1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2008 Red Hat. All rights reserved.
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>
16 #include "free-space-cache.h"
17 #include "transaction.h"
19 #include "extent_io.h"
21 #include "space-info.h"
22 #include "delalloc-space.h"
23 #include "block-group.h"
26 #define BITS_PER_BITMAP (PAGE_SIZE * 8UL)
27 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
28 #define FORCE_EXTENT_THRESHOLD SZ_1M
30 struct btrfs_trim_range {
33 struct list_head list;
36 static int link_free_space(struct btrfs_free_space_ctl *ctl,
37 struct btrfs_free_space *info);
38 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
39 struct btrfs_free_space *info);
40 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
41 struct btrfs_free_space *bitmap_info, u64 *offset,
42 u64 *bytes, bool for_alloc);
43 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
44 struct btrfs_free_space *bitmap_info);
45 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
46 struct btrfs_free_space *info, u64 offset,
49 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
50 struct btrfs_path *path,
53 struct btrfs_fs_info *fs_info = root->fs_info;
55 struct btrfs_key location;
56 struct btrfs_disk_key disk_key;
57 struct btrfs_free_space_header *header;
58 struct extent_buffer *leaf;
59 struct inode *inode = NULL;
63 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
67 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
71 btrfs_release_path(path);
72 return ERR_PTR(-ENOENT);
75 leaf = path->nodes[0];
76 header = btrfs_item_ptr(leaf, path->slots[0],
77 struct btrfs_free_space_header);
78 btrfs_free_space_key(leaf, header, &disk_key);
79 btrfs_disk_key_to_cpu(&location, &disk_key);
80 btrfs_release_path(path);
83 * We are often under a trans handle at this point, so we need to make
84 * sure NOFS is set to keep us from deadlocking.
86 nofs_flag = memalloc_nofs_save();
87 inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
88 btrfs_release_path(path);
89 memalloc_nofs_restore(nofs_flag);
93 mapping_set_gfp_mask(inode->i_mapping,
94 mapping_gfp_constraint(inode->i_mapping,
95 ~(__GFP_FS | __GFP_HIGHMEM)));
100 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
101 struct btrfs_path *path)
103 struct btrfs_fs_info *fs_info = block_group->fs_info;
104 struct inode *inode = NULL;
105 u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
107 spin_lock(&block_group->lock);
108 if (block_group->inode)
109 inode = igrab(block_group->inode);
110 spin_unlock(&block_group->lock);
114 inode = __lookup_free_space_inode(fs_info->tree_root, path,
119 spin_lock(&block_group->lock);
120 if (!((BTRFS_I(inode)->flags & flags) == flags)) {
121 btrfs_info(fs_info, "Old style space inode found, converting.");
122 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
123 BTRFS_INODE_NODATACOW;
124 block_group->disk_cache_state = BTRFS_DC_CLEAR;
127 if (!block_group->iref) {
128 block_group->inode = igrab(inode);
129 block_group->iref = 1;
131 spin_unlock(&block_group->lock);
136 static int __create_free_space_inode(struct btrfs_root *root,
137 struct btrfs_trans_handle *trans,
138 struct btrfs_path *path,
141 struct btrfs_key key;
142 struct btrfs_disk_key disk_key;
143 struct btrfs_free_space_header *header;
144 struct btrfs_inode_item *inode_item;
145 struct extent_buffer *leaf;
146 /* We inline CRCs for the free disk space cache */
147 const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
148 BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
151 ret = btrfs_insert_empty_inode(trans, root, path, ino);
155 leaf = path->nodes[0];
156 inode_item = btrfs_item_ptr(leaf, path->slots[0],
157 struct btrfs_inode_item);
158 btrfs_item_key(leaf, &disk_key, path->slots[0]);
159 memzero_extent_buffer(leaf, (unsigned long)inode_item,
160 sizeof(*inode_item));
161 btrfs_set_inode_generation(leaf, inode_item, trans->transid);
162 btrfs_set_inode_size(leaf, inode_item, 0);
163 btrfs_set_inode_nbytes(leaf, inode_item, 0);
164 btrfs_set_inode_uid(leaf, inode_item, 0);
165 btrfs_set_inode_gid(leaf, inode_item, 0);
166 btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
167 btrfs_set_inode_flags(leaf, inode_item, flags);
168 btrfs_set_inode_nlink(leaf, inode_item, 1);
169 btrfs_set_inode_transid(leaf, inode_item, trans->transid);
170 btrfs_set_inode_block_group(leaf, inode_item, offset);
171 btrfs_mark_buffer_dirty(leaf);
172 btrfs_release_path(path);
174 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
177 ret = btrfs_insert_empty_item(trans, root, path, &key,
178 sizeof(struct btrfs_free_space_header));
180 btrfs_release_path(path);
184 leaf = path->nodes[0];
185 header = btrfs_item_ptr(leaf, path->slots[0],
186 struct btrfs_free_space_header);
187 memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
188 btrfs_set_free_space_key(leaf, header, &disk_key);
189 btrfs_mark_buffer_dirty(leaf);
190 btrfs_release_path(path);
195 int create_free_space_inode(struct btrfs_trans_handle *trans,
196 struct btrfs_block_group *block_group,
197 struct btrfs_path *path)
202 ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
206 return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
207 ino, block_group->start);
211 * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
212 * handles lookup, otherwise it takes ownership and iputs the inode.
213 * Don't reuse an inode pointer after passing it into this function.
215 int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
217 struct btrfs_block_group *block_group)
219 struct btrfs_path *path;
220 struct btrfs_key key;
223 path = btrfs_alloc_path();
228 inode = lookup_free_space_inode(block_group, path);
230 if (PTR_ERR(inode) != -ENOENT)
231 ret = PTR_ERR(inode);
234 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
236 btrfs_add_delayed_iput(inode);
240 /* One for the block groups ref */
241 spin_lock(&block_group->lock);
242 if (block_group->iref) {
243 block_group->iref = 0;
244 block_group->inode = NULL;
245 spin_unlock(&block_group->lock);
248 spin_unlock(&block_group->lock);
250 /* One for the lookup ref */
251 btrfs_add_delayed_iput(inode);
253 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
255 key.offset = block_group->start;
256 ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
263 ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
265 btrfs_free_path(path);
269 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
270 struct btrfs_block_rsv *rsv)
275 /* 1 for slack space, 1 for updating the inode */
276 needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
277 btrfs_calc_metadata_size(fs_info, 1);
279 spin_lock(&rsv->lock);
280 if (rsv->reserved < needed_bytes)
284 spin_unlock(&rsv->lock);
288 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
289 struct btrfs_block_group *block_group,
292 struct btrfs_root *root = BTRFS_I(inode)->root;
297 struct btrfs_path *path = btrfs_alloc_path();
304 mutex_lock(&trans->transaction->cache_write_mutex);
305 if (!list_empty(&block_group->io_list)) {
306 list_del_init(&block_group->io_list);
308 btrfs_wait_cache_io(trans, block_group, path);
309 btrfs_put_block_group(block_group);
313 * now that we've truncated the cache away, its no longer
316 spin_lock(&block_group->lock);
317 block_group->disk_cache_state = BTRFS_DC_CLEAR;
318 spin_unlock(&block_group->lock);
319 btrfs_free_path(path);
322 btrfs_i_size_write(BTRFS_I(inode), 0);
323 truncate_pagecache(inode, 0);
326 * We skip the throttling logic for free space cache inodes, so we don't
327 * need to check for -EAGAIN.
329 ret = btrfs_truncate_inode_items(trans, root, BTRFS_I(inode),
330 0, BTRFS_EXTENT_DATA_KEY);
334 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
338 mutex_unlock(&trans->transaction->cache_write_mutex);
340 btrfs_abort_transaction(trans, ret);
345 static void readahead_cache(struct inode *inode)
347 struct file_ra_state *ra;
348 unsigned long last_index;
350 ra = kzalloc(sizeof(*ra), GFP_NOFS);
354 file_ra_state_init(ra, inode->i_mapping);
355 last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
357 page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
362 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
367 num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
369 /* Make sure we can fit our crcs and generation into the first page */
370 if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
373 memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
375 io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
379 io_ctl->num_pages = num_pages;
380 io_ctl->fs_info = btrfs_sb(inode->i_sb);
381 io_ctl->inode = inode;
385 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
387 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
389 kfree(io_ctl->pages);
390 io_ctl->pages = NULL;
393 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
401 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
403 ASSERT(io_ctl->index < io_ctl->num_pages);
404 io_ctl->page = io_ctl->pages[io_ctl->index++];
405 io_ctl->cur = page_address(io_ctl->page);
406 io_ctl->orig = io_ctl->cur;
407 io_ctl->size = PAGE_SIZE;
409 clear_page(io_ctl->cur);
412 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
416 io_ctl_unmap_page(io_ctl);
418 for (i = 0; i < io_ctl->num_pages; i++) {
419 if (io_ctl->pages[i]) {
420 ClearPageChecked(io_ctl->pages[i]);
421 unlock_page(io_ctl->pages[i]);
422 put_page(io_ctl->pages[i]);
427 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
430 struct inode *inode = io_ctl->inode;
431 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
434 for (i = 0; i < io_ctl->num_pages; i++) {
437 page = find_or_create_page(inode->i_mapping, i, mask);
439 io_ctl_drop_pages(io_ctl);
443 ret = set_page_extent_mapped(page);
447 io_ctl_drop_pages(io_ctl);
451 io_ctl->pages[i] = page;
452 if (uptodate && !PageUptodate(page)) {
453 btrfs_readpage(NULL, page);
455 if (page->mapping != inode->i_mapping) {
456 btrfs_err(BTRFS_I(inode)->root->fs_info,
457 "free space cache page truncated");
458 io_ctl_drop_pages(io_ctl);
461 if (!PageUptodate(page)) {
462 btrfs_err(BTRFS_I(inode)->root->fs_info,
463 "error reading free space cache");
464 io_ctl_drop_pages(io_ctl);
470 for (i = 0; i < io_ctl->num_pages; i++)
471 clear_page_dirty_for_io(io_ctl->pages[i]);
476 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
478 io_ctl_map_page(io_ctl, 1);
481 * Skip the csum areas. If we don't check crcs then we just have a
482 * 64bit chunk at the front of the first page.
484 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
485 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
487 put_unaligned_le64(generation, io_ctl->cur);
488 io_ctl->cur += sizeof(u64);
491 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
496 * Skip the crc area. If we don't check crcs then we just have a 64bit
497 * chunk at the front of the first page.
499 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
500 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
502 cache_gen = get_unaligned_le64(io_ctl->cur);
503 if (cache_gen != generation) {
504 btrfs_err_rl(io_ctl->fs_info,
505 "space cache generation (%llu) does not match inode (%llu)",
506 cache_gen, generation);
507 io_ctl_unmap_page(io_ctl);
510 io_ctl->cur += sizeof(u64);
514 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
521 offset = sizeof(u32) * io_ctl->num_pages;
523 crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
524 btrfs_crc32c_final(crc, (u8 *)&crc);
525 io_ctl_unmap_page(io_ctl);
526 tmp = page_address(io_ctl->pages[0]);
531 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
538 offset = sizeof(u32) * io_ctl->num_pages;
540 tmp = page_address(io_ctl->pages[0]);
544 io_ctl_map_page(io_ctl, 0);
545 crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
546 btrfs_crc32c_final(crc, (u8 *)&crc);
548 btrfs_err_rl(io_ctl->fs_info,
549 "csum mismatch on free space cache");
550 io_ctl_unmap_page(io_ctl);
557 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
560 struct btrfs_free_space_entry *entry;
566 put_unaligned_le64(offset, &entry->offset);
567 put_unaligned_le64(bytes, &entry->bytes);
568 entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
569 BTRFS_FREE_SPACE_EXTENT;
570 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
571 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
573 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
576 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
578 /* No more pages to map */
579 if (io_ctl->index >= io_ctl->num_pages)
582 /* map the next page */
583 io_ctl_map_page(io_ctl, 1);
587 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
593 * If we aren't at the start of the current page, unmap this one and
594 * map the next one if there is any left.
596 if (io_ctl->cur != io_ctl->orig) {
597 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
598 if (io_ctl->index >= io_ctl->num_pages)
600 io_ctl_map_page(io_ctl, 0);
603 copy_page(io_ctl->cur, bitmap);
604 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
605 if (io_ctl->index < io_ctl->num_pages)
606 io_ctl_map_page(io_ctl, 0);
610 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
613 * If we're not on the boundary we know we've modified the page and we
614 * need to crc the page.
616 if (io_ctl->cur != io_ctl->orig)
617 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
619 io_ctl_unmap_page(io_ctl);
621 while (io_ctl->index < io_ctl->num_pages) {
622 io_ctl_map_page(io_ctl, 1);
623 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
627 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
628 struct btrfs_free_space *entry, u8 *type)
630 struct btrfs_free_space_entry *e;
634 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
640 entry->offset = get_unaligned_le64(&e->offset);
641 entry->bytes = get_unaligned_le64(&e->bytes);
643 io_ctl->cur += sizeof(struct btrfs_free_space_entry);
644 io_ctl->size -= sizeof(struct btrfs_free_space_entry);
646 if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
649 io_ctl_unmap_page(io_ctl);
654 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
655 struct btrfs_free_space *entry)
659 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
663 copy_page(entry->bitmap, io_ctl->cur);
664 io_ctl_unmap_page(io_ctl);
669 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
671 struct btrfs_block_group *block_group = ctl->private;
675 u64 size = block_group->length;
676 u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
677 u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
679 max_bitmaps = max_t(u64, max_bitmaps, 1);
681 ASSERT(ctl->total_bitmaps <= max_bitmaps);
684 * We are trying to keep the total amount of memory used per 1GiB of
685 * space to be MAX_CACHE_BYTES_PER_GIG. However, with a reclamation
686 * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
687 * bitmaps, we may end up using more memory than this.
690 max_bytes = MAX_CACHE_BYTES_PER_GIG;
692 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
694 bitmap_bytes = ctl->total_bitmaps * ctl->unit;
697 * we want the extent entry threshold to always be at most 1/2 the max
698 * bytes we can have, or whatever is less than that.
700 extent_bytes = max_bytes - bitmap_bytes;
701 extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
703 ctl->extents_thresh =
704 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
707 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
708 struct btrfs_free_space_ctl *ctl,
709 struct btrfs_path *path, u64 offset)
711 struct btrfs_fs_info *fs_info = root->fs_info;
712 struct btrfs_free_space_header *header;
713 struct extent_buffer *leaf;
714 struct btrfs_io_ctl io_ctl;
715 struct btrfs_key key;
716 struct btrfs_free_space *e, *n;
724 /* Nothing in the space cache, goodbye */
725 if (!i_size_read(inode))
728 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
732 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
736 btrfs_release_path(path);
742 leaf = path->nodes[0];
743 header = btrfs_item_ptr(leaf, path->slots[0],
744 struct btrfs_free_space_header);
745 num_entries = btrfs_free_space_entries(leaf, header);
746 num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
747 generation = btrfs_free_space_generation(leaf, header);
748 btrfs_release_path(path);
750 if (!BTRFS_I(inode)->generation) {
752 "the free space cache file (%llu) is invalid, skip it",
757 if (BTRFS_I(inode)->generation != generation) {
759 "free space inode generation (%llu) did not match free space cache generation (%llu)",
760 BTRFS_I(inode)->generation, generation);
767 ret = io_ctl_init(&io_ctl, inode, 0);
771 readahead_cache(inode);
773 ret = io_ctl_prepare_pages(&io_ctl, true);
777 ret = io_ctl_check_crc(&io_ctl, 0);
781 ret = io_ctl_check_generation(&io_ctl, generation);
785 while (num_entries) {
786 e = kmem_cache_zalloc(btrfs_free_space_cachep,
793 ret = io_ctl_read_entry(&io_ctl, e, &type);
795 kmem_cache_free(btrfs_free_space_cachep, e);
801 kmem_cache_free(btrfs_free_space_cachep, e);
805 if (type == BTRFS_FREE_SPACE_EXTENT) {
806 spin_lock(&ctl->tree_lock);
807 ret = link_free_space(ctl, e);
808 spin_unlock(&ctl->tree_lock);
811 "Duplicate entries in free space cache, dumping");
812 kmem_cache_free(btrfs_free_space_cachep, e);
818 e->bitmap = kmem_cache_zalloc(
819 btrfs_free_space_bitmap_cachep, GFP_NOFS);
823 btrfs_free_space_cachep, e);
826 spin_lock(&ctl->tree_lock);
827 ret = link_free_space(ctl, e);
828 ctl->total_bitmaps++;
829 recalculate_thresholds(ctl);
830 spin_unlock(&ctl->tree_lock);
833 "Duplicate entries in free space cache, dumping");
834 kmem_cache_free(btrfs_free_space_cachep, e);
837 list_add_tail(&e->list, &bitmaps);
843 io_ctl_unmap_page(&io_ctl);
846 * We add the bitmaps at the end of the entries in order that
847 * the bitmap entries are added to the cache.
849 list_for_each_entry_safe(e, n, &bitmaps, list) {
850 list_del_init(&e->list);
851 ret = io_ctl_read_bitmap(&io_ctl, e);
856 io_ctl_drop_pages(&io_ctl);
859 io_ctl_free(&io_ctl);
862 io_ctl_drop_pages(&io_ctl);
863 __btrfs_remove_free_space_cache(ctl);
867 static int copy_free_space_cache(struct btrfs_block_group *block_group,
868 struct btrfs_free_space_ctl *ctl)
870 struct btrfs_free_space *info;
874 while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
875 info = rb_entry(n, struct btrfs_free_space, offset_index);
877 unlink_free_space(ctl, info);
878 ret = btrfs_add_free_space(block_group, info->offset,
880 kmem_cache_free(btrfs_free_space_cachep, info);
882 u64 offset = info->offset;
883 u64 bytes = ctl->unit;
885 while (search_bitmap(ctl, info, &offset, &bytes,
887 ret = btrfs_add_free_space(block_group, offset,
891 bitmap_clear_bits(ctl, info, offset, bytes);
892 offset = info->offset;
895 free_bitmap(ctl, info);
902 int load_free_space_cache(struct btrfs_block_group *block_group)
904 struct btrfs_fs_info *fs_info = block_group->fs_info;
905 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
906 struct btrfs_free_space_ctl tmp_ctl = {};
908 struct btrfs_path *path;
911 u64 used = block_group->used;
914 * Because we could potentially discard our loaded free space, we want
915 * to load everything into a temporary structure first, and then if it's
916 * valid copy it all into the actual free space ctl.
918 btrfs_init_free_space_ctl(block_group, &tmp_ctl);
921 * If this block group has been marked to be cleared for one reason or
922 * another then we can't trust the on disk cache, so just return.
924 spin_lock(&block_group->lock);
925 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
926 spin_unlock(&block_group->lock);
929 spin_unlock(&block_group->lock);
931 path = btrfs_alloc_path();
934 path->search_commit_root = 1;
935 path->skip_locking = 1;
938 * We must pass a path with search_commit_root set to btrfs_iget in
939 * order to avoid a deadlock when allocating extents for the tree root.
941 * When we are COWing an extent buffer from the tree root, when looking
942 * for a free extent, at extent-tree.c:find_free_extent(), we can find
943 * block group without its free space cache loaded. When we find one
944 * we must load its space cache which requires reading its free space
945 * cache's inode item from the root tree. If this inode item is located
946 * in the same leaf that we started COWing before, then we end up in
947 * deadlock on the extent buffer (trying to read lock it when we
948 * previously write locked it).
950 * It's safe to read the inode item using the commit root because
951 * block groups, once loaded, stay in memory forever (until they are
952 * removed) as well as their space caches once loaded. New block groups
953 * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
954 * we will never try to read their inode item while the fs is mounted.
956 inode = lookup_free_space_inode(block_group, path);
958 btrfs_free_path(path);
962 /* We may have converted the inode and made the cache invalid. */
963 spin_lock(&block_group->lock);
964 if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
965 spin_unlock(&block_group->lock);
966 btrfs_free_path(path);
969 spin_unlock(&block_group->lock);
971 ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
972 path, block_group->start);
973 btrfs_free_path(path);
977 matched = (tmp_ctl.free_space == (block_group->length - used -
978 block_group->bytes_super));
981 ret = copy_free_space_cache(block_group, &tmp_ctl);
983 * ret == 1 means we successfully loaded the free space cache,
984 * so we need to re-set it here.
989 __btrfs_remove_free_space_cache(&tmp_ctl);
991 "block group %llu has wrong amount of free space",
997 /* This cache is bogus, make sure it gets cleared */
998 spin_lock(&block_group->lock);
999 block_group->disk_cache_state = BTRFS_DC_CLEAR;
1000 spin_unlock(&block_group->lock);
1004 "failed to load free space cache for block group %llu, rebuilding it now",
1005 block_group->start);
1008 spin_lock(&ctl->tree_lock);
1009 btrfs_discard_update_discardable(block_group);
1010 spin_unlock(&ctl->tree_lock);
1015 static noinline_for_stack
1016 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1017 struct btrfs_free_space_ctl *ctl,
1018 struct btrfs_block_group *block_group,
1019 int *entries, int *bitmaps,
1020 struct list_head *bitmap_list)
1023 struct btrfs_free_cluster *cluster = NULL;
1024 struct btrfs_free_cluster *cluster_locked = NULL;
1025 struct rb_node *node = rb_first(&ctl->free_space_offset);
1026 struct btrfs_trim_range *trim_entry;
1028 /* Get the cluster for this block_group if it exists */
1029 if (block_group && !list_empty(&block_group->cluster_list)) {
1030 cluster = list_entry(block_group->cluster_list.next,
1031 struct btrfs_free_cluster,
1035 if (!node && cluster) {
1036 cluster_locked = cluster;
1037 spin_lock(&cluster_locked->lock);
1038 node = rb_first(&cluster->root);
1042 /* Write out the extent entries */
1044 struct btrfs_free_space *e;
1046 e = rb_entry(node, struct btrfs_free_space, offset_index);
1049 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1055 list_add_tail(&e->list, bitmap_list);
1058 node = rb_next(node);
1059 if (!node && cluster) {
1060 node = rb_first(&cluster->root);
1061 cluster_locked = cluster;
1062 spin_lock(&cluster_locked->lock);
1066 if (cluster_locked) {
1067 spin_unlock(&cluster_locked->lock);
1068 cluster_locked = NULL;
1072 * Make sure we don't miss any range that was removed from our rbtree
1073 * because trimming is running. Otherwise after a umount+mount (or crash
1074 * after committing the transaction) we would leak free space and get
1075 * an inconsistent free space cache report from fsck.
1077 list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1078 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1079 trim_entry->bytes, NULL);
1088 spin_unlock(&cluster_locked->lock);
1092 static noinline_for_stack int
1093 update_cache_item(struct btrfs_trans_handle *trans,
1094 struct btrfs_root *root,
1095 struct inode *inode,
1096 struct btrfs_path *path, u64 offset,
1097 int entries, int bitmaps)
1099 struct btrfs_key key;
1100 struct btrfs_free_space_header *header;
1101 struct extent_buffer *leaf;
1104 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1105 key.offset = offset;
1108 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1110 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1111 EXTENT_DELALLOC, 0, 0, NULL);
1114 leaf = path->nodes[0];
1116 struct btrfs_key found_key;
1117 ASSERT(path->slots[0]);
1119 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1120 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1121 found_key.offset != offset) {
1122 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1123 inode->i_size - 1, EXTENT_DELALLOC, 0,
1125 btrfs_release_path(path);
1130 BTRFS_I(inode)->generation = trans->transid;
1131 header = btrfs_item_ptr(leaf, path->slots[0],
1132 struct btrfs_free_space_header);
1133 btrfs_set_free_space_entries(leaf, header, entries);
1134 btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1135 btrfs_set_free_space_generation(leaf, header, trans->transid);
1136 btrfs_mark_buffer_dirty(leaf);
1137 btrfs_release_path(path);
1145 static noinline_for_stack int write_pinned_extent_entries(
1146 struct btrfs_trans_handle *trans,
1147 struct btrfs_block_group *block_group,
1148 struct btrfs_io_ctl *io_ctl,
1151 u64 start, extent_start, extent_end, len;
1152 struct extent_io_tree *unpin = NULL;
1159 * We want to add any pinned extents to our free space cache
1160 * so we don't leak the space
1162 * We shouldn't have switched the pinned extents yet so this is the
1165 unpin = &trans->transaction->pinned_extents;
1167 start = block_group->start;
1169 while (start < block_group->start + block_group->length) {
1170 ret = find_first_extent_bit(unpin, start,
1171 &extent_start, &extent_end,
1172 EXTENT_DIRTY, NULL);
1176 /* This pinned extent is out of our range */
1177 if (extent_start >= block_group->start + block_group->length)
1180 extent_start = max(extent_start, start);
1181 extent_end = min(block_group->start + block_group->length,
1183 len = extent_end - extent_start;
1186 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1196 static noinline_for_stack int
1197 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1199 struct btrfs_free_space *entry, *next;
1202 /* Write out the bitmaps */
1203 list_for_each_entry_safe(entry, next, bitmap_list, list) {
1204 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1207 list_del_init(&entry->list);
1213 static int flush_dirty_cache(struct inode *inode)
1217 ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1219 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1220 EXTENT_DELALLOC, 0, 0, NULL);
1225 static void noinline_for_stack
1226 cleanup_bitmap_list(struct list_head *bitmap_list)
1228 struct btrfs_free_space *entry, *next;
1230 list_for_each_entry_safe(entry, next, bitmap_list, list)
1231 list_del_init(&entry->list);
1234 static void noinline_for_stack
1235 cleanup_write_cache_enospc(struct inode *inode,
1236 struct btrfs_io_ctl *io_ctl,
1237 struct extent_state **cached_state)
1239 io_ctl_drop_pages(io_ctl);
1240 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1241 i_size_read(inode) - 1, cached_state);
1244 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1245 struct btrfs_trans_handle *trans,
1246 struct btrfs_block_group *block_group,
1247 struct btrfs_io_ctl *io_ctl,
1248 struct btrfs_path *path, u64 offset)
1251 struct inode *inode = io_ctl->inode;
1256 /* Flush the dirty pages in the cache file. */
1257 ret = flush_dirty_cache(inode);
1261 /* Update the cache item to tell everyone this cache file is valid. */
1262 ret = update_cache_item(trans, root, inode, path, offset,
1263 io_ctl->entries, io_ctl->bitmaps);
1266 invalidate_inode_pages2(inode->i_mapping);
1267 BTRFS_I(inode)->generation = 0;
1269 btrfs_debug(root->fs_info,
1270 "failed to write free space cache for block group %llu error %d",
1271 block_group->start, ret);
1273 btrfs_update_inode(trans, root, BTRFS_I(inode));
1276 /* the dirty list is protected by the dirty_bgs_lock */
1277 spin_lock(&trans->transaction->dirty_bgs_lock);
1279 /* the disk_cache_state is protected by the block group lock */
1280 spin_lock(&block_group->lock);
1283 * only mark this as written if we didn't get put back on
1284 * the dirty list while waiting for IO. Otherwise our
1285 * cache state won't be right, and we won't get written again
1287 if (!ret && list_empty(&block_group->dirty_list))
1288 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1290 block_group->disk_cache_state = BTRFS_DC_ERROR;
1292 spin_unlock(&block_group->lock);
1293 spin_unlock(&trans->transaction->dirty_bgs_lock);
1294 io_ctl->inode = NULL;
1302 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1303 struct btrfs_block_group *block_group,
1304 struct btrfs_path *path)
1306 return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1307 block_group, &block_group->io_ctl,
1308 path, block_group->start);
1312 * Write out cached info to an inode
1314 * @root: root the inode belongs to
1315 * @inode: freespace inode we are writing out
1316 * @ctl: free space cache we are going to write out
1317 * @block_group: block_group for this cache if it belongs to a block_group
1318 * @io_ctl: holds context for the io
1319 * @trans: the trans handle
1321 * This function writes out a free space cache struct to disk for quick recovery
1322 * on mount. This will return 0 if it was successful in writing the cache out,
1323 * or an errno if it was not.
1325 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1326 struct btrfs_free_space_ctl *ctl,
1327 struct btrfs_block_group *block_group,
1328 struct btrfs_io_ctl *io_ctl,
1329 struct btrfs_trans_handle *trans)
1331 struct extent_state *cached_state = NULL;
1332 LIST_HEAD(bitmap_list);
1338 if (!i_size_read(inode))
1341 WARN_ON(io_ctl->pages);
1342 ret = io_ctl_init(io_ctl, inode, 1);
1346 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1347 down_write(&block_group->data_rwsem);
1348 spin_lock(&block_group->lock);
1349 if (block_group->delalloc_bytes) {
1350 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1351 spin_unlock(&block_group->lock);
1352 up_write(&block_group->data_rwsem);
1353 BTRFS_I(inode)->generation = 0;
1358 spin_unlock(&block_group->lock);
1361 /* Lock all pages first so we can lock the extent safely. */
1362 ret = io_ctl_prepare_pages(io_ctl, false);
1366 lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1369 io_ctl_set_generation(io_ctl, trans->transid);
1371 mutex_lock(&ctl->cache_writeout_mutex);
1372 /* Write out the extent entries in the free space cache */
1373 spin_lock(&ctl->tree_lock);
1374 ret = write_cache_extent_entries(io_ctl, ctl,
1375 block_group, &entries, &bitmaps,
1378 goto out_nospc_locked;
1381 * Some spaces that are freed in the current transaction are pinned,
1382 * they will be added into free space cache after the transaction is
1383 * committed, we shouldn't lose them.
1385 * If this changes while we are working we'll get added back to
1386 * the dirty list and redo it. No locking needed
1388 ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1390 goto out_nospc_locked;
1393 * At last, we write out all the bitmaps and keep cache_writeout_mutex
1394 * locked while doing it because a concurrent trim can be manipulating
1395 * or freeing the bitmap.
1397 ret = write_bitmap_entries(io_ctl, &bitmap_list);
1398 spin_unlock(&ctl->tree_lock);
1399 mutex_unlock(&ctl->cache_writeout_mutex);
1403 /* Zero out the rest of the pages just to make sure */
1404 io_ctl_zero_remaining_pages(io_ctl);
1406 /* Everything is written out, now we dirty the pages in the file. */
1407 ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1408 io_ctl->num_pages, 0, i_size_read(inode),
1409 &cached_state, false);
1413 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1414 up_write(&block_group->data_rwsem);
1416 * Release the pages and unlock the extent, we will flush
1419 io_ctl_drop_pages(io_ctl);
1420 io_ctl_free(io_ctl);
1422 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1423 i_size_read(inode) - 1, &cached_state);
1426 * at this point the pages are under IO and we're happy,
1427 * The caller is responsible for waiting on them and updating
1428 * the cache and the inode
1430 io_ctl->entries = entries;
1431 io_ctl->bitmaps = bitmaps;
1433 ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1440 cleanup_bitmap_list(&bitmap_list);
1441 spin_unlock(&ctl->tree_lock);
1442 mutex_unlock(&ctl->cache_writeout_mutex);
1445 cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1448 if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1449 up_write(&block_group->data_rwsem);
1452 io_ctl->inode = NULL;
1453 io_ctl_free(io_ctl);
1455 invalidate_inode_pages2(inode->i_mapping);
1456 BTRFS_I(inode)->generation = 0;
1458 btrfs_update_inode(trans, root, BTRFS_I(inode));
1464 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1465 struct btrfs_block_group *block_group,
1466 struct btrfs_path *path)
1468 struct btrfs_fs_info *fs_info = trans->fs_info;
1469 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1470 struct inode *inode;
1473 spin_lock(&block_group->lock);
1474 if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1475 spin_unlock(&block_group->lock);
1478 spin_unlock(&block_group->lock);
1480 inode = lookup_free_space_inode(block_group, path);
1484 ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1485 block_group, &block_group->io_ctl, trans);
1487 btrfs_debug(fs_info,
1488 "failed to write free space cache for block group %llu error %d",
1489 block_group->start, ret);
1490 spin_lock(&block_group->lock);
1491 block_group->disk_cache_state = BTRFS_DC_ERROR;
1492 spin_unlock(&block_group->lock);
1494 block_group->io_ctl.inode = NULL;
1499 * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1500 * to wait for IO and put the inode
1506 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1509 ASSERT(offset >= bitmap_start);
1510 offset -= bitmap_start;
1511 return (unsigned long)(div_u64(offset, unit));
1514 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1516 return (unsigned long)(div_u64(bytes, unit));
1519 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1523 u64 bytes_per_bitmap;
1525 bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1526 bitmap_start = offset - ctl->start;
1527 bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1528 bitmap_start *= bytes_per_bitmap;
1529 bitmap_start += ctl->start;
1531 return bitmap_start;
1534 static int tree_insert_offset(struct rb_root *root, u64 offset,
1535 struct rb_node *node, int bitmap)
1537 struct rb_node **p = &root->rb_node;
1538 struct rb_node *parent = NULL;
1539 struct btrfs_free_space *info;
1543 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1545 if (offset < info->offset) {
1547 } else if (offset > info->offset) {
1548 p = &(*p)->rb_right;
1551 * we could have a bitmap entry and an extent entry
1552 * share the same offset. If this is the case, we want
1553 * the extent entry to always be found first if we do a
1554 * linear search through the tree, since we want to have
1555 * the quickest allocation time, and allocating from an
1556 * extent is faster than allocating from a bitmap. So
1557 * if we're inserting a bitmap and we find an entry at
1558 * this offset, we want to go right, or after this entry
1559 * logically. If we are inserting an extent and we've
1560 * found a bitmap, we want to go left, or before
1568 p = &(*p)->rb_right;
1570 if (!info->bitmap) {
1579 rb_link_node(node, parent, p);
1580 rb_insert_color(node, root);
1586 * searches the tree for the given offset.
1588 * fuzzy - If this is set, then we are trying to make an allocation, and we just
1589 * want a section that has at least bytes size and comes at or after the given
1592 static struct btrfs_free_space *
1593 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1594 u64 offset, int bitmap_only, int fuzzy)
1596 struct rb_node *n = ctl->free_space_offset.rb_node;
1597 struct btrfs_free_space *entry, *prev = NULL;
1599 /* find entry that is closest to the 'offset' */
1606 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1609 if (offset < entry->offset)
1611 else if (offset > entry->offset)
1624 * bitmap entry and extent entry may share same offset,
1625 * in that case, bitmap entry comes after extent entry.
1630 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1631 if (entry->offset != offset)
1634 WARN_ON(!entry->bitmap);
1637 if (entry->bitmap) {
1639 * if previous extent entry covers the offset,
1640 * we should return it instead of the bitmap entry
1642 n = rb_prev(&entry->offset_index);
1644 prev = rb_entry(n, struct btrfs_free_space,
1646 if (!prev->bitmap &&
1647 prev->offset + prev->bytes > offset)
1657 /* find last entry before the 'offset' */
1659 if (entry->offset > offset) {
1660 n = rb_prev(&entry->offset_index);
1662 entry = rb_entry(n, struct btrfs_free_space,
1664 ASSERT(entry->offset <= offset);
1673 if (entry->bitmap) {
1674 n = rb_prev(&entry->offset_index);
1676 prev = rb_entry(n, struct btrfs_free_space,
1678 if (!prev->bitmap &&
1679 prev->offset + prev->bytes > offset)
1682 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1684 } else if (entry->offset + entry->bytes > offset)
1691 if (entry->bitmap) {
1692 if (entry->offset + BITS_PER_BITMAP *
1696 if (entry->offset + entry->bytes > offset)
1700 n = rb_next(&entry->offset_index);
1703 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1709 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1710 struct btrfs_free_space *info)
1712 rb_erase(&info->offset_index, &ctl->free_space_offset);
1713 ctl->free_extents--;
1715 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1716 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1717 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1721 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1722 struct btrfs_free_space *info)
1724 __unlink_free_space(ctl, info);
1725 ctl->free_space -= info->bytes;
1728 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1729 struct btrfs_free_space *info)
1733 ASSERT(info->bytes || info->bitmap);
1734 ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1735 &info->offset_index, (info->bitmap != NULL));
1739 if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1740 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1741 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1744 ctl->free_space += info->bytes;
1745 ctl->free_extents++;
1749 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1750 struct btrfs_free_space *info,
1751 u64 offset, u64 bytes)
1753 unsigned long start, count, end;
1754 int extent_delta = -1;
1756 start = offset_to_bit(info->offset, ctl->unit, offset);
1757 count = bytes_to_bits(bytes, ctl->unit);
1758 end = start + count;
1759 ASSERT(end <= BITS_PER_BITMAP);
1761 bitmap_clear(info->bitmap, start, count);
1763 info->bytes -= bytes;
1764 if (info->max_extent_size > ctl->unit)
1765 info->max_extent_size = 0;
1767 if (start && test_bit(start - 1, info->bitmap))
1770 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1773 info->bitmap_extents += extent_delta;
1774 if (!btrfs_free_space_trimmed(info)) {
1775 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1776 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1780 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1781 struct btrfs_free_space *info, u64 offset,
1784 __bitmap_clear_bits(ctl, info, offset, bytes);
1785 ctl->free_space -= bytes;
1788 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1789 struct btrfs_free_space *info, u64 offset,
1792 unsigned long start, count, end;
1793 int extent_delta = 1;
1795 start = offset_to_bit(info->offset, ctl->unit, offset);
1796 count = bytes_to_bits(bytes, ctl->unit);
1797 end = start + count;
1798 ASSERT(end <= BITS_PER_BITMAP);
1800 bitmap_set(info->bitmap, start, count);
1802 info->bytes += bytes;
1803 ctl->free_space += bytes;
1805 if (start && test_bit(start - 1, info->bitmap))
1808 if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1811 info->bitmap_extents += extent_delta;
1812 if (!btrfs_free_space_trimmed(info)) {
1813 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1814 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1819 * If we can not find suitable extent, we will use bytes to record
1820 * the size of the max extent.
1822 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1823 struct btrfs_free_space *bitmap_info, u64 *offset,
1824 u64 *bytes, bool for_alloc)
1826 unsigned long found_bits = 0;
1827 unsigned long max_bits = 0;
1828 unsigned long bits, i;
1829 unsigned long next_zero;
1830 unsigned long extent_bits;
1833 * Skip searching the bitmap if we don't have a contiguous section that
1834 * is large enough for this allocation.
1837 bitmap_info->max_extent_size &&
1838 bitmap_info->max_extent_size < *bytes) {
1839 *bytes = bitmap_info->max_extent_size;
1843 i = offset_to_bit(bitmap_info->offset, ctl->unit,
1844 max_t(u64, *offset, bitmap_info->offset));
1845 bits = bytes_to_bits(*bytes, ctl->unit);
1847 for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1848 if (for_alloc && bits == 1) {
1852 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1853 BITS_PER_BITMAP, i);
1854 extent_bits = next_zero - i;
1855 if (extent_bits >= bits) {
1856 found_bits = extent_bits;
1858 } else if (extent_bits > max_bits) {
1859 max_bits = extent_bits;
1865 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1866 *bytes = (u64)(found_bits) * ctl->unit;
1870 *bytes = (u64)(max_bits) * ctl->unit;
1871 bitmap_info->max_extent_size = *bytes;
1875 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1878 return entry->max_extent_size;
1879 return entry->bytes;
1882 /* Cache the size of the max extent in bytes */
1883 static struct btrfs_free_space *
1884 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1885 unsigned long align, u64 *max_extent_size)
1887 struct btrfs_free_space *entry;
1888 struct rb_node *node;
1893 if (!ctl->free_space_offset.rb_node)
1896 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1900 for (node = &entry->offset_index; node; node = rb_next(node)) {
1901 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1902 if (entry->bytes < *bytes) {
1903 *max_extent_size = max(get_max_extent_size(entry),
1908 /* make sure the space returned is big enough
1909 * to match our requested alignment
1911 if (*bytes >= align) {
1912 tmp = entry->offset - ctl->start + align - 1;
1913 tmp = div64_u64(tmp, align);
1914 tmp = tmp * align + ctl->start;
1915 align_off = tmp - entry->offset;
1918 tmp = entry->offset;
1921 if (entry->bytes < *bytes + align_off) {
1922 *max_extent_size = max(get_max_extent_size(entry),
1927 if (entry->bitmap) {
1930 ret = search_bitmap(ctl, entry, &tmp, &size, true);
1937 max(get_max_extent_size(entry),
1944 *bytes = entry->bytes - align_off;
1951 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1952 struct btrfs_free_space *info, u64 offset)
1954 info->offset = offset_to_bitmap(ctl, offset);
1956 info->bitmap_extents = 0;
1957 INIT_LIST_HEAD(&info->list);
1958 link_free_space(ctl, info);
1959 ctl->total_bitmaps++;
1960 recalculate_thresholds(ctl);
1963 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1964 struct btrfs_free_space *bitmap_info)
1967 * Normally when this is called, the bitmap is completely empty. However,
1968 * if we are blowing up the free space cache for one reason or another
1969 * via __btrfs_remove_free_space_cache(), then it may not be freed and
1970 * we may leave stats on the table.
1972 if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
1973 ctl->discardable_extents[BTRFS_STAT_CURR] -=
1974 bitmap_info->bitmap_extents;
1975 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
1978 unlink_free_space(ctl, bitmap_info);
1979 kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1980 kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1981 ctl->total_bitmaps--;
1982 recalculate_thresholds(ctl);
1985 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1986 struct btrfs_free_space *bitmap_info,
1987 u64 *offset, u64 *bytes)
1990 u64 search_start, search_bytes;
1994 end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1997 * We need to search for bits in this bitmap. We could only cover some
1998 * of the extent in this bitmap thanks to how we add space, so we need
1999 * to search for as much as it as we can and clear that amount, and then
2000 * go searching for the next bit.
2002 search_start = *offset;
2003 search_bytes = ctl->unit;
2004 search_bytes = min(search_bytes, end - search_start + 1);
2005 ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2007 if (ret < 0 || search_start != *offset)
2010 /* We may have found more bits than what we need */
2011 search_bytes = min(search_bytes, *bytes);
2013 /* Cannot clear past the end of the bitmap */
2014 search_bytes = min(search_bytes, end - search_start + 1);
2016 bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
2017 *offset += search_bytes;
2018 *bytes -= search_bytes;
2021 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2022 if (!bitmap_info->bytes)
2023 free_bitmap(ctl, bitmap_info);
2026 * no entry after this bitmap, but we still have bytes to
2027 * remove, so something has gone wrong.
2032 bitmap_info = rb_entry(next, struct btrfs_free_space,
2036 * if the next entry isn't a bitmap we need to return to let the
2037 * extent stuff do its work.
2039 if (!bitmap_info->bitmap)
2043 * Ok the next item is a bitmap, but it may not actually hold
2044 * the information for the rest of this free space stuff, so
2045 * look for it, and if we don't find it return so we can try
2046 * everything over again.
2048 search_start = *offset;
2049 search_bytes = ctl->unit;
2050 ret = search_bitmap(ctl, bitmap_info, &search_start,
2051 &search_bytes, false);
2052 if (ret < 0 || search_start != *offset)
2056 } else if (!bitmap_info->bytes)
2057 free_bitmap(ctl, bitmap_info);
2062 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2063 struct btrfs_free_space *info, u64 offset,
2064 u64 bytes, enum btrfs_trim_state trim_state)
2066 u64 bytes_to_set = 0;
2070 * This is a tradeoff to make bitmap trim state minimal. We mark the
2071 * whole bitmap untrimmed if at any point we add untrimmed regions.
2073 if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2074 if (btrfs_free_space_trimmed(info)) {
2075 ctl->discardable_extents[BTRFS_STAT_CURR] +=
2076 info->bitmap_extents;
2077 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2079 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2082 end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2084 bytes_to_set = min(end - offset, bytes);
2086 bitmap_set_bits(ctl, info, offset, bytes_to_set);
2089 * We set some bytes, we have no idea what the max extent size is
2092 info->max_extent_size = 0;
2094 return bytes_to_set;
2098 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2099 struct btrfs_free_space *info)
2101 struct btrfs_block_group *block_group = ctl->private;
2102 struct btrfs_fs_info *fs_info = block_group->fs_info;
2103 bool forced = false;
2105 #ifdef CONFIG_BTRFS_DEBUG
2106 if (btrfs_should_fragment_free_space(block_group))
2110 /* This is a way to reclaim large regions from the bitmaps. */
2111 if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2115 * If we are below the extents threshold then we can add this as an
2116 * extent, and don't have to deal with the bitmap
2118 if (!forced && ctl->free_extents < ctl->extents_thresh) {
2120 * If this block group has some small extents we don't want to
2121 * use up all of our free slots in the cache with them, we want
2122 * to reserve them to larger extents, however if we have plenty
2123 * of cache left then go ahead an dadd them, no sense in adding
2124 * the overhead of a bitmap if we don't have to.
2126 if (info->bytes <= fs_info->sectorsize * 8) {
2127 if (ctl->free_extents * 3 <= ctl->extents_thresh)
2135 * The original block groups from mkfs can be really small, like 8
2136 * megabytes, so don't bother with a bitmap for those entries. However
2137 * some block groups can be smaller than what a bitmap would cover but
2138 * are still large enough that they could overflow the 32k memory limit,
2139 * so allow those block groups to still be allowed to have a bitmap
2142 if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2148 static const struct btrfs_free_space_op free_space_op = {
2149 .use_bitmap = use_bitmap,
2152 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2153 struct btrfs_free_space *info)
2155 struct btrfs_free_space *bitmap_info;
2156 struct btrfs_block_group *block_group = NULL;
2158 u64 bytes, offset, bytes_added;
2159 enum btrfs_trim_state trim_state;
2162 bytes = info->bytes;
2163 offset = info->offset;
2164 trim_state = info->trim_state;
2166 if (!ctl->op->use_bitmap(ctl, info))
2169 if (ctl->op == &free_space_op)
2170 block_group = ctl->private;
2173 * Since we link bitmaps right into the cluster we need to see if we
2174 * have a cluster here, and if so and it has our bitmap we need to add
2175 * the free space to that bitmap.
2177 if (block_group && !list_empty(&block_group->cluster_list)) {
2178 struct btrfs_free_cluster *cluster;
2179 struct rb_node *node;
2180 struct btrfs_free_space *entry;
2182 cluster = list_entry(block_group->cluster_list.next,
2183 struct btrfs_free_cluster,
2185 spin_lock(&cluster->lock);
2186 node = rb_first(&cluster->root);
2188 spin_unlock(&cluster->lock);
2189 goto no_cluster_bitmap;
2192 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2193 if (!entry->bitmap) {
2194 spin_unlock(&cluster->lock);
2195 goto no_cluster_bitmap;
2198 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2199 bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2201 bytes -= bytes_added;
2202 offset += bytes_added;
2204 spin_unlock(&cluster->lock);
2212 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2219 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2221 bytes -= bytes_added;
2222 offset += bytes_added;
2232 if (info && info->bitmap) {
2233 add_new_bitmap(ctl, info, offset);
2238 spin_unlock(&ctl->tree_lock);
2240 /* no pre-allocated info, allocate a new one */
2242 info = kmem_cache_zalloc(btrfs_free_space_cachep,
2245 spin_lock(&ctl->tree_lock);
2251 /* allocate the bitmap */
2252 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2254 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2255 spin_lock(&ctl->tree_lock);
2256 if (!info->bitmap) {
2266 kmem_cache_free(btrfs_free_space_bitmap_cachep,
2268 kmem_cache_free(btrfs_free_space_cachep, info);
2275 * Free space merging rules:
2276 * 1) Merge trimmed areas together
2277 * 2) Let untrimmed areas coalesce with trimmed areas
2278 * 3) Always pull neighboring regions from bitmaps
2280 * The above rules are for when we merge free space based on btrfs_trim_state.
2281 * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2282 * same reason: to promote larger extent regions which makes life easier for
2283 * find_free_extent(). Rule 2 enables coalescing based on the common path
2284 * being returning free space from btrfs_finish_extent_commit(). So when free
2285 * space is trimmed, it will prevent aggregating trimmed new region and
2286 * untrimmed regions in the rb_tree. Rule 3 is purely to obtain larger extents
2287 * and provide find_free_extent() with the largest extents possible hoping for
2290 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2291 struct btrfs_free_space *info, bool update_stat)
2293 struct btrfs_free_space *left_info = NULL;
2294 struct btrfs_free_space *right_info;
2295 bool merged = false;
2296 u64 offset = info->offset;
2297 u64 bytes = info->bytes;
2298 const bool is_trimmed = btrfs_free_space_trimmed(info);
2301 * first we want to see if there is free space adjacent to the range we
2302 * are adding, if there is remove that struct and add a new one to
2303 * cover the entire range
2305 right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2306 if (right_info && rb_prev(&right_info->offset_index))
2307 left_info = rb_entry(rb_prev(&right_info->offset_index),
2308 struct btrfs_free_space, offset_index);
2309 else if (!right_info)
2310 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2312 /* See try_merge_free_space() comment. */
2313 if (right_info && !right_info->bitmap &&
2314 (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2316 unlink_free_space(ctl, right_info);
2318 __unlink_free_space(ctl, right_info);
2319 info->bytes += right_info->bytes;
2320 kmem_cache_free(btrfs_free_space_cachep, right_info);
2324 /* See try_merge_free_space() comment. */
2325 if (left_info && !left_info->bitmap &&
2326 left_info->offset + left_info->bytes == offset &&
2327 (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2329 unlink_free_space(ctl, left_info);
2331 __unlink_free_space(ctl, left_info);
2332 info->offset = left_info->offset;
2333 info->bytes += left_info->bytes;
2334 kmem_cache_free(btrfs_free_space_cachep, left_info);
2341 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2342 struct btrfs_free_space *info,
2345 struct btrfs_free_space *bitmap;
2348 const u64 end = info->offset + info->bytes;
2349 const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2352 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2356 i = offset_to_bit(bitmap->offset, ctl->unit, end);
2357 j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2360 bytes = (j - i) * ctl->unit;
2361 info->bytes += bytes;
2363 /* See try_merge_free_space() comment. */
2364 if (!btrfs_free_space_trimmed(bitmap))
2365 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2368 bitmap_clear_bits(ctl, bitmap, end, bytes);
2370 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2373 free_bitmap(ctl, bitmap);
2378 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2379 struct btrfs_free_space *info,
2382 struct btrfs_free_space *bitmap;
2386 unsigned long prev_j;
2389 bitmap_offset = offset_to_bitmap(ctl, info->offset);
2390 /* If we're on a boundary, try the previous logical bitmap. */
2391 if (bitmap_offset == info->offset) {
2392 if (info->offset == 0)
2394 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2397 bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2401 i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2403 prev_j = (unsigned long)-1;
2404 for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2412 if (prev_j == (unsigned long)-1)
2413 bytes = (i + 1) * ctl->unit;
2415 bytes = (i - prev_j) * ctl->unit;
2417 info->offset -= bytes;
2418 info->bytes += bytes;
2420 /* See try_merge_free_space() comment. */
2421 if (!btrfs_free_space_trimmed(bitmap))
2422 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2425 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2427 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2430 free_bitmap(ctl, bitmap);
2436 * We prefer always to allocate from extent entries, both for clustered and
2437 * non-clustered allocation requests. So when attempting to add a new extent
2438 * entry, try to see if there's adjacent free space in bitmap entries, and if
2439 * there is, migrate that space from the bitmaps to the extent.
2440 * Like this we get better chances of satisfying space allocation requests
2441 * because we attempt to satisfy them based on a single cache entry, and never
2442 * on 2 or more entries - even if the entries represent a contiguous free space
2443 * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2446 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2447 struct btrfs_free_space *info,
2451 * Only work with disconnected entries, as we can change their offset,
2452 * and must be extent entries.
2454 ASSERT(!info->bitmap);
2455 ASSERT(RB_EMPTY_NODE(&info->offset_index));
2457 if (ctl->total_bitmaps > 0) {
2459 bool stole_front = false;
2461 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2462 if (ctl->total_bitmaps > 0)
2463 stole_front = steal_from_bitmap_to_front(ctl, info,
2466 if (stole_end || stole_front)
2467 try_merge_free_space(ctl, info, update_stat);
2471 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2472 struct btrfs_free_space_ctl *ctl,
2473 u64 offset, u64 bytes,
2474 enum btrfs_trim_state trim_state)
2476 struct btrfs_block_group *block_group = ctl->private;
2477 struct btrfs_free_space *info;
2479 u64 filter_bytes = bytes;
2481 ASSERT(!btrfs_is_zoned(fs_info));
2483 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2487 info->offset = offset;
2488 info->bytes = bytes;
2489 info->trim_state = trim_state;
2490 RB_CLEAR_NODE(&info->offset_index);
2492 spin_lock(&ctl->tree_lock);
2494 if (try_merge_free_space(ctl, info, true))
2498 * There was no extent directly to the left or right of this new
2499 * extent then we know we're going to have to allocate a new extent, so
2500 * before we do that see if we need to drop this into a bitmap
2502 ret = insert_into_bitmap(ctl, info);
2511 * Only steal free space from adjacent bitmaps if we're sure we're not
2512 * going to add the new free space to existing bitmap entries - because
2513 * that would mean unnecessary work that would be reverted. Therefore
2514 * attempt to steal space from bitmaps if we're adding an extent entry.
2516 steal_from_bitmap(ctl, info, true);
2518 filter_bytes = max(filter_bytes, info->bytes);
2520 ret = link_free_space(ctl, info);
2522 kmem_cache_free(btrfs_free_space_cachep, info);
2524 btrfs_discard_update_discardable(block_group);
2525 spin_unlock(&ctl->tree_lock);
2528 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2529 ASSERT(ret != -EEXIST);
2532 if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2533 btrfs_discard_check_filter(block_group, filter_bytes);
2534 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2540 static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2541 u64 bytenr, u64 size, bool used)
2543 struct btrfs_fs_info *fs_info = block_group->fs_info;
2544 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2545 u64 offset = bytenr - block_group->start;
2546 u64 to_free, to_unusable;
2548 spin_lock(&ctl->tree_lock);
2551 else if (offset >= block_group->alloc_offset)
2553 else if (offset + size <= block_group->alloc_offset)
2556 to_free = offset + size - block_group->alloc_offset;
2557 to_unusable = size - to_free;
2559 ctl->free_space += to_free;
2561 * If the block group is read-only, we should account freed space into
2564 if (!block_group->ro)
2565 block_group->zone_unusable += to_unusable;
2566 spin_unlock(&ctl->tree_lock);
2568 spin_lock(&block_group->lock);
2569 block_group->alloc_offset -= size;
2570 spin_unlock(&block_group->lock);
2573 /* All the region is now unusable. Mark it as unused and reclaim */
2574 if (block_group->zone_unusable == block_group->length) {
2575 btrfs_mark_bg_unused(block_group);
2576 } else if (block_group->zone_unusable >=
2577 div_factor_fine(block_group->length,
2578 fs_info->bg_reclaim_threshold)) {
2579 btrfs_mark_bg_to_reclaim(block_group);
2585 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2586 u64 bytenr, u64 size)
2588 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2590 if (btrfs_is_zoned(block_group->fs_info))
2591 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2594 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2595 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2597 return __btrfs_add_free_space(block_group->fs_info,
2598 block_group->free_space_ctl,
2599 bytenr, size, trim_state);
2602 int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2603 u64 bytenr, u64 size)
2605 if (btrfs_is_zoned(block_group->fs_info))
2606 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2609 return btrfs_add_free_space(block_group, bytenr, size);
2613 * This is a subtle distinction because when adding free space back in general,
2614 * we want it to be added as untrimmed for async. But in the case where we add
2615 * it on loading of a block group, we want to consider it trimmed.
2617 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2618 u64 bytenr, u64 size)
2620 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2622 if (btrfs_is_zoned(block_group->fs_info))
2623 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2626 if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2627 btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2628 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2630 return __btrfs_add_free_space(block_group->fs_info,
2631 block_group->free_space_ctl,
2632 bytenr, size, trim_state);
2635 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2636 u64 offset, u64 bytes)
2638 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2639 struct btrfs_free_space *info;
2641 bool re_search = false;
2643 if (btrfs_is_zoned(block_group->fs_info)) {
2645 * This can happen with conventional zones when replaying log.
2646 * Since the allocation info of tree-log nodes are not recorded
2647 * to the extent-tree, calculate_alloc_pointer() failed to
2648 * advance the allocation pointer after last allocated tree log
2651 * This function is called from
2652 * btrfs_pin_extent_for_log_replay() when replaying the log.
2653 * Advance the pointer not to overwrite the tree-log nodes.
2655 if (block_group->alloc_offset < offset + bytes)
2656 block_group->alloc_offset = offset + bytes;
2660 spin_lock(&ctl->tree_lock);
2667 info = tree_search_offset(ctl, offset, 0, 0);
2670 * oops didn't find an extent that matched the space we wanted
2671 * to remove, look for a bitmap instead
2673 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2677 * If we found a partial bit of our free space in a
2678 * bitmap but then couldn't find the other part this may
2679 * be a problem, so WARN about it.
2687 if (!info->bitmap) {
2688 unlink_free_space(ctl, info);
2689 if (offset == info->offset) {
2690 u64 to_free = min(bytes, info->bytes);
2692 info->bytes -= to_free;
2693 info->offset += to_free;
2695 ret = link_free_space(ctl, info);
2698 kmem_cache_free(btrfs_free_space_cachep, info);
2705 u64 old_end = info->bytes + info->offset;
2707 info->bytes = offset - info->offset;
2708 ret = link_free_space(ctl, info);
2713 /* Not enough bytes in this entry to satisfy us */
2714 if (old_end < offset + bytes) {
2715 bytes -= old_end - offset;
2718 } else if (old_end == offset + bytes) {
2722 spin_unlock(&ctl->tree_lock);
2724 ret = __btrfs_add_free_space(block_group->fs_info, ctl,
2726 old_end - (offset + bytes),
2733 ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2734 if (ret == -EAGAIN) {
2739 btrfs_discard_update_discardable(block_group);
2740 spin_unlock(&ctl->tree_lock);
2745 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2748 struct btrfs_fs_info *fs_info = block_group->fs_info;
2749 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2750 struct btrfs_free_space *info;
2755 * Zoned btrfs does not use free space tree and cluster. Just print
2756 * out the free space after the allocation offset.
2758 if (btrfs_is_zoned(fs_info)) {
2759 btrfs_info(fs_info, "free space %llu",
2760 block_group->length - block_group->alloc_offset);
2764 spin_lock(&ctl->tree_lock);
2765 for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2766 info = rb_entry(n, struct btrfs_free_space, offset_index);
2767 if (info->bytes >= bytes && !block_group->ro)
2769 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2770 info->offset, info->bytes,
2771 (info->bitmap) ? "yes" : "no");
2773 spin_unlock(&ctl->tree_lock);
2774 btrfs_info(fs_info, "block group has cluster?: %s",
2775 list_empty(&block_group->cluster_list) ? "no" : "yes");
2777 "%d blocks of free space at or bigger than bytes is", count);
2780 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2781 struct btrfs_free_space_ctl *ctl)
2783 struct btrfs_fs_info *fs_info = block_group->fs_info;
2785 spin_lock_init(&ctl->tree_lock);
2786 ctl->unit = fs_info->sectorsize;
2787 ctl->start = block_group->start;
2788 ctl->private = block_group;
2789 ctl->op = &free_space_op;
2790 INIT_LIST_HEAD(&ctl->trimming_ranges);
2791 mutex_init(&ctl->cache_writeout_mutex);
2794 * we only want to have 32k of ram per block group for keeping
2795 * track of free space, and if we pass 1/2 of that we want to
2796 * start converting things over to using bitmaps
2798 ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2802 * for a given cluster, put all of its extents back into the free
2803 * space cache. If the block group passed doesn't match the block group
2804 * pointed to by the cluster, someone else raced in and freed the
2805 * cluster already. In that case, we just return without changing anything
2807 static void __btrfs_return_cluster_to_free_space(
2808 struct btrfs_block_group *block_group,
2809 struct btrfs_free_cluster *cluster)
2811 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2812 struct btrfs_free_space *entry;
2813 struct rb_node *node;
2815 spin_lock(&cluster->lock);
2816 if (cluster->block_group != block_group) {
2817 spin_unlock(&cluster->lock);
2821 cluster->block_group = NULL;
2822 cluster->window_start = 0;
2823 list_del_init(&cluster->block_group_list);
2825 node = rb_first(&cluster->root);
2829 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2830 node = rb_next(&entry->offset_index);
2831 rb_erase(&entry->offset_index, &cluster->root);
2832 RB_CLEAR_NODE(&entry->offset_index);
2834 bitmap = (entry->bitmap != NULL);
2836 /* Merging treats extents as if they were new */
2837 if (!btrfs_free_space_trimmed(entry)) {
2838 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2839 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2843 try_merge_free_space(ctl, entry, false);
2844 steal_from_bitmap(ctl, entry, false);
2846 /* As we insert directly, update these statistics */
2847 if (!btrfs_free_space_trimmed(entry)) {
2848 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2849 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2853 tree_insert_offset(&ctl->free_space_offset,
2854 entry->offset, &entry->offset_index, bitmap);
2856 cluster->root = RB_ROOT;
2857 spin_unlock(&cluster->lock);
2858 btrfs_put_block_group(block_group);
2861 static void __btrfs_remove_free_space_cache_locked(
2862 struct btrfs_free_space_ctl *ctl)
2864 struct btrfs_free_space *info;
2865 struct rb_node *node;
2867 while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2868 info = rb_entry(node, struct btrfs_free_space, offset_index);
2869 if (!info->bitmap) {
2870 unlink_free_space(ctl, info);
2871 kmem_cache_free(btrfs_free_space_cachep, info);
2873 free_bitmap(ctl, info);
2876 cond_resched_lock(&ctl->tree_lock);
2880 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2882 spin_lock(&ctl->tree_lock);
2883 __btrfs_remove_free_space_cache_locked(ctl);
2885 btrfs_discard_update_discardable(ctl->private);
2886 spin_unlock(&ctl->tree_lock);
2889 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2891 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2892 struct btrfs_free_cluster *cluster;
2893 struct list_head *head;
2895 spin_lock(&ctl->tree_lock);
2896 while ((head = block_group->cluster_list.next) !=
2897 &block_group->cluster_list) {
2898 cluster = list_entry(head, struct btrfs_free_cluster,
2901 WARN_ON(cluster->block_group != block_group);
2902 __btrfs_return_cluster_to_free_space(block_group, cluster);
2904 cond_resched_lock(&ctl->tree_lock);
2906 __btrfs_remove_free_space_cache_locked(ctl);
2907 btrfs_discard_update_discardable(block_group);
2908 spin_unlock(&ctl->tree_lock);
2913 * btrfs_is_free_space_trimmed - see if everything is trimmed
2914 * @block_group: block_group of interest
2916 * Walk @block_group's free space rb_tree to determine if everything is trimmed.
2918 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
2920 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2921 struct btrfs_free_space *info;
2922 struct rb_node *node;
2925 spin_lock(&ctl->tree_lock);
2926 node = rb_first(&ctl->free_space_offset);
2929 info = rb_entry(node, struct btrfs_free_space, offset_index);
2931 if (!btrfs_free_space_trimmed(info)) {
2936 node = rb_next(node);
2939 spin_unlock(&ctl->tree_lock);
2943 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2944 u64 offset, u64 bytes, u64 empty_size,
2945 u64 *max_extent_size)
2947 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2948 struct btrfs_discard_ctl *discard_ctl =
2949 &block_group->fs_info->discard_ctl;
2950 struct btrfs_free_space *entry = NULL;
2951 u64 bytes_search = bytes + empty_size;
2954 u64 align_gap_len = 0;
2955 enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2957 ASSERT(!btrfs_is_zoned(block_group->fs_info));
2959 spin_lock(&ctl->tree_lock);
2960 entry = find_free_space(ctl, &offset, &bytes_search,
2961 block_group->full_stripe_len, max_extent_size);
2966 if (entry->bitmap) {
2967 bitmap_clear_bits(ctl, entry, offset, bytes);
2969 if (!btrfs_free_space_trimmed(entry))
2970 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2973 free_bitmap(ctl, entry);
2975 unlink_free_space(ctl, entry);
2976 align_gap_len = offset - entry->offset;
2977 align_gap = entry->offset;
2978 align_gap_trim_state = entry->trim_state;
2980 if (!btrfs_free_space_trimmed(entry))
2981 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2983 entry->offset = offset + bytes;
2984 WARN_ON(entry->bytes < bytes + align_gap_len);
2986 entry->bytes -= bytes + align_gap_len;
2988 kmem_cache_free(btrfs_free_space_cachep, entry);
2990 link_free_space(ctl, entry);
2993 btrfs_discard_update_discardable(block_group);
2994 spin_unlock(&ctl->tree_lock);
2997 __btrfs_add_free_space(block_group->fs_info, ctl,
2998 align_gap, align_gap_len,
2999 align_gap_trim_state);
3004 * given a cluster, put all of its extents back into the free space
3005 * cache. If a block group is passed, this function will only free
3006 * a cluster that belongs to the passed block group.
3008 * Otherwise, it'll get a reference on the block group pointed to by the
3009 * cluster and remove the cluster from it.
3011 void btrfs_return_cluster_to_free_space(
3012 struct btrfs_block_group *block_group,
3013 struct btrfs_free_cluster *cluster)
3015 struct btrfs_free_space_ctl *ctl;
3017 /* first, get a safe pointer to the block group */
3018 spin_lock(&cluster->lock);
3020 block_group = cluster->block_group;
3022 spin_unlock(&cluster->lock);
3025 } else if (cluster->block_group != block_group) {
3026 /* someone else has already freed it don't redo their work */
3027 spin_unlock(&cluster->lock);
3030 btrfs_get_block_group(block_group);
3031 spin_unlock(&cluster->lock);
3033 ctl = block_group->free_space_ctl;
3035 /* now return any extents the cluster had on it */
3036 spin_lock(&ctl->tree_lock);
3037 __btrfs_return_cluster_to_free_space(block_group, cluster);
3038 spin_unlock(&ctl->tree_lock);
3040 btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3042 /* finally drop our ref */
3043 btrfs_put_block_group(block_group);
3046 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3047 struct btrfs_free_cluster *cluster,
3048 struct btrfs_free_space *entry,
3049 u64 bytes, u64 min_start,
3050 u64 *max_extent_size)
3052 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3054 u64 search_start = cluster->window_start;
3055 u64 search_bytes = bytes;
3058 search_start = min_start;
3059 search_bytes = bytes;
3061 err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3063 *max_extent_size = max(get_max_extent_size(entry),
3069 __bitmap_clear_bits(ctl, entry, ret, bytes);
3075 * given a cluster, try to allocate 'bytes' from it, returns 0
3076 * if it couldn't find anything suitably large, or a logical disk offset
3077 * if things worked out
3079 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3080 struct btrfs_free_cluster *cluster, u64 bytes,
3081 u64 min_start, u64 *max_extent_size)
3083 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3084 struct btrfs_discard_ctl *discard_ctl =
3085 &block_group->fs_info->discard_ctl;
3086 struct btrfs_free_space *entry = NULL;
3087 struct rb_node *node;
3090 ASSERT(!btrfs_is_zoned(block_group->fs_info));
3092 spin_lock(&cluster->lock);
3093 if (bytes > cluster->max_size)
3096 if (cluster->block_group != block_group)
3099 node = rb_first(&cluster->root);
3103 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3105 if (entry->bytes < bytes)
3106 *max_extent_size = max(get_max_extent_size(entry),
3109 if (entry->bytes < bytes ||
3110 (!entry->bitmap && entry->offset < min_start)) {
3111 node = rb_next(&entry->offset_index);
3114 entry = rb_entry(node, struct btrfs_free_space,
3119 if (entry->bitmap) {
3120 ret = btrfs_alloc_from_bitmap(block_group,
3121 cluster, entry, bytes,
3122 cluster->window_start,
3125 node = rb_next(&entry->offset_index);
3128 entry = rb_entry(node, struct btrfs_free_space,
3132 cluster->window_start += bytes;
3134 ret = entry->offset;
3136 entry->offset += bytes;
3137 entry->bytes -= bytes;
3143 spin_unlock(&cluster->lock);
3148 spin_lock(&ctl->tree_lock);
3150 if (!btrfs_free_space_trimmed(entry))
3151 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3153 ctl->free_space -= bytes;
3154 if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3155 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3157 spin_lock(&cluster->lock);
3158 if (entry->bytes == 0) {
3159 rb_erase(&entry->offset_index, &cluster->root);
3160 ctl->free_extents--;
3161 if (entry->bitmap) {
3162 kmem_cache_free(btrfs_free_space_bitmap_cachep,
3164 ctl->total_bitmaps--;
3165 recalculate_thresholds(ctl);
3166 } else if (!btrfs_free_space_trimmed(entry)) {
3167 ctl->discardable_extents[BTRFS_STAT_CURR]--;
3169 kmem_cache_free(btrfs_free_space_cachep, entry);
3172 spin_unlock(&cluster->lock);
3173 spin_unlock(&ctl->tree_lock);
3178 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3179 struct btrfs_free_space *entry,
3180 struct btrfs_free_cluster *cluster,
3181 u64 offset, u64 bytes,
3182 u64 cont1_bytes, u64 min_bytes)
3184 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3185 unsigned long next_zero;
3187 unsigned long want_bits;
3188 unsigned long min_bits;
3189 unsigned long found_bits;
3190 unsigned long max_bits = 0;
3191 unsigned long start = 0;
3192 unsigned long total_found = 0;
3195 i = offset_to_bit(entry->offset, ctl->unit,
3196 max_t(u64, offset, entry->offset));
3197 want_bits = bytes_to_bits(bytes, ctl->unit);
3198 min_bits = bytes_to_bits(min_bytes, ctl->unit);
3201 * Don't bother looking for a cluster in this bitmap if it's heavily
3204 if (entry->max_extent_size &&
3205 entry->max_extent_size < cont1_bytes)
3209 for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3210 next_zero = find_next_zero_bit(entry->bitmap,
3211 BITS_PER_BITMAP, i);
3212 if (next_zero - i >= min_bits) {
3213 found_bits = next_zero - i;
3214 if (found_bits > max_bits)
3215 max_bits = found_bits;
3218 if (next_zero - i > max_bits)
3219 max_bits = next_zero - i;
3224 entry->max_extent_size = (u64)max_bits * ctl->unit;
3230 cluster->max_size = 0;
3233 total_found += found_bits;
3235 if (cluster->max_size < found_bits * ctl->unit)
3236 cluster->max_size = found_bits * ctl->unit;
3238 if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3243 cluster->window_start = start * ctl->unit + entry->offset;
3244 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3245 ret = tree_insert_offset(&cluster->root, entry->offset,
3246 &entry->offset_index, 1);
3247 ASSERT(!ret); /* -EEXIST; Logic error */
3249 trace_btrfs_setup_cluster(block_group, cluster,
3250 total_found * ctl->unit, 1);
3255 * This searches the block group for just extents to fill the cluster with.
3256 * Try to find a cluster with at least bytes total bytes, at least one
3257 * extent of cont1_bytes, and other clusters of at least min_bytes.
3260 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3261 struct btrfs_free_cluster *cluster,
3262 struct list_head *bitmaps, u64 offset, u64 bytes,
3263 u64 cont1_bytes, u64 min_bytes)
3265 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3266 struct btrfs_free_space *first = NULL;
3267 struct btrfs_free_space *entry = NULL;
3268 struct btrfs_free_space *last;
3269 struct rb_node *node;
3274 entry = tree_search_offset(ctl, offset, 0, 1);
3279 * We don't want bitmaps, so just move along until we find a normal
3282 while (entry->bitmap || entry->bytes < min_bytes) {
3283 if (entry->bitmap && list_empty(&entry->list))
3284 list_add_tail(&entry->list, bitmaps);
3285 node = rb_next(&entry->offset_index);
3288 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3291 window_free = entry->bytes;
3292 max_extent = entry->bytes;
3296 for (node = rb_next(&entry->offset_index); node;
3297 node = rb_next(&entry->offset_index)) {
3298 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3300 if (entry->bitmap) {
3301 if (list_empty(&entry->list))
3302 list_add_tail(&entry->list, bitmaps);
3306 if (entry->bytes < min_bytes)
3310 window_free += entry->bytes;
3311 if (entry->bytes > max_extent)
3312 max_extent = entry->bytes;
3315 if (window_free < bytes || max_extent < cont1_bytes)
3318 cluster->window_start = first->offset;
3320 node = &first->offset_index;
3323 * now we've found our entries, pull them out of the free space
3324 * cache and put them into the cluster rbtree
3329 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3330 node = rb_next(&entry->offset_index);
3331 if (entry->bitmap || entry->bytes < min_bytes)
3334 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3335 ret = tree_insert_offset(&cluster->root, entry->offset,
3336 &entry->offset_index, 0);
3337 total_size += entry->bytes;
3338 ASSERT(!ret); /* -EEXIST; Logic error */
3339 } while (node && entry != last);
3341 cluster->max_size = max_extent;
3342 trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3347 * This specifically looks for bitmaps that may work in the cluster, we assume
3348 * that we have already failed to find extents that will work.
3351 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3352 struct btrfs_free_cluster *cluster,
3353 struct list_head *bitmaps, u64 offset, u64 bytes,
3354 u64 cont1_bytes, u64 min_bytes)
3356 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3357 struct btrfs_free_space *entry = NULL;
3359 u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3361 if (ctl->total_bitmaps == 0)
3365 * The bitmap that covers offset won't be in the list unless offset
3366 * is just its start offset.
3368 if (!list_empty(bitmaps))
3369 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3371 if (!entry || entry->offset != bitmap_offset) {
3372 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3373 if (entry && list_empty(&entry->list))
3374 list_add(&entry->list, bitmaps);
3377 list_for_each_entry(entry, bitmaps, list) {
3378 if (entry->bytes < bytes)
3380 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3381 bytes, cont1_bytes, min_bytes);
3387 * The bitmaps list has all the bitmaps that record free space
3388 * starting after offset, so no more search is required.
3394 * here we try to find a cluster of blocks in a block group. The goal
3395 * is to find at least bytes+empty_size.
3396 * We might not find them all in one contiguous area.
3398 * returns zero and sets up cluster if things worked out, otherwise
3399 * it returns -enospc
3401 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3402 struct btrfs_free_cluster *cluster,
3403 u64 offset, u64 bytes, u64 empty_size)
3405 struct btrfs_fs_info *fs_info = block_group->fs_info;
3406 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3407 struct btrfs_free_space *entry, *tmp;
3414 * Choose the minimum extent size we'll require for this
3415 * cluster. For SSD_SPREAD, don't allow any fragmentation.
3416 * For metadata, allow allocates with smaller extents. For
3417 * data, keep it dense.
3419 if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3420 cont1_bytes = min_bytes = bytes + empty_size;
3421 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3422 cont1_bytes = bytes;
3423 min_bytes = fs_info->sectorsize;
3425 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3426 min_bytes = fs_info->sectorsize;
3429 spin_lock(&ctl->tree_lock);
3432 * If we know we don't have enough space to make a cluster don't even
3433 * bother doing all the work to try and find one.
3435 if (ctl->free_space < bytes) {
3436 spin_unlock(&ctl->tree_lock);
3440 spin_lock(&cluster->lock);
3442 /* someone already found a cluster, hooray */
3443 if (cluster->block_group) {
3448 trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3451 ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3453 cont1_bytes, min_bytes);
3455 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3456 offset, bytes + empty_size,
3457 cont1_bytes, min_bytes);
3459 /* Clear our temporary list */
3460 list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3461 list_del_init(&entry->list);
3464 btrfs_get_block_group(block_group);
3465 list_add_tail(&cluster->block_group_list,
3466 &block_group->cluster_list);
3467 cluster->block_group = block_group;
3469 trace_btrfs_failed_cluster_setup(block_group);
3472 spin_unlock(&cluster->lock);
3473 spin_unlock(&ctl->tree_lock);
3479 * simple code to zero out a cluster
3481 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3483 spin_lock_init(&cluster->lock);
3484 spin_lock_init(&cluster->refill_lock);
3485 cluster->root = RB_ROOT;
3486 cluster->max_size = 0;
3487 cluster->fragmented = false;
3488 INIT_LIST_HEAD(&cluster->block_group_list);
3489 cluster->block_group = NULL;
3492 static int do_trimming(struct btrfs_block_group *block_group,
3493 u64 *total_trimmed, u64 start, u64 bytes,
3494 u64 reserved_start, u64 reserved_bytes,
3495 enum btrfs_trim_state reserved_trim_state,
3496 struct btrfs_trim_range *trim_entry)
3498 struct btrfs_space_info *space_info = block_group->space_info;
3499 struct btrfs_fs_info *fs_info = block_group->fs_info;
3500 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3503 const u64 end = start + bytes;
3504 const u64 reserved_end = reserved_start + reserved_bytes;
3505 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3508 spin_lock(&space_info->lock);
3509 spin_lock(&block_group->lock);
3510 if (!block_group->ro) {
3511 block_group->reserved += reserved_bytes;
3512 space_info->bytes_reserved += reserved_bytes;
3515 spin_unlock(&block_group->lock);
3516 spin_unlock(&space_info->lock);
3518 ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3520 *total_trimmed += trimmed;
3521 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3524 mutex_lock(&ctl->cache_writeout_mutex);
3525 if (reserved_start < start)
3526 __btrfs_add_free_space(fs_info, ctl, reserved_start,
3527 start - reserved_start,
3528 reserved_trim_state);
3529 if (start + bytes < reserved_start + reserved_bytes)
3530 __btrfs_add_free_space(fs_info, ctl, end, reserved_end - end,
3531 reserved_trim_state);
3532 __btrfs_add_free_space(fs_info, ctl, start, bytes, trim_state);
3533 list_del(&trim_entry->list);
3534 mutex_unlock(&ctl->cache_writeout_mutex);
3537 spin_lock(&space_info->lock);
3538 spin_lock(&block_group->lock);
3539 if (block_group->ro)
3540 space_info->bytes_readonly += reserved_bytes;
3541 block_group->reserved -= reserved_bytes;
3542 space_info->bytes_reserved -= reserved_bytes;
3543 spin_unlock(&block_group->lock);
3544 spin_unlock(&space_info->lock);
3551 * If @async is set, then we will trim 1 region and return.
3553 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3554 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3557 struct btrfs_discard_ctl *discard_ctl =
3558 &block_group->fs_info->discard_ctl;
3559 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3560 struct btrfs_free_space *entry;
3561 struct rb_node *node;
3565 enum btrfs_trim_state extent_trim_state;
3567 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3569 while (start < end) {
3570 struct btrfs_trim_range trim_entry;
3572 mutex_lock(&ctl->cache_writeout_mutex);
3573 spin_lock(&ctl->tree_lock);
3575 if (ctl->free_space < minlen)
3578 entry = tree_search_offset(ctl, start, 0, 1);
3582 /* Skip bitmaps and if async, already trimmed entries */
3583 while (entry->bitmap ||
3584 (async && btrfs_free_space_trimmed(entry))) {
3585 node = rb_next(&entry->offset_index);
3588 entry = rb_entry(node, struct btrfs_free_space,
3592 if (entry->offset >= end)
3595 extent_start = entry->offset;
3596 extent_bytes = entry->bytes;
3597 extent_trim_state = entry->trim_state;
3599 start = entry->offset;
3600 bytes = entry->bytes;
3601 if (bytes < minlen) {
3602 spin_unlock(&ctl->tree_lock);
3603 mutex_unlock(&ctl->cache_writeout_mutex);
3606 unlink_free_space(ctl, entry);
3608 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3609 * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3610 * X when we come back around. So trim it now.
3612 if (max_discard_size &&
3613 bytes >= (max_discard_size +
3614 BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3615 bytes = max_discard_size;
3616 extent_bytes = max_discard_size;
3617 entry->offset += max_discard_size;
3618 entry->bytes -= max_discard_size;
3619 link_free_space(ctl, entry);
3621 kmem_cache_free(btrfs_free_space_cachep, entry);
3624 start = max(start, extent_start);
3625 bytes = min(extent_start + extent_bytes, end) - start;
3626 if (bytes < minlen) {
3627 spin_unlock(&ctl->tree_lock);
3628 mutex_unlock(&ctl->cache_writeout_mutex);
3632 unlink_free_space(ctl, entry);
3633 kmem_cache_free(btrfs_free_space_cachep, entry);
3636 spin_unlock(&ctl->tree_lock);
3637 trim_entry.start = extent_start;
3638 trim_entry.bytes = extent_bytes;
3639 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3640 mutex_unlock(&ctl->cache_writeout_mutex);
3642 ret = do_trimming(block_group, total_trimmed, start, bytes,
3643 extent_start, extent_bytes, extent_trim_state,
3646 block_group->discard_cursor = start + bytes;
3651 block_group->discard_cursor = start;
3652 if (async && *total_trimmed)
3655 if (fatal_signal_pending(current)) {
3666 block_group->discard_cursor = btrfs_block_group_end(block_group);
3667 spin_unlock(&ctl->tree_lock);
3668 mutex_unlock(&ctl->cache_writeout_mutex);
3674 * If we break out of trimming a bitmap prematurely, we should reset the
3675 * trimming bit. In a rather contrieved case, it's possible to race here so
3676 * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3678 * start = start of bitmap
3679 * end = near end of bitmap
3681 * Thread 1: Thread 2:
3682 * trim_bitmaps(start)
3684 * end_trimming_bitmap()
3685 * reset_trimming_bitmap()
3687 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3689 struct btrfs_free_space *entry;
3691 spin_lock(&ctl->tree_lock);
3692 entry = tree_search_offset(ctl, offset, 1, 0);
3694 if (btrfs_free_space_trimmed(entry)) {
3695 ctl->discardable_extents[BTRFS_STAT_CURR] +=
3696 entry->bitmap_extents;
3697 ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3699 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3702 spin_unlock(&ctl->tree_lock);
3705 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3706 struct btrfs_free_space *entry)
3708 if (btrfs_free_space_trimming_bitmap(entry)) {
3709 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3710 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3711 entry->bitmap_extents;
3712 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3717 * If @async is set, then we will trim 1 region and return.
3719 static int trim_bitmaps(struct btrfs_block_group *block_group,
3720 u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3721 u64 maxlen, bool async)
3723 struct btrfs_discard_ctl *discard_ctl =
3724 &block_group->fs_info->discard_ctl;
3725 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3726 struct btrfs_free_space *entry;
3730 u64 offset = offset_to_bitmap(ctl, start);
3731 const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3733 while (offset < end) {
3734 bool next_bitmap = false;
3735 struct btrfs_trim_range trim_entry;
3737 mutex_lock(&ctl->cache_writeout_mutex);
3738 spin_lock(&ctl->tree_lock);
3740 if (ctl->free_space < minlen) {
3741 block_group->discard_cursor =
3742 btrfs_block_group_end(block_group);
3743 spin_unlock(&ctl->tree_lock);
3744 mutex_unlock(&ctl->cache_writeout_mutex);
3748 entry = tree_search_offset(ctl, offset, 1, 0);
3750 * Bitmaps are marked trimmed lossily now to prevent constant
3751 * discarding of the same bitmap (the reason why we are bound
3752 * by the filters). So, retrim the block group bitmaps when we
3753 * are preparing to punt to the unused_bgs list. This uses
3754 * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3755 * which is the only discard index which sets minlen to 0.
3757 if (!entry || (async && minlen && start == offset &&
3758 btrfs_free_space_trimmed(entry))) {
3759 spin_unlock(&ctl->tree_lock);
3760 mutex_unlock(&ctl->cache_writeout_mutex);
3766 * Async discard bitmap trimming begins at by setting the start
3767 * to be key.objectid and the offset_to_bitmap() aligns to the
3768 * start of the bitmap. This lets us know we are fully
3769 * scanning the bitmap rather than only some portion of it.
3771 if (start == offset)
3772 entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3775 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3776 if (ret2 || start >= end) {
3778 * We lossily consider a bitmap trimmed if we only skip
3779 * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3781 if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3782 end_trimming_bitmap(ctl, entry);
3784 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3785 spin_unlock(&ctl->tree_lock);
3786 mutex_unlock(&ctl->cache_writeout_mutex);
3792 * We already trimmed a region, but are using the locking above
3793 * to reset the trim_state.
3795 if (async && *total_trimmed) {
3796 spin_unlock(&ctl->tree_lock);
3797 mutex_unlock(&ctl->cache_writeout_mutex);
3801 bytes = min(bytes, end - start);
3802 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3803 spin_unlock(&ctl->tree_lock);
3804 mutex_unlock(&ctl->cache_writeout_mutex);
3809 * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3810 * If X < @minlen, we won't trim X when we come back around.
3811 * So trim it now. We differ here from trimming extents as we
3812 * don't keep individual state per bit.
3816 bytes > (max_discard_size + minlen))
3817 bytes = max_discard_size;
3819 bitmap_clear_bits(ctl, entry, start, bytes);
3820 if (entry->bytes == 0)
3821 free_bitmap(ctl, entry);
3823 spin_unlock(&ctl->tree_lock);
3824 trim_entry.start = start;
3825 trim_entry.bytes = bytes;
3826 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3827 mutex_unlock(&ctl->cache_writeout_mutex);
3829 ret = do_trimming(block_group, total_trimmed, start, bytes,
3830 start, bytes, 0, &trim_entry);
3832 reset_trimming_bitmap(ctl, offset);
3833 block_group->discard_cursor =
3834 btrfs_block_group_end(block_group);
3839 offset += BITS_PER_BITMAP * ctl->unit;
3844 block_group->discard_cursor = start;
3846 if (fatal_signal_pending(current)) {
3847 if (start != offset)
3848 reset_trimming_bitmap(ctl, offset);
3857 block_group->discard_cursor = end;
3863 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3864 u64 *trimmed, u64 start, u64 end, u64 minlen)
3866 struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3870 ASSERT(!btrfs_is_zoned(block_group->fs_info));
3874 spin_lock(&block_group->lock);
3875 if (block_group->removed) {
3876 spin_unlock(&block_group->lock);
3879 btrfs_freeze_block_group(block_group);
3880 spin_unlock(&block_group->lock);
3882 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
3886 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
3887 div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
3888 /* If we ended in the middle of a bitmap, reset the trimming flag */
3890 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
3892 btrfs_unfreeze_block_group(block_group);
3896 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
3897 u64 *trimmed, u64 start, u64 end, u64 minlen,
3904 spin_lock(&block_group->lock);
3905 if (block_group->removed) {
3906 spin_unlock(&block_group->lock);
3909 btrfs_freeze_block_group(block_group);
3910 spin_unlock(&block_group->lock);
3912 ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
3913 btrfs_unfreeze_block_group(block_group);
3918 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
3919 u64 *trimmed, u64 start, u64 end, u64 minlen,
3920 u64 maxlen, bool async)
3926 spin_lock(&block_group->lock);
3927 if (block_group->removed) {
3928 spin_unlock(&block_group->lock);
3931 btrfs_freeze_block_group(block_group);
3932 spin_unlock(&block_group->lock);
3934 ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
3937 btrfs_unfreeze_block_group(block_group);
3942 bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
3944 return btrfs_super_cache_generation(fs_info->super_copy);
3947 static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
3948 struct btrfs_trans_handle *trans)
3950 struct btrfs_block_group *block_group;
3951 struct rb_node *node;
3954 btrfs_info(fs_info, "cleaning free space cache v1");
3956 node = rb_first(&fs_info->block_group_cache_tree);
3958 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
3959 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
3962 node = rb_next(node);
3968 int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
3970 struct btrfs_trans_handle *trans;
3974 * update_super_roots will appropriately set or unset
3975 * super_copy->cache_generation based on SPACE_CACHE and
3976 * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
3977 * transaction commit whether we are enabling space cache v1 and don't
3978 * have any other work to do, or are disabling it and removing free
3981 trans = btrfs_start_transaction(fs_info->tree_root, 0);
3983 return PTR_ERR(trans);
3986 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
3987 ret = cleanup_free_space_cache_v1(fs_info, trans);
3989 btrfs_abort_transaction(trans, ret);
3990 btrfs_end_transaction(trans);
3995 ret = btrfs_commit_transaction(trans);
3997 clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4002 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4004 * Use this if you need to make a bitmap or extent entry specifically, it
4005 * doesn't do any of the merging that add_free_space does, this acts a lot like
4006 * how the free space cache loading stuff works, so you can get really weird
4009 int test_add_free_space_entry(struct btrfs_block_group *cache,
4010 u64 offset, u64 bytes, bool bitmap)
4012 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4013 struct btrfs_free_space *info = NULL, *bitmap_info;
4015 enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4021 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4027 spin_lock(&ctl->tree_lock);
4028 info->offset = offset;
4029 info->bytes = bytes;
4030 info->max_extent_size = 0;
4031 ret = link_free_space(ctl, info);
4032 spin_unlock(&ctl->tree_lock);
4034 kmem_cache_free(btrfs_free_space_cachep, info);
4039 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4041 kmem_cache_free(btrfs_free_space_cachep, info);
4046 spin_lock(&ctl->tree_lock);
4047 bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4052 add_new_bitmap(ctl, info, offset);
4057 bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4060 bytes -= bytes_added;
4061 offset += bytes_added;
4062 spin_unlock(&ctl->tree_lock);
4068 kmem_cache_free(btrfs_free_space_cachep, info);
4070 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4075 * Checks to see if the given range is in the free space cache. This is really
4076 * just used to check the absence of space, so if there is free space in the
4077 * range at all we will return 1.
4079 int test_check_exists(struct btrfs_block_group *cache,
4080 u64 offset, u64 bytes)
4082 struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4083 struct btrfs_free_space *info;
4086 spin_lock(&ctl->tree_lock);
4087 info = tree_search_offset(ctl, offset, 0, 0);
4089 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4097 u64 bit_off, bit_bytes;
4099 struct btrfs_free_space *tmp;
4102 bit_bytes = ctl->unit;
4103 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4105 if (bit_off == offset) {
4108 } else if (bit_off > offset &&
4109 offset + bytes > bit_off) {
4115 n = rb_prev(&info->offset_index);
4117 tmp = rb_entry(n, struct btrfs_free_space,
4119 if (tmp->offset + tmp->bytes < offset)
4121 if (offset + bytes < tmp->offset) {
4122 n = rb_prev(&tmp->offset_index);
4129 n = rb_next(&info->offset_index);
4131 tmp = rb_entry(n, struct btrfs_free_space,
4133 if (offset + bytes < tmp->offset)
4135 if (tmp->offset + tmp->bytes < offset) {
4136 n = rb_next(&tmp->offset_index);
4147 if (info->offset == offset) {
4152 if (offset > info->offset && offset < info->offset + info->bytes)
4155 spin_unlock(&ctl->tree_lock);
4158 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */