 if (ret < 0) {
470 unlock_page(page);
471 put_page(page);
472 io_ctl_drop_pages(io_ctl);
473 return ret;
474 }
475
476 io_ctl->pages[i] = page;
477 if (uptodate && !PageUptodate(page)) {
478 btrfs_read_folio(NULL, page_folio(page));
479 lock_page(page);
480 if (page->mapping != inode->i_mapping) {
481 btrfs_err(BTRFS_I(inode)->root->fs_info,
482 "free space cache page truncated");
483 io_ctl_drop_pages(io_ctl);
484 return -EIO;
485 }
486 if (!PageUptodate(page)) {
487 btrfs_err(BTRFS_I(inode)->root->fs_info,
488 "error reading free space cache");
489 io_ctl_drop_pages(io_ctl);
490 return -EIO;
491 }
492 }
493 }
494
495 for (i = 0; i < io_ctl->num_pages; i++)
496 clear_page_dirty_for_io(io_ctl->pages[i]);
497
498 return 0;
499}
500
501static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
502{
503 io_ctl_map_page(io_ctl, 1);
504
505 /*
506 * Skip the csum areas. If we don't check crcs then we just have a
507 * 64bit chunk at the front of the first page.
508 */
509 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
510 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
511
512 put_unaligned_le64(generation, io_ctl->cur);
513 io_ctl->cur += sizeof(u64);
514}
515
516static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
517{
518 u64 cache_gen;
519
520 /*
521 * Skip the crc area. If we don't check crcs then we just have a 64bit
522 * chunk at the front of the first page.
523 */
524 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
525 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
526
527 cache_gen = get_unaligned_le64(io_ctl->cur);
528 if (cache_gen != generation) {
529 btrfs_err_rl(io_ctl->fs_info,
530 "space cache generation (%llu) does not match inode (%llu)",
531 cache_gen, generation);
532 io_ctl_unmap_page(io_ctl);
533 return -EIO;
534 }
535 io_ctl->cur += sizeof(u64);
536 return 0;
537}
538
539static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
540{
541 u32 *tmp;
542 u32 crc = ~(u32)0;
543 unsigned offset = 0;
544
545 if (index == 0)
546 offset = sizeof(u32) * io_ctl->num_pages;
547
548 crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
549 btrfs_crc32c_final(crc, (u8 *)&crc);
550 io_ctl_unmap_page(io_ctl);
551 tmp = page_address(io_ctl->pages[0]);
552 tmp += index;
553 *tmp = crc;
554}
555
556static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
557{
558 u32 *tmp, val;
559 u32 crc = ~(u32)0;
560 unsigned offset = 0;
561
562 if (index == 0)
563 offset = sizeof(u32) * io_ctl->num_pages;
564
565 tmp = page_address(io_ctl->pages[0]);
566 tmp += index;
567 val = *tmp;
568
569 io_ctl_map_page(io_ctl, 0);
570 crc = crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
571 btrfs_crc32c_final(crc, (u8 *)&crc);
572 if (val != crc) {
573 btrfs_err_rl(io_ctl->fs_info,
574 "csum mismatch on free space cache");
575 io_ctl_unmap_page(io_ctl);
576 return -EIO;
577 }
578
579 return 0;
580}
581
582static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
583 void *bitmap)
584{
585 struct btrfs_free_space_entry *entry;
586
587 if (!io_ctl->cur)
588 return -ENOSPC;
589
590 entry = io_ctl->cur;
591 put_unaligned_le64(offset, &entry->offset);
592 put_unaligned_le64(bytes, &entry->bytes);
593 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
594 BTRFS_FREE_SPACE_EXTENT;
595 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
596 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
597
598 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
599 return 0;
600
601 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
602
603 /* No more pages to map */
604 if (io_ctl->index >= io_ctl->num_pages)
605 return 0;
606
607 /* map the next page */
608 io_ctl_map_page(io_ctl, 1);
609 return 0;
610}
611
612static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
613{
614 if (!io_ctl->cur)
615 return -ENOSPC;
616
617 /*
618 * If we aren't at the start of the current page, unmap this one and
619 * map the next one if there is any left.
620 */
621 if (io_ctl->cur != io_ctl->orig) {
622 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
623 if (io_ctl->index >= io_ctl->num_pages)
624 return -ENOSPC;
625 io_ctl_map_page(io_ctl, 0);
626 }
627
628 copy_page(io_ctl->cur, bitmap);
629 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
630 if (io_ctl->index < io_ctl->num_pages)
631 io_ctl_map_page(io_ctl, 0);
632 return 0;
633}
634
635static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
636{
637 /*
638 * If we're not on the boundary we know we've modified the page and we
639 * need to crc the page.
640 */
641 if (io_ctl->cur != io_ctl->orig)
642 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
643 else
644 io_ctl_unmap_page(io_ctl);
645
646 while (io_ctl->index < io_ctl->num_pages) {
647 io_ctl_map_page(io_ctl, 1);
648 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
649 }
650}
651
652static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
653 struct btrfs_free_space *entry, u8 *type)
654{
655 struct btrfs_free_space_entry *e;
656 int ret;
657
658 if (!io_ctl->cur) {
659 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
660 if (ret)
661 return ret;
662 }
663
664 e = io_ctl->cur;
665 entry->offset = get_unaligned_le64(&e->offset);
666 entry->bytes = get_unaligned_le64(&e->bytes);
667 *type = e->type;
668 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
669 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
670
671 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
672 return 0;
673
674 io_ctl_unmap_page(io_ctl);
675
676 return 0;
677}
678
679static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
680 struct btrfs_free_space *entry)
681{
682 int ret;
683
684 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
685 if (ret)
686 return ret;
687
688 copy_page(entry->bitmap, io_ctl->cur);
689 io_ctl_unmap_page(io_ctl);
690
691 return 0;
692}
693
694static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
695{
696 struct btrfs_block_group *block_group = ctl->block_group;
697 u64 max_bytes;
698 u64 bitmap_bytes;
699 u64 extent_bytes;
700 u64 size = block_group->length;
701 u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
702 u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
703
704 max_bitmaps = max_t(u64, max_bitmaps, 1);
705
706 if (ctl->total_bitmaps > max_bitmaps)
707 btrfs_err(block_group->fs_info,
708"invalid free space control: bg start=%llu len=%llu total_bitmaps=%u unit=%u max_bitmaps=%llu bytes_per_bg=%llu",
709 block_group->start, block_group->length,
710 ctl->total_bitmaps, ctl->unit, max_bitmaps,
711 bytes_per_bg);
712 ASSERT(ctl->total_bitmaps <= max_bitmaps);
713
714 /*
715 * We are trying to keep the total amount of memory used per 1GiB of
716 * space to be MAX_CACHE_BYTES_PER_GIG. However, with a reclamation
717 * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
718 * bitmaps, we may end up using more memory than this.
719 */
720 if (size < SZ_1G)
721 max_bytes = MAX_CACHE_BYTES_PER_GIG;
722 else
723 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
724
725 bitmap_bytes = ctl->total_bitmaps * ctl->unit;
726
727 /*
728 * we want the extent entry threshold to always be at most 1/2 the max
729 * bytes we can have, or whatever is less than that.
730 */
731 extent_bytes = max_bytes - bitmap_bytes;
732 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
733
734 ctl->extents_thresh =
735 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
736}
737
738static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
739 struct btrfs_free_space_ctl *ctl,
740 struct btrfs_path *path, u64 offset)
741{
742 struct btrfs_fs_info *fs_info = root->fs_info;
743 struct btrfs_free_space_header *header;
744 struct extent_buffer *leaf;
745 struct btrfs_io_ctl io_ctl;
746 struct btrfs_key key;
747 struct btrfs_free_space *e, *n;
748 LIST_HEAD(bitmaps);
749 u64 num_entries;
750 u64 num_bitmaps;
751 u64 generation;
752 u8 type;
753 int ret = 0;
754
755 /* Nothing in the space cache, goodbye */
756 if (!i_size_read(inode))
757 return 0;
758
759 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
760 key.offset = offset;
761 key.type = 0;
762
763 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
764 if (ret < 0)
765 return 0;
766 else if (ret > 0) {
767 btrfs_release_path(path);
768 return 0;
769 }
770
771 ret = -1;
772
773 leaf = path->nodes[0];
774 header = btrfs_item_ptr(leaf, path->slots[0],
775 struct btrfs_free_space_header);
776 num_entries = btrfs_free_space_entries(leaf, header);
777 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
778 generation = btrfs_free_space_generation(leaf, header);
779 btrfs_release_path(path);
780
781 if (!BTRFS_I(inode)->generation) {
782 btrfs_info(fs_info,
783 "the free space cache file (%llu) is invalid, skip it",
784 offset);
785 return 0;
786 }
787
788 if (BTRFS_I(inode)->generation != generation) {
789 btrfs_err(fs_info,
790 "free space inode generation (%llu) did not match free space cache generation (%llu)",
791 BTRFS_I(inode)->generation, generation);
792 return 0;
793 }
794
795 if (!num_entries)
796 return 0;
797
798 ret = io_ctl_init(&io_ctl, inode, 0);
799 if (ret)
800 return ret;
801
802 readahead_cache(inode);
803
804 ret = io_ctl_prepare_pages(&io_ctl, true);
805 if (ret)
806 goto out;
807
808 ret = io_ctl_check_crc(&io_ctl, 0);
809 if (ret)
810 goto free_cache;
811
812 ret = io_ctl_check_generation(&io_ctl, generation);
813 if (ret)
814 goto free_cache;
815
816 while (num_entries) {
817 e = kmem_cache_zalloc(btrfs_free_space_cachep,
818 GFP_NOFS);
819 if (!e) {
820 ret = -ENOMEM;
821 goto free_cache;
822 }
823
824 ret = io_ctl_read_entry(&io_ctl, e, &type);
825 if (ret) {
826 kmem_cache_free(btrfs_free_space_cachep, e);
827 goto free_cache;
828 }
829
830 if (!e->bytes) {
831 ret = -1;
832 kmem_cache_free(btrfs_free_space_cachep, e);
833 goto free_cache;
834 }
835
836 if (type == BTRFS_FREE_SPACE_EXTENT) {
837 spin_lock(&ctl->tree_lock);
838 ret = link_free_space(ctl, e);
839 spin_unlock(&ctl->tree_lock);
840 if (ret) {
841 btrfs_err(fs_info,
842 "Duplicate entries in free space cache, dumping");
843 kmem_cache_free(btrfs_free_space_cachep, e);
844 goto free_cache;
845 }
846 } else {
847 ASSERT(num_bitmaps);
848 num_bitmaps--;
849 e->bitmap = kmem_cache_zalloc(
850 btrfs_free_space_bitmap_cachep, GFP_NOFS);
851 if (!e->bitmap) {
852 ret = -ENOMEM;
853 kmem_cache_free(
854 btrfs_free_space_cachep, e);
855 goto free_cache;
856 }
857 spin_lock(&ctl->tree_lock);
858 ret = link_free_space(ctl, e);
859 if (ret) {
860 spin_unlock(&ctl->tree_lock);
861 btrfs_err(fs_info,
862 "Duplicate entries in free space cache, dumping");
863 kmem_cache_free(btrfs_free_space_cachep, e);
864 goto free_cache;
865 }
866 ctl->total_bitmaps++;
867 recalculate_thresholds(ctl);
868 spin_unlock(&ctl->tree_lock);
869 list_add_tail(&e->list, &bitmaps);
870 }
871
872 num_entries--;
873 }
874
875 io_ctl_unmap_page(&io_ctl);
876
877 /*
878 * We add the bitmaps at the end of the entries in order that
879 * the bitmap entries are added to the cache.
880 */
881 list_for_each_entry_safe(e, n, &bitmaps, list) {
882 list_del_init(&e->list);
883 ret = io_ctl_read_bitmap(&io_ctl, e);
884 if (ret)
885 goto free_cache;
886 }
887
888 io_ctl_drop_pages(&io_ctl);
889 ret = 1;
890out:
891 io_ctl_free(&io_ctl);
892 return ret;
893free_cache:
894 io_ctl_drop_pages(&io_ctl);
895
896 spin_lock(&ctl->tree_lock);
897 __btrfs_remove_free_space_cache(ctl);
898 spin_unlock(&ctl->tree_lock);
899 goto out;
900}
901
902static int copy_free_space_cache(struct btrfs_block_group *block_group,
903 struct btrfs_free_space_ctl *ctl)
904{
905 struct btrfs_free_space *info;
906 struct rb_node *n;
907 int ret = 0;
908
909 while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
910 info = rb_entry(n, struct btrfs_free_space, offset_index);
911 if (!info->bitmap) {
912 const u64 offset = info->offset;
913 const u64 bytes = info->bytes;
914
915 unlink_free_space(ctl, info, true);
916 spin_unlock(&ctl->tree_lock);
917 kmem_cache_free(btrfs_free_space_cachep, info);
918 ret = btrfs_add_free_space(block_group, offset, bytes);
919 spin_lock(&ctl->tree_lock);
920 } else {
921 u64 offset = info->offset;
922 u64 bytes = ctl->unit;
923
924 ret = search_bitmap(ctl, info, &offset, &bytes, false);
925 if (ret == 0) {
926 bitmap_clear_bits(ctl, info, offset, bytes, true);
927 spin_unlock(&ctl->tree_lock);
928 ret = btrfs_add_free_space(block_group, offset,
929 bytes);
930 spin_lock(&ctl->tree_lock);
931 } else {
932 free_bitmap(ctl, info);
933 ret = 0;
934 }
935 }
936 cond_resched_lock(&ctl->tree_lock);
937 }
938 return ret;
939}
940
941static struct lock_class_key btrfs_free_space_inode_key;
942
943int load_free_space_cache(struct btrfs_block_group *block_group)
944{
945 struct btrfs_fs_info *fs_info = block_group->fs_info;
946 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
947 struct btrfs_free_space_ctl tmp_ctl = {};
948 struct inode *inode;
949 struct btrfs_path *path;
950 int ret = 0;
951 bool matched;
952 u64 used = block_group->used;
953
954 /*
955 * Because we could potentially discard our loaded free space, we want
956 * to load everything into a temporary structure first, and then if it's
957 * valid copy it all into the actual free space ctl.
958 */
959 btrfs_init_free_space_ctl(block_group, &tmp_ctl);
960
961 /*
962 * If this block group has been marked to be cleared for one reason or
963 * another then we can't trust the on disk cache, so just return.
964 */
965 spin_lock(&block_group->lock);
966 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
967 spin_unlock(&block_group->lock);
968 return 0;
969 }
970 spin_unlock(&block_group->lock);
971
972 path = btrfs_alloc_path();
973 if (!path)
974 return 0;
975 path->search_commit_root = 1;
976 path->skip_locking = 1;
977
978 /*
979 * We must pass a path with search_commit_root set to btrfs_iget in
980 * order to avoid a deadlock when allocating extents for the tree root.
981 *
982 * When we are COWing an extent buffer from the tree root, when looking
983 * for a free extent, at extent-tree.c:find_free_extent(), we can find
984 * block group without its free space cache loaded. When we find one
985 * we must load its space cache which requires reading its free space
986 * cache's inode item from the root tree. If this inode item is located
987 * in the same leaf that we started COWing before, then we end up in
988 * deadlock on the extent buffer (trying to read lock it when we
989 * previously write locked it).
990 *
991 * It's safe to read the inode item using the commit root because
992 * block groups, once loaded, stay in memory forever (until they are
993 * removed) as well as their space caches once loaded. New block groups
994 * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
995 * we will never try to read their inode item while the fs is mounted.
996 */
997 inode = lookup_free_space_inode(block_group, path);
998 if (IS_ERR(inode)) {
999 btrfs_free_path(path);
1000 return 0;
1001 }
1002
1003 /* We may have converted the inode and made the cache invalid. */
1004 spin_lock(&block_group->lock);
1005 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
1006 spin_unlock(&block_group->lock);
1007 btrfs_free_path(path);
1008 goto out;
1009 }
1010 spin_unlock(&block_group->lock);
1011
1012 /*
1013 * Reinitialize the class of struct inode's mapping->invalidate_lock for
1014 * free space inodes to prevent false positives related to locks for normal
1015 * inodes.
1016 */
1017 lockdep_set_class(&(&inode->i_data)->invalidate_lock,
1018 &btrfs_free_space_inode_key);
1019
1020 ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
1021 path, block_group->start);
1022 btrfs_free_path(path);
1023 if (ret <= 0)
1024 goto out;
1025
1026 matched = (tmp_ctl.free_space == (block_group->length - used -
1027 block_group->bytes_super));
1028
1029 if (matched) {
1030 spin_lock(&tmp_ctl.tree_lock);
1031 ret = copy_free_space_cache(block_group, &tmp_ctl);
1032 spin_unlock(&tmp_ctl.tree_lock);
1033 /*
1034 * ret == 1 means we successfully loaded the free space cache,
1035 * so we need to re-set it here.
1036 */
1037 if (ret == 0)
1038 ret = 1;
1039 } else {
1040 /*
1041 * We need to call the _locked variant so we don't try to update
1042 * the discard counters.
1043 */
1044 spin_lock(&tmp_ctl.tree_lock);
1045 __btrfs_remove_free_space_cache(&tmp_ctl);
1046 spin_unlock(&tmp_ctl.tree_lock);
1047 btrfs_warn(fs_info,
1048 "block group %llu has wrong amount of free space",
1049 block_group->start);
1050 ret = -1;
1051 }
1052out:
1053 if (ret < 0) {
1054 /* This cache is bogus, make sure it gets cleared */
1055 spin_lock(&block_group->lock);
1056 block_group->disk_cache_state = BTRFS_DC_CLEAR;
1057 spin_unlock(&block_group->lock);
1058 ret = 0;
1059
1060 btrfs_warn(fs_info,
1061 "failed to load free space cache for block group %llu, rebuilding it now",
1062 block_group->start);
1063 }
1064
1065 spin_lock(&ctl->tree_lock);
1066 btrfs_discard_update_discardable(block_group);
1067 spin_unlock(&ctl->tree_lock);
1068 iput(inode);
1069 return ret;
1070}
1071
1072static noinline_for_stack
1073int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1074 struct btrfs_free_space_ctl *ctl,
1075 struct btrfs_block_group *block_group,
1076 int *entries, int *bitmaps,
1077 struct list_head *bitmap_list)
1078{
1079 int ret;
1080 struct btrfs_free_cluster *cluster = NULL;
1081 struct btrfs_free_cluster *cluster_locked = NULL;
1082 struct rb_node *node = rb_first(&ctl->free_space_offset);
1083 struct btrfs_trim_range *trim_entry;
1084
1085 /* Get the cluster for this block_group if it exists */
1086 if (block_group && !list_empty(&block_group->cluster_list)) {
1087 cluster = list_entry(block_group->cluster_list.next,
1088 struct btrfs_free_cluster,
1089 block_group_list);
1090 }
1091
1092 if (!node && cluster) {
1093 cluster_locked = cluster;
1094 spin_lock(&cluster_locked->lock);
1095 node = rb_first(&cluster->root);
1096 cluster = NULL;
1097 }
1098
1099 /* Write out the extent entries */
1100 while (node) {
1101 struct btrfs_free_space *e;
1102
1103 e = rb_entry(node, struct btrfs_free_space, offset_index);
1104 *entries += 1;
1105
1106 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1107 e->bitmap);
1108 if (ret)
1109 goto fail;
1110
1111 if (e->bitmap) {
1112 list_add_tail(&e->list, bitmap_list);
1113 *bitmaps += 1;
1114 }
1115 node = rb_next(node);
1116 if (!node && cluster) {
1117 node = rb_first(&cluster->root);
1118 cluster_locked = cluster;
1119 spin_lock(&cluster_locked->lock);
1120 cluster = NULL;
1121 }
1122 }
1123 if (cluster_locked) {
1124 spin_unlock(&cluster_locked->lock);
1125 cluster_locked = NULL;
1126 }
1127
1128 /*
1129 * Make sure we don't miss any range that was removed from our rbtree
1130 * because trimming is running. Otherwise after a umount+mount (or crash
1131 * after committing the transaction) we would leak free space and get
1132 * an inconsistent free space cache report from fsck.
1133 */
1134 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1135 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1136 trim_entry->bytes, NULL);
1137 if (ret)
1138 goto fail;
1139 *entries += 1;
1140 }
1141
1142 return 0;
1143fail:
1144 if (cluster_locked)
1145 spin_unlock(&cluster_locked->lock);
1146 return -ENOSPC;
1147}
1148
1149static noinline_for_stack int
1150update_cache_item(struct btrfs_trans_handle *trans,
1151 struct btrfs_root *root,
1152 struct inode *inode,
1153 struct btrfs_path *path, u64 offset,
1154 int entries, int bitmaps)
1155{
1156 struct btrfs_key key;
1157 struct btrfs_free_space_header *header;
1158 struct extent_buffer *leaf;
1159 int ret;
1160
1161 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1162 key.offset = offset;
1163 key.type = 0;
1164
1165 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1166 if (ret < 0) {
1167 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1168 EXTENT_DELALLOC, NULL);
1169 goto fail;
1170 }
1171 leaf = path->nodes[0];
1172 if (ret > 0) {
1173 struct btrfs_key found_key;
1174 ASSERT(path->slots[0]);
1175 path->slots[0]--;
1176 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1177 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1178 found_key.offset != offset) {
1179 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1180 inode->i_size - 1, EXTENT_DELALLOC,
1181 NULL);
1182 btrfs_release_path(path);
1183 goto fail;
1184 }
1185 }
1186
1187 BTRFS_I(inode)->generation = trans->transid;
1188 header = btrfs_item_ptr(leaf, path->slots[0],
1189 struct btrfs_free_space_header);
1190 btrfs_set_free_space_entries(leaf, header, entries);
1191 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1192 btrfs_set_free_space_generation(leaf, header, trans->transid);
1193 btrfs_mark_buffer_dirty(trans, leaf);
1194 btrfs_release_path(path);
1195
1196 return 0;
1197
1198fail:
1199 return -1;
1200}
1201
1202static noinline_for_stack int write_pinned_extent_entries(
1203 struct btrfs_trans_handle *trans,
1204 struct btrfs_block_group *block_group,
1205 struct btrfs_io_ctl *io_ctl,
1206 int *entries)
1207{
1208 u64 start, extent_start, extent_end, len;
1209 struct extent_io_tree *unpin = NULL;
1210 int ret;
1211
1212 if (!block_group)
1213 return 0;
1214
1215 /*
1216 * We want to add any pinned extents to our free space cache
1217 * so we don't leak the space
1218 *
1219 * We shouldn't have switched the pinned extents yet so this is the
1220 * right one
1221 */
1222 unpin = &trans->transaction->pinned_extents;
1223
1224 start = block_group->start;
1225
1226 while (start < block_group->start + block_group->length) {
1227 if (!find_first_extent_bit(unpin, start,
1228 &extent_start, &extent_end,
1229 EXTENT_DIRTY, NULL))
1230 return 0;
1231
1232 /* This pinned extent is out of our range */
1233 if (extent_start >= block_group->start + block_group->length)
1234 return 0;
1235
1236 extent_start = max(extent_start, start);
1237 extent_end = min(block_group->start + block_group->length,
1238 extent_end + 1);
1239 len = extent_end - extent_start;
1240
1241 *entries += 1;
1242 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1243 if (ret)
1244 return -ENOSPC;
1245
1246 start = extent_end;
1247 }
1248
1249 return 0;
1250}
1251
1252static noinline_for_stack int
1253write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1254{
1255 struct btrfs_free_space *entry, *next;
1256 int ret;
1257
1258 /* Write out the bitmaps */
1259 list_for_each_entry_safe(entry, next, bitmap_list, list) {
1260 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1261 if (ret)
1262 return -ENOSPC;
1263 list_del_init(&entry->list);
1264 }
1265
1266 return 0;
1267}
1268
1269static int flush_dirty_cache(struct inode *inode)
1270{
1271 int ret;
1272
1273 ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1274 if (ret)
1275 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1276 EXTENT_DELALLOC, NULL);
1277
1278 return ret;
1279}
1280
1281static void noinline_for_stack
1282cleanup_bitmap_list(struct list_head *bitmap_list)
1283{
1284 struct btrfs_free_space *entry, *next;
1285
1286 list_for_each_entry_safe(entry, next, bitmap_list, list)
1287 list_del_init(&entry->list);
1288}
1289
1290static void noinline_for_stack
1291cleanup_write_cache_enospc(struct inode *inode,
1292 struct btrfs_io_ctl *io_ctl,
1293 struct extent_state **cached_state)
1294{
1295 io_ctl_drop_pages(io_ctl);
1296 unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1297 cached_state);
1298}
1299
1300static int __btrfs_wait_cache_io(struct btrfs_root *root,
1301 struct btrfs_trans_handle *trans,
1302 struct btrfs_block_group *block_group,
1303 struct btrfs_io_ctl *io_ctl,
1304 struct btrfs_path *path, u64 offset)
1305{
1306 int ret;
1307 struct inode *inode = io_ctl->inode;
1308
1309 if (!inode)
1310 return 0;
1311
1312 /* Flush the dirty pages in the cache file. */
1313 ret = flush_dirty_cache(inode);
1314 if (ret)
1315 goto out;
1316
1317 /* Update the cache item to tell everyone this cache file is valid. */
1318 ret = update_cache_item(trans, root, inode, path, offset,
1319 io_ctl->entries, io_ctl->bitmaps);
1320out:
1321 if (ret) {
1322 invalidate_inode_pages2(inode->i_mapping);
1323 BTRFS_I(inode)->generation = 0;
1324 if (block_group)
1325 btrfs_debug(root->fs_info,
1326 "failed to write free space cache for block group %llu error %d",
1327 block_group->start, ret);
1328 }
1329 btrfs_update_inode(trans, BTRFS_I(inode));
1330
1331 if (block_group) {
1332 /* the dirty list is protected by the dirty_bgs_lock */
1333 spin_lock(&trans->transaction->dirty_bgs_lock);
1334
1335 /* the disk_cache_state is protected by the block group lock */
1336 spin_lock(&block_group->lock);
1337
1338 /*
1339 * only mark this as written if we didn't get put back on
1340 * the dirty list while waiting for IO. Otherwise our
1341 * cache state won't be right, and we won't get written again
1342 */
1343 if (!ret && list_empty(&block_group->dirty_list))
1344 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1345 else if (ret)
1346 block_group->disk_cache_state = BTRFS_DC_ERROR;
1347
1348 spin_unlock(&block_group->lock);
1349 spin_unlock(&trans->transaction->dirty_bgs_lock);
1350 io_ctl->inode = NULL;
1351 iput(inode);
1352 }
1353
1354 return ret;
1355
1356}
1357
1358int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1359 struct btrfs_block_group *block_group,
1360 struct btrfs_path *path)
1361{
1362 return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1363 block_group, &block_group->io_ctl,
1364 path, block_group->start);
1365}
1366
1367/*
1368 * Write out cached info to an inode.
1369 *
1370 * @inode: freespace inode we are writing out
1371 * @ctl: free space cache we are going to write out
1372 * @block_group: block_group for this cache if it belongs to a block_group
1373 * @io_ctl: holds context for the io
1374 * @trans: the trans handle
1375 *
1376 * This function writes out a free space cache struct to disk for quick recovery
1377 * on mount. This will return 0 if it was successful in writing the cache out,
1378 * or an errno if it was not.
1379 */
1380static int __btrfs_write_out_cache(struct inode *inode,
1381 struct btrfs_free_space_ctl *ctl,
1382 struct btrfs_block_group *block_group,
1383 struct btrfs_io_ctl *io_ctl,
1384 struct btrfs_trans_handle *trans)
1385{
1386 struct extent_state *cached_state = NULL;
1387 LIST_HEAD(bitmap_list);
1388 int entries = 0;
1389 int bitmaps = 0;
1390 int ret;
1391 int must_iput = 0;
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 ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1463 io_ctl->num_pages, 0, i_size_read(inode),
1464 &cached_state, false);
1465 if (ret)
1466 goto out_nospc;
1467
1468 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1469 up_write(&block_group->data_rwsem);
1470 /*
1471 * Release the pages and unlock the extent, we will flush
1472 * them out later
1473 */
1474 io_ctl_drop_pages(io_ctl);
1475 io_ctl_free(io_ctl);
1476
1477 unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1478 &cached_state);
1479
1480 /*
1481 * at this point the pages are under IO and we're happy,
1482 * The caller is responsible for waiting on them and updating
1483 * the cache and the inode
1484 */
1485 io_ctl->entries = entries;
1486 io_ctl->bitmaps = bitmaps;
1487
1488 ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1489 if (ret)
1490 goto out;
1491
1492 return 0;
1493
1494out_nospc_locked:
1495 cleanup_bitmap_list(&bitmap_list);
1496 spin_unlock(&ctl->tree_lock);
1497 mutex_unlock(&ctl->cache_writeout_mutex);
1498
1499out_nospc:
1500 cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1501
1502out_unlock:
1503 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1504 up_write(&block_group->data_rwsem);
1505
1506out:
1507 io_ctl->inode = NULL;
1508 io_ctl_free(io_ctl);
1509 if (ret) {
1510 invalidate_inode_pages2(inode->i_mapping);
1511 BTRFS_I(inode)->generation = 0;
1512 }
1513 btrfs_update_inode(trans, BTRFS_I(inode));
1514 if (must_iput)
1515 iput(inode);
1516 return ret;
1517}
1518
1519int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1520 struct btrfs_block_group *block_group,
1521 struct btrfs_path *path)
1522{
1523 struct btrfs_fs_info *fs_info = trans->fs_info;
1524 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1525 struct inode *inode;
1526 int ret = 0;
1527
1528 spin_lock(&block_group->lock);
1529 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1530 spin_unlock(&block_group->lock);
1531 return 0;
1532 }
1533 spin_unlock(&block_group->lock);
1534
1535 inode = lookup_free_space_inode(block_group, path);
1536 if (IS_ERR(inode))
1537 return 0;
1538
1539 ret = __btrfs_write_out_cache(inode, ctl, block_group,
1540 &block_group->io_ctl, trans);
1541 if (ret) {
1542 btrfs_debug(fs_info,
1543 "failed to write free space cache for block group %llu error %d",
1544 block_group->start, ret);
1545 spin_lock(&block_group->lock);
1546 block_group->disk_cache_state = BTRFS_DC_ERROR;
1547 spin_unlock(&block_group->lock);
1548
1549 block_group->io_ctl.inode = NULL;
1550 iput(inode);
1551 }
1552
1553 /*
1554 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1555 * to wait for IO and put the inode
1556 */
1557
1558 return ret;
1559}
1560
1561static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1562 u64 offset)
1563{
1564 ASSERT(offset >= bitmap_start);
1565 offset -= bitmap_start;
1566 return (unsigned long)(div_u64(offset, unit));
1567}
1568
1569static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1570{
1571 return (unsigned long)(div_u64(bytes, unit));
1572}
1573
1574static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1575 u64 offset)
1576{
1577 u64 bitmap_start;
1578 u64 bytes_per_bitmap;
1579
1580 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1581 bitmap_start = offset - ctl->start;
1582 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1583 bitmap_start *= bytes_per_bitmap;
1584 bitmap_start += ctl->start;
1585
1586 return bitmap_start;
1587}
1588
1589static int tree_insert_offset(struct btrfs_free_space_ctl *ctl,
1590 struct btrfs_free_cluster *cluster,
1591 struct btrfs_free_space *new_entry)
1592{
1593 struct rb_root *root;
1594 struct rb_node **p;
1595 struct rb_node *parent = NULL;
1596
1597 lockdep_assert_held(&ctl->tree_lock);
1598
1599 if (cluster) {
1600 lockdep_assert_held(&cluster->lock);
1601 root = &cluster->root;
1602 } else {
1603 root = &ctl->free_space_offset;
1604 }
1605
1606 p = &root->rb_node;
1607
1608 while (*p) {
1609 struct btrfs_free_space *info;
1610
1611 parent = *p;
1612 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1613
1614 if (new_entry->offset < info->offset) {
1615 p = &(*p)->rb_left;
1616 } else if (new_entry->offset > info->offset) {
1617 p = &(*p)->rb_right;
1618 } else {
1619 /*
1620 * we could have a bitmap entry and an extent entry
1621 * share the same offset. If this is the case, we want
1622 * the extent entry to always be found first if we do a
1623 * linear search through the tree, since we want to have
1624 * the quickest allocation time, and allocating from an
1625 * extent is faster than allocating from a bitmap. So
1626 * if we're inserting a bitmap and we find an entry at
1627 * this offset, we want to go right, or after this entry
1628 * logically. If we are inserting an extent and we've
1629 * found a bitmap, we want to go left, or before
1630 * logically.
1631 */
1632 if (new_entry->bitmap) {
1633 if (info->bitmap) {
1634 WARN_ON_ONCE(1);
1635 return -EEXIST;
1636 }
1637 p = &(*p)->rb_right;
1638 } else {
1639 if (!info->bitmap) {
1640 WARN_ON_ONCE(1);
1641 return -EEXIST;
1642 }
1643 p = &(*p)->rb_left;
1644 }
1645 }
1646 }
1647
1648 rb_link_node(&new_entry->offset_index, parent, p);
1649 rb_insert_color(&new_entry->offset_index, root);
1650
1651 return 0;
1652}
1653
1654/*
1655 * This is a little subtle. We *only* have ->max_extent_size set if we actually
1656 * searched through the bitmap and figured out the largest ->max_extent_size,
1657 * otherwise it's 0. In the case that it's 0 we don't want to tell the
1658 * allocator the wrong thing, we want to use the actual real max_extent_size
1659 * we've found already if it's larger, or we want to use ->bytes.
1660 *
1661 * This matters because find_free_space() will skip entries who's ->bytes is
1662 * less than the required bytes. So if we didn't search down this bitmap, we
1663 * may pick some previous entry that has a smaller ->max_extent_size than we
1664 * have. For example, assume we have two entries, one that has
1665 * ->max_extent_size set to 4K and ->bytes set to 1M. A second entry hasn't set
1666 * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous. We will
1667 * call into find_free_space(), and return with max_extent_size == 4K, because
1668 * that first bitmap entry had ->max_extent_size set, but the second one did
1669 * not. If instead we returned 8K we'd come in searching for 8K, and find the
1670 * 8K contiguous range.
1671 *
1672 * Consider the other case, we have 2 8K chunks in that second entry and still
1673 * don't have ->max_extent_size set. We'll return 16K, and the next time the
1674 * allocator comes in it'll fully search our second bitmap, and this time it'll
1675 * get an uptodate value of 8K as the maximum chunk size. Then we'll get the
1676 * right allocation the next loop through.
1677 */
1678static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1679{
1680 if (entry->bitmap && entry->max_extent_size)
1681 return entry->max_extent_size;
1682 return entry->bytes;
1683}
1684
1685/*
1686 * We want the largest entry to be leftmost, so this is inverted from what you'd
1687 * normally expect.
1688 */
1689static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1690{
1691 const struct btrfs_free_space *entry, *exist;
1692
1693 entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1694 exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1695 return get_max_extent_size(exist) < get_max_extent_size(entry);
1696}
1697
1698/*
1699 * searches the tree for the given offset.
1700 *
1701 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1702 * want a section that has at least bytes size and comes at or after the given
1703 * offset.
1704 */
1705static struct btrfs_free_space *
1706tree_search_offset(struct btrfs_free_space_ctl *ctl,
1707 u64 offset, int bitmap_only, int fuzzy)
1708{
1709 struct rb_node *n = ctl->free_space_offset.rb_node;
1710 struct btrfs_free_space *entry = NULL, *prev = NULL;
1711
1712 lockdep_assert_held(&ctl->tree_lock);
1713
1714 /* find entry that is closest to the 'offset' */
1715 while (n) {
1716 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1717 prev = entry;
1718
1719 if (offset < entry->offset)
1720 n = n->rb_left;
1721 else if (offset > entry->offset)
1722 n = n->rb_right;
1723 else
1724 break;
1725
1726 entry = NULL;
1727 }
1728
1729 if (bitmap_only) {
1730 if (!entry)
1731 return NULL;
1732 if (entry->bitmap)
1733 return entry;
1734
1735 /*
1736 * bitmap entry and extent entry may share same offset,
1737 * in that case, bitmap entry comes after extent entry.
1738 */
1739 n = rb_next(n);
1740 if (!n)
1741 return NULL;
1742 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1743 if (entry->offset != offset)
1744 return NULL;
1745
1746 WARN_ON(!entry->bitmap);
1747 return entry;
1748 } else if (entry) {
1749 if (entry->bitmap) {
1750 /*
1751 * if previous extent entry covers the offset,
1752 * we should return it instead of the bitmap entry
1753 */
1754 n = rb_prev(&entry->offset_index);
1755 if (n) {
1756 prev = rb_entry(n, struct btrfs_free_space,
1757 offset_index);
1758 if (!prev->bitmap &&
1759 prev->offset + prev->bytes > offset)
1760 entry = prev;
1761 }
1762 }
1763 return entry;
1764 }
1765
1766 if (!prev)
1767 return NULL;
1768
1769 /* find last entry before the 'offset' */
1770 entry = prev;
1771 if (entry->offset > offset) {
1772 n = rb_prev(&entry->offset_index);
1773 if (n) {
1774 entry = rb_entry(n, struct btrfs_free_space,
1775 offset_index);
1776 ASSERT(entry->offset <= offset);
1777 } else {
1778 if (fuzzy)
1779 return entry;
1780 else
1781 return NULL;
1782 }
1783 }
1784
1785 if (entry->bitmap) {
1786 n = rb_prev(&entry->offset_index);
1787 if (n) {
1788 prev = rb_entry(n, struct btrfs_free_space,
1789 offset_index);
1790 if (!prev->bitmap &&
1791 prev->offset + prev->bytes > offset)
1792 return prev;
1793 }
1794 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1795 return entry;
1796 } else if (entry->offset + entry->bytes > offset)
1797 return entry;
1798
1799 if (!fuzzy)
1800 return NULL;
1801
1802 while (1) {
1803 n = rb_next(&entry->offset_index);
1804 if (!n)
1805 return NULL;
1806 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1807 if (entry->bitmap) {
1808 if (entry->offset + BITS_PER_BITMAP *
1809 ctl->unit > offset)
1810 break;
1811 } else {
1812 if (entry->offset + entry->bytes > offset)
1813 break;
1814 }
1815 }
1816 return entry;
1817}
1818
1819static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1820 struct btrfs_free_space *info,
1821 bool update_stat)
1822{
1823 lockdep_assert_held(&ctl->tree_lock);
1824
1825 rb_erase(&info->offset_index, &ctl->free_space_offset);
1826 rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1827 ctl->free_extents--;
1828
1829 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1830 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1831 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1832 }
1833
1834 if (update_stat)
1835 ctl->free_space -= info->bytes;
1836}
1837
1838static int link_free_space(struct btrfs_free_space_ctl *ctl,
1839 struct btrfs_free_space *info)
1840{
1841 int ret = 0;
1842
1843 lockdep_assert_held(&ctl->tree_lock);
1844
1845 ASSERT(info->bytes || info->bitmap);
1846 ret = tree_insert_offset(ctl, NULL, info);
1847 if (ret)
1848 return ret;
1849
1850 rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1851
1852 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1853 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1854 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1855 }
1856
1857 ctl->free_space += info->bytes;
1858 ctl->free_extents++;
1859 return ret;
1860}
1861
1862static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1863 struct btrfs_free_space *info)
1864{
1865 ASSERT(info->bitmap);
1866
1867 /*
1868 * If our entry is empty it's because we're on a cluster and we don't
1869 * want to re-link it into our ctl bytes index.
1870 */
1871 if (RB_EMPTY_NODE(&info->bytes_index))
1872 return;
1873
1874 lockdep_assert_held(&ctl->tree_lock);
1875
1876 rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1877 rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1878}
1879
1880static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1881 struct btrfs_free_space *info,
1882 u64 offset, u64 bytes, bool update_stat)
1883{
1884 unsigned long start, count, end;
1885 int extent_delta = -1;
1886
1887 start = offset_to_bit(info->offset, ctl->unit, offset);
1888 count = bytes_to_bits(bytes, ctl->unit);
1889 end = start + count;
1890 ASSERT(end <= BITS_PER_BITMAP);
1891
1892 bitmap_clear(info->bitmap, start, count);
1893
1894 info->bytes -= bytes;
1895 if (info->max_extent_size > ctl->unit)
1896 info->max_extent_size = 0;
1897
1898 relink_bitmap_entry(ctl, info);
1899
1900 if (start && test_bit(start - 1, info->bitmap))
1901 extent_delta++;
1902
1903 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1904 extent_delta++;
1905
1906 info->bitmap_extents += extent_delta;
1907 if (!btrfs_free_space_trimmed(info)) {
1908 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1909 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1910 }
1911
1912 if (update_stat)
1913 ctl->free_space -= bytes;
1914}
1915
1916static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1917 struct btrfs_free_space *info, u64 offset,
1918 u64 bytes)
1919{
1920 unsigned long start, count, end;
1921 int extent_delta = 1;
1922
1923 start = offset_to_bit(info->offset, ctl->unit, offset);
1924 count = bytes_to_bits(bytes, ctl->unit);
1925 end = start + count;
1926 ASSERT(end <= BITS_PER_BITMAP);
1927
1928 bitmap_set(info->bitmap, start, count);
1929
1930 /*
1931 * We set some bytes, we have no idea what the max extent size is
1932 * anymore.
1933 */
1934 info->max_extent_size = 0;
1935 info->bytes += bytes;
1936 ctl->free_space += bytes;
1937
1938 relink_bitmap_entry(ctl, info);
1939
1940 if (start && test_bit(start - 1, info->bitmap))
1941 extent_delta--;
1942
1943 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1944 extent_delta--;
1945
1946 info->bitmap_extents += extent_delta;
1947 if (!btrfs_free_space_trimmed(info)) {
1948 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1949 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1950 }
1951}
1952
1953/*
1954 * If we can not find suitable extent, we will use bytes to record
1955 * the size of the max extent.
1956 */
1957static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1958 struct btrfs_free_space *bitmap_info, u64 *offset,
1959 u64 *bytes, bool for_alloc)
1960{
1961 unsigned long found_bits = 0;
1962 unsigned long max_bits = 0;
1963 unsigned long bits, i;
1964 unsigned long next_zero;
1965 unsigned long extent_bits;
1966
1967 /*
1968 * Skip searching the bitmap if we don't have a contiguous section that
1969 * is large enough for this allocation.
1970 */
1971 if (for_alloc &&
1972 bitmap_info->max_extent_size &&
1973 bitmap_info->max_extent_size < *bytes) {
1974 *bytes = bitmap_info->max_extent_size;
1975 return -1;
1976 }
1977
1978 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1979 max_t(u64, *offset, bitmap_info->offset));
1980 bits = bytes_to_bits(*bytes, ctl->unit);
1981
1982 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1983 if (for_alloc && bits == 1) {
1984 found_bits = 1;
1985 break;
1986 }
1987 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1988 BITS_PER_BITMAP, i);
1989 extent_bits = next_zero - i;
1990 if (extent_bits >= bits) {
1991 found_bits = extent_bits;
1992 break;
1993 } else if (extent_bits > max_bits) {
1994 max_bits = extent_bits;
1995 }
1996 i = next_zero;
1997 }
1998
1999 if (found_bits) {
2000 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
2001 *bytes = (u64)(found_bits) * ctl->unit;
2002 return 0;
2003 }
2004
2005 *bytes = (u64)(max_bits) * ctl->unit;
2006 bitmap_info->max_extent_size = *bytes;
2007 relink_bitmap_entry(ctl, bitmap_info);
2008 return -1;
2009}
2010
2011/* Cache the size of the max extent in bytes */
2012static struct btrfs_free_space *
2013find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
2014 unsigned long align, u64 *max_extent_size, bool use_bytes_index)
2015{
2016 struct btrfs_free_space *entry;
2017 struct rb_node *node;
2018 u64 tmp;
2019 u64 align_off;
2020 int ret;
2021
2022 if (!ctl->free_space_offset.rb_node)
2023 goto out;
2024again:
2025 if (use_bytes_index) {
2026 node = rb_first_cached(&ctl->free_space_bytes);
2027 } else {
2028 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
2029 0, 1);
2030 if (!entry)
2031 goto out;
2032 node = &entry->offset_index;
2033 }
2034
2035 for (; node; node = rb_next(node)) {
2036 if (use_bytes_index)
2037 entry = rb_entry(node, struct btrfs_free_space,
2038 bytes_index);
2039 else
2040 entry = rb_entry(node, struct btrfs_free_space,
2041 offset_index);
2042
2043 /*
2044 * If we are using the bytes index then all subsequent entries
2045 * in this tree are going to be < bytes, so simply set the max
2046 * extent size and exit the loop.
2047 *
2048 * If we're using the offset index then we need to keep going
2049 * through the rest of the tree.
2050 */
2051 if (entry->bytes < *bytes) {
2052 *max_extent_size = max(get_max_extent_size(entry),
2053 *max_extent_size);
2054 if (use_bytes_index)
2055 break;
2056 continue;
2057 }
2058
2059 /* make sure the space returned is big enough
2060 * to match our requested alignment
2061 */
2062 if (*bytes >= align) {
2063 tmp = entry->offset - ctl->start + align - 1;
2064 tmp = div64_u64(tmp, align);
2065 tmp = tmp * align + ctl->start;
2066 align_off = tmp - entry->offset;
2067 } else {
2068 align_off = 0;
2069 tmp = entry->offset;
2070 }
2071
2072 /*
2073 * We don't break here if we're using the bytes index because we
2074 * may have another entry that has the correct alignment that is
2075 * the right size, so we don't want to miss that possibility.
2076 * At worst this adds another loop through the logic, but if we
2077 * broke here we could prematurely ENOSPC.
2078 */
2079 if (entry->bytes < *bytes + align_off) {
2080 *max_extent_size = max(get_max_extent_size(entry),
2081 *max_extent_size);
2082 continue;
2083 }
2084
2085 if (entry->bitmap) {
2086 struct rb_node *old_next = rb_next(node);
2087 u64 size = *bytes;
2088
2089 ret = search_bitmap(ctl, entry, &tmp, &size, true);
2090 if (!ret) {
2091 *offset = tmp;
2092 *bytes = size;
2093 return entry;
2094 } else {
2095 *max_extent_size =
2096 max(get_max_extent_size(entry),
2097 *max_extent_size);
2098 }
2099
2100 /*
2101 * The bitmap may have gotten re-arranged in the space
2102 * index here because the max_extent_size may have been
2103 * updated. Start from the beginning again if this
2104 * happened.
2105 */
2106 if (use_bytes_index && old_next != rb_next(node))
2107 goto again;
2108 continue;
2109 }
2110
2111 *offset = tmp;
2112 *bytes = entry->bytes - align_off;
2113 return entry;
2114 }
2115out:
2116 return NULL;
2117}
2118
2119static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2120 struct btrfs_free_space *info, u64 offset)
2121{
2122 info->offset = offset_to_bitmap(ctl, offset);
2123 info->bytes = 0;
2124 info->bitmap_extents = 0;
2125 INIT_LIST_HEAD(&info->list);
2126 link_free_space(ctl, info);
2127 ctl->total_bitmaps++;
2128 recalculate_thresholds(ctl);
2129}
2130
2131static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2132 struct btrfs_free_space *bitmap_info)
2133{
2134 /*
2135 * Normally when this is called, the bitmap is completely empty. However,
2136 * if we are blowing up the free space cache for one reason or another
2137 * via __btrfs_remove_free_space_cache(), then it may not be freed and
2138 * we may leave stats on the table.
2139 */
2140 if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2141 ctl->discardable_extents[BTRFS_STAT_CURR] -=
2142 bitmap_info->bitmap_extents;
2143 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2144
2145 }
2146 unlink_free_space(ctl, bitmap_info, true);
2147 kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2148 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2149 ctl->total_bitmaps--;
2150 recalculate_thresholds(ctl);
2151}
2152
2153static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2154 struct btrfs_free_space *bitmap_info,
2155 u64 *offset, u64 *bytes)
2156{
2157 u64 end;
2158 u64 search_start, search_bytes;
2159 int ret;
2160
2161again:
2162 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2163
2164 /*
2165 * We need to search for bits in this bitmap. We could only cover some
2166 * of the extent in this bitmap thanks to how we add space, so we need
2167 * to search for as much as it as we can and clear that amount, and then
2168 * go searching for the next bit.
2169 */
2170 search_start = *offset;
2171 search_bytes = ctl->unit;
2172 search_bytes = min(search_bytes, end - search_start + 1);
2173 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2174 false);
2175 if (ret < 0 || search_start != *offset)
2176 return -EINVAL;
2177
2178 /* We may have found more bits than what we need */
2179 search_bytes = min(search_bytes, *bytes);
2180
2181 /* Cannot clear past the end of the bitmap */
2182 search_bytes = min(search_bytes, end - search_start + 1);
2183
2184 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2185 *offset += search_bytes;
2186 *bytes -= search_bytes;
2187
2188 if (*bytes) {
2189 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2190 if (!bitmap_info->bytes)
2191 free_bitmap(ctl, bitmap_info);
2192
2193 /*
2194 * no entry after this bitmap, but we still have bytes to
2195 * remove, so something has gone wrong.
2196 */
2197 if (!next)
2198 return -EINVAL;
2199
2200 bitmap_info = rb_entry(next, struct btrfs_free_space,
2201 offset_index);
2202
2203 /*
2204 * if the next entry isn't a bitmap we need to return to let the
2205 * extent stuff do its work.
2206 */
2207 if (!bitmap_info->bitmap)
2208 return -EAGAIN;
2209
2210 /*
2211 * Ok the next item is a bitmap, but it may not actually hold
2212 * the information for the rest of this free space stuff, so
2213 * look for it, and if we don't find it return so we can try
2214 * everything over again.
2215 */
2216 search_start = *offset;
2217 search_bytes = ctl->unit;
2218 ret = search_bitmap(ctl, bitmap_info, &search_start,
2219 &search_bytes, false);
2220 if (ret < 0 || search_start != *offset)
2221 return -EAGAIN;
2222
2223 goto again;
2224 } else if (!bitmap_info->bytes)
2225 free_bitmap(ctl, bitmap_info);
2226
2227 return 0;
2228}
2229
2230static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2231 struct btrfs_free_space *info, u64 offset,
2232 u64 bytes, enum btrfs_trim_state trim_state)
2233{
2234 u64 bytes_to_set = 0;
2235 u64 end;
2236
2237 /*
2238 * This is a tradeoff to make bitmap trim state minimal. We mark the
2239 * whole bitmap untrimmed if at any point we add untrimmed regions.
2240 */
2241 if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2242 if (btrfs_free_space_trimmed(info)) {
2243 ctl->discardable_extents[BTRFS_STAT_CURR] +=
2244 info->bitmap_extents;
2245 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2246 }
2247 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2248 }
2249
2250 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2251
2252 bytes_to_set = min(end - offset, bytes);
2253
2254 bitmap_set_bits(ctl, info, offset, bytes_to_set);
2255
2256 return bytes_to_set;
2257
2258}
2259
2260static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2261 struct btrfs_free_space *info)
2262{
2263 struct btrfs_block_group *block_group = ctl->block_group;
2264 struct btrfs_fs_info *fs_info = block_group->fs_info;
2265 bool forced = false;
2266
2267#ifdef CONFIG_BTRFS_DEBUG
2268 if (btrfs_should_fragment_free_space(block_group))
2269 forced = true;
2270#endif
2271
2272 /* This is a way to reclaim large regions from the bitmaps. */
2273 if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2274 return false;
2275
2276 /*
2277 * If we are below the extents threshold then we can add this as an
2278 * extent, and don't have to deal with the bitmap
2279 */
2280 if (!forced && ctl->free_extents < ctl->extents_thresh) {
2281 /*
2282 * If this block group has some small extents we don't want to
2283 * use up all of our free slots in the cache with them, we want
2284 * to reserve them to larger extents, however if we have plenty
2285 * of cache left then go ahead an dadd them, no sense in adding
2286 * the overhead of a bitmap if we don't have to.
2287 */
2288 if (info->bytes <= fs_info->sectorsize * 8) {
2289 if (ctl->free_extents * 3 <= ctl->extents_thresh)
2290 return false;
2291 } else {
2292 return false;
2293 }
2294 }
2295
2296 /*
2297 * The original block groups from mkfs can be really small, like 8
2298 * megabytes, so don't bother with a bitmap for those entries. However
2299 * some block groups can be smaller than what a bitmap would cover but
2300 * are still large enough that they could overflow the 32k memory limit,
2301 * so allow those block groups to still be allowed to have a bitmap
2302 * entry.
2303 */
2304 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2305 return false;
2306
2307 return true;
2308}
2309
2310static const struct btrfs_free_space_op free_space_op = {
2311 .use_bitmap = use_bitmap,
2312};
2313
2314static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2315 struct btrfs_free_space *info)
2316{
2317 struct btrfs_free_space *bitmap_info;
2318 struct btrfs_block_group *block_group = NULL;
2319 int added = 0;
2320 u64 bytes, offset, bytes_added;
2321 enum btrfs_trim_state trim_state;
2322 int ret;
2323
2324 bytes = info->bytes;
2325 offset = info->offset;
2326 trim_state = info->trim_state;
2327
2328 if (!ctl->op->use_bitmap(ctl, info))
2329 return 0;
2330
2331 if (ctl->op == &free_space_op)
2332 block_group = ctl->block_group;
2333again:
2334 /*
2335 * Since we link bitmaps right into the cluster we need to see if we
2336 * have a cluster here, and if so and it has our bitmap we need to add
2337 * the free space to that bitmap.
2338 */
2339 if (block_group && !list_empty(&block_group->cluster_list)) {
2340 struct btrfs_free_cluster *cluster;
2341 struct rb_node *node;
2342 struct btrfs_free_space *entry;
2343
2344 cluster = list_entry(block_group->cluster_list.next,
2345 struct btrfs_free_cluster,
2346 block_group_list);
2347 spin_lock(&cluster->lock);
2348 node = rb_first(&cluster->root);
2349 if (!node) {
2350 spin_unlock(&cluster->lock);
2351 goto no_cluster_bitmap;
2352 }
2353
2354 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2355 if (!entry->bitmap) {
2356 spin_unlock(&cluster->lock);
2357 goto no_cluster_bitmap;
2358 }
2359
2360 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2361 bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2362 bytes, trim_state);
2363 bytes -= bytes_added;
2364 offset += bytes_added;
2365 }
2366 spin_unlock(&cluster->lock);
2367 if (!bytes) {
2368 ret = 1;
2369 goto out;
2370 }
2371 }
2372
2373no_cluster_bitmap:
2374 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2375 1, 0);
2376 if (!bitmap_info) {
2377 ASSERT(added == 0);
2378 goto new_bitmap;
2379 }
2380
2381 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2382 trim_state);
2383 bytes -= bytes_added;
2384 offset += bytes_added;
2385 added = 0;
2386
2387 if (!bytes) {
2388 ret = 1;
2389 goto out;
2390 } else
2391 goto again;
2392
2393new_bitmap:
2394 if (info && info->bitmap) {
2395 add_new_bitmap(ctl, info, offset);
2396 added = 1;
2397 info = NULL;
2398 goto again;
2399 } else {
2400 spin_unlock(&ctl->tree_lock);
2401
2402 /* no pre-allocated info, allocate a new one */
2403 if (!info) {
2404 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2405 GFP_NOFS);
2406 if (!info) {
2407 spin_lock(&ctl->tree_lock);
2408 ret = -ENOMEM;
2409 goto out;
2410 }
2411 }
2412
2413 /* allocate the bitmap */
2414 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2415 GFP_NOFS);
2416 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2417 spin_lock(&ctl->tree_lock);
2418 if (!info->bitmap) {
2419 ret = -ENOMEM;
2420 goto out;
2421 }
2422 goto again;
2423 }
2424
2425out:
2426 if (info) {
2427 if (info->bitmap)
2428 kmem_cache_free(btrfs_free_space_bitmap_cachep,
2429 info->bitmap);
2430 kmem_cache_free(btrfs_free_space_cachep, info);
2431 }
2432
2433 return ret;
2434}
2435
2436/*
2437 * Free space merging rules:
2438 * 1) Merge trimmed areas together
2439 * 2) Let untrimmed areas coalesce with trimmed areas
2440 * 3) Always pull neighboring regions from bitmaps
2441 *
2442 * The above rules are for when we merge free space based on btrfs_trim_state.
2443 * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2444 * same reason: to promote larger extent regions which makes life easier for
2445 * find_free_extent(). Rule 2 enables coalescing based on the common path
2446 * being returning free space from btrfs_finish_extent_commit(). So when free
2447 * space is trimmed, it will prevent aggregating trimmed new region and
2448 * untrimmed regions in the rb_tree. Rule 3 is purely to obtain larger extents
2449 * and provide find_free_extent() with the largest extents possible hoping for
2450 * the reuse path.
2451 */
2452static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2453 struct btrfs_free_space *info, bool update_stat)
2454{
2455 struct btrfs_free_space *left_info = NULL;
2456 struct btrfs_free_space *right_info;
2457 bool merged = false;
2458 u64 offset = info->offset;
2459 u64 bytes = info->bytes;
2460 const bool is_trimmed = btrfs_free_space_trimmed(info);
2461 struct rb_node *right_prev = NULL;
2462
2463 /*
2464 * first we want to see if there is free space adjacent to the range we
2465 * are adding, if there is remove that struct and add a new one to
2466 * cover the entire range
2467 */
2468 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2469 if (right_info)
2470 right_prev = rb_prev(&right_info->offset_index);
2471
2472 if (right_prev)
2473 left_info = rb_entry(right_prev, struct btrfs_free_space, offset_index);
2474 else if (!right_info)
2475 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2476
2477 /* See try_merge_free_space() comment. */
2478 if (right_info && !right_info->bitmap &&
2479 (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2480 unlink_free_space(ctl, right_info, update_stat);
2481 info->bytes += right_info->bytes;
2482 kmem_cache_free(btrfs_free_space_cachep, right_info);
2483 merged = true;
2484 }
2485
2486 /* See try_merge_free_space() comment. */
2487 if (left_info && !left_info->bitmap &&
2488 left_info->offset + left_info->bytes == offset &&
2489 (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2490 unlink_free_space(ctl, left_info, update_stat);
2491 info->offset = left_info->offset;
2492 info->bytes += left_info->bytes;
2493 kmem_cache_free(btrfs_free_space_cachep, left_info);
2494 merged = true;
2495 }
2496
2497 return merged;
2498}
2499
2500static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2501 struct btrfs_free_space *info,
2502 bool update_stat)
2503{
2504 struct btrfs_free_space *bitmap;
2505 unsigned long i;
2506 unsigned long j;
2507 const u64 end = info->offset + info->bytes;
2508 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2509 u64 bytes;
2510
2511 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2512 if (!bitmap)
2513 return false;
2514
2515 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2516 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2517 if (j == i)
2518 return false;
2519 bytes = (j - i) * ctl->unit;
2520 info->bytes += bytes;
2521
2522 /* See try_merge_free_space() comment. */
2523 if (!btrfs_free_space_trimmed(bitmap))
2524 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2525
2526 bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2527
2528 if (!bitmap->bytes)
2529 free_bitmap(ctl, bitmap);
2530
2531 return true;
2532}
2533
2534static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2535 struct btrfs_free_space *info,
2536 bool update_stat)
2537{
2538 struct btrfs_free_space *bitmap;
2539 u64 bitmap_offset;
2540 unsigned long i;
2541 unsigned long j;
2542 unsigned long prev_j;
2543 u64 bytes;
2544
2545 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2546 /* If we're on a boundary, try the previous logical bitmap. */
2547 if (bitmap_offset == info->offset) {
2548 if (info->offset == 0)
2549 return false;
2550 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2551 }
2552
2553 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2554 if (!bitmap)
2555 return false;
2556
2557 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2558 j = 0;
2559 prev_j = (unsigned long)-1;
2560 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2561 if (j > i)
2562 break;
2563 prev_j = j;
2564 }
2565 if (prev_j == i)
2566 return false;
2567
2568 if (prev_j == (unsigned long)-1)
2569 bytes = (i + 1) * ctl->unit;
2570 else
2571 bytes = (i - prev_j) * ctl->unit;
2572
2573 info->offset -= bytes;
2574 info->bytes += bytes;
2575
2576 /* See try_merge_free_space() comment. */
2577 if (!btrfs_free_space_trimmed(bitmap))
2578 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2579
2580 bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2581
2582 if (!bitmap->bytes)
2583 free_bitmap(ctl, bitmap);
2584
2585 return true;
2586}
2587
2588/*
2589 * We prefer always to allocate from extent entries, both for clustered and
2590 * non-clustered allocation requests. So when attempting to add a new extent
2591 * entry, try to see if there's adjacent free space in bitmap entries, and if
2592 * there is, migrate that space from the bitmaps to the extent.
2593 * Like this we get better chances of satisfying space allocation requests
2594 * because we attempt to satisfy them based on a single cache entry, and never
2595 * on 2 or more entries - even if the entries represent a contiguous free space
2596 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2597 * ends).
2598 */
2599static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2600 struct btrfs_free_space *info,
2601 bool update_stat)
2602{
2603 /*
2604 * Only work with disconnected entries, as we can change their offset,
2605 * and must be extent entries.
2606 */
2607 ASSERT(!info->bitmap);
2608 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2609
2610 if (ctl->total_bitmaps > 0) {
2611 bool stole_end;
2612 bool stole_front = false;
2613
2614 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2615 if (ctl->total_bitmaps > 0)
2616 stole_front = steal_from_bitmap_to_front(ctl, info,
2617 update_stat);
2618
2619 if (stole_end || stole_front)
2620 try_merge_free_space(ctl, info, update_stat);
2621 }
2622}
2623
2624int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2625 u64 offset, u64 bytes,
2626 enum btrfs_trim_state trim_state)
2627{
2628 struct btrfs_fs_info *fs_info = block_group->fs_info;
2629 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2630 struct btrfs_free_space *info;
2631 int ret = 0;
2632 u64 filter_bytes = bytes;
2633
2634 ASSERT(!btrfs_is_zoned(fs_info));
2635
2636 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2637 if (!info)
2638 return -ENOMEM;
2639
2640 info->offset = offset;
2641 info->bytes = bytes;
2642 info->trim_state = trim_state;
2643 RB_CLEAR_NODE(&info->offset_index);
2644 RB_CLEAR_NODE(&info->bytes_index);
2645
2646 spin_lock(&ctl->tree_lock);
2647
2648 if (try_merge_free_space(ctl, info, true))
2649 goto link;
2650
2651 /*
2652 * There was no extent directly to the left or right of this new
2653 * extent then we know we're going to have to allocate a new extent, so
2654 * before we do that see if we need to drop this into a bitmap
2655 */
2656 ret = insert_into_bitmap(ctl, info);
2657 if (ret < 0) {
2658 goto out;
2659 } else if (ret) {
2660 ret = 0;
2661 goto out;
2662 }
2663link:
2664 /*
2665 * Only steal free space from adjacent bitmaps if we're sure we're not
2666 * going to add the new free space to existing bitmap entries - because
2667 * that would mean unnecessary work that would be reverted. Therefore
2668 * attempt to steal space from bitmaps if we're adding an extent entry.
2669 */
2670 steal_from_bitmap(ctl, info, true);
2671
2672 filter_bytes = max(filter_bytes, info->bytes);
2673
2674 ret = link_free_space(ctl, info);
2675 if (ret)
2676 kmem_cache_free(btrfs_free_space_cachep, info);
2677out:
2678 btrfs_discard_update_discardable(block_group);
2679 spin_unlock(&ctl->tree_lock);
2680
2681 if (ret) {
2682 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2683 ASSERT(ret != -EEXIST);
2684 }
2685
2686 if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2687 btrfs_discard_check_filter(block_group, filter_bytes);
2688 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2689 }
2690
2691 return ret;
2692}
2693
2694static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2695 u64 bytenr, u64 size, bool used)
2696{
2697 struct btrfs_space_info *sinfo = block_group->space_info;
2698 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2699 u64 offset = bytenr - block_group->start;
2700 u64 to_free, to_unusable;
2701 int bg_reclaim_threshold = 0;
2702 bool initial = (size == block_group->length);
2703 u64 reclaimable_unusable;
2704
2705 WARN_ON(!initial && offset + size > block_group->zone_capacity);
2706
2707 if (!initial)
2708 bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2709
2710 spin_lock(&ctl->tree_lock);
2711 if (!used)
2712 to_free = size;
2713 else if (initial)
2714 to_free = block_group->zone_capacity;
2715 else if (offset >= block_group->alloc_offset)
2716 to_free = size;
2717 else if (offset + size <= block_group->alloc_offset)
2718 to_free = 0;
2719 else
2720 to_free = offset + size - block_group->alloc_offset;
2721 to_unusable = size - to_free;
2722
2723 ctl->free_space += to_free;
2724 /*
2725 * If the block group is read-only, we should account freed space into
2726 * bytes_readonly.
2727 */
2728 if (!block_group->ro)
2729 block_group->zone_unusable += to_unusable;
2730 spin_unlock(&ctl->tree_lock);
2731 if (!used) {
2732 spin_lock(&block_group->lock);
2733 block_group->alloc_offset -= size;
2734 spin_unlock(&block_group->lock);
2735 }
2736
2737 reclaimable_unusable = block_group->zone_unusable -
2738 (block_group->length - block_group->zone_capacity);
2739 /* All the region is now unusable. Mark it as unused and reclaim */
2740 if (block_group->zone_unusable == block_group->length) {
2741 btrfs_mark_bg_unused(block_group);
2742 } else if (bg_reclaim_threshold &&
2743 reclaimable_unusable >=
2744 mult_perc(block_group->zone_capacity, bg_reclaim_threshold)) {
2745 btrfs_mark_bg_to_reclaim(block_group);
2746 }
2747
2748 return 0;
2749}
2750
2751int btrfs_add_free_space(struct btrfs_block_group *block_group,
2752 u64 bytenr, u64 size)
2753{
2754 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2755
2756 if (btrfs_is_zoned(block_group->fs_info))
2757 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2758 true);
2759
2760 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2761 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2762
2763 return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2764}
2765
2766int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2767 u64 bytenr, u64 size)
2768{
2769 if (btrfs_is_zoned(block_group->fs_info))
2770 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2771 false);
2772
2773 return btrfs_add_free_space(block_group, bytenr, size);
2774}
2775
2776/*
2777 * This is a subtle distinction because when adding free space back in general,
2778 * we want it to be added as untrimmed for async. But in the case where we add
2779 * it on loading of a block group, we want to consider it trimmed.
2780 */
2781int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2782 u64 bytenr, u64 size)
2783{
2784 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2785
2786 if (btrfs_is_zoned(block_group->fs_info))
2787 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2788 true);
2789
2790 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2791 btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2792 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2793
2794 return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2795}
2796
2797int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2798 u64 offset, u64 bytes)
2799{
2800 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2801 struct btrfs_free_space *info;
2802 int ret;
2803 bool re_search = false;
2804
2805 if (btrfs_is_zoned(block_group->fs_info)) {
2806 /*
2807 * This can happen with conventional zones when replaying log.
2808 * Since the allocation info of tree-log nodes are not recorded
2809 * to the extent-tree, calculate_alloc_pointer() failed to
2810 * advance the allocation pointer after last allocated tree log
2811 * node blocks.
2812 *
2813 * This function is called from
2814 * btrfs_pin_extent_for_log_replay() when replaying the log.
2815 * Advance the pointer not to overwrite the tree-log nodes.
2816 */
2817 if (block_group->start + block_group->alloc_offset <
2818 offset + bytes) {
2819 block_group->alloc_offset =
2820 offset + bytes - block_group->start;
2821 }
2822 return 0;
2823 }
2824
2825 spin_lock(&ctl->tree_lock);
2826
2827again:
2828 ret = 0;
2829 if (!bytes)
2830 goto out_lock;
2831
2832 info = tree_search_offset(ctl, offset, 0, 0);
2833 if (!info) {
2834 /*
2835 * oops didn't find an extent that matched the space we wanted
2836 * to remove, look for a bitmap instead
2837 */
2838 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2839 1, 0);
2840 if (!info) {
2841 /*
2842 * If we found a partial bit of our free space in a
2843 * bitmap but then couldn't find the other part this may
2844 * be a problem, so WARN about it.
2845 */
2846 WARN_ON(re_search);
2847 goto out_lock;
2848 }
2849 }
2850
2851 re_search = false;
2852 if (!info->bitmap) {
2853 unlink_free_space(ctl, info, true);
2854 if (offset == info->offset) {
2855 u64 to_free = min(bytes, info->bytes);
2856
2857 info->bytes -= to_free;
2858 info->offset += to_free;
2859 if (info->bytes) {
2860 ret = link_free_space(ctl, info);
2861 WARN_ON(ret);
2862 } else {
2863 kmem_cache_free(btrfs_free_space_cachep, info);
2864 }
2865
2866 offset += to_free;
2867 bytes -= to_free;
2868 goto again;
2869 } else {
2870 u64 old_end = info->bytes + info->offset;
2871
2872 info->bytes = offset - info->offset;
2873 ret = link_free_space(ctl, info);
2874 WARN_ON(ret);
2875 if (ret)
2876 goto out_lock;
2877
2878 /* Not enough bytes in this entry to satisfy us */
2879 if (old_end < offset + bytes) {
2880 bytes -= old_end - offset;
2881 offset = old_end;
2882 goto again;
2883 } else if (old_end == offset + bytes) {
2884 /* all done */
2885 goto out_lock;
2886 }
2887 spin_unlock(&ctl->tree_lock);
2888
2889 ret = __btrfs_add_free_space(block_group,
2890 offset + bytes,
2891 old_end - (offset + bytes),
2892 info->trim_state);
2893 WARN_ON(ret);
2894 goto out;
2895 }
2896 }
2897
2898 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2899 if (ret == -EAGAIN) {
2900 re_search = true;
2901 goto again;
2902 }
2903out_lock:
2904 btrfs_discard_update_discardable(block_group);
2905 spin_unlock(&ctl->tree_lock);
2906out:
2907 return ret;
2908}
2909
2910void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2911 u64 bytes)
2912{
2913 struct btrfs_fs_info *fs_info = block_group->fs_info;
2914 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2915 struct btrfs_free_space *info;
2916 struct rb_node *n;
2917 int count = 0;
2918
2919 /*
2920 * Zoned btrfs does not use free space tree and cluster. Just print
2921 * out the free space after the allocation offset.
2922 */
2923 if (btrfs_is_zoned(fs_info)) {
2924 btrfs_info(fs_info, "free space %llu active %d",
2925 block_group->zone_capacity - block_group->alloc_offset,
2926 test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2927 &block_group->runtime_flags));
2928 return;
2929 }
2930
2931 spin_lock(&ctl->tree_lock);
2932 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2933 info = rb_entry(n, struct btrfs_free_space, offset_index);
2934 if (info->bytes >= bytes && !block_group->ro)
2935 count++;
2936 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2937 info->offset, info->bytes,
2938 (info->bitmap) ? "yes" : "no");
2939 }
2940 spin_unlock(&ctl->tree_lock);
2941 btrfs_info(fs_info, "block group has cluster?: %s",
2942 list_empty(&block_group->cluster_list) ? "no" : "yes");
2943 btrfs_info(fs_info,
2944 "%d free space entries at or bigger than %llu bytes",
2945 count, bytes);
2946}
2947
2948void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2949 struct btrfs_free_space_ctl *ctl)
2950{
2951 struct btrfs_fs_info *fs_info = block_group->fs_info;
2952
2953 spin_lock_init(&ctl->tree_lock);
2954 ctl->unit = fs_info->sectorsize;
2955 ctl->start = block_group->start;
2956 ctl->block_group = block_group;
2957 ctl->op = &free_space_op;
2958 ctl->free_space_bytes = RB_ROOT_CACHED;
2959 INIT_LIST_HEAD(&ctl->trimming_ranges);
2960 mutex_init(&ctl->cache_writeout_mutex);
2961
2962 /*
2963 * we only want to have 32k of ram per block group for keeping
2964 * track of free space, and if we pass 1/2 of that we want to
2965 * start converting things over to using bitmaps
2966 */
2967 ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2968}
2969
2970/*
2971 * for a given cluster, put all of its extents back into the free
2972 * space cache. If the block group passed doesn't match the block group
2973 * pointed to by the cluster, someone else raced in and freed the
2974 * cluster already. In that case, we just return without changing anything
2975 */
2976static void __btrfs_return_cluster_to_free_space(
2977 struct btrfs_block_group *block_group,
2978 struct btrfs_free_cluster *cluster)
2979{
2980 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2981 struct rb_node *node;
2982
2983 lockdep_assert_held(&ctl->tree_lock);
2984
2985 spin_lock(&cluster->lock);
2986 if (cluster->block_group != block_group) {
2987 spin_unlock(&cluster->lock);
2988 return;
2989 }
2990
2991 cluster->block_group = NULL;
2992 cluster->window_start = 0;
2993 list_del_init(&cluster->block_group_list);
2994
2995 node = rb_first(&cluster->root);
2996 while (node) {
2997 struct btrfs_free_space *entry;
2998
2999 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3000 node = rb_next(&entry->offset_index);
3001 rb_erase(&entry->offset_index, &cluster->root);
3002 RB_CLEAR_NODE(&entry->offset_index);
3003
3004 if (!entry->bitmap) {
3005 /* Merging treats extents as if they were new */
3006 if (!btrfs_free_space_trimmed(entry)) {
3007 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3008 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
3009 entry->bytes;
3010 }
3011
3012 try_merge_free_space(ctl, entry, false);
3013 steal_from_bitmap(ctl, entry, false);
3014
3015 /* As we insert directly, update these statistics */
3016 if (!btrfs_free_space_trimmed(entry)) {
3017 ctl->discardable_extents[BTRFS_STAT_CURR]++;
3018 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
3019 entry->bytes;
3020 }
3021 }
3022 tree_insert_offset(ctl, NULL, entry);
3023 rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
3024 entry_less);
3025 }
3026 cluster->root = RB_ROOT;
3027 spin_unlock(&cluster->lock);
3028 btrfs_put_block_group(block_group);
3029}
3030
3031void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
3032{
3033 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3034 struct btrfs_free_cluster *cluster;
3035 struct list_head *head;
3036
3037 spin_lock(&ctl->tree_lock);
3038 while ((head = block_group->cluster_list.next) !=
3039 &block_group->cluster_list) {
3040 cluster = list_entry(head, struct btrfs_free_cluster,
3041 block_group_list);
3042
3043 WARN_ON(cluster->block_group != block_group);
3044 __btrfs_return_cluster_to_free_space(block_group, cluster);
3045
3046 cond_resched_lock(&ctl->tree_lock);
3047 }
3048 __btrfs_remove_free_space_cache(ctl);
3049 btrfs_discard_update_discardable(block_group);
3050 spin_unlock(&ctl->tree_lock);
3051
3052}
3053
3054/*
3055 * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3056 */
3057bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3058{
3059 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3060 struct btrfs_free_space *info;
3061 struct rb_node *node;
3062 bool ret = true;
3063
3064 spin_lock(&ctl->tree_lock);
3065 node = rb_first(&ctl->free_space_offset);
3066
3067 while (node) {
3068 info = rb_entry(node, struct btrfs_free_space, offset_index);
3069
3070 if (!btrfs_free_space_trimmed(info)) {
3071 ret = false;
3072 break;
3073 }
3074
3075 node = rb_next(node);
3076 }
3077
3078 spin_unlock(&ctl->tree_lock);
3079 return ret;
3080}
3081
3082u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3083 u64 offset, u64 bytes, u64 empty_size,
3084 u64 *max_extent_size)
3085{
3086 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3087 struct btrfs_discard_ctl *discard_ctl =
3088 &block_group->fs_info->discard_ctl;
3089 struct btrfs_free_space *entry = NULL;
3090 u64 bytes_search = bytes + empty_size;
3091 u64 ret = 0;
3092 u64 align_gap = 0;
3093 u64 align_gap_len = 0;
3094 enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3095 bool use_bytes_index = (offset == block_group->start);
3096
3097 ASSERT(!btrfs_is_zoned(block_group->fs_info));
3098
3099 spin_lock(&ctl->tree_lock);
3100 entry = find_free_space(ctl, &offset, &bytes_search,
3101 block_group->full_stripe_len, max_extent_size,
3102 use_bytes_index);
3103 if (!entry)
3104 goto out;
3105
3106 ret = offset;
3107 if (entry->bitmap) {
3108 bitmap_clear_bits(ctl, entry, offset, bytes, true);
3109
3110 if (!btrfs_free_space_trimmed(entry))
3111 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3112
3113 if (!entry->bytes)
3114 free_bitmap(ctl, entry);
3115 } else {
3116 unlink_free_space(ctl, entry, true);
3117 align_gap_len = offset - entry->offset;
3118 align_gap = entry->offset;
3119 align_gap_trim_state = entry->trim_state;
3120
3121 if (!btrfs_free_space_trimmed(entry))
3122 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3123
3124 entry->offset = offset + bytes;
3125 WARN_ON(entry->bytes < bytes + align_gap_len);
3126
3127 entry->bytes -= bytes + align_gap_len;
3128 if (!entry->bytes)
3129 kmem_cache_free(btrfs_free_space_cachep, entry);
3130 else
3131 link_free_space(ctl, entry);
3132 }
3133out:
3134 btrfs_discard_update_discardable(block_group);
3135 spin_unlock(&ctl->tree_lock);
3136
3137 if (align_gap_len)
3138 __btrfs_add_free_space(block_group, align_gap, align_gap_len,
3139 align_gap_trim_state);
3140 return ret;
3141}
3142
3143/*
3144 * given a cluster, put all of its extents back into the free space
3145 * cache. If a block group is passed, this function will only free
3146 * a cluster that belongs to the passed block group.
3147 *
3148 * Otherwise, it'll get a reference on the block group pointed to by the
3149 * cluster and remove the cluster from it.
3150 */
3151void btrfs_return_cluster_to_free_space(
3152 struct btrfs_block_group *block_group,
3153 struct btrfs_free_cluster *cluster)
3154{
3155 struct btrfs_free_space_ctl *ctl;
3156
3157 /* first, get a safe pointer to the block group */
3158 spin_lock(&cluster->lock);
3159 if (!block_group) {
3160 block_group = cluster->block_group;
3161 if (!block_group) {
3162 spin_unlock(&cluster->lock);
3163 return;
3164 }
3165 } else if (cluster->block_group != block_group) {
3166 /* someone else has already freed it don't redo their work */
3167 spin_unlock(&cluster->lock);
3168 return;
3169 }
3170 btrfs_get_block_group(block_group);
3171 spin_unlock(&cluster->lock);
3172
3173 ctl = block_group->free_space_ctl;
3174
3175 /* now return any extents the cluster had on it */
3176 spin_lock(&ctl->tree_lock);
3177 __btrfs_return_cluster_to_free_space(block_group, cluster);
3178 spin_unlock(&ctl->tree_lock);
3179
3180 btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3181
3182 /* finally drop our ref */
3183 btrfs_put_block_group(block_group);
3184}
3185
3186static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3187 struct btrfs_free_cluster *cluster,
3188 struct btrfs_free_space *entry,
3189 u64 bytes, u64 min_start,
3190 u64 *max_extent_size)
3191{
3192 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3193 int err;
3194 u64 search_start = cluster->window_start;
3195 u64 search_bytes = bytes;
3196 u64 ret = 0;
3197
3198 search_start = min_start;
3199 search_bytes = bytes;
3200
3201 err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3202 if (err) {
3203 *max_extent_size = max(get_max_extent_size(entry),
3204 *max_extent_size);
3205 return 0;
3206 }
3207
3208 ret = search_start;
3209 bitmap_clear_bits(ctl, entry, ret, bytes, false);
3210
3211 return ret;
3212}
3213
3214/*
3215 * given a cluster, try to allocate 'bytes' from it, returns 0
3216 * if it couldn't find anything suitably large, or a logical disk offset
3217 * if things worked out
3218 */
3219u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3220 struct btrfs_free_cluster *cluster, u64 bytes,
3221 u64 min_start, u64 *max_extent_size)
3222{
3223 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3224 struct btrfs_discard_ctl *discard_ctl =
3225 &block_group->fs_info->discard_ctl;
3226 struct btrfs_free_space *entry = NULL;
3227 struct rb_node *node;
3228 u64 ret = 0;
3229
3230 ASSERT(!btrfs_is_zoned(block_group->fs_info));
3231
3232 spin_lock(&cluster->lock);
3233 if (bytes > cluster->max_size)
3234 goto out;
3235
3236 if (cluster->block_group != block_group)
3237 goto out;
3238
3239 node = rb_first(&cluster->root);
3240 if (!node)
3241 goto out;
3242
3243 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3244 while (1) {
3245 if (entry->bytes < bytes)
3246 *max_extent_size = max(get_max_extent_size(entry),
3247 *max_extent_size);
3248
3249 if (entry->bytes < bytes ||
3250 (!entry->bitmap && entry->offset < min_start)) {
3251 node = rb_next(&entry->offset_index);
3252 if (!node)
3253 break;
3254 entry = rb_entry(node, struct btrfs_free_space,
3255 offset_index);
3256 continue;
3257 }
3258
3259 if (entry->bitmap) {
3260 ret = btrfs_alloc_from_bitmap(block_group,
3261 cluster, entry, bytes,
3262 cluster->window_start,
3263 max_extent_size);
3264 if (ret == 0) {
3265 node = rb_next(&entry->offset_index);
3266 if (!node)
3267 break;
3268 entry = rb_entry(node, struct btrfs_free_space,
3269 offset_index);
3270 continue;
3271 }
3272 cluster->window_start += bytes;
3273 } else {
3274 ret = entry->offset;
3275
3276 entry->offset += bytes;
3277 entry->bytes -= bytes;
3278 }
3279
3280 break;
3281 }
3282out:
3283 spin_unlock(&cluster->lock);
3284
3285 if (!ret)
3286 return 0;
3287
3288 spin_lock(&ctl->tree_lock);
3289
3290 if (!btrfs_free_space_trimmed(entry))
3291 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3292
3293 ctl->free_space -= bytes;
3294 if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3295 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3296
3297 spin_lock(&cluster->lock);
3298 if (entry->bytes == 0) {
3299 rb_erase(&entry->offset_index, &cluster->root);
3300 ctl->free_extents--;
3301 if (entry->bitmap) {
3302 kmem_cache_free(btrfs_free_space_bitmap_cachep,
3303 entry->bitmap);
3304 ctl->total_bitmaps--;
3305 recalculate_thresholds(ctl);
3306 } else if (!btrfs_free_space_trimmed(entry)) {
3307 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3308 }
3309 kmem_cache_free(btrfs_free_space_cachep, entry);
3310 }
3311
3312 spin_unlock(&cluster->lock);
3313 spin_unlock(&ctl->tree_lock);
3314
3315 return ret;
3316}
3317
3318static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3319 struct btrfs_free_space *entry,
3320 struct btrfs_free_cluster *cluster,
3321 u64 offset, u64 bytes,
3322 u64 cont1_bytes, u64 min_bytes)
3323{
3324 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3325 unsigned long next_zero;
3326 unsigned long i;
3327 unsigned long want_bits;
3328 unsigned long min_bits;
3329 unsigned long found_bits;
3330 unsigned long max_bits = 0;
3331 unsigned long start = 0;
3332 unsigned long total_found = 0;
3333 int ret;
3334
3335 lockdep_assert_held(&ctl->tree_lock);
3336
3337 i = offset_to_bit(entry->offset, ctl->unit,
3338 max_t(u64, offset, entry->offset));
3339 want_bits = bytes_to_bits(bytes, ctl->unit);
3340 min_bits = bytes_to_bits(min_bytes, ctl->unit);
3341
3342 /*
3343 * Don't bother looking for a cluster in this bitmap if it's heavily
3344 * fragmented.
3345 */
3346 if (entry->max_extent_size &&
3347 entry->max_extent_size < cont1_bytes)
3348 return -ENOSPC;
3349again:
3350 found_bits = 0;
3351 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3352 next_zero = find_next_zero_bit(entry->bitmap,
3353 BITS_PER_BITMAP, i);
3354 if (next_zero - i >= min_bits) {
3355 found_bits = next_zero - i;
3356 if (found_bits > max_bits)
3357 max_bits = found_bits;
3358 break;
3359 }
3360 if (next_zero - i > max_bits)
3361 max_bits = next_zero - i;
3362 i = next_zero;
3363 }
3364
3365 if (!found_bits) {
3366 entry->max_extent_size = (u64)max_bits * ctl->unit;
3367 return -ENOSPC;
3368 }
3369
3370 if (!total_found) {
3371 start = i;
3372 cluster->max_size = 0;
3373 }
3374
3375 total_found += found_bits;
3376
3377 if (cluster->max_size < found_bits * ctl->unit)
3378 cluster->max_size = found_bits * ctl->unit;
3379
3380 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3381 i = next_zero + 1;
3382 goto again;
3383 }
3384
3385 cluster->window_start = start * ctl->unit + entry->offset;
3386 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3387 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3388
3389 /*
3390 * We need to know if we're currently on the normal space index when we
3391 * manipulate the bitmap so that we know we need to remove and re-insert
3392 * it into the space_index tree. Clear the bytes_index node here so the
3393 * bitmap manipulation helpers know not to mess with the space_index
3394 * until this bitmap entry is added back into the normal cache.
3395 */
3396 RB_CLEAR_NODE(&entry->bytes_index);
3397
3398 ret = tree_insert_offset(ctl, cluster, entry);
3399 ASSERT(!ret); /* -EEXIST; Logic error */
3400
3401 trace_btrfs_setup_cluster(block_group, cluster,
3402 total_found * ctl->unit, 1);
3403 return 0;
3404}
3405
3406/*
3407 * This searches the block group for just extents to fill the cluster with.
3408 * Try to find a cluster with at least bytes total bytes, at least one
3409 * extent of cont1_bytes, and other clusters of at least min_bytes.
3410 */
3411static noinline int
3412setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3413 struct btrfs_free_cluster *cluster,
3414 struct list_head *bitmaps, u64 offset, u64 bytes,
3415 u64 cont1_bytes, u64 min_bytes)
3416{
3417 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3418 struct btrfs_free_space *first = NULL;
3419 struct btrfs_free_space *entry = NULL;
3420 struct btrfs_free_space *last;
3421 struct rb_node *node;
3422 u64 window_free;
3423 u64 max_extent;
3424 u64 total_size = 0;
3425
3426 lockdep_assert_held(&ctl->tree_lock);
3427
3428 entry = tree_search_offset(ctl, offset, 0, 1);
3429 if (!entry)
3430 return -ENOSPC;
3431
3432 /*
3433 * We don't want bitmaps, so just move along until we find a normal
3434 * extent entry.
3435 */
3436 while (entry->bitmap || entry->bytes < min_bytes) {
3437 if (entry->bitmap && list_empty(&entry->list))
3438 list_add_tail(&entry->list, bitmaps);
3439 node = rb_next(&entry->offset_index);
3440 if (!node)
3441 return -ENOSPC;
3442 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3443 }
3444
3445 window_free = entry->bytes;
3446 max_extent = entry->bytes;
3447 first = entry;
3448 last = entry;
3449
3450 for (node = rb_next(&entry->offset_index); node;
3451 node = rb_next(&entry->offset_index)) {
3452 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3453
3454 if (entry->bitmap) {
3455 if (list_empty(&entry->list))
3456 list_add_tail(&entry->list, bitmaps);
3457 continue;
3458 }
3459
3460 if (entry->bytes < min_bytes)
3461 continue;
3462
3463 last = entry;
3464 window_free += entry->bytes;
3465 if (entry->bytes > max_extent)
3466 max_extent = entry->bytes;
3467 }
3468
3469 if (window_free < bytes || max_extent < cont1_bytes)
3470 return -ENOSPC;
3471
3472 cluster->window_start = first->offset;
3473
3474 node = &first->offset_index;
3475
3476 /*
3477 * now we've found our entries, pull them out of the free space
3478 * cache and put them into the cluster rbtree
3479 */
3480 do {
3481 int ret;
3482
3483 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3484 node = rb_next(&entry->offset_index);
3485 if (entry->bitmap || entry->bytes < min_bytes)
3486 continue;
3487
3488 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3489 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3490 ret = tree_insert_offset(ctl, cluster, entry);
3491 total_size += entry->bytes;
3492 ASSERT(!ret); /* -EEXIST; Logic error */
3493 } while (node && entry != last);
3494
3495 cluster->max_size = max_extent;
3496 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3497 return 0;
3498}
3499
3500/*
3501 * This specifically looks for bitmaps that may work in the cluster, we assume
3502 * that we have already failed to find extents that will work.
3503 */
3504static noinline int
3505setup_cluster_bitmap(struct btrfs_block_group *block_group,
3506 struct btrfs_free_cluster *cluster,
3507 struct list_head *bitmaps, u64 offset, u64 bytes,
3508 u64 cont1_bytes, u64 min_bytes)
3509{
3510 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3511 struct btrfs_free_space *entry = NULL;
3512 int ret = -ENOSPC;
3513 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3514
3515 if (ctl->total_bitmaps == 0)
3516 return -ENOSPC;
3517
3518 /*
3519 * The bitmap that covers offset won't be in the list unless offset
3520 * is just its start offset.
3521 */
3522 if (!list_empty(bitmaps))
3523 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3524
3525 if (!entry || entry->offset != bitmap_offset) {
3526 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3527 if (entry && list_empty(&entry->list))
3528 list_add(&entry->list, bitmaps);
3529 }
3530
3531 list_for_each_entry(entry, bitmaps, list) {
3532 if (entry->bytes < bytes)
3533 continue;
3534 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3535 bytes, cont1_bytes, min_bytes);
3536 if (!ret)
3537 return 0;
3538 }
3539
3540 /*
3541 * The bitmaps list has all the bitmaps that record free space
3542 * starting after offset, so no more search is required.
3543 */
3544 return -ENOSPC;
3545}
3546
3547/*
3548 * here we try to find a cluster of blocks in a block group. The goal
3549 * is to find at least bytes+empty_size.
3550 * We might not find them all in one contiguous area.
3551 *
3552 * returns zero and sets up cluster if things worked out, otherwise
3553 * it returns -enospc
3554 */
3555int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3556 struct btrfs_free_cluster *cluster,
3557 u64 offset, u64 bytes, u64 empty_size)
3558{
3559 struct btrfs_fs_info *fs_info = block_group->fs_info;
3560 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3561 struct btrfs_free_space *entry, *tmp;
3562 LIST_HEAD(bitmaps);
3563 u64 min_bytes;
3564 u64 cont1_bytes;
3565 int ret;
3566
3567 /*
3568 * Choose the minimum extent size we'll require for this
3569 * cluster. For SSD_SPREAD, don't allow any fragmentation.
3570 * For metadata, allow allocates with smaller extents. For
3571 * data, keep it dense.
3572 */
3573 if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3574 cont1_bytes = bytes + empty_size;
3575 min_bytes = cont1_bytes;
3576 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3577 cont1_bytes = bytes;
3578 min_bytes = fs_info->sectorsize;
3579 } else {
3580 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3581 min_bytes = fs_info->sectorsize;
3582 }
3583
3584 spin_lock(&ctl->tree_lock);
3585
3586 /*
3587 * If we know we don't have enough space to make a cluster don't even
3588 * bother doing all the work to try and find one.
3589 */
3590 if (ctl->free_space < bytes) {
3591 spin_unlock(&ctl->tree_lock);
3592 return -ENOSPC;
3593 }
3594
3595 spin_lock(&cluster->lock);
3596
3597 /* someone already found a cluster, hooray */
3598 if (cluster->block_group) {
3599 ret = 0;
3600 goto out;
3601 }
3602
3603 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3604 min_bytes);
3605
3606 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3607 bytes + empty_size,
3608 cont1_bytes, min_bytes);
3609 if (ret)
3610 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3611 offset, bytes + empty_size,
3612 cont1_bytes, min_bytes);
3613
3614 /* Clear our temporary list */
3615 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3616 list_del_init(&entry->list);
3617
3618 if (!ret) {
3619 btrfs_get_block_group(block_group);
3620 list_add_tail(&cluster->block_group_list,
3621 &block_group->cluster_list);
3622 cluster->block_group = block_group;
3623 } else {
3624 trace_btrfs_failed_cluster_setup(block_group);
3625 }
3626out:
3627 spin_unlock(&cluster->lock);
3628 spin_unlock(&ctl->tree_lock);
3629
3630 return ret;
3631}
3632
3633/*
3634 * simple code to zero out a cluster
3635 */
3636void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3637{
3638 spin_lock_init(&cluster->lock);
3639 spin_lock_init(&cluster->refill_lock);
3640 cluster->root = RB_ROOT;
3641 cluster->max_size = 0;
3642 cluster->fragmented = false;
3643 INIT_LIST_HEAD(&cluster->block_group_list);
3644 cluster->block_group = NULL;
3645}
3646
3647static int do_trimming(struct btrfs_block_group *block_group,
3648 u64 *total_trimmed, u64 start, u64 bytes,
3649 u64 reserved_start, u64 reserved_bytes,
3650 enum btrfs_trim_state reserved_trim_state,
3651 struct btrfs_trim_range *trim_entry)
3652{
3653 struct btrfs_space_info *space_info = block_group->space_info;
3654 struct btrfs_fs_info *fs_info = block_group->fs_info;
3655 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3656 int ret;
3657 int update = 0;
3658 const u64 end = start + bytes;
3659 const u64 reserved_end = reserved_start + reserved_bytes;
3660 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3661 u64 trimmed = 0;
3662
3663 spin_lock(&space_info->lock);
3664 spin_lock(&block_group->lock);
3665 if (!block_group->ro) {
3666 block_group->reserved += reserved_bytes;
3667 space_info->bytes_reserved += reserved_bytes;
3668 update = 1;
3669 }
3670 spin_unlock(&block_group->lock);
3671 spin_unlock(&space_info->lock);
3672
3673 ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3674 if (!ret) {
3675 *total_trimmed += trimmed;
3676 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3677 }
3678
3679 mutex_lock(&ctl->cache_writeout_mutex);
3680 if (reserved_start < start)
3681 __btrfs_add_free_space(block_group, reserved_start,
3682 start - reserved_start,
3683 reserved_trim_state);
3684 if (end < reserved_end)
3685 __btrfs_add_free_space(block_group, end, reserved_end - end,
3686 reserved_trim_state);
3687 __btrfs_add_free_space(block_group, start, bytes, trim_state);
3688 list_del(&trim_entry->list);
3689 mutex_unlock(&ctl->cache_writeout_mutex);
3690
3691 if (update) {
3692 spin_lock(&space_info->lock);
3693 spin_lock(&block_group->lock);
3694 if (block_group->ro)
3695 space_info->bytes_readonly += reserved_bytes;
3696 block_group->reserved -= reserved_bytes;
3697 space_info->bytes_reserved -= reserved_bytes;
3698 spin_unlock(&block_group->lock);
3699 spin_unlock(&space_info->lock);
3700 }
3701
3702 return ret;
3703}
3704
3705/*
3706 * If @async is set, then we will trim 1 region and return.
3707 */
3708static int trim_no_bitmap(struct btrfs_block_group *block_group,
3709 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3710 bool async)
3711{
3712 struct btrfs_discard_ctl *discard_ctl =
3713 &block_group->fs_info->discard_ctl;
3714 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3715 struct btrfs_free_space *entry;
3716 struct rb_node *node;
3717 int ret = 0;
3718 u64 extent_start;
3719 u64 extent_bytes;
3720 enum btrfs_trim_state extent_trim_state;
3721 u64 bytes;
3722 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3723
3724 while (start < end) {
3725 struct btrfs_trim_range trim_entry;
3726
3727 mutex_lock(&ctl->cache_writeout_mutex);
3728 spin_lock(&ctl->tree_lock);
3729
3730 if (ctl->free_space < minlen)
3731 goto out_unlock;
3732
3733 entry = tree_search_offset(ctl, start, 0, 1);
3734 if (!entry)
3735 goto out_unlock;
3736
3737 /* Skip bitmaps and if async, already trimmed entries */
3738 while (entry->bitmap ||
3739 (async && btrfs_free_space_trimmed(entry))) {
3740 node = rb_next(&entry->offset_index);
3741 if (!node)
3742 goto out_unlock;
3743 entry = rb_entry(node, struct btrfs_free_space,
3744 offset_index);
3745 }
3746
3747 if (entry->offset >= end)
3748 goto out_unlock;
3749
3750 extent_start = entry->offset;
3751 extent_bytes = entry->bytes;
3752 extent_trim_state = entry->trim_state;
3753 if (async) {
3754 start = entry->offset;
3755 bytes = entry->bytes;
3756 if (bytes < minlen) {
3757 spin_unlock(&ctl->tree_lock);
3758 mutex_unlock(&ctl->cache_writeout_mutex);
3759 goto next;
3760 }
3761 unlink_free_space(ctl, entry, true);
3762 /*
3763 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3764 * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3765 * X when we come back around. So trim it now.
3766 */
3767 if (max_discard_size &&
3768 bytes >= (max_discard_size +
3769 BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3770 bytes = max_discard_size;
3771 extent_bytes = max_discard_size;
3772 entry->offset += max_discard_size;
3773 entry->bytes -= max_discard_size;
3774 link_free_space(ctl, entry);
3775 } else {
3776 kmem_cache_free(btrfs_free_space_cachep, entry);
3777 }
3778 } else {
3779 start = max(start, extent_start);
3780 bytes = min(extent_start + extent_bytes, end) - start;
3781 if (bytes < minlen) {
3782 spin_unlock(&ctl->tree_lock);
3783 mutex_unlock(&ctl->cache_writeout_mutex);
3784 goto next;
3785 }
3786
3787 unlink_free_space(ctl, entry, true);
3788 kmem_cache_free(btrfs_free_space_cachep, entry);
3789 }
3790
3791 spin_unlock(&ctl->tree_lock);
3792 trim_entry.start = extent_start;
3793 trim_entry.bytes = extent_bytes;
3794 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3795 mutex_unlock(&ctl->cache_writeout_mutex);
3796
3797 ret = do_trimming(block_group, total_trimmed, start, bytes,
3798 extent_start, extent_bytes, extent_trim_state,
3799 &trim_entry);
3800 if (ret) {
3801 block_group->discard_cursor = start + bytes;
3802 break;
3803 }
3804next:
3805 start += bytes;
3806 block_group->discard_cursor = start;
3807 if (async && *total_trimmed)
3808 break;
3809
3810 if (fatal_signal_pending(current)) {
3811 ret = -ERESTARTSYS;
3812 break;
3813 }
3814
3815 cond_resched();
3816 }
3817
3818 return ret;
3819
3820out_unlock:
3821 block_group->discard_cursor = btrfs_block_group_end(block_group);
3822 spin_unlock(&ctl->tree_lock);
3823 mutex_unlock(&ctl->cache_writeout_mutex);
3824
3825 return ret;
3826}
3827
3828/*
3829 * If we break out of trimming a bitmap prematurely, we should reset the
3830 * trimming bit. In a rather contrieved case, it's possible to race here so
3831 * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3832 *
3833 * start = start of bitmap
3834 * end = near end of bitmap
3835 *
3836 * Thread 1: Thread 2:
3837 * trim_bitmaps(start)
3838 * trim_bitmaps(end)
3839 * end_trimming_bitmap()
3840 * reset_trimming_bitmap()
3841 */
3842static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3843{
3844 struct btrfs_free_space *entry;
3845
3846 spin_lock(&ctl->tree_lock);
3847 entry = tree_search_offset(ctl, offset, 1, 0);
3848 if (entry) {
3849 if (btrfs_free_space_trimmed(entry)) {
3850 ctl->discardable_extents[BTRFS_STAT_CURR] +=
3851 entry->bitmap_extents;
3852 ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3853 }
3854 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3855 }
3856
3857 spin_unlock(&ctl->tree_lock);
3858}
3859
3860static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3861 struct btrfs_free_space *entry)
3862{
3863 if (btrfs_free_space_trimming_bitmap(entry)) {
3864 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3865 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3866 entry->bitmap_extents;
3867 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3868 }
3869}
3870
3871/*
3872 * If @async is set, then we will trim 1 region and return.
3873 */
3874static int trim_bitmaps(struct btrfs_block_group *block_group,
3875 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3876 u64 maxlen, bool async)
3877{
3878 struct btrfs_discard_ctl *discard_ctl =
3879 &block_group->fs_info->discard_ctl;
3880 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3881 struct btrfs_free_space *entry;
3882 int ret = 0;
3883 int ret2;
3884 u64 bytes;
3885 u64 offset = offset_to_bitmap(ctl, start);
3886 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3887
3888 while (offset < end) {
3889 bool next_bitmap = false;
3890 struct btrfs_trim_range trim_entry;
3891
3892 mutex_lock(&ctl->cache_writeout_mutex);
3893 spin_lock(&ctl->tree_lock);
3894
3895 if (ctl->free_space < minlen) {
3896 block_group->discard_cursor =
3897 btrfs_block_group_end(block_group);
3898 spin_unlock(&ctl->tree_lock);
3899 mutex_unlock(&ctl->cache_writeout_mutex);
3900 break;
3901 }
3902
3903 entry = tree_search_offset(ctl, offset, 1, 0);
3904 /*
3905 * Bitmaps are marked trimmed lossily now to prevent constant
3906 * discarding of the same bitmap (the reason why we are bound
3907 * by the filters). So, retrim the block group bitmaps when we
3908 * are preparing to punt to the unused_bgs list. This uses
3909 * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3910 * which is the only discard index which sets minlen to 0.
3911 */
3912 if (!entry || (async && minlen && start == offset &&
3913 btrfs_free_space_trimmed(entry))) {
3914 spin_unlock(&ctl->tree_lock);
3915 mutex_unlock(&ctl->cache_writeout_mutex);
3916 next_bitmap = true;
3917 goto next;
3918 }
3919
3920 /*
3921 * Async discard bitmap trimming begins at by setting the start
3922 * to be key.objectid and the offset_to_bitmap() aligns to the
3923 * start of the bitmap. This lets us know we are fully
3924 * scanning the bitmap rather than only some portion of it.
3925 */
3926 if (start == offset)
3927 entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3928
3929 bytes = minlen;
3930 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3931 if (ret2 || start >= end) {
3932 /*
3933 * We lossily consider a bitmap trimmed if we only skip
3934 * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3935 */
3936 if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3937 end_trimming_bitmap(ctl, entry);
3938 else
3939 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3940 spin_unlock(&ctl->tree_lock);
3941 mutex_unlock(&ctl->cache_writeout_mutex);
3942 next_bitmap = true;
3943 goto next;
3944 }
3945
3946 /*
3947 * We already trimmed a region, but are using the locking above
3948 * to reset the trim_state.
3949 */
3950 if (async && *total_trimmed) {
3951 spin_unlock(&ctl->tree_lock);
3952 mutex_unlock(&ctl->cache_writeout_mutex);
3953 goto out;
3954 }
3955
3956 bytes = min(bytes, end - start);
3957 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3958 spin_unlock(&ctl->tree_lock);
3959 mutex_unlock(&ctl->cache_writeout_mutex);
3960 goto next;
3961 }
3962
3963 /*
3964 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3965 * If X < @minlen, we won't trim X when we come back around.
3966 * So trim it now. We differ here from trimming extents as we
3967 * don't keep individual state per bit.
3968 */
3969 if (async &&
3970 max_discard_size &&
3971 bytes > (max_discard_size + minlen))
3972 bytes = max_discard_size;
3973
3974 bitmap_clear_bits(ctl, entry, start, bytes, true);
3975 if (entry->bytes == 0)
3976 free_bitmap(ctl, entry);
3977
3978 spin_unlock(&ctl->tree_lock);
3979 trim_entry.start = start;
3980 trim_entry.bytes = bytes;
3981 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3982 mutex_unlock(&ctl->cache_writeout_mutex);
3983
3984 ret = do_trimming(block_group, total_trimmed, start, bytes,
3985 start, bytes, 0, &trim_entry);
3986 if (ret) {
3987 reset_trimming_bitmap(ctl, offset);
3988 block_group->discard_cursor =
3989 btrfs_block_group_end(block_group);
3990 break;
3991 }
3992next:
3993 if (next_bitmap) {
3994 offset += BITS_PER_BITMAP * ctl->unit;
3995 start = offset;
3996 } else {
3997 start += bytes;
3998 }
3999 block_group->discard_cursor = start;
4000
4001 if (fatal_signal_pending(current)) {
4002 if (start != offset)
4003 reset_trimming_bitmap(ctl, offset);
4004 ret = -ERESTARTSYS;
4005 break;
4006 }
4007
4008 cond_resched();
4009 }
4010
4011 if (offset >= end)
4012 block_group->discard_cursor = end;
4013
4014out:
4015 return ret;
4016}
4017
4018int btrfs_trim_block_group(struct btrfs_block_group *block_group,
4019 u64 *trimmed, u64 start, u64 end, u64 minlen)
4020{
4021 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
4022 int ret;
4023 u64 rem = 0;
4024
4025 ASSERT(!btrfs_is_zoned(block_group->fs_info));
4026
4027 *trimmed = 0;
4028
4029 spin_lock(&block_group->lock);
4030 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4031 spin_unlock(&block_group->lock);
4032 return 0;
4033 }
4034 btrfs_freeze_block_group(block_group);
4035 spin_unlock(&block_group->lock);
4036
4037 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4038 if (ret)
4039 goto out;
4040
4041 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4042 div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4043 /* If we ended in the middle of a bitmap, reset the trimming flag */
4044 if (rem)
4045 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4046out:
4047 btrfs_unfreeze_block_group(block_group);
4048 return ret;
4049}
4050
4051int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4052 u64 *trimmed, u64 start, u64 end, u64 minlen,
4053 bool async)
4054{
4055 int ret;
4056
4057 *trimmed = 0;
4058
4059 spin_lock(&block_group->lock);
4060 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4061 spin_unlock(&block_group->lock);
4062 return 0;
4063 }
4064 btrfs_freeze_block_group(block_group);
4065 spin_unlock(&block_group->lock);
4066
4067 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4068 btrfs_unfreeze_block_group(block_group);
4069
4070 return ret;
4071}
4072
4073int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4074 u64 *trimmed, u64 start, u64 end, u64 minlen,
4075 u64 maxlen, bool async)
4076{
4077 int ret;
4078
4079 *trimmed = 0;
4080
4081 spin_lock(&block_group->lock);
4082 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4083 spin_unlock(&block_group->lock);
4084 return 0;
4085 }
4086 btrfs_freeze_block_group(block_group);
4087 spin_unlock(&block_group->lock);
4088
4089 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4090 async);
4091
4092 btrfs_unfreeze_block_group(block_group);
4093
4094 return ret;
4095}
4096
4097bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4098{
4099 return btrfs_super_cache_generation(fs_info->super_copy);
4100}
4101
4102static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4103 struct btrfs_trans_handle *trans)
4104{
4105 struct btrfs_block_group *block_group;
4106 struct rb_node *node;
4107 int ret = 0;
4108
4109 btrfs_info(fs_info, "cleaning free space cache v1");
4110
4111 node = rb_first_cached(&fs_info->block_group_cache_tree);
4112 while (node) {
4113 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4114 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4115 if (ret)
4116 goto out;
4117 node = rb_next(node);
4118 }
4119out:
4120 return ret;
4121}
4122
4123int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4124{
4125 struct btrfs_trans_handle *trans;
4126 int ret;
4127
4128 /*
4129 * update_super_roots will appropriately set or unset
4130 * super_copy->cache_generation based on SPACE_CACHE and
4131 * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4132 * transaction commit whether we are enabling space cache v1 and don't
4133 * have any other work to do, or are disabling it and removing free
4134 * space inodes.
4135 */
4136 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4137 if (IS_ERR(trans))
4138 return PTR_ERR(trans);
4139
4140 if (!active) {
4141 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4142 ret = cleanup_free_space_cache_v1(fs_info, trans);
4143 if (ret) {
4144 btrfs_abort_transaction(trans, ret);
4145 btrfs_end_transaction(trans);
4146 goto out;
4147 }
4148 }
4149
4150 ret = btrfs_commit_transaction(trans);
4151out:
4152 clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4153
4154 return ret;
4155}
4156
4157int __init btrfs_free_space_init(void)
4158{
4159 btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
4160 sizeof(struct btrfs_free_space), 0,
4161 SLAB_MEM_SPREAD, NULL);
4162 if (!btrfs_free_space_cachep)
4163 return -ENOMEM;
4164
4165 btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
4166 PAGE_SIZE, PAGE_SIZE,
4167 SLAB_MEM_SPREAD, NULL);
4168 if (!btrfs_free_space_bitmap_cachep) {
4169 kmem_cache_destroy(btrfs_free_space_cachep);
4170 return -ENOMEM;
4171 }
4172
4173 return 0;
4174}
4175
4176void __cold btrfs_free_space_exit(void)
4177{
4178 kmem_cache_destroy(btrfs_free_space_cachep);
4179 kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
4180}
4181
4182#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4183/*
4184 * Use this if you need to make a bitmap or extent entry specifically, it
4185 * doesn't do any of the merging that add_free_space does, this acts a lot like
4186 * how the free space cache loading stuff works, so you can get really weird
4187 * configurations.
4188 */
4189int test_add_free_space_entry(struct btrfs_block_group *cache,
4190 u64 offset, u64 bytes, bool bitmap)
4191{
4192 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4193 struct btrfs_free_space *info = NULL, *bitmap_info;
4194 void *map = NULL;
4195 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4196 u64 bytes_added;
4197 int ret;
4198
4199again:
4200 if (!info) {
4201 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4202 if (!info)
4203 return -ENOMEM;
4204 }
4205
4206 if (!bitmap) {
4207 spin_lock(&ctl->tree_lock);
4208 info->offset = offset;
4209 info->bytes = bytes;
4210 info->max_extent_size = 0;
4211 ret = link_free_space(ctl, info);
4212 spin_unlock(&ctl->tree_lock);
4213 if (ret)
4214 kmem_cache_free(btrfs_free_space_cachep, info);
4215 return ret;
4216 }
4217
4218 if (!map) {
4219 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4220 if (!map) {
4221 kmem_cache_free(btrfs_free_space_cachep, info);
4222 return -ENOMEM;
4223 }
4224 }
4225
4226 spin_lock(&ctl->tree_lock);
4227 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4228 1, 0);
4229 if (!bitmap_info) {
4230 info->bitmap = map;
4231 map = NULL;
4232 add_new_bitmap(ctl, info, offset);
4233 bitmap_info = info;
4234 info = NULL;
4235 }
4236
4237 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4238 trim_state);
4239
4240 bytes -= bytes_added;
4241 offset += bytes_added;
4242 spin_unlock(&ctl->tree_lock);
4243
4244 if (bytes)
4245 goto again;
4246
4247 if (info)
4248 kmem_cache_free(btrfs_free_space_cachep, info);
4249 if (map)
4250 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4251 return 0;
4252}
4253
4254/*
4255 * Checks to see if the given range is in the free space cache. This is really
4256 * just used to check the absence of space, so if there is free space in the
4257 * range at all we will return 1.
4258 */
4259int test_check_exists(struct btrfs_block_group *cache,
4260 u64 offset, u64 bytes)
4261{
4262 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4263 struct btrfs_free_space *info;
4264 int ret = 0;
4265
4266 spin_lock(&ctl->tree_lock);
4267 info = tree_search_offset(ctl, offset, 0, 0);
4268 if (!info) {
4269 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4270 1, 0);
4271 if (!info)
4272 goto out;
4273 }
4274
4275have_info:
4276 if (info->bitmap) {
4277 u64 bit_off, bit_bytes;
4278 struct rb_node *n;
4279 struct btrfs_free_space *tmp;
4280
4281 bit_off = offset;
4282 bit_bytes = ctl->unit;
4283 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4284 if (!ret) {
4285 if (bit_off == offset) {
4286 ret = 1;
4287 goto out;
4288 } else if (bit_off > offset &&
4289 offset + bytes > bit_off) {
4290 ret = 1;
4291 goto out;
4292 }
4293 }
4294
4295 n = rb_prev(&info->offset_index);
4296 while (n) {
4297 tmp = rb_entry(n, struct btrfs_free_space,
4298 offset_index);
4299 if (tmp->offset + tmp->bytes < offset)
4300 break;
4301 if (offset + bytes < tmp->offset) {
4302 n = rb_prev(&tmp->offset_index);
4303 continue;
4304 }
4305 info = tmp;
4306 goto have_info;
4307 }
4308
4309 n = rb_next(&info->offset_index);
4310 while (n) {
4311 tmp = rb_entry(n, struct btrfs_free_space,
4312 offset_index);
4313 if (offset + bytes < tmp->offset)
4314 break;
4315 if (tmp->offset + tmp->bytes < offset) {
4316 n = rb_next(&tmp->offset_index);
4317 continue;
4318 }
4319 info = tmp;
4320 goto have_info;
4321 }
4322
4323 ret = 0;
4324 goto out;
4325 }
4326
4327 if (info->offset == offset) {
4328 ret = 1;
4329 goto out;
4330 }
4331
4332 if (offset > info->offset && offset < info->offset + info->bytes)
4333 ret = 1;
4334out:
4335 spin_unlock(&ctl->tree_lock);
4336 return ret;
4337}
4338#endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */
1/*
2 * Copyright (C) 2008 Red Hat. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/pagemap.h>
20#include <linux/sched.h>
21#include <linux/slab.h>
22#include <linux/math64.h>
23#include "ctree.h"
24#include "free-space-cache.h"
25#include "transaction.h"
26#include "disk-io.h"
27#include "extent_io.h"
28#include "inode-map.h"
29
30#define BITS_PER_BITMAP (PAGE_CACHE_SIZE * 8)
31#define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
32
33static int link_free_space(struct btrfs_free_space_ctl *ctl,
34 struct btrfs_free_space *info);
35
36static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
37 struct btrfs_path *path,
38 u64 offset)
39{
40 struct btrfs_key key;
41 struct btrfs_key location;
42 struct btrfs_disk_key disk_key;
43 struct btrfs_free_space_header *header;
44 struct extent_buffer *leaf;
45 struct inode *inode = NULL;
46 int ret;
47
48 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
49 key.offset = offset;
50 key.type = 0;
51
52 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
53 if (ret < 0)
54 return ERR_PTR(ret);
55 if (ret > 0) {
56 btrfs_release_path(path);
57 return ERR_PTR(-ENOENT);
58 }
59
60 leaf = path->nodes[0];
61 header = btrfs_item_ptr(leaf, path->slots[0],
62 struct btrfs_free_space_header);
63 btrfs_free_space_key(leaf, header, &disk_key);
64 btrfs_disk_key_to_cpu(&location, &disk_key);
65 btrfs_release_path(path);
66
67 inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
68 if (!inode)
69 return ERR_PTR(-ENOENT);
70 if (IS_ERR(inode))
71 return inode;
72 if (is_bad_inode(inode)) {
73 iput(inode);
74 return ERR_PTR(-ENOENT);
75 }
76
77 inode->i_mapping->flags &= ~__GFP_FS;
78
79 return inode;
80}
81
82struct inode *lookup_free_space_inode(struct btrfs_root *root,
83 struct btrfs_block_group_cache
84 *block_group, struct btrfs_path *path)
85{
86 struct inode *inode = NULL;
87
88 spin_lock(&block_group->lock);
89 if (block_group->inode)
90 inode = igrab(block_group->inode);
91 spin_unlock(&block_group->lock);
92 if (inode)
93 return inode;
94
95 inode = __lookup_free_space_inode(root, path,
96 block_group->key.objectid);
97 if (IS_ERR(inode))
98 return inode;
99
100 spin_lock(&block_group->lock);
101 if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM) {
102 printk(KERN_INFO "Old style space inode found, converting.\n");
103 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NODATASUM;
104 block_group->disk_cache_state = BTRFS_DC_CLEAR;
105 }
106
107 if (!btrfs_fs_closing(root->fs_info)) {
108 block_group->inode = igrab(inode);
109 block_group->iref = 1;
110 }
111 spin_unlock(&block_group->lock);
112
113 return inode;
114}
115
116int __create_free_space_inode(struct btrfs_root *root,
117 struct btrfs_trans_handle *trans,
118 struct btrfs_path *path, u64 ino, u64 offset)
119{
120 struct btrfs_key key;
121 struct btrfs_disk_key disk_key;
122 struct btrfs_free_space_header *header;
123 struct btrfs_inode_item *inode_item;
124 struct extent_buffer *leaf;
125 int ret;
126
127 ret = btrfs_insert_empty_inode(trans, root, path, ino);
128 if (ret)
129 return ret;
130
131 leaf = path->nodes[0];
132 inode_item = btrfs_item_ptr(leaf, path->slots[0],
133 struct btrfs_inode_item);
134 btrfs_item_key(leaf, &disk_key, path->slots[0]);
135 memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
136 sizeof(*inode_item));
137 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
138 btrfs_set_inode_size(leaf, inode_item, 0);
139 btrfs_set_inode_nbytes(leaf, inode_item, 0);
140 btrfs_set_inode_uid(leaf, inode_item, 0);
141 btrfs_set_inode_gid(leaf, inode_item, 0);
142 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
143 btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
144 BTRFS_INODE_PREALLOC);
145 btrfs_set_inode_nlink(leaf, inode_item, 1);
146 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
147 btrfs_set_inode_block_group(leaf, inode_item, offset);
148 btrfs_mark_buffer_dirty(leaf);
149 btrfs_release_path(path);
150
151 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
152 key.offset = offset;
153 key.type = 0;
154
155 ret = btrfs_insert_empty_item(trans, root, path, &key,
156 sizeof(struct btrfs_free_space_header));
157 if (ret < 0) {
158 btrfs_release_path(path);
159 return ret;
160 }
161 leaf = path->nodes[0];
162 header = btrfs_item_ptr(leaf, path->slots[0],
163 struct btrfs_free_space_header);
164 memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
165 btrfs_set_free_space_key(leaf, header, &disk_key);
166 btrfs_mark_buffer_dirty(leaf);
167 btrfs_release_path(path);
168
169 return 0;
170}
171
172int create_free_space_inode(struct btrfs_root *root,
173 struct btrfs_trans_handle *trans,
174 struct btrfs_block_group_cache *block_group,
175 struct btrfs_path *path)
176{
177 int ret;
178 u64 ino;
179
180 ret = btrfs_find_free_objectid(root, &ino);
181 if (ret < 0)
182 return ret;
183
184 return __create_free_space_inode(root, trans, path, ino,
185 block_group->key.objectid);
186}
187
188int btrfs_truncate_free_space_cache(struct btrfs_root *root,
189 struct btrfs_trans_handle *trans,
190 struct btrfs_path *path,
191 struct inode *inode)
192{
193 struct btrfs_block_rsv *rsv;
194 loff_t oldsize;
195 int ret = 0;
196
197 rsv = trans->block_rsv;
198 trans->block_rsv = root->orphan_block_rsv;
199 ret = btrfs_block_rsv_check(trans, root,
200 root->orphan_block_rsv,
201 0, 5);
202 if (ret)
203 return ret;
204
205 oldsize = i_size_read(inode);
206 btrfs_i_size_write(inode, 0);
207 truncate_pagecache(inode, oldsize, 0);
208
209 /*
210 * We don't need an orphan item because truncating the free space cache
211 * will never be split across transactions.
212 */
213 ret = btrfs_truncate_inode_items(trans, root, inode,
214 0, BTRFS_EXTENT_DATA_KEY);
215
216 trans->block_rsv = rsv;
217 if (ret) {
218 WARN_ON(1);
219 return ret;
220 }
221
222 ret = btrfs_update_inode(trans, root, inode);
223 return ret;
224}
225
226static int readahead_cache(struct inode *inode)
227{
228 struct file_ra_state *ra;
229 unsigned long last_index;
230
231 ra = kzalloc(sizeof(*ra), GFP_NOFS);
232 if (!ra)
233 return -ENOMEM;
234
235 file_ra_state_init(ra, inode->i_mapping);
236 last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
237
238 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
239
240 kfree(ra);
241
242 return 0;
243}
244
245int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
246 struct btrfs_free_space_ctl *ctl,
247 struct btrfs_path *path, u64 offset)
248{
249 struct btrfs_free_space_header *header;
250 struct extent_buffer *leaf;
251 struct page *page;
252 struct btrfs_key key;
253 struct list_head bitmaps;
254 u64 num_entries;
255 u64 num_bitmaps;
256 u64 generation;
257 pgoff_t index = 0;
258 int ret = 0;
259
260 INIT_LIST_HEAD(&bitmaps);
261
262 /* Nothing in the space cache, goodbye */
263 if (!i_size_read(inode))
264 goto out;
265
266 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
267 key.offset = offset;
268 key.type = 0;
269
270 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
271 if (ret < 0)
272 goto out;
273 else if (ret > 0) {
274 btrfs_release_path(path);
275 ret = 0;
276 goto out;
277 }
278
279 ret = -1;
280
281 leaf = path->nodes[0];
282 header = btrfs_item_ptr(leaf, path->slots[0],
283 struct btrfs_free_space_header);
284 num_entries = btrfs_free_space_entries(leaf, header);
285 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
286 generation = btrfs_free_space_generation(leaf, header);
287 btrfs_release_path(path);
288
289 if (BTRFS_I(inode)->generation != generation) {
290 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
291 " not match free space cache generation (%llu)\n",
292 (unsigned long long)BTRFS_I(inode)->generation,
293 (unsigned long long)generation);
294 goto out;
295 }
296
297 if (!num_entries)
298 goto out;
299
300 ret = readahead_cache(inode);
301 if (ret)
302 goto out;
303
304 while (1) {
305 struct btrfs_free_space_entry *entry;
306 struct btrfs_free_space *e;
307 void *addr;
308 unsigned long offset = 0;
309 int need_loop = 0;
310
311 if (!num_entries && !num_bitmaps)
312 break;
313
314 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
315 if (!page)
316 goto free_cache;
317
318 if (!PageUptodate(page)) {
319 btrfs_readpage(NULL, page);
320 lock_page(page);
321 if (!PageUptodate(page)) {
322 unlock_page(page);
323 page_cache_release(page);
324 printk(KERN_ERR "btrfs: error reading free "
325 "space cache\n");
326 goto free_cache;
327 }
328 }
329 addr = kmap(page);
330
331 if (index == 0) {
332 u64 *gen;
333
334 /*
335 * We put a bogus crc in the front of the first page in
336 * case old kernels try to mount a fs with the new
337 * format to make sure they discard the cache.
338 */
339 addr += sizeof(u64);
340 offset += sizeof(u64);
341
342 gen = addr;
343 if (*gen != BTRFS_I(inode)->generation) {
344 printk(KERN_ERR "btrfs: space cache generation"
345 " (%llu) does not match inode (%llu)\n",
346 (unsigned long long)*gen,
347 (unsigned long long)
348 BTRFS_I(inode)->generation);
349 kunmap(page);
350 unlock_page(page);
351 page_cache_release(page);
352 goto free_cache;
353 }
354 addr += sizeof(u64);
355 offset += sizeof(u64);
356 }
357 entry = addr;
358
359 while (1) {
360 if (!num_entries)
361 break;
362
363 need_loop = 1;
364 e = kmem_cache_zalloc(btrfs_free_space_cachep,
365 GFP_NOFS);
366 if (!e) {
367 kunmap(page);
368 unlock_page(page);
369 page_cache_release(page);
370 goto free_cache;
371 }
372
373 e->offset = le64_to_cpu(entry->offset);
374 e->bytes = le64_to_cpu(entry->bytes);
375 if (!e->bytes) {
376 kunmap(page);
377 kmem_cache_free(btrfs_free_space_cachep, e);
378 unlock_page(page);
379 page_cache_release(page);
380 goto free_cache;
381 }
382
383 if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
384 spin_lock(&ctl->tree_lock);
385 ret = link_free_space(ctl, e);
386 spin_unlock(&ctl->tree_lock);
387 if (ret) {
388 printk(KERN_ERR "Duplicate entries in "
389 "free space cache, dumping\n");
390 kunmap(page);
391 unlock_page(page);
392 page_cache_release(page);
393 goto free_cache;
394 }
395 } else {
396 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
397 if (!e->bitmap) {
398 kunmap(page);
399 kmem_cache_free(
400 btrfs_free_space_cachep, e);
401 unlock_page(page);
402 page_cache_release(page);
403 goto free_cache;
404 }
405 spin_lock(&ctl->tree_lock);
406 ret = link_free_space(ctl, e);
407 ctl->total_bitmaps++;
408 ctl->op->recalc_thresholds(ctl);
409 spin_unlock(&ctl->tree_lock);
410 if (ret) {
411 printk(KERN_ERR "Duplicate entries in "
412 "free space cache, dumping\n");
413 kunmap(page);
414 unlock_page(page);
415 page_cache_release(page);
416 goto free_cache;
417 }
418 list_add_tail(&e->list, &bitmaps);
419 }
420
421 num_entries--;
422 offset += sizeof(struct btrfs_free_space_entry);
423 if (offset + sizeof(struct btrfs_free_space_entry) >=
424 PAGE_CACHE_SIZE)
425 break;
426 entry++;
427 }
428
429 /*
430 * We read an entry out of this page, we need to move on to the
431 * next page.
432 */
433 if (need_loop) {
434 kunmap(page);
435 goto next;
436 }
437
438 /*
439 * We add the bitmaps at the end of the entries in order that
440 * the bitmap entries are added to the cache.
441 */
442 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
443 list_del_init(&e->list);
444 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
445 kunmap(page);
446 num_bitmaps--;
447next:
448 unlock_page(page);
449 page_cache_release(page);
450 index++;
451 }
452
453 ret = 1;
454out:
455 return ret;
456free_cache:
457 __btrfs_remove_free_space_cache(ctl);
458 goto out;
459}
460
461int load_free_space_cache(struct btrfs_fs_info *fs_info,
462 struct btrfs_block_group_cache *block_group)
463{
464 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
465 struct btrfs_root *root = fs_info->tree_root;
466 struct inode *inode;
467 struct btrfs_path *path;
468 int ret;
469 bool matched;
470 u64 used = btrfs_block_group_used(&block_group->item);
471
472 /*
473 * If we're unmounting then just return, since this does a search on the
474 * normal root and not the commit root and we could deadlock.
475 */
476 if (btrfs_fs_closing(fs_info))
477 return 0;
478
479 /*
480 * If this block group has been marked to be cleared for one reason or
481 * another then we can't trust the on disk cache, so just return.
482 */
483 spin_lock(&block_group->lock);
484 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
485 spin_unlock(&block_group->lock);
486 return 0;
487 }
488 spin_unlock(&block_group->lock);
489
490 path = btrfs_alloc_path();
491 if (!path)
492 return 0;
493
494 inode = lookup_free_space_inode(root, block_group, path);
495 if (IS_ERR(inode)) {
496 btrfs_free_path(path);
497 return 0;
498 }
499
500 ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
501 path, block_group->key.objectid);
502 btrfs_free_path(path);
503 if (ret <= 0)
504 goto out;
505
506 spin_lock(&ctl->tree_lock);
507 matched = (ctl->free_space == (block_group->key.offset - used -
508 block_group->bytes_super));
509 spin_unlock(&ctl->tree_lock);
510
511 if (!matched) {
512 __btrfs_remove_free_space_cache(ctl);
513 printk(KERN_ERR "block group %llu has an wrong amount of free "
514 "space\n", block_group->key.objectid);
515 ret = -1;
516 }
517out:
518 if (ret < 0) {
519 /* This cache is bogus, make sure it gets cleared */
520 spin_lock(&block_group->lock);
521 block_group->disk_cache_state = BTRFS_DC_CLEAR;
522 spin_unlock(&block_group->lock);
523 ret = 0;
524
525 printk(KERN_ERR "btrfs: failed to load free space cache "
526 "for block group %llu\n", block_group->key.objectid);
527 }
528
529 iput(inode);
530 return ret;
531}
532
533int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
534 struct btrfs_free_space_ctl *ctl,
535 struct btrfs_block_group_cache *block_group,
536 struct btrfs_trans_handle *trans,
537 struct btrfs_path *path, u64 offset)
538{
539 struct btrfs_free_space_header *header;
540 struct extent_buffer *leaf;
541 struct rb_node *node;
542 struct list_head *pos, *n;
543 struct page **pages;
544 struct page *page;
545 struct extent_state *cached_state = NULL;
546 struct btrfs_free_cluster *cluster = NULL;
547 struct extent_io_tree *unpin = NULL;
548 struct list_head bitmap_list;
549 struct btrfs_key key;
550 u64 start, end, len;
551 u64 bytes = 0;
552 u32 crc = ~(u32)0;
553 int index = 0, num_pages = 0;
554 int entries = 0;
555 int bitmaps = 0;
556 int ret = -1;
557 bool next_page = false;
558 bool out_of_space = false;
559
560 INIT_LIST_HEAD(&bitmap_list);
561
562 node = rb_first(&ctl->free_space_offset);
563 if (!node)
564 return 0;
565
566 if (!i_size_read(inode))
567 return -1;
568
569 num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
570 PAGE_CACHE_SHIFT;
571
572 filemap_write_and_wait(inode->i_mapping);
573 btrfs_wait_ordered_range(inode, inode->i_size &
574 ~(root->sectorsize - 1), (u64)-1);
575
576 pages = kzalloc(sizeof(struct page *) * num_pages, GFP_NOFS);
577 if (!pages)
578 return -1;
579
580 /* Get the cluster for this block_group if it exists */
581 if (block_group && !list_empty(&block_group->cluster_list))
582 cluster = list_entry(block_group->cluster_list.next,
583 struct btrfs_free_cluster,
584 block_group_list);
585
586 /*
587 * We shouldn't have switched the pinned extents yet so this is the
588 * right one
589 */
590 unpin = root->fs_info->pinned_extents;
591
592 /*
593 * Lock all pages first so we can lock the extent safely.
594 *
595 * NOTE: Because we hold the ref the entire time we're going to write to
596 * the page find_get_page should never fail, so we don't do a check
597 * after find_get_page at this point. Just putting this here so people
598 * know and don't freak out.
599 */
600 while (index < num_pages) {
601 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
602 if (!page) {
603 int i;
604
605 for (i = 0; i < num_pages; i++) {
606 unlock_page(pages[i]);
607 page_cache_release(pages[i]);
608 }
609 goto out;
610 }
611 pages[index] = page;
612 index++;
613 }
614
615 index = 0;
616 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
617 0, &cached_state, GFP_NOFS);
618
619 /*
620 * When searching for pinned extents, we need to start at our start
621 * offset.
622 */
623 if (block_group)
624 start = block_group->key.objectid;
625
626 /* Write out the extent entries */
627 do {
628 struct btrfs_free_space_entry *entry;
629 void *addr, *orig;
630 unsigned long offset = 0;
631
632 next_page = false;
633
634 if (index >= num_pages) {
635 out_of_space = true;
636 break;
637 }
638
639 page = pages[index];
640
641 orig = addr = kmap(page);
642 if (index == 0) {
643 u64 *gen;
644
645 /*
646 * We're going to put in a bogus crc for this page to
647 * make sure that old kernels who aren't aware of this
648 * format will be sure to discard the cache.
649 */
650 addr += sizeof(u64);
651 offset += sizeof(u64);
652
653 gen = addr;
654 *gen = trans->transid;
655 addr += sizeof(u64);
656 offset += sizeof(u64);
657 }
658 entry = addr;
659
660 memset(addr, 0, PAGE_CACHE_SIZE - offset);
661 while (node && !next_page) {
662 struct btrfs_free_space *e;
663
664 e = rb_entry(node, struct btrfs_free_space, offset_index);
665 entries++;
666
667 entry->offset = cpu_to_le64(e->offset);
668 entry->bytes = cpu_to_le64(e->bytes);
669 if (e->bitmap) {
670 entry->type = BTRFS_FREE_SPACE_BITMAP;
671 list_add_tail(&e->list, &bitmap_list);
672 bitmaps++;
673 } else {
674 entry->type = BTRFS_FREE_SPACE_EXTENT;
675 }
676 node = rb_next(node);
677 if (!node && cluster) {
678 node = rb_first(&cluster->root);
679 cluster = NULL;
680 }
681 offset += sizeof(struct btrfs_free_space_entry);
682 if (offset + sizeof(struct btrfs_free_space_entry) >=
683 PAGE_CACHE_SIZE)
684 next_page = true;
685 entry++;
686 }
687
688 /*
689 * We want to add any pinned extents to our free space cache
690 * so we don't leak the space
691 */
692 while (block_group && !next_page &&
693 (start < block_group->key.objectid +
694 block_group->key.offset)) {
695 ret = find_first_extent_bit(unpin, start, &start, &end,
696 EXTENT_DIRTY);
697 if (ret) {
698 ret = 0;
699 break;
700 }
701
702 /* This pinned extent is out of our range */
703 if (start >= block_group->key.objectid +
704 block_group->key.offset)
705 break;
706
707 len = block_group->key.objectid +
708 block_group->key.offset - start;
709 len = min(len, end + 1 - start);
710
711 entries++;
712 entry->offset = cpu_to_le64(start);
713 entry->bytes = cpu_to_le64(len);
714 entry->type = BTRFS_FREE_SPACE_EXTENT;
715
716 start = end + 1;
717 offset += sizeof(struct btrfs_free_space_entry);
718 if (offset + sizeof(struct btrfs_free_space_entry) >=
719 PAGE_CACHE_SIZE)
720 next_page = true;
721 entry++;
722 }
723
724 /* Generate bogus crc value */
725 if (index == 0) {
726 u32 *tmp;
727 crc = btrfs_csum_data(root, orig + sizeof(u64), crc,
728 PAGE_CACHE_SIZE - sizeof(u64));
729 btrfs_csum_final(crc, (char *)&crc);
730 crc++;
731 tmp = orig;
732 *tmp = crc;
733 }
734
735 kunmap(page);
736
737 bytes += PAGE_CACHE_SIZE;
738
739 index++;
740 } while (node || next_page);
741
742 /* Write out the bitmaps */
743 list_for_each_safe(pos, n, &bitmap_list) {
744 void *addr;
745 struct btrfs_free_space *entry =
746 list_entry(pos, struct btrfs_free_space, list);
747
748 if (index >= num_pages) {
749 out_of_space = true;
750 break;
751 }
752 page = pages[index];
753
754 addr = kmap(page);
755 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
756 kunmap(page);
757 bytes += PAGE_CACHE_SIZE;
758
759 list_del_init(&entry->list);
760 index++;
761 }
762
763 if (out_of_space) {
764 btrfs_drop_pages(pages, num_pages);
765 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
766 i_size_read(inode) - 1, &cached_state,
767 GFP_NOFS);
768 ret = 0;
769 goto out;
770 }
771
772 /* Zero out the rest of the pages just to make sure */
773 while (index < num_pages) {
774 void *addr;
775
776 page = pages[index];
777 addr = kmap(page);
778 memset(addr, 0, PAGE_CACHE_SIZE);
779 kunmap(page);
780 bytes += PAGE_CACHE_SIZE;
781 index++;
782 }
783
784 ret = btrfs_dirty_pages(root, inode, pages, num_pages, 0,
785 bytes, &cached_state);
786 btrfs_drop_pages(pages, num_pages);
787 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
788 i_size_read(inode) - 1, &cached_state, GFP_NOFS);
789
790 if (ret) {
791 ret = 0;
792 goto out;
793 }
794
795 BTRFS_I(inode)->generation = trans->transid;
796
797 filemap_write_and_wait(inode->i_mapping);
798
799 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
800 key.offset = offset;
801 key.type = 0;
802
803 ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
804 if (ret < 0) {
805 ret = -1;
806 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
807 EXTENT_DIRTY | EXTENT_DELALLOC |
808 EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
809 goto out;
810 }
811 leaf = path->nodes[0];
812 if (ret > 0) {
813 struct btrfs_key found_key;
814 BUG_ON(!path->slots[0]);
815 path->slots[0]--;
816 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
817 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
818 found_key.offset != offset) {
819 ret = -1;
820 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
821 EXTENT_DIRTY | EXTENT_DELALLOC |
822 EXTENT_DO_ACCOUNTING, 0, 0, NULL,
823 GFP_NOFS);
824 btrfs_release_path(path);
825 goto out;
826 }
827 }
828 header = btrfs_item_ptr(leaf, path->slots[0],
829 struct btrfs_free_space_header);
830 btrfs_set_free_space_entries(leaf, header, entries);
831 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
832 btrfs_set_free_space_generation(leaf, header, trans->transid);
833 btrfs_mark_buffer_dirty(leaf);
834 btrfs_release_path(path);
835
836 ret = 1;
837
838out:
839 kfree(pages);
840 if (ret != 1) {
841 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
842 BTRFS_I(inode)->generation = 0;
843 }
844 btrfs_update_inode(trans, root, inode);
845 return ret;
846}
847
848int btrfs_write_out_cache(struct btrfs_root *root,
849 struct btrfs_trans_handle *trans,
850 struct btrfs_block_group_cache *block_group,
851 struct btrfs_path *path)
852{
853 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
854 struct inode *inode;
855 int ret = 0;
856
857 root = root->fs_info->tree_root;
858
859 spin_lock(&block_group->lock);
860 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
861 spin_unlock(&block_group->lock);
862 return 0;
863 }
864 spin_unlock(&block_group->lock);
865
866 inode = lookup_free_space_inode(root, block_group, path);
867 if (IS_ERR(inode))
868 return 0;
869
870 ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
871 path, block_group->key.objectid);
872 if (ret < 0) {
873 spin_lock(&block_group->lock);
874 block_group->disk_cache_state = BTRFS_DC_ERROR;
875 spin_unlock(&block_group->lock);
876 ret = 0;
877
878 printk(KERN_ERR "btrfs: failed to write free space cace "
879 "for block group %llu\n", block_group->key.objectid);
880 }
881
882 iput(inode);
883 return ret;
884}
885
886static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
887 u64 offset)
888{
889 BUG_ON(offset < bitmap_start);
890 offset -= bitmap_start;
891 return (unsigned long)(div_u64(offset, unit));
892}
893
894static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
895{
896 return (unsigned long)(div_u64(bytes, unit));
897}
898
899static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
900 u64 offset)
901{
902 u64 bitmap_start;
903 u64 bytes_per_bitmap;
904
905 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
906 bitmap_start = offset - ctl->start;
907 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
908 bitmap_start *= bytes_per_bitmap;
909 bitmap_start += ctl->start;
910
911 return bitmap_start;
912}
913
914static int tree_insert_offset(struct rb_root *root, u64 offset,
915 struct rb_node *node, int bitmap)
916{
917 struct rb_node **p = &root->rb_node;
918 struct rb_node *parent = NULL;
919 struct btrfs_free_space *info;
920
921 while (*p) {
922 parent = *p;
923 info = rb_entry(parent, struct btrfs_free_space, offset_index);
924
925 if (offset < info->offset) {
926 p = &(*p)->rb_left;
927 } else if (offset > info->offset) {
928 p = &(*p)->rb_right;
929 } else {
930 /*
931 * we could have a bitmap entry and an extent entry
932 * share the same offset. If this is the case, we want
933 * the extent entry to always be found first if we do a
934 * linear search through the tree, since we want to have
935 * the quickest allocation time, and allocating from an
936 * extent is faster than allocating from a bitmap. So
937 * if we're inserting a bitmap and we find an entry at
938 * this offset, we want to go right, or after this entry
939 * logically. If we are inserting an extent and we've
940 * found a bitmap, we want to go left, or before
941 * logically.
942 */
943 if (bitmap) {
944 if (info->bitmap) {
945 WARN_ON_ONCE(1);
946 return -EEXIST;
947 }
948 p = &(*p)->rb_right;
949 } else {
950 if (!info->bitmap) {
951 WARN_ON_ONCE(1);
952 return -EEXIST;
953 }
954 p = &(*p)->rb_left;
955 }
956 }
957 }
958
959 rb_link_node(node, parent, p);
960 rb_insert_color(node, root);
961
962 return 0;
963}
964
965/*
966 * searches the tree for the given offset.
967 *
968 * fuzzy - If this is set, then we are trying to make an allocation, and we just
969 * want a section that has at least bytes size and comes at or after the given
970 * offset.
971 */
972static struct btrfs_free_space *
973tree_search_offset(struct btrfs_free_space_ctl *ctl,
974 u64 offset, int bitmap_only, int fuzzy)
975{
976 struct rb_node *n = ctl->free_space_offset.rb_node;
977 struct btrfs_free_space *entry, *prev = NULL;
978
979 /* find entry that is closest to the 'offset' */
980 while (1) {
981 if (!n) {
982 entry = NULL;
983 break;
984 }
985
986 entry = rb_entry(n, struct btrfs_free_space, offset_index);
987 prev = entry;
988
989 if (offset < entry->offset)
990 n = n->rb_left;
991 else if (offset > entry->offset)
992 n = n->rb_right;
993 else
994 break;
995 }
996
997 if (bitmap_only) {
998 if (!entry)
999 return NULL;
1000 if (entry->bitmap)
1001 return entry;
1002
1003 /*
1004 * bitmap entry and extent entry may share same offset,
1005 * in that case, bitmap entry comes after extent entry.
1006 */
1007 n = rb_next(n);
1008 if (!n)
1009 return NULL;
1010 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1011 if (entry->offset != offset)
1012 return NULL;
1013
1014 WARN_ON(!entry->bitmap);
1015 return entry;
1016 } else if (entry) {
1017 if (entry->bitmap) {
1018 /*
1019 * if previous extent entry covers the offset,
1020 * we should return it instead of the bitmap entry
1021 */
1022 n = &entry->offset_index;
1023 while (1) {
1024 n = rb_prev(n);
1025 if (!n)
1026 break;
1027 prev = rb_entry(n, struct btrfs_free_space,
1028 offset_index);
1029 if (!prev->bitmap) {
1030 if (prev->offset + prev->bytes > offset)
1031 entry = prev;
1032 break;
1033 }
1034 }
1035 }
1036 return entry;
1037 }
1038
1039 if (!prev)
1040 return NULL;
1041
1042 /* find last entry before the 'offset' */
1043 entry = prev;
1044 if (entry->offset > offset) {
1045 n = rb_prev(&entry->offset_index);
1046 if (n) {
1047 entry = rb_entry(n, struct btrfs_free_space,
1048 offset_index);
1049 BUG_ON(entry->offset > offset);
1050 } else {
1051 if (fuzzy)
1052 return entry;
1053 else
1054 return NULL;
1055 }
1056 }
1057
1058 if (entry->bitmap) {
1059 n = &entry->offset_index;
1060 while (1) {
1061 n = rb_prev(n);
1062 if (!n)
1063 break;
1064 prev = rb_entry(n, struct btrfs_free_space,
1065 offset_index);
1066 if (!prev->bitmap) {
1067 if (prev->offset + prev->bytes > offset)
1068 return prev;
1069 break;
1070 }
1071 }
1072 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1073 return entry;
1074 } else if (entry->offset + entry->bytes > offset)
1075 return entry;
1076
1077 if (!fuzzy)
1078 return NULL;
1079
1080 while (1) {
1081 if (entry->bitmap) {
1082 if (entry->offset + BITS_PER_BITMAP *
1083 ctl->unit > offset)
1084 break;
1085 } else {
1086 if (entry->offset + entry->bytes > offset)
1087 break;
1088 }
1089
1090 n = rb_next(&entry->offset_index);
1091 if (!n)
1092 return NULL;
1093 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1094 }
1095 return entry;
1096}
1097
1098static inline void
1099__unlink_free_space(struct btrfs_free_space_ctl *ctl,
1100 struct btrfs_free_space *info)
1101{
1102 rb_erase(&info->offset_index, &ctl->free_space_offset);
1103 ctl->free_extents--;
1104}
1105
1106static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1107 struct btrfs_free_space *info)
1108{
1109 __unlink_free_space(ctl, info);
1110 ctl->free_space -= info->bytes;
1111}
1112
1113static int link_free_space(struct btrfs_free_space_ctl *ctl,
1114 struct btrfs_free_space *info)
1115{
1116 int ret = 0;
1117
1118 BUG_ON(!info->bitmap && !info->bytes);
1119 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1120 &info->offset_index, (info->bitmap != NULL));
1121 if (ret)
1122 return ret;
1123
1124 ctl->free_space += info->bytes;
1125 ctl->free_extents++;
1126 return ret;
1127}
1128
1129static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1130{
1131 struct btrfs_block_group_cache *block_group = ctl->private;
1132 u64 max_bytes;
1133 u64 bitmap_bytes;
1134 u64 extent_bytes;
1135 u64 size = block_group->key.offset;
1136 u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1137 int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1138
1139 BUG_ON(ctl->total_bitmaps > max_bitmaps);
1140
1141 /*
1142 * The goal is to keep the total amount of memory used per 1gb of space
1143 * at or below 32k, so we need to adjust how much memory we allow to be
1144 * used by extent based free space tracking
1145 */
1146 if (size < 1024 * 1024 * 1024)
1147 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1148 else
1149 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1150 div64_u64(size, 1024 * 1024 * 1024);
1151
1152 /*
1153 * we want to account for 1 more bitmap than what we have so we can make
1154 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1155 * we add more bitmaps.
1156 */
1157 bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1158
1159 if (bitmap_bytes >= max_bytes) {
1160 ctl->extents_thresh = 0;
1161 return;
1162 }
1163
1164 /*
1165 * we want the extent entry threshold to always be at most 1/2 the maxw
1166 * bytes we can have, or whatever is less than that.
1167 */
1168 extent_bytes = max_bytes - bitmap_bytes;
1169 extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1170
1171 ctl->extents_thresh =
1172 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1173}
1174
1175static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1176 struct btrfs_free_space *info,
1177 u64 offset, u64 bytes)
1178{
1179 unsigned long start, count;
1180
1181 start = offset_to_bit(info->offset, ctl->unit, offset);
1182 count = bytes_to_bits(bytes, ctl->unit);
1183 BUG_ON(start + count > BITS_PER_BITMAP);
1184
1185 bitmap_clear(info->bitmap, start, count);
1186
1187 info->bytes -= bytes;
1188}
1189
1190static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1191 struct btrfs_free_space *info, u64 offset,
1192 u64 bytes)
1193{
1194 __bitmap_clear_bits(ctl, info, offset, bytes);
1195 ctl->free_space -= bytes;
1196}
1197
1198static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1199 struct btrfs_free_space *info, u64 offset,
1200 u64 bytes)
1201{
1202 unsigned long start, count;
1203
1204 start = offset_to_bit(info->offset, ctl->unit, offset);
1205 count = bytes_to_bits(bytes, ctl->unit);
1206 BUG_ON(start + count > BITS_PER_BITMAP);
1207
1208 bitmap_set(info->bitmap, start, count);
1209
1210 info->bytes += bytes;
1211 ctl->free_space += bytes;
1212}
1213
1214static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1215 struct btrfs_free_space *bitmap_info, u64 *offset,
1216 u64 *bytes)
1217{
1218 unsigned long found_bits = 0;
1219 unsigned long bits, i;
1220 unsigned long next_zero;
1221
1222 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1223 max_t(u64, *offset, bitmap_info->offset));
1224 bits = bytes_to_bits(*bytes, ctl->unit);
1225
1226 for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1227 i < BITS_PER_BITMAP;
1228 i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1229 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1230 BITS_PER_BITMAP, i);
1231 if ((next_zero - i) >= bits) {
1232 found_bits = next_zero - i;
1233 break;
1234 }
1235 i = next_zero;
1236 }
1237
1238 if (found_bits) {
1239 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1240 *bytes = (u64)(found_bits) * ctl->unit;
1241 return 0;
1242 }
1243
1244 return -1;
1245}
1246
1247static struct btrfs_free_space *
1248find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
1249{
1250 struct btrfs_free_space *entry;
1251 struct rb_node *node;
1252 int ret;
1253
1254 if (!ctl->free_space_offset.rb_node)
1255 return NULL;
1256
1257 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1258 if (!entry)
1259 return NULL;
1260
1261 for (node = &entry->offset_index; node; node = rb_next(node)) {
1262 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1263 if (entry->bytes < *bytes)
1264 continue;
1265
1266 if (entry->bitmap) {
1267 ret = search_bitmap(ctl, entry, offset, bytes);
1268 if (!ret)
1269 return entry;
1270 continue;
1271 }
1272
1273 *offset = entry->offset;
1274 *bytes = entry->bytes;
1275 return entry;
1276 }
1277
1278 return NULL;
1279}
1280
1281static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1282 struct btrfs_free_space *info, u64 offset)
1283{
1284 info->offset = offset_to_bitmap(ctl, offset);
1285 info->bytes = 0;
1286 link_free_space(ctl, info);
1287 ctl->total_bitmaps++;
1288
1289 ctl->op->recalc_thresholds(ctl);
1290}
1291
1292static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1293 struct btrfs_free_space *bitmap_info)
1294{
1295 unlink_free_space(ctl, bitmap_info);
1296 kfree(bitmap_info->bitmap);
1297 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1298 ctl->total_bitmaps--;
1299 ctl->op->recalc_thresholds(ctl);
1300}
1301
1302static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1303 struct btrfs_free_space *bitmap_info,
1304 u64 *offset, u64 *bytes)
1305{
1306 u64 end;
1307 u64 search_start, search_bytes;
1308 int ret;
1309
1310again:
1311 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1312
1313 /*
1314 * XXX - this can go away after a few releases.
1315 *
1316 * since the only user of btrfs_remove_free_space is the tree logging
1317 * stuff, and the only way to test that is under crash conditions, we
1318 * want to have this debug stuff here just in case somethings not
1319 * working. Search the bitmap for the space we are trying to use to
1320 * make sure its actually there. If its not there then we need to stop
1321 * because something has gone wrong.
1322 */
1323 search_start = *offset;
1324 search_bytes = *bytes;
1325 search_bytes = min(search_bytes, end - search_start + 1);
1326 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1327 BUG_ON(ret < 0 || search_start != *offset);
1328
1329 if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1330 bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
1331 *bytes -= end - *offset + 1;
1332 *offset = end + 1;
1333 } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1334 bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
1335 *bytes = 0;
1336 }
1337
1338 if (*bytes) {
1339 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1340 if (!bitmap_info->bytes)
1341 free_bitmap(ctl, bitmap_info);
1342
1343 /*
1344 * no entry after this bitmap, but we still have bytes to
1345 * remove, so something has gone wrong.
1346 */
1347 if (!next)
1348 return -EINVAL;
1349
1350 bitmap_info = rb_entry(next, struct btrfs_free_space,
1351 offset_index);
1352
1353 /*
1354 * if the next entry isn't a bitmap we need to return to let the
1355 * extent stuff do its work.
1356 */
1357 if (!bitmap_info->bitmap)
1358 return -EAGAIN;
1359
1360 /*
1361 * Ok the next item is a bitmap, but it may not actually hold
1362 * the information for the rest of this free space stuff, so
1363 * look for it, and if we don't find it return so we can try
1364 * everything over again.
1365 */
1366 search_start = *offset;
1367 search_bytes = *bytes;
1368 ret = search_bitmap(ctl, bitmap_info, &search_start,
1369 &search_bytes);
1370 if (ret < 0 || search_start != *offset)
1371 return -EAGAIN;
1372
1373 goto again;
1374 } else if (!bitmap_info->bytes)
1375 free_bitmap(ctl, bitmap_info);
1376
1377 return 0;
1378}
1379
1380static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1381 struct btrfs_free_space *info, u64 offset,
1382 u64 bytes)
1383{
1384 u64 bytes_to_set = 0;
1385 u64 end;
1386
1387 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1388
1389 bytes_to_set = min(end - offset, bytes);
1390
1391 bitmap_set_bits(ctl, info, offset, bytes_to_set);
1392
1393 return bytes_to_set;
1394
1395}
1396
1397static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1398 struct btrfs_free_space *info)
1399{
1400 struct btrfs_block_group_cache *block_group = ctl->private;
1401
1402 /*
1403 * If we are below the extents threshold then we can add this as an
1404 * extent, and don't have to deal with the bitmap
1405 */
1406 if (ctl->free_extents < ctl->extents_thresh) {
1407 /*
1408 * If this block group has some small extents we don't want to
1409 * use up all of our free slots in the cache with them, we want
1410 * to reserve them to larger extents, however if we have plent
1411 * of cache left then go ahead an dadd them, no sense in adding
1412 * the overhead of a bitmap if we don't have to.
1413 */
1414 if (info->bytes <= block_group->sectorsize * 4) {
1415 if (ctl->free_extents * 2 <= ctl->extents_thresh)
1416 return false;
1417 } else {
1418 return false;
1419 }
1420 }
1421
1422 /*
1423 * some block groups are so tiny they can't be enveloped by a bitmap, so
1424 * don't even bother to create a bitmap for this
1425 */
1426 if (BITS_PER_BITMAP * block_group->sectorsize >
1427 block_group->key.offset)
1428 return false;
1429
1430 return true;
1431}
1432
1433static struct btrfs_free_space_op free_space_op = {
1434 .recalc_thresholds = recalculate_thresholds,
1435 .use_bitmap = use_bitmap,
1436};
1437
1438static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1439 struct btrfs_free_space *info)
1440{
1441 struct btrfs_free_space *bitmap_info;
1442 struct btrfs_block_group_cache *block_group = NULL;
1443 int added = 0;
1444 u64 bytes, offset, bytes_added;
1445 int ret;
1446
1447 bytes = info->bytes;
1448 offset = info->offset;
1449
1450 if (!ctl->op->use_bitmap(ctl, info))
1451 return 0;
1452
1453 if (ctl->op == &free_space_op)
1454 block_group = ctl->private;
1455again:
1456 /*
1457 * Since we link bitmaps right into the cluster we need to see if we
1458 * have a cluster here, and if so and it has our bitmap we need to add
1459 * the free space to that bitmap.
1460 */
1461 if (block_group && !list_empty(&block_group->cluster_list)) {
1462 struct btrfs_free_cluster *cluster;
1463 struct rb_node *node;
1464 struct btrfs_free_space *entry;
1465
1466 cluster = list_entry(block_group->cluster_list.next,
1467 struct btrfs_free_cluster,
1468 block_group_list);
1469 spin_lock(&cluster->lock);
1470 node = rb_first(&cluster->root);
1471 if (!node) {
1472 spin_unlock(&cluster->lock);
1473 goto no_cluster_bitmap;
1474 }
1475
1476 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1477 if (!entry->bitmap) {
1478 spin_unlock(&cluster->lock);
1479 goto no_cluster_bitmap;
1480 }
1481
1482 if (entry->offset == offset_to_bitmap(ctl, offset)) {
1483 bytes_added = add_bytes_to_bitmap(ctl, entry,
1484 offset, bytes);
1485 bytes -= bytes_added;
1486 offset += bytes_added;
1487 }
1488 spin_unlock(&cluster->lock);
1489 if (!bytes) {
1490 ret = 1;
1491 goto out;
1492 }
1493 }
1494
1495no_cluster_bitmap:
1496 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1497 1, 0);
1498 if (!bitmap_info) {
1499 BUG_ON(added);
1500 goto new_bitmap;
1501 }
1502
1503 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1504 bytes -= bytes_added;
1505 offset += bytes_added;
1506 added = 0;
1507
1508 if (!bytes) {
1509 ret = 1;
1510 goto out;
1511 } else
1512 goto again;
1513
1514new_bitmap:
1515 if (info && info->bitmap) {
1516 add_new_bitmap(ctl, info, offset);
1517 added = 1;
1518 info = NULL;
1519 goto again;
1520 } else {
1521 spin_unlock(&ctl->tree_lock);
1522
1523 /* no pre-allocated info, allocate a new one */
1524 if (!info) {
1525 info = kmem_cache_zalloc(btrfs_free_space_cachep,
1526 GFP_NOFS);
1527 if (!info) {
1528 spin_lock(&ctl->tree_lock);
1529 ret = -ENOMEM;
1530 goto out;
1531 }
1532 }
1533
1534 /* allocate the bitmap */
1535 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1536 spin_lock(&ctl->tree_lock);
1537 if (!info->bitmap) {
1538 ret = -ENOMEM;
1539 goto out;
1540 }
1541 goto again;
1542 }
1543
1544out:
1545 if (info) {
1546 if (info->bitmap)
1547 kfree(info->bitmap);
1548 kmem_cache_free(btrfs_free_space_cachep, info);
1549 }
1550
1551 return ret;
1552}
1553
1554static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1555 struct btrfs_free_space *info, bool update_stat)
1556{
1557 struct btrfs_free_space *left_info;
1558 struct btrfs_free_space *right_info;
1559 bool merged = false;
1560 u64 offset = info->offset;
1561 u64 bytes = info->bytes;
1562
1563 /*
1564 * first we want to see if there is free space adjacent to the range we
1565 * are adding, if there is remove that struct and add a new one to
1566 * cover the entire range
1567 */
1568 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1569 if (right_info && rb_prev(&right_info->offset_index))
1570 left_info = rb_entry(rb_prev(&right_info->offset_index),
1571 struct btrfs_free_space, offset_index);
1572 else
1573 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1574
1575 if (right_info && !right_info->bitmap) {
1576 if (update_stat)
1577 unlink_free_space(ctl, right_info);
1578 else
1579 __unlink_free_space(ctl, right_info);
1580 info->bytes += right_info->bytes;
1581 kmem_cache_free(btrfs_free_space_cachep, right_info);
1582 merged = true;
1583 }
1584
1585 if (left_info && !left_info->bitmap &&
1586 left_info->offset + left_info->bytes == offset) {
1587 if (update_stat)
1588 unlink_free_space(ctl, left_info);
1589 else
1590 __unlink_free_space(ctl, left_info);
1591 info->offset = left_info->offset;
1592 info->bytes += left_info->bytes;
1593 kmem_cache_free(btrfs_free_space_cachep, left_info);
1594 merged = true;
1595 }
1596
1597 return merged;
1598}
1599
1600int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1601 u64 offset, u64 bytes)
1602{
1603 struct btrfs_free_space *info;
1604 int ret = 0;
1605
1606 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1607 if (!info)
1608 return -ENOMEM;
1609
1610 info->offset = offset;
1611 info->bytes = bytes;
1612
1613 spin_lock(&ctl->tree_lock);
1614
1615 if (try_merge_free_space(ctl, info, true))
1616 goto link;
1617
1618 /*
1619 * There was no extent directly to the left or right of this new
1620 * extent then we know we're going to have to allocate a new extent, so
1621 * before we do that see if we need to drop this into a bitmap
1622 */
1623 ret = insert_into_bitmap(ctl, info);
1624 if (ret < 0) {
1625 goto out;
1626 } else if (ret) {
1627 ret = 0;
1628 goto out;
1629 }
1630link:
1631 ret = link_free_space(ctl, info);
1632 if (ret)
1633 kmem_cache_free(btrfs_free_space_cachep, info);
1634out:
1635 spin_unlock(&ctl->tree_lock);
1636
1637 if (ret) {
1638 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1639 BUG_ON(ret == -EEXIST);
1640 }
1641
1642 return ret;
1643}
1644
1645int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1646 u64 offset, u64 bytes)
1647{
1648 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1649 struct btrfs_free_space *info;
1650 struct btrfs_free_space *next_info = NULL;
1651 int ret = 0;
1652
1653 spin_lock(&ctl->tree_lock);
1654
1655again:
1656 info = tree_search_offset(ctl, offset, 0, 0);
1657 if (!info) {
1658 /*
1659 * oops didn't find an extent that matched the space we wanted
1660 * to remove, look for a bitmap instead
1661 */
1662 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1663 1, 0);
1664 if (!info) {
1665 WARN_ON(1);
1666 goto out_lock;
1667 }
1668 }
1669
1670 if (info->bytes < bytes && rb_next(&info->offset_index)) {
1671 u64 end;
1672 next_info = rb_entry(rb_next(&info->offset_index),
1673 struct btrfs_free_space,
1674 offset_index);
1675
1676 if (next_info->bitmap)
1677 end = next_info->offset +
1678 BITS_PER_BITMAP * ctl->unit - 1;
1679 else
1680 end = next_info->offset + next_info->bytes;
1681
1682 if (next_info->bytes < bytes ||
1683 next_info->offset > offset || offset > end) {
1684 printk(KERN_CRIT "Found free space at %llu, size %llu,"
1685 " trying to use %llu\n",
1686 (unsigned long long)info->offset,
1687 (unsigned long long)info->bytes,
1688 (unsigned long long)bytes);
1689 WARN_ON(1);
1690 ret = -EINVAL;
1691 goto out_lock;
1692 }
1693
1694 info = next_info;
1695 }
1696
1697 if (info->bytes == bytes) {
1698 unlink_free_space(ctl, info);
1699 if (info->bitmap) {
1700 kfree(info->bitmap);
1701 ctl->total_bitmaps--;
1702 }
1703 kmem_cache_free(btrfs_free_space_cachep, info);
1704 goto out_lock;
1705 }
1706
1707 if (!info->bitmap && info->offset == offset) {
1708 unlink_free_space(ctl, info);
1709 info->offset += bytes;
1710 info->bytes -= bytes;
1711 link_free_space(ctl, info);
1712 goto out_lock;
1713 }
1714
1715 if (!info->bitmap && info->offset <= offset &&
1716 info->offset + info->bytes >= offset + bytes) {
1717 u64 old_start = info->offset;
1718 /*
1719 * we're freeing space in the middle of the info,
1720 * this can happen during tree log replay
1721 *
1722 * first unlink the old info and then
1723 * insert it again after the hole we're creating
1724 */
1725 unlink_free_space(ctl, info);
1726 if (offset + bytes < info->offset + info->bytes) {
1727 u64 old_end = info->offset + info->bytes;
1728
1729 info->offset = offset + bytes;
1730 info->bytes = old_end - info->offset;
1731 ret = link_free_space(ctl, info);
1732 WARN_ON(ret);
1733 if (ret)
1734 goto out_lock;
1735 } else {
1736 /* the hole we're creating ends at the end
1737 * of the info struct, just free the info
1738 */
1739 kmem_cache_free(btrfs_free_space_cachep, info);
1740 }
1741 spin_unlock(&ctl->tree_lock);
1742
1743 /* step two, insert a new info struct to cover
1744 * anything before the hole
1745 */
1746 ret = btrfs_add_free_space(block_group, old_start,
1747 offset - old_start);
1748 WARN_ON(ret);
1749 goto out;
1750 }
1751
1752 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1753 if (ret == -EAGAIN)
1754 goto again;
1755 BUG_ON(ret);
1756out_lock:
1757 spin_unlock(&ctl->tree_lock);
1758out:
1759 return ret;
1760}
1761
1762void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1763 u64 bytes)
1764{
1765 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1766 struct btrfs_free_space *info;
1767 struct rb_node *n;
1768 int count = 0;
1769
1770 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1771 info = rb_entry(n, struct btrfs_free_space, offset_index);
1772 if (info->bytes >= bytes)
1773 count++;
1774 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1775 (unsigned long long)info->offset,
1776 (unsigned long long)info->bytes,
1777 (info->bitmap) ? "yes" : "no");
1778 }
1779 printk(KERN_INFO "block group has cluster?: %s\n",
1780 list_empty(&block_group->cluster_list) ? "no" : "yes");
1781 printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1782 "\n", count);
1783}
1784
1785void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
1786{
1787 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1788
1789 spin_lock_init(&ctl->tree_lock);
1790 ctl->unit = block_group->sectorsize;
1791 ctl->start = block_group->key.objectid;
1792 ctl->private = block_group;
1793 ctl->op = &free_space_op;
1794
1795 /*
1796 * we only want to have 32k of ram per block group for keeping
1797 * track of free space, and if we pass 1/2 of that we want to
1798 * start converting things over to using bitmaps
1799 */
1800 ctl->extents_thresh = ((1024 * 32) / 2) /
1801 sizeof(struct btrfs_free_space);
1802}
1803
1804/*
1805 * for a given cluster, put all of its extents back into the free
1806 * space cache. If the block group passed doesn't match the block group
1807 * pointed to by the cluster, someone else raced in and freed the
1808 * cluster already. In that case, we just return without changing anything
1809 */
1810static int
1811__btrfs_return_cluster_to_free_space(
1812 struct btrfs_block_group_cache *block_group,
1813 struct btrfs_free_cluster *cluster)
1814{
1815 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1816 struct btrfs_free_space *entry;
1817 struct rb_node *node;
1818
1819 spin_lock(&cluster->lock);
1820 if (cluster->block_group != block_group)
1821 goto out;
1822
1823 cluster->block_group = NULL;
1824 cluster->window_start = 0;
1825 list_del_init(&cluster->block_group_list);
1826
1827 node = rb_first(&cluster->root);
1828 while (node) {
1829 bool bitmap;
1830
1831 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1832 node = rb_next(&entry->offset_index);
1833 rb_erase(&entry->offset_index, &cluster->root);
1834
1835 bitmap = (entry->bitmap != NULL);
1836 if (!bitmap)
1837 try_merge_free_space(ctl, entry, false);
1838 tree_insert_offset(&ctl->free_space_offset,
1839 entry->offset, &entry->offset_index, bitmap);
1840 }
1841 cluster->root = RB_ROOT;
1842
1843out:
1844 spin_unlock(&cluster->lock);
1845 btrfs_put_block_group(block_group);
1846 return 0;
1847}
1848
1849void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
1850{
1851 struct btrfs_free_space *info;
1852 struct rb_node *node;
1853
1854 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
1855 info = rb_entry(node, struct btrfs_free_space, offset_index);
1856 if (!info->bitmap) {
1857 unlink_free_space(ctl, info);
1858 kmem_cache_free(btrfs_free_space_cachep, info);
1859 } else {
1860 free_bitmap(ctl, info);
1861 }
1862 if (need_resched()) {
1863 spin_unlock(&ctl->tree_lock);
1864 cond_resched();
1865 spin_lock(&ctl->tree_lock);
1866 }
1867 }
1868}
1869
1870void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
1871{
1872 spin_lock(&ctl->tree_lock);
1873 __btrfs_remove_free_space_cache_locked(ctl);
1874 spin_unlock(&ctl->tree_lock);
1875}
1876
1877void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1878{
1879 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1880 struct btrfs_free_cluster *cluster;
1881 struct list_head *head;
1882
1883 spin_lock(&ctl->tree_lock);
1884 while ((head = block_group->cluster_list.next) !=
1885 &block_group->cluster_list) {
1886 cluster = list_entry(head, struct btrfs_free_cluster,
1887 block_group_list);
1888
1889 WARN_ON(cluster->block_group != block_group);
1890 __btrfs_return_cluster_to_free_space(block_group, cluster);
1891 if (need_resched()) {
1892 spin_unlock(&ctl->tree_lock);
1893 cond_resched();
1894 spin_lock(&ctl->tree_lock);
1895 }
1896 }
1897 __btrfs_remove_free_space_cache_locked(ctl);
1898 spin_unlock(&ctl->tree_lock);
1899
1900}
1901
1902u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1903 u64 offset, u64 bytes, u64 empty_size)
1904{
1905 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1906 struct btrfs_free_space *entry = NULL;
1907 u64 bytes_search = bytes + empty_size;
1908 u64 ret = 0;
1909
1910 spin_lock(&ctl->tree_lock);
1911 entry = find_free_space(ctl, &offset, &bytes_search);
1912 if (!entry)
1913 goto out;
1914
1915 ret = offset;
1916 if (entry->bitmap) {
1917 bitmap_clear_bits(ctl, entry, offset, bytes);
1918 if (!entry->bytes)
1919 free_bitmap(ctl, entry);
1920 } else {
1921 unlink_free_space(ctl, entry);
1922 entry->offset += bytes;
1923 entry->bytes -= bytes;
1924 if (!entry->bytes)
1925 kmem_cache_free(btrfs_free_space_cachep, entry);
1926 else
1927 link_free_space(ctl, entry);
1928 }
1929
1930out:
1931 spin_unlock(&ctl->tree_lock);
1932
1933 return ret;
1934}
1935
1936/*
1937 * given a cluster, put all of its extents back into the free space
1938 * cache. If a block group is passed, this function will only free
1939 * a cluster that belongs to the passed block group.
1940 *
1941 * Otherwise, it'll get a reference on the block group pointed to by the
1942 * cluster and remove the cluster from it.
1943 */
1944int btrfs_return_cluster_to_free_space(
1945 struct btrfs_block_group_cache *block_group,
1946 struct btrfs_free_cluster *cluster)
1947{
1948 struct btrfs_free_space_ctl *ctl;
1949 int ret;
1950
1951 /* first, get a safe pointer to the block group */
1952 spin_lock(&cluster->lock);
1953 if (!block_group) {
1954 block_group = cluster->block_group;
1955 if (!block_group) {
1956 spin_unlock(&cluster->lock);
1957 return 0;
1958 }
1959 } else if (cluster->block_group != block_group) {
1960 /* someone else has already freed it don't redo their work */
1961 spin_unlock(&cluster->lock);
1962 return 0;
1963 }
1964 atomic_inc(&block_group->count);
1965 spin_unlock(&cluster->lock);
1966
1967 ctl = block_group->free_space_ctl;
1968
1969 /* now return any extents the cluster had on it */
1970 spin_lock(&ctl->tree_lock);
1971 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1972 spin_unlock(&ctl->tree_lock);
1973
1974 /* finally drop our ref */
1975 btrfs_put_block_group(block_group);
1976 return ret;
1977}
1978
1979static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1980 struct btrfs_free_cluster *cluster,
1981 struct btrfs_free_space *entry,
1982 u64 bytes, u64 min_start)
1983{
1984 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1985 int err;
1986 u64 search_start = cluster->window_start;
1987 u64 search_bytes = bytes;
1988 u64 ret = 0;
1989
1990 search_start = min_start;
1991 search_bytes = bytes;
1992
1993 err = search_bitmap(ctl, entry, &search_start, &search_bytes);
1994 if (err)
1995 return 0;
1996
1997 ret = search_start;
1998 __bitmap_clear_bits(ctl, entry, ret, bytes);
1999
2000 return ret;
2001}
2002
2003/*
2004 * given a cluster, try to allocate 'bytes' from it, returns 0
2005 * if it couldn't find anything suitably large, or a logical disk offset
2006 * if things worked out
2007 */
2008u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2009 struct btrfs_free_cluster *cluster, u64 bytes,
2010 u64 min_start)
2011{
2012 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2013 struct btrfs_free_space *entry = NULL;
2014 struct rb_node *node;
2015 u64 ret = 0;
2016
2017 spin_lock(&cluster->lock);
2018 if (bytes > cluster->max_size)
2019 goto out;
2020
2021 if (cluster->block_group != block_group)
2022 goto out;
2023
2024 node = rb_first(&cluster->root);
2025 if (!node)
2026 goto out;
2027
2028 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2029 while(1) {
2030 if (entry->bytes < bytes ||
2031 (!entry->bitmap && entry->offset < min_start)) {
2032 node = rb_next(&entry->offset_index);
2033 if (!node)
2034 break;
2035 entry = rb_entry(node, struct btrfs_free_space,
2036 offset_index);
2037 continue;
2038 }
2039
2040 if (entry->bitmap) {
2041 ret = btrfs_alloc_from_bitmap(block_group,
2042 cluster, entry, bytes,
2043 min_start);
2044 if (ret == 0) {
2045 node = rb_next(&entry->offset_index);
2046 if (!node)
2047 break;
2048 entry = rb_entry(node, struct btrfs_free_space,
2049 offset_index);
2050 continue;
2051 }
2052 } else {
2053 ret = entry->offset;
2054
2055 entry->offset += bytes;
2056 entry->bytes -= bytes;
2057 }
2058
2059 if (entry->bytes == 0)
2060 rb_erase(&entry->offset_index, &cluster->root);
2061 break;
2062 }
2063out:
2064 spin_unlock(&cluster->lock);
2065
2066 if (!ret)
2067 return 0;
2068
2069 spin_lock(&ctl->tree_lock);
2070
2071 ctl->free_space -= bytes;
2072 if (entry->bytes == 0) {
2073 ctl->free_extents--;
2074 if (entry->bitmap) {
2075 kfree(entry->bitmap);
2076 ctl->total_bitmaps--;
2077 ctl->op->recalc_thresholds(ctl);
2078 }
2079 kmem_cache_free(btrfs_free_space_cachep, entry);
2080 }
2081
2082 spin_unlock(&ctl->tree_lock);
2083
2084 return ret;
2085}
2086
2087static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2088 struct btrfs_free_space *entry,
2089 struct btrfs_free_cluster *cluster,
2090 u64 offset, u64 bytes, u64 min_bytes)
2091{
2092 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2093 unsigned long next_zero;
2094 unsigned long i;
2095 unsigned long search_bits;
2096 unsigned long total_bits;
2097 unsigned long found_bits;
2098 unsigned long start = 0;
2099 unsigned long total_found = 0;
2100 int ret;
2101 bool found = false;
2102
2103 i = offset_to_bit(entry->offset, block_group->sectorsize,
2104 max_t(u64, offset, entry->offset));
2105 search_bits = bytes_to_bits(bytes, block_group->sectorsize);
2106 total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
2107
2108again:
2109 found_bits = 0;
2110 for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
2111 i < BITS_PER_BITMAP;
2112 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
2113 next_zero = find_next_zero_bit(entry->bitmap,
2114 BITS_PER_BITMAP, i);
2115 if (next_zero - i >= search_bits) {
2116 found_bits = next_zero - i;
2117 break;
2118 }
2119 i = next_zero;
2120 }
2121
2122 if (!found_bits)
2123 return -ENOSPC;
2124
2125 if (!found) {
2126 start = i;
2127 found = true;
2128 }
2129
2130 total_found += found_bits;
2131
2132 if (cluster->max_size < found_bits * block_group->sectorsize)
2133 cluster->max_size = found_bits * block_group->sectorsize;
2134
2135 if (total_found < total_bits) {
2136 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2137 if (i - start > total_bits * 2) {
2138 total_found = 0;
2139 cluster->max_size = 0;
2140 found = false;
2141 }
2142 goto again;
2143 }
2144
2145 cluster->window_start = start * block_group->sectorsize +
2146 entry->offset;
2147 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2148 ret = tree_insert_offset(&cluster->root, entry->offset,
2149 &entry->offset_index, 1);
2150 BUG_ON(ret);
2151
2152 return 0;
2153}
2154
2155/*
2156 * This searches the block group for just extents to fill the cluster with.
2157 */
2158static noinline int
2159setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2160 struct btrfs_free_cluster *cluster,
2161 struct list_head *bitmaps, u64 offset, u64 bytes,
2162 u64 min_bytes)
2163{
2164 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2165 struct btrfs_free_space *first = NULL;
2166 struct btrfs_free_space *entry = NULL;
2167 struct btrfs_free_space *prev = NULL;
2168 struct btrfs_free_space *last;
2169 struct rb_node *node;
2170 u64 window_start;
2171 u64 window_free;
2172 u64 max_extent;
2173 u64 max_gap = 128 * 1024;
2174
2175 entry = tree_search_offset(ctl, offset, 0, 1);
2176 if (!entry)
2177 return -ENOSPC;
2178
2179 /*
2180 * We don't want bitmaps, so just move along until we find a normal
2181 * extent entry.
2182 */
2183 while (entry->bitmap) {
2184 if (list_empty(&entry->list))
2185 list_add_tail(&entry->list, bitmaps);
2186 node = rb_next(&entry->offset_index);
2187 if (!node)
2188 return -ENOSPC;
2189 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2190 }
2191
2192 window_start = entry->offset;
2193 window_free = entry->bytes;
2194 max_extent = entry->bytes;
2195 first = entry;
2196 last = entry;
2197 prev = entry;
2198
2199 while (window_free <= min_bytes) {
2200 node = rb_next(&entry->offset_index);
2201 if (!node)
2202 return -ENOSPC;
2203 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2204
2205 if (entry->bitmap) {
2206 if (list_empty(&entry->list))
2207 list_add_tail(&entry->list, bitmaps);
2208 continue;
2209 }
2210
2211 /*
2212 * we haven't filled the empty size and the window is
2213 * very large. reset and try again
2214 */
2215 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2216 entry->offset - window_start > (min_bytes * 2)) {
2217 first = entry;
2218 window_start = entry->offset;
2219 window_free = entry->bytes;
2220 last = entry;
2221 max_extent = entry->bytes;
2222 } else {
2223 last = entry;
2224 window_free += entry->bytes;
2225 if (entry->bytes > max_extent)
2226 max_extent = entry->bytes;
2227 }
2228 prev = entry;
2229 }
2230
2231 cluster->window_start = first->offset;
2232
2233 node = &first->offset_index;
2234
2235 /*
2236 * now we've found our entries, pull them out of the free space
2237 * cache and put them into the cluster rbtree
2238 */
2239 do {
2240 int ret;
2241
2242 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2243 node = rb_next(&entry->offset_index);
2244 if (entry->bitmap)
2245 continue;
2246
2247 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2248 ret = tree_insert_offset(&cluster->root, entry->offset,
2249 &entry->offset_index, 0);
2250 BUG_ON(ret);
2251 } while (node && entry != last);
2252
2253 cluster->max_size = max_extent;
2254
2255 return 0;
2256}
2257
2258/*
2259 * This specifically looks for bitmaps that may work in the cluster, we assume
2260 * that we have already failed to find extents that will work.
2261 */
2262static noinline int
2263setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2264 struct btrfs_free_cluster *cluster,
2265 struct list_head *bitmaps, u64 offset, u64 bytes,
2266 u64 min_bytes)
2267{
2268 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2269 struct btrfs_free_space *entry;
2270 struct rb_node *node;
2271 int ret = -ENOSPC;
2272
2273 if (ctl->total_bitmaps == 0)
2274 return -ENOSPC;
2275
2276 /*
2277 * First check our cached list of bitmaps and see if there is an entry
2278 * here that will work.
2279 */
2280 list_for_each_entry(entry, bitmaps, list) {
2281 if (entry->bytes < min_bytes)
2282 continue;
2283 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2284 bytes, min_bytes);
2285 if (!ret)
2286 return 0;
2287 }
2288
2289 /*
2290 * If we do have entries on our list and we are here then we didn't find
2291 * anything, so go ahead and get the next entry after the last entry in
2292 * this list and start the search from there.
2293 */
2294 if (!list_empty(bitmaps)) {
2295 entry = list_entry(bitmaps->prev, struct btrfs_free_space,
2296 list);
2297 node = rb_next(&entry->offset_index);
2298 if (!node)
2299 return -ENOSPC;
2300 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2301 goto search;
2302 }
2303
2304 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, offset), 0, 1);
2305 if (!entry)
2306 return -ENOSPC;
2307
2308search:
2309 node = &entry->offset_index;
2310 do {
2311 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2312 node = rb_next(&entry->offset_index);
2313 if (!entry->bitmap)
2314 continue;
2315 if (entry->bytes < min_bytes)
2316 continue;
2317 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2318 bytes, min_bytes);
2319 } while (ret && node);
2320
2321 return ret;
2322}
2323
2324/*
2325 * here we try to find a cluster of blocks in a block group. The goal
2326 * is to find at least bytes free and up to empty_size + bytes free.
2327 * We might not find them all in one contiguous area.
2328 *
2329 * returns zero and sets up cluster if things worked out, otherwise
2330 * it returns -enospc
2331 */
2332int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2333 struct btrfs_root *root,
2334 struct btrfs_block_group_cache *block_group,
2335 struct btrfs_free_cluster *cluster,
2336 u64 offset, u64 bytes, u64 empty_size)
2337{
2338 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2339 struct list_head bitmaps;
2340 struct btrfs_free_space *entry, *tmp;
2341 u64 min_bytes;
2342 int ret;
2343
2344 /* for metadata, allow allocates with more holes */
2345 if (btrfs_test_opt(root, SSD_SPREAD)) {
2346 min_bytes = bytes + empty_size;
2347 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2348 /*
2349 * we want to do larger allocations when we are
2350 * flushing out the delayed refs, it helps prevent
2351 * making more work as we go along.
2352 */
2353 if (trans->transaction->delayed_refs.flushing)
2354 min_bytes = max(bytes, (bytes + empty_size) >> 1);
2355 else
2356 min_bytes = max(bytes, (bytes + empty_size) >> 4);
2357 } else
2358 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2359
2360 spin_lock(&ctl->tree_lock);
2361
2362 /*
2363 * If we know we don't have enough space to make a cluster don't even
2364 * bother doing all the work to try and find one.
2365 */
2366 if (ctl->free_space < min_bytes) {
2367 spin_unlock(&ctl->tree_lock);
2368 return -ENOSPC;
2369 }
2370
2371 spin_lock(&cluster->lock);
2372
2373 /* someone already found a cluster, hooray */
2374 if (cluster->block_group) {
2375 ret = 0;
2376 goto out;
2377 }
2378
2379 INIT_LIST_HEAD(&bitmaps);
2380 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2381 bytes, min_bytes);
2382 if (ret)
2383 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2384 offset, bytes, min_bytes);
2385
2386 /* Clear our temporary list */
2387 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2388 list_del_init(&entry->list);
2389
2390 if (!ret) {
2391 atomic_inc(&block_group->count);
2392 list_add_tail(&cluster->block_group_list,
2393 &block_group->cluster_list);
2394 cluster->block_group = block_group;
2395 }
2396out:
2397 spin_unlock(&cluster->lock);
2398 spin_unlock(&ctl->tree_lock);
2399
2400 return ret;
2401}
2402
2403/*
2404 * simple code to zero out a cluster
2405 */
2406void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2407{
2408 spin_lock_init(&cluster->lock);
2409 spin_lock_init(&cluster->refill_lock);
2410 cluster->root = RB_ROOT;
2411 cluster->max_size = 0;
2412 INIT_LIST_HEAD(&cluster->block_group_list);
2413 cluster->block_group = NULL;
2414}
2415
2416int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2417 u64 *trimmed, u64 start, u64 end, u64 minlen)
2418{
2419 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2420 struct btrfs_free_space *entry = NULL;
2421 struct btrfs_fs_info *fs_info = block_group->fs_info;
2422 u64 bytes = 0;
2423 u64 actually_trimmed;
2424 int ret = 0;
2425
2426 *trimmed = 0;
2427
2428 while (start < end) {
2429 spin_lock(&ctl->tree_lock);
2430
2431 if (ctl->free_space < minlen) {
2432 spin_unlock(&ctl->tree_lock);
2433 break;
2434 }
2435
2436 entry = tree_search_offset(ctl, start, 0, 1);
2437 if (!entry)
2438 entry = tree_search_offset(ctl,
2439 offset_to_bitmap(ctl, start),
2440 1, 1);
2441
2442 if (!entry || entry->offset >= end) {
2443 spin_unlock(&ctl->tree_lock);
2444 break;
2445 }
2446
2447 if (entry->bitmap) {
2448 ret = search_bitmap(ctl, entry, &start, &bytes);
2449 if (!ret) {
2450 if (start >= end) {
2451 spin_unlock(&ctl->tree_lock);
2452 break;
2453 }
2454 bytes = min(bytes, end - start);
2455 bitmap_clear_bits(ctl, entry, start, bytes);
2456 if (entry->bytes == 0)
2457 free_bitmap(ctl, entry);
2458 } else {
2459 start = entry->offset + BITS_PER_BITMAP *
2460 block_group->sectorsize;
2461 spin_unlock(&ctl->tree_lock);
2462 ret = 0;
2463 continue;
2464 }
2465 } else {
2466 start = entry->offset;
2467 bytes = min(entry->bytes, end - start);
2468 unlink_free_space(ctl, entry);
2469 kmem_cache_free(btrfs_free_space_cachep, entry);
2470 }
2471
2472 spin_unlock(&ctl->tree_lock);
2473
2474 if (bytes >= minlen) {
2475 int update_ret;
2476 update_ret = btrfs_update_reserved_bytes(block_group,
2477 bytes, 1, 1);
2478
2479 ret = btrfs_error_discard_extent(fs_info->extent_root,
2480 start,
2481 bytes,
2482 &actually_trimmed);
2483
2484 btrfs_add_free_space(block_group, start, bytes);
2485 if (!update_ret)
2486 btrfs_update_reserved_bytes(block_group,
2487 bytes, 0, 1);
2488
2489 if (ret)
2490 break;
2491 *trimmed += actually_trimmed;
2492 }
2493 start += bytes;
2494 bytes = 0;
2495
2496 if (fatal_signal_pending(current)) {
2497 ret = -ERESTARTSYS;
2498 break;
2499 }
2500
2501 cond_resched();
2502 }
2503
2504 return ret;
2505}
2506
2507/*
2508 * Find the left-most item in the cache tree, and then return the
2509 * smallest inode number in the item.
2510 *
2511 * Note: the returned inode number may not be the smallest one in
2512 * the tree, if the left-most item is a bitmap.
2513 */
2514u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2515{
2516 struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2517 struct btrfs_free_space *entry = NULL;
2518 u64 ino = 0;
2519
2520 spin_lock(&ctl->tree_lock);
2521
2522 if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2523 goto out;
2524
2525 entry = rb_entry(rb_first(&ctl->free_space_offset),
2526 struct btrfs_free_space, offset_index);
2527
2528 if (!entry->bitmap) {
2529 ino = entry->offset;
2530
2531 unlink_free_space(ctl, entry);
2532 entry->offset++;
2533 entry->bytes--;
2534 if (!entry->bytes)
2535 kmem_cache_free(btrfs_free_space_cachep, entry);
2536 else
2537 link_free_space(ctl, entry);
2538 } else {
2539 u64 offset = 0;
2540 u64 count = 1;
2541 int ret;
2542
2543 ret = search_bitmap(ctl, entry, &offset, &count);
2544 BUG_ON(ret);
2545
2546 ino = offset;
2547 bitmap_clear_bits(ctl, entry, offset, 1);
2548 if (entry->bytes == 0)
2549 free_bitmap(ctl, entry);
2550 }
2551out:
2552 spin_unlock(&ctl->tree_lock);
2553
2554 return ino;
2555}
2556
2557struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2558 struct btrfs_path *path)
2559{
2560 struct inode *inode = NULL;
2561
2562 spin_lock(&root->cache_lock);
2563 if (root->cache_inode)
2564 inode = igrab(root->cache_inode);
2565 spin_unlock(&root->cache_lock);
2566 if (inode)
2567 return inode;
2568
2569 inode = __lookup_free_space_inode(root, path, 0);
2570 if (IS_ERR(inode))
2571 return inode;
2572
2573 spin_lock(&root->cache_lock);
2574 if (!btrfs_fs_closing(root->fs_info))
2575 root->cache_inode = igrab(inode);
2576 spin_unlock(&root->cache_lock);
2577
2578 return inode;
2579}
2580
2581int create_free_ino_inode(struct btrfs_root *root,
2582 struct btrfs_trans_handle *trans,
2583 struct btrfs_path *path)
2584{
2585 return __create_free_space_inode(root, trans, path,
2586 BTRFS_FREE_INO_OBJECTID, 0);
2587}
2588
2589int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2590{
2591 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2592 struct btrfs_path *path;
2593 struct inode *inode;
2594 int ret = 0;
2595 u64 root_gen = btrfs_root_generation(&root->root_item);
2596
2597 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2598 return 0;
2599
2600 /*
2601 * If we're unmounting then just return, since this does a search on the
2602 * normal root and not the commit root and we could deadlock.
2603 */
2604 if (btrfs_fs_closing(fs_info))
2605 return 0;
2606
2607 path = btrfs_alloc_path();
2608 if (!path)
2609 return 0;
2610
2611 inode = lookup_free_ino_inode(root, path);
2612 if (IS_ERR(inode))
2613 goto out;
2614
2615 if (root_gen != BTRFS_I(inode)->generation)
2616 goto out_put;
2617
2618 ret = __load_free_space_cache(root, inode, ctl, path, 0);
2619
2620 if (ret < 0)
2621 printk(KERN_ERR "btrfs: failed to load free ino cache for "
2622 "root %llu\n", root->root_key.objectid);
2623out_put:
2624 iput(inode);
2625out:
2626 btrfs_free_path(path);
2627 return ret;
2628}
2629
2630int btrfs_write_out_ino_cache(struct btrfs_root *root,
2631 struct btrfs_trans_handle *trans,
2632 struct btrfs_path *path)
2633{
2634 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2635 struct inode *inode;
2636 int ret;
2637
2638 if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2639 return 0;
2640
2641 inode = lookup_free_ino_inode(root, path);
2642 if (IS_ERR(inode))
2643 return 0;
2644
2645 ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2646 if (ret < 0)
2647 printk(KERN_ERR "btrfs: failed to write free ino cache "
2648 "for root %llu\n", root->root_key.objectid);
2649
2650 iput(inode);
2651 return ret;
2652}