1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
6 #include <linux/kthread.h>
7 #include <linux/pagemap.h>
11 #include "free-space-cache.h"
12 #include "inode-map.h"
13 #include "transaction.h"
14 #include "delalloc-space.h"
16 static int caching_kthread(void *data)
18 struct btrfs_root *root = data;
19 struct btrfs_fs_info *fs_info = root->fs_info;
20 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
22 struct btrfs_path *path;
23 struct extent_buffer *leaf;
28 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
31 path = btrfs_alloc_path();
35 /* Since the commit root is read-only, we can safely skip locking. */
36 path->skip_locking = 1;
37 path->search_commit_root = 1;
38 path->reada = READA_FORWARD;
40 key.objectid = BTRFS_FIRST_FREE_OBJECTID;
42 key.type = BTRFS_INODE_ITEM_KEY;
44 /* need to make sure the commit_root doesn't disappear */
45 down_read(&fs_info->commit_root_sem);
47 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
52 if (btrfs_fs_closing(fs_info))
55 leaf = path->nodes[0];
56 slot = path->slots[0];
57 if (slot >= btrfs_header_nritems(leaf)) {
58 ret = btrfs_next_leaf(root, path);
65 btrfs_transaction_in_commit(fs_info)) {
66 leaf = path->nodes[0];
68 if (WARN_ON(btrfs_header_nritems(leaf) == 0))
72 * Save the key so we can advances forward
75 btrfs_item_key_to_cpu(leaf, &key, 0);
76 btrfs_release_path(path);
77 root->ino_cache_progress = last;
78 up_read(&fs_info->commit_root_sem);
85 btrfs_item_key_to_cpu(leaf, &key, slot);
87 if (key.type != BTRFS_INODE_ITEM_KEY)
90 if (key.objectid >= root->highest_objectid)
93 if (last != (u64)-1 && last + 1 != key.objectid) {
94 __btrfs_add_free_space(fs_info, ctl, last + 1,
95 key.objectid - last - 1);
96 wake_up(&root->ino_cache_wait);
104 if (last < root->highest_objectid - 1) {
105 __btrfs_add_free_space(fs_info, ctl, last + 1,
106 root->highest_objectid - last - 1);
109 spin_lock(&root->ino_cache_lock);
110 root->ino_cache_state = BTRFS_CACHE_FINISHED;
111 spin_unlock(&root->ino_cache_lock);
113 root->ino_cache_progress = (u64)-1;
114 btrfs_unpin_free_ino(root);
116 wake_up(&root->ino_cache_wait);
117 up_read(&fs_info->commit_root_sem);
119 btrfs_free_path(path);
124 static void start_caching(struct btrfs_root *root)
126 struct btrfs_fs_info *fs_info = root->fs_info;
127 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
128 struct task_struct *tsk;
132 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
135 spin_lock(&root->ino_cache_lock);
136 if (root->ino_cache_state != BTRFS_CACHE_NO) {
137 spin_unlock(&root->ino_cache_lock);
141 root->ino_cache_state = BTRFS_CACHE_STARTED;
142 spin_unlock(&root->ino_cache_lock);
144 ret = load_free_ino_cache(fs_info, root);
146 spin_lock(&root->ino_cache_lock);
147 root->ino_cache_state = BTRFS_CACHE_FINISHED;
148 spin_unlock(&root->ino_cache_lock);
153 * It can be quite time-consuming to fill the cache by searching
154 * through the extent tree, and this can keep ino allocation path
155 * waiting. Therefore at start we quickly find out the highest
156 * inode number and we know we can use inode numbers which fall in
157 * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
159 ret = btrfs_find_free_objectid(root, &objectid);
160 if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
161 __btrfs_add_free_space(fs_info, ctl, objectid,
162 BTRFS_LAST_FREE_OBJECTID - objectid + 1);
165 tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
166 root->root_key.objectid);
168 btrfs_warn(fs_info, "failed to start inode caching task");
169 btrfs_clear_pending_and_info(fs_info, INODE_MAP_CACHE,
170 "disabling inode map caching");
174 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
176 if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
177 return btrfs_find_free_objectid(root, objectid);
180 *objectid = btrfs_find_ino_for_alloc(root);
187 wait_event(root->ino_cache_wait,
188 root->ino_cache_state == BTRFS_CACHE_FINISHED ||
189 root->free_ino_ctl->free_space > 0);
191 if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
192 root->free_ino_ctl->free_space == 0)
198 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
200 struct btrfs_fs_info *fs_info = root->fs_info;
201 struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
203 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
206 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
207 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
209 down_write(&fs_info->commit_root_sem);
210 spin_lock(&root->ino_cache_lock);
211 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
212 spin_unlock(&root->ino_cache_lock);
213 up_write(&fs_info->commit_root_sem);
216 spin_unlock(&root->ino_cache_lock);
220 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
222 up_write(&fs_info->commit_root_sem);
227 * When a transaction is committed, we'll move those inode numbers which are
228 * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
229 * others will just be dropped, because the commit root we were searching has
232 * Must be called with root->fs_info->commit_root_sem held
234 void btrfs_unpin_free_ino(struct btrfs_root *root)
236 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
237 struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
238 spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
239 struct btrfs_free_space *info;
243 if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
247 spin_lock(rbroot_lock);
248 n = rb_first(rbroot);
250 spin_unlock(rbroot_lock);
254 info = rb_entry(n, struct btrfs_free_space, offset_index);
255 BUG_ON(info->bitmap); /* Logic error */
257 if (info->offset > root->ino_cache_progress)
260 count = min(root->ino_cache_progress - info->offset + 1,
263 rb_erase(&info->offset_index, rbroot);
264 spin_unlock(rbroot_lock);
266 __btrfs_add_free_space(root->fs_info, ctl,
267 info->offset, count);
268 kmem_cache_free(btrfs_free_space_cachep, info);
272 #define INIT_THRESHOLD ((SZ_32K / 2) / sizeof(struct btrfs_free_space))
273 #define INODES_PER_BITMAP (PAGE_SIZE * 8)
276 * The goal is to keep the memory used by the free_ino tree won't
277 * exceed the memory if we use bitmaps only.
279 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
281 struct btrfs_free_space *info;
286 n = rb_last(&ctl->free_space_offset);
288 ctl->extents_thresh = INIT_THRESHOLD;
291 info = rb_entry(n, struct btrfs_free_space, offset_index);
294 * Find the maximum inode number in the filesystem. Note we
295 * ignore the fact that this can be a bitmap, because we are
296 * not doing precise calculation.
298 max_ino = info->bytes - 1;
300 max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
301 if (max_bitmaps <= ctl->total_bitmaps) {
302 ctl->extents_thresh = 0;
306 ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
307 PAGE_SIZE / sizeof(*info);
311 * We don't fall back to bitmap, if we are below the extents threshold
312 * or this chunk of inode numbers is a big one.
314 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
315 struct btrfs_free_space *info)
317 if (ctl->free_extents < ctl->extents_thresh ||
318 info->bytes > INODES_PER_BITMAP / 10)
324 static const struct btrfs_free_space_op free_ino_op = {
325 .recalc_thresholds = recalculate_thresholds,
326 .use_bitmap = use_bitmap,
329 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
333 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
334 struct btrfs_free_space *info)
337 * We always use extents for two reasons:
339 * - The pinned tree is only used during the process of caching
341 * - Make code simpler. See btrfs_unpin_free_ino().
346 static const struct btrfs_free_space_op pinned_free_ino_op = {
347 .recalc_thresholds = pinned_recalc_thresholds,
348 .use_bitmap = pinned_use_bitmap,
351 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
353 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
354 struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
356 spin_lock_init(&ctl->tree_lock);
360 ctl->op = &free_ino_op;
361 INIT_LIST_HEAD(&ctl->trimming_ranges);
362 mutex_init(&ctl->cache_writeout_mutex);
365 * Initially we allow to use 16K of ram to cache chunks of
366 * inode numbers before we resort to bitmaps. This is somewhat
367 * arbitrary, but it will be adjusted in runtime.
369 ctl->extents_thresh = INIT_THRESHOLD;
371 spin_lock_init(&pinned->tree_lock);
374 pinned->private = NULL;
375 pinned->extents_thresh = 0;
376 pinned->op = &pinned_free_ino_op;
379 int btrfs_save_ino_cache(struct btrfs_root *root,
380 struct btrfs_trans_handle *trans)
382 struct btrfs_fs_info *fs_info = root->fs_info;
383 struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
384 struct btrfs_path *path;
386 struct btrfs_block_rsv *rsv;
387 struct extent_changeset *data_reserved = NULL;
394 /* only fs tree and subvol/snap needs ino cache */
395 if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
396 (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
397 root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
400 /* Don't save inode cache if we are deleting this root */
401 if (btrfs_root_refs(&root->root_item) == 0)
404 if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
407 path = btrfs_alloc_path();
411 rsv = trans->block_rsv;
412 trans->block_rsv = &fs_info->trans_block_rsv;
414 num_bytes = trans->bytes_reserved;
416 * 1 item for inode item insertion if need
417 * 4 items for inode item update (in the worst case)
418 * 1 items for slack space if we need do truncation
419 * 1 item for free space object
420 * 3 items for pre-allocation
422 trans->bytes_reserved = btrfs_calc_trans_metadata_size(fs_info, 10);
423 ret = btrfs_block_rsv_add(root, trans->block_rsv,
424 trans->bytes_reserved,
425 BTRFS_RESERVE_NO_FLUSH);
428 trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
429 trans->bytes_reserved, 1);
431 inode = lookup_free_ino_inode(root, path);
432 if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
433 ret = PTR_ERR(inode);
438 BUG_ON(retry); /* Logic error */
441 ret = create_free_ino_inode(root, trans, path);
447 BTRFS_I(inode)->generation = 0;
448 ret = btrfs_update_inode(trans, root, inode);
450 btrfs_abort_transaction(trans, ret);
454 if (i_size_read(inode) > 0) {
455 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
458 btrfs_abort_transaction(trans, ret);
463 spin_lock(&root->ino_cache_lock);
464 if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
466 spin_unlock(&root->ino_cache_lock);
469 spin_unlock(&root->ino_cache_lock);
471 spin_lock(&ctl->tree_lock);
472 prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
473 prealloc = ALIGN(prealloc, PAGE_SIZE);
474 prealloc += ctl->total_bitmaps * PAGE_SIZE;
475 spin_unlock(&ctl->tree_lock);
477 /* Just to make sure we have enough space */
478 prealloc += 8 * PAGE_SIZE;
480 ret = btrfs_delalloc_reserve_space(inode, &data_reserved, 0, prealloc);
484 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
485 prealloc, prealloc, &alloc_hint);
487 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc, true);
491 ret = btrfs_write_out_ino_cache(root, trans, path, inode);
492 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc, false);
496 trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
497 trans->bytes_reserved, 0);
498 btrfs_block_rsv_release(fs_info, trans->block_rsv,
499 trans->bytes_reserved);
501 trans->block_rsv = rsv;
502 trans->bytes_reserved = num_bytes;
504 btrfs_free_path(path);
505 extent_changeset_free(data_reserved);
509 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
511 struct btrfs_path *path;
513 struct extent_buffer *l;
514 struct btrfs_key search_key;
515 struct btrfs_key found_key;
518 path = btrfs_alloc_path();
522 search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
523 search_key.type = -1;
524 search_key.offset = (u64)-1;
525 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
528 BUG_ON(ret == 0); /* Corruption */
529 if (path->slots[0] > 0) {
530 slot = path->slots[0] - 1;
532 btrfs_item_key_to_cpu(l, &found_key, slot);
533 *objectid = max_t(u64, found_key.objectid,
534 BTRFS_FIRST_FREE_OBJECTID - 1);
536 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
540 btrfs_free_path(path);
544 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
547 mutex_lock(&root->objectid_mutex);
549 if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
550 btrfs_warn(root->fs_info,
551 "the objectid of root %llu reaches its highest value",
552 root->root_key.objectid);
557 *objectid = ++root->highest_objectid;
560 mutex_unlock(&root->objectid_mutex);