2 * fs/ext4/extents_status.c
4 * Written by Yongqiang Yang <xiaoqiangnk@gmail.com>
6 * Allison Henderson <achender@linux.vnet.ibm.com>
7 * Hugh Dickins <hughd@google.com>
8 * Zheng Liu <wenqing.lz@taobao.com>
10 * Ext4 extents status tree core functions.
12 #include <linux/rbtree.h>
14 #include "extents_status.h"
15 #include "ext4_extents.h"
17 #include <trace/events/ext4.h>
20 * According to previous discussion in Ext4 Developer Workshop, we
21 * will introduce a new structure called io tree to track all extent
22 * status in order to solve some problems that we have met
23 * (e.g. Reservation space warning), and provide extent-level locking.
24 * Delay extent tree is the first step to achieve this goal. It is
25 * original built by Yongqiang Yang. At that time it is called delay
26 * extent tree, whose goal is only track delayed extents in memory to
27 * simplify the implementation of fiemap and bigalloc, and introduce
28 * lseek SEEK_DATA/SEEK_HOLE support. That is why it is still called
29 * delay extent tree at the first commit. But for better understand
30 * what it does, it has been rename to extent status tree.
33 * Currently the first step has been done. All delayed extents are
34 * tracked in the tree. It maintains the delayed extent when a delayed
35 * allocation is issued, and the delayed extent is written out or
36 * invalidated. Therefore the implementation of fiemap and bigalloc
37 * are simplified, and SEEK_DATA/SEEK_HOLE are introduced.
39 * The following comment describes the implemenmtation of extent
40 * status tree and future works.
43 * In this step all extent status are tracked by extent status tree.
44 * Thus, we can first try to lookup a block mapping in this tree before
45 * finding it in extent tree. Hence, single extent cache can be removed
46 * because extent status tree can do a better job. Extents in status
47 * tree are loaded on-demand. Therefore, the extent status tree may not
48 * contain all of the extents in a file. Meanwhile we define a shrinker
49 * to reclaim memory from extent status tree because fragmented extent
50 * tree will make status tree cost too much memory. written/unwritten/-
51 * hole extents in the tree will be reclaimed by this shrinker when we
52 * are under high memory pressure. Delayed extents will not be
53 * reclimed because fiemap, bigalloc, and seek_data/hole need it.
57 * Extent status tree implementation for ext4.
60 * ==========================================================================
61 * Extent status tree tracks all extent status.
63 * 1. Why we need to implement extent status tree?
65 * Without extent status tree, ext4 identifies a delayed extent by looking
66 * up page cache, this has several deficiencies - complicated, buggy,
67 * and inefficient code.
69 * FIEMAP, SEEK_HOLE/DATA, bigalloc, and writeout all need to know if a
70 * block or a range of blocks are belonged to a delayed extent.
72 * Let us have a look at how they do without extent status tree.
74 * FIEMAP looks up page cache to identify delayed allocations from holes.
77 * SEEK_HOLE/DATA has the same problem as FIEMAP.
80 * bigalloc looks up page cache to figure out if a block is
81 * already under delayed allocation or not to determine whether
82 * quota reserving is needed for the cluster.
85 * Writeout looks up whole page cache to see if a buffer is
86 * mapped, If there are not very many delayed buffers, then it is
89 * With extent status tree implementation, FIEMAP, SEEK_HOLE/DATA,
90 * bigalloc and writeout can figure out if a block or a range of
91 * blocks is under delayed allocation(belonged to a delayed extent) or
92 * not by searching the extent tree.
95 * ==========================================================================
96 * 2. Ext4 extent status tree impelmentation
99 * A extent is a range of blocks which are contiguous logically and
100 * physically. Unlike extent in extent tree, this extent in ext4 is
101 * a in-memory struct, there is no corresponding on-disk data. There
102 * is no limit on length of extent, so an extent can contain as many
103 * blocks as they are contiguous logically and physically.
105 * -- extent status tree
106 * Every inode has an extent status tree and all allocation blocks
107 * are added to the tree with different status. The extent in the
108 * tree are ordered by logical block no.
110 * -- operations on a extent status tree
111 * There are three important operations on a delayed extent tree: find
112 * next extent, adding a extent(a range of blocks) and removing a extent.
114 * -- race on a extent status tree
115 * Extent status tree is protected by inode->i_es_lock.
117 * -- memory consumption
118 * Fragmented extent tree will make extent status tree cost too much
119 * memory. Hence, we will reclaim written/unwritten/hole extents from
120 * the tree under a heavy memory pressure.
123 * ==========================================================================
124 * 3. Performance analysis
127 * 1. There is a cache extent for write access, so if writes are
128 * not very random, adding space operaions are in O(1) time.
131 * 2. Code is much simpler, more readable, more maintainable and
135 * ==========================================================================
138 * -- Refactor delayed space reservation
140 * -- Extent-level locking
143 static struct kmem_cache *ext4_es_cachep;
145 static int __es_insert_extent(struct inode *inode, struct extent_status *newes);
146 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
148 static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
151 int __init ext4_init_es(void)
153 ext4_es_cachep = kmem_cache_create("ext4_extent_status",
154 sizeof(struct extent_status),
155 0, (SLAB_RECLAIM_ACCOUNT), NULL);
156 if (ext4_es_cachep == NULL)
161 void ext4_exit_es(void)
164 kmem_cache_destroy(ext4_es_cachep);
167 void ext4_es_init_tree(struct ext4_es_tree *tree)
169 tree->root = RB_ROOT;
170 tree->cache_es = NULL;
174 static void ext4_es_print_tree(struct inode *inode)
176 struct ext4_es_tree *tree;
177 struct rb_node *node;
179 printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
180 tree = &EXT4_I(inode)->i_es_tree;
181 node = rb_first(&tree->root);
183 struct extent_status *es;
184 es = rb_entry(node, struct extent_status, rb_node);
185 printk(KERN_DEBUG " [%u/%u) %llu %llx",
186 es->es_lblk, es->es_len,
187 ext4_es_pblock(es), ext4_es_status(es));
188 node = rb_next(node);
190 printk(KERN_DEBUG "\n");
193 #define ext4_es_print_tree(inode)
196 static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
198 BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
199 return es->es_lblk + es->es_len - 1;
203 * search through the tree for an delayed extent with a given offset. If
204 * it can't be found, try to find next extent.
206 static struct extent_status *__es_tree_search(struct rb_root *root,
209 struct rb_node *node = root->rb_node;
210 struct extent_status *es = NULL;
213 es = rb_entry(node, struct extent_status, rb_node);
214 if (lblk < es->es_lblk)
215 node = node->rb_left;
216 else if (lblk > ext4_es_end(es))
217 node = node->rb_right;
222 if (es && lblk < es->es_lblk)
225 if (es && lblk > ext4_es_end(es)) {
226 node = rb_next(&es->rb_node);
227 return node ? rb_entry(node, struct extent_status, rb_node) :
235 * ext4_es_find_delayed_extent: find the 1st delayed extent covering @es->lblk
236 * if it exists, otherwise, the next extent after @es->lblk.
238 * @inode: the inode which owns delayed extents
239 * @lblk: the offset where we start to search
240 * @es: delayed extent that we found
242 void ext4_es_find_delayed_extent(struct inode *inode, ext4_lblk_t lblk,
243 struct extent_status *es)
245 struct ext4_es_tree *tree = NULL;
246 struct extent_status *es1 = NULL;
247 struct rb_node *node;
250 trace_ext4_es_find_delayed_extent_enter(inode, lblk);
252 read_lock(&EXT4_I(inode)->i_es_lock);
253 tree = &EXT4_I(inode)->i_es_tree;
255 /* find extent in cache firstly */
256 es->es_lblk = es->es_len = es->es_pblk = 0;
257 if (tree->cache_es) {
258 es1 = tree->cache_es;
259 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
260 es_debug("%u cached by [%u/%u) %llu %llx\n",
261 lblk, es1->es_lblk, es1->es_len,
262 ext4_es_pblock(es1), ext4_es_status(es1));
267 es1 = __es_tree_search(&tree->root, lblk);
270 if (es1 && !ext4_es_is_delayed(es1)) {
271 while ((node = rb_next(&es1->rb_node)) != NULL) {
272 es1 = rb_entry(node, struct extent_status, rb_node);
273 if (ext4_es_is_delayed(es1))
278 if (es1 && ext4_es_is_delayed(es1)) {
279 tree->cache_es = es1;
280 es->es_lblk = es1->es_lblk;
281 es->es_len = es1->es_len;
282 es->es_pblk = es1->es_pblk;
285 read_unlock(&EXT4_I(inode)->i_es_lock);
287 ext4_es_lru_add(inode);
288 trace_ext4_es_find_delayed_extent_exit(inode, es);
291 static struct extent_status *
292 ext4_es_alloc_extent(struct inode *inode, ext4_lblk_t lblk, ext4_lblk_t len,
295 struct extent_status *es;
296 es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
304 * We don't count delayed extent because we never try to reclaim them
306 if (!ext4_es_is_delayed(es)) {
307 EXT4_I(inode)->i_es_lru_nr++;
308 atomic_inc(&EXT4_SB(inode->i_sb)->s_extent_cache_cnt);
314 static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
316 /* Decrease the lru counter when this es is not delayed */
317 if (!ext4_es_is_delayed(es)) {
318 BUG_ON(EXT4_I(inode)->i_es_lru_nr == 0);
319 EXT4_I(inode)->i_es_lru_nr--;
320 atomic_dec(&EXT4_SB(inode->i_sb)->s_extent_cache_cnt);
323 kmem_cache_free(ext4_es_cachep, es);
327 * Check whether or not two extents can be merged
329 * - logical block number is contiguous
330 * - physical block number is contiguous
333 static int ext4_es_can_be_merged(struct extent_status *es1,
334 struct extent_status *es2)
336 if (es1->es_lblk + es1->es_len != es2->es_lblk)
339 if (ext4_es_status(es1) != ext4_es_status(es2))
342 if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
343 (ext4_es_pblock(es1) + es1->es_len != ext4_es_pblock(es2)))
349 static struct extent_status *
350 ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
352 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
353 struct extent_status *es1;
354 struct rb_node *node;
356 node = rb_prev(&es->rb_node);
360 es1 = rb_entry(node, struct extent_status, rb_node);
361 if (ext4_es_can_be_merged(es1, es)) {
362 es1->es_len += es->es_len;
363 rb_erase(&es->rb_node, &tree->root);
364 ext4_es_free_extent(inode, es);
371 static struct extent_status *
372 ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
374 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
375 struct extent_status *es1;
376 struct rb_node *node;
378 node = rb_next(&es->rb_node);
382 es1 = rb_entry(node, struct extent_status, rb_node);
383 if (ext4_es_can_be_merged(es, es1)) {
384 es->es_len += es1->es_len;
385 rb_erase(node, &tree->root);
386 ext4_es_free_extent(inode, es1);
392 static int __es_insert_extent(struct inode *inode, struct extent_status *newes)
394 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
395 struct rb_node **p = &tree->root.rb_node;
396 struct rb_node *parent = NULL;
397 struct extent_status *es;
401 es = rb_entry(parent, struct extent_status, rb_node);
403 if (newes->es_lblk < es->es_lblk) {
404 if (ext4_es_can_be_merged(newes, es)) {
406 * Here we can modify es_lblk directly
407 * because it isn't overlapped.
409 es->es_lblk = newes->es_lblk;
410 es->es_len += newes->es_len;
411 if (ext4_es_is_written(es) ||
412 ext4_es_is_unwritten(es))
413 ext4_es_store_pblock(es,
415 es = ext4_es_try_to_merge_left(inode, es);
419 } else if (newes->es_lblk > ext4_es_end(es)) {
420 if (ext4_es_can_be_merged(es, newes)) {
421 es->es_len += newes->es_len;
422 es = ext4_es_try_to_merge_right(inode, es);
432 es = ext4_es_alloc_extent(inode, newes->es_lblk, newes->es_len,
436 rb_link_node(&es->rb_node, parent, p);
437 rb_insert_color(&es->rb_node, &tree->root);
445 * ext4_es_insert_extent() adds a space to a extent status tree.
447 * ext4_es_insert_extent is called by ext4_da_write_begin and
448 * ext4_es_remove_extent.
450 * Return 0 on success, error code on failure.
452 int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk,
453 ext4_lblk_t len, ext4_fsblk_t pblk,
454 unsigned long long status)
456 struct extent_status newes;
457 ext4_lblk_t end = lblk + len - 1;
460 es_debug("add [%u/%u) %llu %llx to extent status tree of inode %lu\n",
461 lblk, len, pblk, status, inode->i_ino);
468 newes.es_lblk = lblk;
470 ext4_es_store_pblock(&newes, pblk);
471 ext4_es_store_status(&newes, status);
472 trace_ext4_es_insert_extent(inode, &newes);
474 write_lock(&EXT4_I(inode)->i_es_lock);
475 err = __es_remove_extent(inode, lblk, end);
478 err = __es_insert_extent(inode, &newes);
481 write_unlock(&EXT4_I(inode)->i_es_lock);
483 ext4_es_lru_add(inode);
484 ext4_es_print_tree(inode);
490 * ext4_es_lookup_extent() looks up an extent in extent status tree.
492 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
494 * Return: 1 on found, 0 on not
496 int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
497 struct extent_status *es)
499 struct ext4_es_tree *tree;
500 struct extent_status *es1 = NULL;
501 struct rb_node *node;
504 trace_ext4_es_lookup_extent_enter(inode, lblk);
505 es_debug("lookup extent in block %u\n", lblk);
507 tree = &EXT4_I(inode)->i_es_tree;
508 read_lock(&EXT4_I(inode)->i_es_lock);
510 /* find extent in cache firstly */
511 es->es_lblk = es->es_len = es->es_pblk = 0;
512 if (tree->cache_es) {
513 es1 = tree->cache_es;
514 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
515 es_debug("%u cached by [%u/%u)\n",
516 lblk, es1->es_lblk, es1->es_len);
522 node = tree->root.rb_node;
524 es1 = rb_entry(node, struct extent_status, rb_node);
525 if (lblk < es1->es_lblk)
526 node = node->rb_left;
527 else if (lblk > ext4_es_end(es1))
528 node = node->rb_right;
538 es->es_lblk = es1->es_lblk;
539 es->es_len = es1->es_len;
540 es->es_pblk = es1->es_pblk;
543 read_unlock(&EXT4_I(inode)->i_es_lock);
545 ext4_es_lru_add(inode);
546 trace_ext4_es_lookup_extent_exit(inode, es, found);
550 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
553 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
554 struct rb_node *node;
555 struct extent_status *es;
556 struct extent_status orig_es;
557 ext4_lblk_t len1, len2;
561 es = __es_tree_search(&tree->root, lblk);
564 if (es->es_lblk > end)
567 /* Simply invalidate cache_es. */
568 tree->cache_es = NULL;
570 orig_es.es_lblk = es->es_lblk;
571 orig_es.es_len = es->es_len;
572 orig_es.es_pblk = es->es_pblk;
574 len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
575 len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
580 struct extent_status newes;
582 newes.es_lblk = end + 1;
584 if (ext4_es_is_written(&orig_es) ||
585 ext4_es_is_unwritten(&orig_es)) {
586 block = ext4_es_pblock(&orig_es) +
587 orig_es.es_len - len2;
588 ext4_es_store_pblock(&newes, block);
590 ext4_es_store_status(&newes, ext4_es_status(&orig_es));
591 err = __es_insert_extent(inode, &newes);
593 es->es_lblk = orig_es.es_lblk;
594 es->es_len = orig_es.es_len;
598 es->es_lblk = end + 1;
600 if (ext4_es_is_written(es) ||
601 ext4_es_is_unwritten(es)) {
602 block = orig_es.es_pblk + orig_es.es_len - len2;
603 ext4_es_store_pblock(es, block);
610 node = rb_next(&es->rb_node);
612 es = rb_entry(node, struct extent_status, rb_node);
617 while (es && ext4_es_end(es) <= end) {
618 node = rb_next(&es->rb_node);
619 rb_erase(&es->rb_node, &tree->root);
620 ext4_es_free_extent(inode, es);
625 es = rb_entry(node, struct extent_status, rb_node);
628 if (es && es->es_lblk < end + 1) {
629 ext4_lblk_t orig_len = es->es_len;
631 len1 = ext4_es_end(es) - end;
632 es->es_lblk = end + 1;
634 if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
635 block = es->es_pblk + orig_len - len1;
636 ext4_es_store_pblock(es, block);
645 * ext4_es_remove_extent() removes a space from a extent status tree.
647 * Return 0 on success, error code on failure.
649 int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
655 trace_ext4_es_remove_extent(inode, lblk, len);
656 es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
657 lblk, len, inode->i_ino);
662 end = lblk + len - 1;
665 write_lock(&EXT4_I(inode)->i_es_lock);
666 err = __es_remove_extent(inode, lblk, end);
667 write_unlock(&EXT4_I(inode)->i_es_lock);
668 ext4_es_print_tree(inode);
672 static int ext4_es_shrink(struct shrinker *shrink, struct shrink_control *sc)
674 struct ext4_sb_info *sbi = container_of(shrink,
675 struct ext4_sb_info, s_es_shrinker);
676 struct ext4_inode_info *ei;
677 struct list_head *cur, *tmp, scanned;
678 int nr_to_scan = sc->nr_to_scan;
679 int ret, nr_shrunk = 0;
681 ret = atomic_read(&sbi->s_extent_cache_cnt);
682 trace_ext4_es_shrink_enter(sbi->s_sb, nr_to_scan, ret);
687 INIT_LIST_HEAD(&scanned);
689 spin_lock(&sbi->s_es_lru_lock);
690 list_for_each_safe(cur, tmp, &sbi->s_es_lru) {
691 list_move_tail(cur, &scanned);
693 ei = list_entry(cur, struct ext4_inode_info, i_es_lru);
695 read_lock(&ei->i_es_lock);
696 if (ei->i_es_lru_nr == 0) {
697 read_unlock(&ei->i_es_lock);
700 read_unlock(&ei->i_es_lock);
702 write_lock(&ei->i_es_lock);
703 ret = __es_try_to_reclaim_extents(ei, nr_to_scan);
704 write_unlock(&ei->i_es_lock);
711 list_splice_tail(&scanned, &sbi->s_es_lru);
712 spin_unlock(&sbi->s_es_lru_lock);
714 ret = atomic_read(&sbi->s_extent_cache_cnt);
715 trace_ext4_es_shrink_exit(sbi->s_sb, nr_shrunk, ret);
719 void ext4_es_register_shrinker(struct super_block *sb)
721 struct ext4_sb_info *sbi;
724 INIT_LIST_HEAD(&sbi->s_es_lru);
725 spin_lock_init(&sbi->s_es_lru_lock);
726 sbi->s_es_shrinker.shrink = ext4_es_shrink;
727 sbi->s_es_shrinker.seeks = DEFAULT_SEEKS;
728 register_shrinker(&sbi->s_es_shrinker);
731 void ext4_es_unregister_shrinker(struct super_block *sb)
733 unregister_shrinker(&EXT4_SB(sb)->s_es_shrinker);
736 void ext4_es_lru_add(struct inode *inode)
738 struct ext4_inode_info *ei = EXT4_I(inode);
739 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
741 spin_lock(&sbi->s_es_lru_lock);
742 if (list_empty(&ei->i_es_lru))
743 list_add_tail(&ei->i_es_lru, &sbi->s_es_lru);
745 list_move_tail(&ei->i_es_lru, &sbi->s_es_lru);
746 spin_unlock(&sbi->s_es_lru_lock);
749 void ext4_es_lru_del(struct inode *inode)
751 struct ext4_inode_info *ei = EXT4_I(inode);
752 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
754 spin_lock(&sbi->s_es_lru_lock);
755 if (!list_empty(&ei->i_es_lru))
756 list_del_init(&ei->i_es_lru);
757 spin_unlock(&sbi->s_es_lru_lock);
760 static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
763 struct inode *inode = &ei->vfs_inode;
764 struct ext4_es_tree *tree = &ei->i_es_tree;
765 struct rb_node *node;
766 struct extent_status *es;
769 if (ei->i_es_lru_nr == 0)
772 node = rb_first(&tree->root);
773 while (node != NULL) {
774 es = rb_entry(node, struct extent_status, rb_node);
775 node = rb_next(&es->rb_node);
777 * We can't reclaim delayed extent from status tree because
778 * fiemap, bigallic, and seek_data/hole need to use it.
780 if (!ext4_es_is_delayed(es)) {
781 rb_erase(&es->rb_node, &tree->root);
782 ext4_es_free_extent(inode, es);
784 if (--nr_to_scan == 0)
788 tree->cache_es = NULL;