2 * Copyright (C) 2008 Red Hat. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
19 #include <linux/sched.h>
21 #include "free-space-cache.h"
22 #include "transaction.h"
24 struct btrfs_free_space {
25 struct rb_node bytes_index;
26 struct rb_node offset_index;
31 static int tree_insert_offset(struct rb_root *root, u64 offset,
34 struct rb_node **p = &root->rb_node;
35 struct rb_node *parent = NULL;
36 struct btrfs_free_space *info;
40 info = rb_entry(parent, struct btrfs_free_space, offset_index);
42 if (offset < info->offset)
44 else if (offset > info->offset)
50 rb_link_node(node, parent, p);
51 rb_insert_color(node, root);
56 static int tree_insert_bytes(struct rb_root *root, u64 bytes,
59 struct rb_node **p = &root->rb_node;
60 struct rb_node *parent = NULL;
61 struct btrfs_free_space *info;
65 info = rb_entry(parent, struct btrfs_free_space, bytes_index);
67 if (bytes < info->bytes)
73 rb_link_node(node, parent, p);
74 rb_insert_color(node, root);
80 * searches the tree for the given offset.
82 * fuzzy == 1: this is used for allocations where we are given a hint of where
83 * to look for free space. Because the hint may not be completely on an offset
84 * mark, or the hint may no longer point to free space we need to fudge our
85 * results a bit. So we look for free space starting at or after offset with at
86 * least bytes size. We prefer to find as close to the given offset as we can.
87 * Also if the offset is within a free space range, then we will return the free
88 * space that contains the given offset, which means we can return a free space
89 * chunk with an offset before the provided offset.
91 * fuzzy == 0: this is just a normal tree search. Give us the free space that
92 * starts at the given offset which is at least bytes size, and if its not there
95 static struct btrfs_free_space *tree_search_offset(struct rb_root *root,
96 u64 offset, u64 bytes,
99 struct rb_node *n = root->rb_node;
100 struct btrfs_free_space *entry, *ret = NULL;
103 entry = rb_entry(n, struct btrfs_free_space, offset_index);
105 if (offset < entry->offset) {
107 (!ret || entry->offset < ret->offset) &&
108 (bytes <= entry->bytes))
111 } else if (offset > entry->offset) {
113 (entry->offset + entry->bytes - 1) >= offset &&
114 bytes <= entry->bytes) {
120 if (bytes > entry->bytes) {
133 * return a chunk at least bytes size, as close to offset that we can get.
135 static struct btrfs_free_space *tree_search_bytes(struct rb_root *root,
136 u64 offset, u64 bytes)
138 struct rb_node *n = root->rb_node;
139 struct btrfs_free_space *entry, *ret = NULL;
142 entry = rb_entry(n, struct btrfs_free_space, bytes_index);
144 if (bytes < entry->bytes) {
146 * We prefer to get a hole size as close to the size we
147 * are asking for so we don't take small slivers out of
148 * huge holes, but we also want to get as close to the
149 * offset as possible so we don't have a whole lot of
152 if (offset <= entry->offset) {
155 else if (entry->bytes < ret->bytes)
157 else if (entry->offset < ret->offset)
161 } else if (bytes > entry->bytes) {
165 * Ok we may have multiple chunks of the wanted size,
166 * so we don't want to take the first one we find, we
167 * want to take the one closest to our given offset, so
168 * keep searching just in case theres a better match.
171 if (offset > entry->offset)
173 else if (!ret || entry->offset < ret->offset)
181 static void unlink_free_space(struct btrfs_block_group_cache *block_group,
182 struct btrfs_free_space *info)
184 rb_erase(&info->offset_index, &block_group->free_space_offset);
185 rb_erase(&info->bytes_index, &block_group->free_space_bytes);
188 static int link_free_space(struct btrfs_block_group_cache *block_group,
189 struct btrfs_free_space *info)
194 BUG_ON(!info->bytes);
195 ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
196 &info->offset_index);
200 ret = tree_insert_bytes(&block_group->free_space_bytes, info->bytes,
208 int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
209 u64 offset, u64 bytes)
211 struct btrfs_free_space *right_info;
212 struct btrfs_free_space *left_info;
213 struct btrfs_free_space *info = NULL;
216 info = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
220 info->offset = offset;
223 spin_lock(&block_group->tree_lock);
226 * first we want to see if there is free space adjacent to the range we
227 * are adding, if there is remove that struct and add a new one to
228 * cover the entire range
230 right_info = tree_search_offset(&block_group->free_space_offset,
232 left_info = tree_search_offset(&block_group->free_space_offset,
236 unlink_free_space(block_group, right_info);
237 info->bytes += right_info->bytes;
241 if (left_info && left_info->offset + left_info->bytes == offset) {
242 unlink_free_space(block_group, left_info);
243 info->offset = left_info->offset;
244 info->bytes += left_info->bytes;
248 ret = link_free_space(block_group, info);
252 spin_unlock(&block_group->tree_lock);
255 printk(KERN_ERR "btrfs: unable to add free space :%d\n", ret);
256 BUG_ON(ret == -EEXIST);
262 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
263 u64 offset, u64 bytes)
265 struct btrfs_free_space *info;
268 spin_lock(&block_group->tree_lock);
270 info = tree_search_offset(&block_group->free_space_offset, offset, 0,
272 if (info && info->offset == offset) {
273 if (info->bytes < bytes) {
274 printk(KERN_ERR "Found free space at %llu, size %llu,"
275 "trying to use %llu\n",
276 (unsigned long long)info->offset,
277 (unsigned long long)info->bytes,
278 (unsigned long long)bytes);
281 spin_unlock(&block_group->tree_lock);
284 unlink_free_space(block_group, info);
286 if (info->bytes == bytes) {
288 spin_unlock(&block_group->tree_lock);
292 info->offset += bytes;
293 info->bytes -= bytes;
295 ret = link_free_space(block_group, info);
296 spin_unlock(&block_group->tree_lock);
298 } else if (info && info->offset < offset &&
299 info->offset + info->bytes >= offset + bytes) {
300 u64 old_start = info->offset;
302 * we're freeing space in the middle of the info,
303 * this can happen during tree log replay
305 * first unlink the old info and then
306 * insert it again after the hole we're creating
308 unlink_free_space(block_group, info);
309 if (offset + bytes < info->offset + info->bytes) {
310 u64 old_end = info->offset + info->bytes;
312 info->offset = offset + bytes;
313 info->bytes = old_end - info->offset;
314 ret = link_free_space(block_group, info);
317 /* the hole we're creating ends at the end
318 * of the info struct, just free the info
322 spin_unlock(&block_group->tree_lock);
323 /* step two, insert a new info struct to cover anything
326 ret = btrfs_add_free_space(block_group, old_start,
330 spin_unlock(&block_group->tree_lock);
332 printk(KERN_ERR "couldn't find space %llu to free\n",
333 (unsigned long long)offset);
334 printk(KERN_ERR "cached is %d, offset %llu bytes %llu\n",
336 (unsigned long long)block_group->key.objectid,
337 (unsigned long long)block_group->key.offset);
338 btrfs_dump_free_space(block_group, bytes);
340 printk(KERN_ERR "hmm, found offset=%llu bytes=%llu, "
341 "but wanted offset=%llu bytes=%llu\n",
342 (unsigned long long)info->offset,
343 (unsigned long long)info->bytes,
344 (unsigned long long)offset,
345 (unsigned long long)bytes);
353 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
356 struct btrfs_free_space *info;
360 for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
361 info = rb_entry(n, struct btrfs_free_space, offset_index);
362 if (info->bytes >= bytes)
364 printk(KERN_ERR "entry offset %llu, bytes %llu\n",
365 (unsigned long long)info->offset,
366 (unsigned long long)info->bytes);
368 printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
372 u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
374 struct btrfs_free_space *info;
378 for (n = rb_first(&block_group->free_space_offset); n;
380 info = rb_entry(n, struct btrfs_free_space, offset_index);
388 * for a given cluster, put all of its extents back into the free
389 * space cache. If the block group passed doesn't match the block group
390 * pointed to by the cluster, someone else raced in and freed the
391 * cluster already. In that case, we just return without changing anything
394 __btrfs_return_cluster_to_free_space(
395 struct btrfs_block_group_cache *block_group,
396 struct btrfs_free_cluster *cluster)
398 struct btrfs_free_space *entry;
399 struct rb_node *node;
401 spin_lock(&cluster->lock);
402 if (cluster->block_group != block_group)
405 cluster->window_start = 0;
406 node = rb_first(&cluster->root);
408 entry = rb_entry(node, struct btrfs_free_space, offset_index);
409 node = rb_next(&entry->offset_index);
410 rb_erase(&entry->offset_index, &cluster->root);
411 link_free_space(block_group, entry);
413 list_del_init(&cluster->block_group_list);
415 btrfs_put_block_group(cluster->block_group);
416 cluster->block_group = NULL;
417 cluster->root.rb_node = NULL;
419 spin_unlock(&cluster->lock);
423 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
425 struct btrfs_free_space *info;
426 struct rb_node *node;
427 struct btrfs_free_cluster *cluster;
428 struct btrfs_free_cluster *safe;
430 spin_lock(&block_group->tree_lock);
432 list_for_each_entry_safe(cluster, safe, &block_group->cluster_list,
435 WARN_ON(cluster->block_group != block_group);
436 __btrfs_return_cluster_to_free_space(block_group, cluster);
439 while ((node = rb_last(&block_group->free_space_bytes)) != NULL) {
440 info = rb_entry(node, struct btrfs_free_space, bytes_index);
441 unlink_free_space(block_group, info);
443 if (need_resched()) {
444 spin_unlock(&block_group->tree_lock);
446 spin_lock(&block_group->tree_lock);
449 spin_unlock(&block_group->tree_lock);
452 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
453 u64 offset, u64 bytes, u64 empty_size)
455 struct btrfs_free_space *entry = NULL;
458 spin_lock(&block_group->tree_lock);
459 entry = tree_search_offset(&block_group->free_space_offset, offset,
460 bytes + empty_size, 1);
462 entry = tree_search_bytes(&block_group->free_space_bytes,
463 offset, bytes + empty_size);
465 unlink_free_space(block_group, entry);
467 entry->offset += bytes;
468 entry->bytes -= bytes;
473 link_free_space(block_group, entry);
475 spin_unlock(&block_group->tree_lock);
481 * given a cluster, put all of its extents back into the free space
482 * cache. If a block group is passed, this function will only free
483 * a cluster that belongs to the passed block group.
485 * Otherwise, it'll get a reference on the block group pointed to by the
486 * cluster and remove the cluster from it.
488 int btrfs_return_cluster_to_free_space(
489 struct btrfs_block_group_cache *block_group,
490 struct btrfs_free_cluster *cluster)
494 /* first, get a safe pointer to the block group */
495 spin_lock(&cluster->lock);
497 block_group = cluster->block_group;
499 spin_unlock(&cluster->lock);
502 } else if (cluster->block_group != block_group) {
503 /* someone else has already freed it don't redo their work */
504 spin_unlock(&cluster->lock);
507 atomic_inc(&block_group->count);
508 spin_unlock(&cluster->lock);
510 /* now return any extents the cluster had on it */
511 spin_lock(&block_group->tree_lock);
512 ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
513 spin_unlock(&block_group->tree_lock);
515 /* finally drop our ref */
516 btrfs_put_block_group(block_group);
521 * given a cluster, try to allocate 'bytes' from it, returns 0
522 * if it couldn't find anything suitably large, or a logical disk offset
523 * if things worked out
525 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
526 struct btrfs_free_cluster *cluster, u64 bytes,
529 struct btrfs_free_space *entry = NULL;
530 struct rb_node *node;
533 spin_lock(&cluster->lock);
534 if (bytes > cluster->max_size)
537 if (cluster->block_group != block_group)
540 node = rb_first(&cluster->root);
544 entry = rb_entry(node, struct btrfs_free_space, offset_index);
547 if (entry->bytes < bytes || entry->offset < min_start) {
548 struct rb_node *node;
550 node = rb_next(&entry->offset_index);
553 entry = rb_entry(node, struct btrfs_free_space,
559 entry->offset += bytes;
560 entry->bytes -= bytes;
562 if (entry->bytes == 0) {
563 rb_erase(&entry->offset_index, &cluster->root);
569 spin_unlock(&cluster->lock);
574 * here we try to find a cluster of blocks in a block group. The goal
575 * is to find at least bytes free and up to empty_size + bytes free.
576 * We might not find them all in one contiguous area.
578 * returns zero and sets up cluster if things worked out, otherwise
581 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
582 struct btrfs_root *root,
583 struct btrfs_block_group_cache *block_group,
584 struct btrfs_free_cluster *cluster,
585 u64 offset, u64 bytes, u64 empty_size)
587 struct btrfs_free_space *entry = NULL;
588 struct rb_node *node;
589 struct btrfs_free_space *next;
590 struct btrfs_free_space *last;
595 int total_retries = 0;
598 /* for metadata, allow allocates with more holes */
599 if (btrfs_test_opt(root, SSD_SPREAD)) {
600 min_bytes = bytes + empty_size;
601 } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
603 * we want to do larger allocations when we are
604 * flushing out the delayed refs, it helps prevent
605 * making more work as we go along.
607 if (trans->transaction->delayed_refs.flushing)
608 min_bytes = max(bytes, (bytes + empty_size) >> 1);
610 min_bytes = max(bytes, (bytes + empty_size) >> 4);
612 min_bytes = max(bytes, (bytes + empty_size) >> 2);
614 spin_lock(&block_group->tree_lock);
615 spin_lock(&cluster->lock);
617 /* someone already found a cluster, hooray */
618 if (cluster->block_group) {
623 min_bytes = min(min_bytes, bytes + empty_size);
624 entry = tree_search_bytes(&block_group->free_space_bytes,
630 window_start = entry->offset;
631 window_free = entry->bytes;
633 max_extent = entry->bytes;
636 /* out window is just right, lets fill it */
637 if (window_free >= bytes + empty_size)
640 node = rb_next(&last->offset_index);
645 next = rb_entry(node, struct btrfs_free_space, offset_index);
648 * we haven't filled the empty size and the window is
649 * very large. reset and try again
651 if (next->offset - (last->offset + last->bytes) > 128 * 1024 ||
652 next->offset - window_start > (bytes + empty_size) * 2) {
654 window_start = entry->offset;
655 window_free = entry->bytes;
659 if (total_retries % 64 == 0) {
660 if (min_bytes >= (bytes + empty_size)) {
665 * grow our allocation a bit, we're not having
673 window_free += next->bytes;
674 if (entry->bytes > max_extent)
675 max_extent = entry->bytes;
679 cluster->window_start = entry->offset;
682 * now we've found our entries, pull them out of the free space
683 * cache and put them into the cluster rbtree
685 * The cluster includes an rbtree, but only uses the offset index
686 * of each free space cache entry.
689 node = rb_next(&entry->offset_index);
690 unlink_free_space(block_group, entry);
691 ret = tree_insert_offset(&cluster->root, entry->offset,
692 &entry->offset_index);
695 if (!node || entry == last)
698 entry = rb_entry(node, struct btrfs_free_space, offset_index);
701 cluster->max_size = max_extent;
702 atomic_inc(&block_group->count);
703 list_add_tail(&cluster->block_group_list, &block_group->cluster_list);
704 cluster->block_group = block_group;
706 spin_unlock(&cluster->lock);
707 spin_unlock(&block_group->tree_lock);
713 * simple code to zero out a cluster
715 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
717 spin_lock_init(&cluster->lock);
718 spin_lock_init(&cluster->refill_lock);
719 cluster->root.rb_node = NULL;
720 cluster->max_size = 0;
721 INIT_LIST_HEAD(&cluster->block_group_list);
722 cluster->block_group = NULL;