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>
13 #include <linux/list_sort.h>
15 #include "extents_status.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,
150 static int __ext4_es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
151 struct ext4_inode_info *locked_ei);
153 int __init ext4_init_es(void)
155 ext4_es_cachep = kmem_cache_create("ext4_extent_status",
156 sizeof(struct extent_status),
157 0, (SLAB_RECLAIM_ACCOUNT), NULL);
158 if (ext4_es_cachep == NULL)
163 void ext4_exit_es(void)
166 kmem_cache_destroy(ext4_es_cachep);
169 void ext4_es_init_tree(struct ext4_es_tree *tree)
171 tree->root = RB_ROOT;
172 tree->cache_es = NULL;
176 static void ext4_es_print_tree(struct inode *inode)
178 struct ext4_es_tree *tree;
179 struct rb_node *node;
181 printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
182 tree = &EXT4_I(inode)->i_es_tree;
183 node = rb_first(&tree->root);
185 struct extent_status *es;
186 es = rb_entry(node, struct extent_status, rb_node);
187 printk(KERN_DEBUG " [%u/%u) %llu %llx",
188 es->es_lblk, es->es_len,
189 ext4_es_pblock(es), ext4_es_status(es));
190 node = rb_next(node);
192 printk(KERN_DEBUG "\n");
195 #define ext4_es_print_tree(inode)
198 static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
200 BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
201 return es->es_lblk + es->es_len - 1;
205 * search through the tree for an delayed extent with a given offset. If
206 * it can't be found, try to find next extent.
208 static struct extent_status *__es_tree_search(struct rb_root *root,
211 struct rb_node *node = root->rb_node;
212 struct extent_status *es = NULL;
215 es = rb_entry(node, struct extent_status, rb_node);
216 if (lblk < es->es_lblk)
217 node = node->rb_left;
218 else if (lblk > ext4_es_end(es))
219 node = node->rb_right;
224 if (es && lblk < es->es_lblk)
227 if (es && lblk > ext4_es_end(es)) {
228 node = rb_next(&es->rb_node);
229 return node ? rb_entry(node, struct extent_status, rb_node) :
237 * ext4_es_find_delayed_extent_range: find the 1st delayed extent covering
238 * @es->lblk if it exists, otherwise, the next extent after @es->lblk.
240 * @inode: the inode which owns delayed extents
241 * @lblk: the offset where we start to search
242 * @end: the offset where we stop to search
243 * @es: delayed extent that we found
245 void ext4_es_find_delayed_extent_range(struct inode *inode,
246 ext4_lblk_t lblk, ext4_lblk_t end,
247 struct extent_status *es)
249 struct ext4_es_tree *tree = NULL;
250 struct extent_status *es1 = NULL;
251 struct rb_node *node;
255 trace_ext4_es_find_delayed_extent_range_enter(inode, lblk);
257 read_lock(&EXT4_I(inode)->i_es_lock);
258 tree = &EXT4_I(inode)->i_es_tree;
260 /* find extent in cache firstly */
261 es->es_lblk = es->es_len = es->es_pblk = 0;
262 if (tree->cache_es) {
263 es1 = tree->cache_es;
264 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
265 es_debug("%u cached by [%u/%u) %llu %x\n",
266 lblk, es1->es_lblk, es1->es_len,
267 ext4_es_pblock(es1), ext4_es_status(es1));
272 es1 = __es_tree_search(&tree->root, lblk);
275 if (es1 && !ext4_es_is_delayed(es1)) {
276 while ((node = rb_next(&es1->rb_node)) != NULL) {
277 es1 = rb_entry(node, struct extent_status, rb_node);
278 if (es1->es_lblk > end) {
282 if (ext4_es_is_delayed(es1))
287 if (es1 && ext4_es_is_delayed(es1)) {
288 tree->cache_es = es1;
289 es->es_lblk = es1->es_lblk;
290 es->es_len = es1->es_len;
291 es->es_pblk = es1->es_pblk;
294 read_unlock(&EXT4_I(inode)->i_es_lock);
296 trace_ext4_es_find_delayed_extent_range_exit(inode, es);
299 static struct extent_status *
300 ext4_es_alloc_extent(struct inode *inode, ext4_lblk_t lblk, ext4_lblk_t len,
303 struct extent_status *es;
304 es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
312 * We don't count delayed extent because we never try to reclaim them
314 if (!ext4_es_is_delayed(es)) {
315 EXT4_I(inode)->i_es_lru_nr++;
316 percpu_counter_inc(&EXT4_SB(inode->i_sb)->s_extent_cache_cnt);
322 static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
324 /* Decrease the lru counter when this es is not delayed */
325 if (!ext4_es_is_delayed(es)) {
326 BUG_ON(EXT4_I(inode)->i_es_lru_nr == 0);
327 EXT4_I(inode)->i_es_lru_nr--;
328 percpu_counter_dec(&EXT4_SB(inode->i_sb)->s_extent_cache_cnt);
331 kmem_cache_free(ext4_es_cachep, es);
335 * Check whether or not two extents can be merged
337 * - logical block number is contiguous
338 * - physical block number is contiguous
341 static int ext4_es_can_be_merged(struct extent_status *es1,
342 struct extent_status *es2)
344 if (ext4_es_status(es1) != ext4_es_status(es2))
347 if (((__u64) es1->es_len) + es2->es_len > 0xFFFFFFFFULL)
350 if (((__u64) es1->es_lblk) + es1->es_len != es2->es_lblk)
353 if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
354 (ext4_es_pblock(es1) + es1->es_len == ext4_es_pblock(es2)))
357 if (ext4_es_is_hole(es1))
360 /* we need to check delayed extent is without unwritten status */
361 if (ext4_es_is_delayed(es1) && !ext4_es_is_unwritten(es1))
367 static struct extent_status *
368 ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
370 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
371 struct extent_status *es1;
372 struct rb_node *node;
374 node = rb_prev(&es->rb_node);
378 es1 = rb_entry(node, struct extent_status, rb_node);
379 if (ext4_es_can_be_merged(es1, es)) {
380 es1->es_len += es->es_len;
381 rb_erase(&es->rb_node, &tree->root);
382 ext4_es_free_extent(inode, es);
389 static struct extent_status *
390 ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
392 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
393 struct extent_status *es1;
394 struct rb_node *node;
396 node = rb_next(&es->rb_node);
400 es1 = rb_entry(node, struct extent_status, rb_node);
401 if (ext4_es_can_be_merged(es, es1)) {
402 es->es_len += es1->es_len;
403 rb_erase(node, &tree->root);
404 ext4_es_free_extent(inode, es1);
410 #ifdef ES_AGGRESSIVE_TEST
411 #include "ext4_extents.h" /* Needed when ES_AGGRESSIVE_TEST is defined */
413 static void ext4_es_insert_extent_ext_check(struct inode *inode,
414 struct extent_status *es)
416 struct ext4_ext_path *path = NULL;
417 struct ext4_extent *ex;
418 ext4_lblk_t ee_block;
419 ext4_fsblk_t ee_start;
420 unsigned short ee_len;
421 int depth, ee_status, es_status;
423 path = ext4_ext_find_extent(inode, es->es_lblk, NULL, EXT4_EX_NOCACHE);
427 depth = ext_depth(inode);
428 ex = path[depth].p_ext;
432 ee_block = le32_to_cpu(ex->ee_block);
433 ee_start = ext4_ext_pblock(ex);
434 ee_len = ext4_ext_get_actual_len(ex);
436 ee_status = ext4_ext_is_uninitialized(ex) ? 1 : 0;
437 es_status = ext4_es_is_unwritten(es) ? 1 : 0;
440 * Make sure ex and es are not overlap when we try to insert
441 * a delayed/hole extent.
443 if (!ext4_es_is_written(es) && !ext4_es_is_unwritten(es)) {
444 if (in_range(es->es_lblk, ee_block, ee_len)) {
445 pr_warn("ES insert assertion failed for "
446 "inode: %lu we can find an extent "
447 "at block [%d/%d/%llu/%c], but we "
448 "want to add an delayed/hole extent "
449 "[%d/%d/%llu/%llx]\n",
450 inode->i_ino, ee_block, ee_len,
451 ee_start, ee_status ? 'u' : 'w',
452 es->es_lblk, es->es_len,
453 ext4_es_pblock(es), ext4_es_status(es));
459 * We don't check ee_block == es->es_lblk, etc. because es
460 * might be a part of whole extent, vice versa.
462 if (es->es_lblk < ee_block ||
463 ext4_es_pblock(es) != ee_start + es->es_lblk - ee_block) {
464 pr_warn("ES insert assertion failed for inode: %lu "
465 "ex_status [%d/%d/%llu/%c] != "
466 "es_status [%d/%d/%llu/%c]\n", inode->i_ino,
467 ee_block, ee_len, ee_start,
468 ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
469 ext4_es_pblock(es), es_status ? 'u' : 'w');
473 if (ee_status ^ es_status) {
474 pr_warn("ES insert assertion failed for inode: %lu "
475 "ex_status [%d/%d/%llu/%c] != "
476 "es_status [%d/%d/%llu/%c]\n", inode->i_ino,
477 ee_block, ee_len, ee_start,
478 ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
479 ext4_es_pblock(es), es_status ? 'u' : 'w');
483 * We can't find an extent on disk. So we need to make sure
484 * that we don't want to add an written/unwritten extent.
486 if (!ext4_es_is_delayed(es) && !ext4_es_is_hole(es)) {
487 pr_warn("ES insert assertion failed for inode: %lu "
488 "can't find an extent at block %d but we want "
489 "to add an written/unwritten extent "
490 "[%d/%d/%llu/%llx]\n", inode->i_ino,
491 es->es_lblk, es->es_lblk, es->es_len,
492 ext4_es_pblock(es), ext4_es_status(es));
497 ext4_ext_drop_refs(path);
502 static void ext4_es_insert_extent_ind_check(struct inode *inode,
503 struct extent_status *es)
505 struct ext4_map_blocks map;
509 * Here we call ext4_ind_map_blocks to lookup a block mapping because
510 * 'Indirect' structure is defined in indirect.c. So we couldn't
511 * access direct/indirect tree from outside. It is too dirty to define
512 * this function in indirect.c file.
515 map.m_lblk = es->es_lblk;
516 map.m_len = es->es_len;
518 retval = ext4_ind_map_blocks(NULL, inode, &map, 0);
520 if (ext4_es_is_delayed(es) || ext4_es_is_hole(es)) {
522 * We want to add a delayed/hole extent but this
523 * block has been allocated.
525 pr_warn("ES insert assertion failed for inode: %lu "
526 "We can find blocks but we want to add a "
527 "delayed/hole extent [%d/%d/%llu/%llx]\n",
528 inode->i_ino, es->es_lblk, es->es_len,
529 ext4_es_pblock(es), ext4_es_status(es));
531 } else if (ext4_es_is_written(es)) {
532 if (retval != es->es_len) {
533 pr_warn("ES insert assertion failed for "
534 "inode: %lu retval %d != es_len %d\n",
535 inode->i_ino, retval, es->es_len);
538 if (map.m_pblk != ext4_es_pblock(es)) {
539 pr_warn("ES insert assertion failed for "
540 "inode: %lu m_pblk %llu != "
542 inode->i_ino, map.m_pblk,
548 * We don't need to check unwritten extent because
549 * indirect-based file doesn't have it.
553 } else if (retval == 0) {
554 if (ext4_es_is_written(es)) {
555 pr_warn("ES insert assertion failed for inode: %lu "
556 "We can't find the block but we want to add "
557 "an written extent [%d/%d/%llu/%llx]\n",
558 inode->i_ino, es->es_lblk, es->es_len,
559 ext4_es_pblock(es), ext4_es_status(es));
565 static inline void ext4_es_insert_extent_check(struct inode *inode,
566 struct extent_status *es)
569 * We don't need to worry about the race condition because
570 * caller takes i_data_sem locking.
572 BUG_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
573 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
574 ext4_es_insert_extent_ext_check(inode, es);
576 ext4_es_insert_extent_ind_check(inode, es);
579 static inline void ext4_es_insert_extent_check(struct inode *inode,
580 struct extent_status *es)
585 static int __es_insert_extent(struct inode *inode, struct extent_status *newes)
587 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
588 struct rb_node **p = &tree->root.rb_node;
589 struct rb_node *parent = NULL;
590 struct extent_status *es;
594 es = rb_entry(parent, struct extent_status, rb_node);
596 if (newes->es_lblk < es->es_lblk) {
597 if (ext4_es_can_be_merged(newes, es)) {
599 * Here we can modify es_lblk directly
600 * because it isn't overlapped.
602 es->es_lblk = newes->es_lblk;
603 es->es_len += newes->es_len;
604 if (ext4_es_is_written(es) ||
605 ext4_es_is_unwritten(es))
606 ext4_es_store_pblock(es,
608 es = ext4_es_try_to_merge_left(inode, es);
612 } else if (newes->es_lblk > ext4_es_end(es)) {
613 if (ext4_es_can_be_merged(es, newes)) {
614 es->es_len += newes->es_len;
615 es = ext4_es_try_to_merge_right(inode, es);
625 es = ext4_es_alloc_extent(inode, newes->es_lblk, newes->es_len,
629 rb_link_node(&es->rb_node, parent, p);
630 rb_insert_color(&es->rb_node, &tree->root);
638 * ext4_es_insert_extent() adds information to an inode's extent
641 * Return 0 on success, error code on failure.
643 int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk,
644 ext4_lblk_t len, ext4_fsblk_t pblk,
647 struct extent_status newes;
648 ext4_lblk_t end = lblk + len - 1;
651 es_debug("add [%u/%u) %llu %x to extent status tree of inode %lu\n",
652 lblk, len, pblk, status, inode->i_ino);
659 newes.es_lblk = lblk;
661 ext4_es_store_pblock(&newes, pblk);
662 ext4_es_store_status(&newes, status);
663 trace_ext4_es_insert_extent(inode, &newes);
665 ext4_es_insert_extent_check(inode, &newes);
667 write_lock(&EXT4_I(inode)->i_es_lock);
668 err = __es_remove_extent(inode, lblk, end);
672 err = __es_insert_extent(inode, &newes);
673 if (err == -ENOMEM && __ext4_es_shrink(EXT4_SB(inode->i_sb), 1,
676 if (err == -ENOMEM && !ext4_es_is_delayed(&newes))
680 write_unlock(&EXT4_I(inode)->i_es_lock);
682 ext4_es_print_tree(inode);
688 * ext4_es_cache_extent() inserts information into the extent status
689 * tree if and only if there isn't information about the range in
692 void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
693 ext4_lblk_t len, ext4_fsblk_t pblk,
696 struct extent_status *es;
697 struct extent_status newes;
698 ext4_lblk_t end = lblk + len - 1;
700 newes.es_lblk = lblk;
702 ext4_es_store_pblock(&newes, pblk);
703 ext4_es_store_status(&newes, status);
704 trace_ext4_es_cache_extent(inode, &newes);
711 write_lock(&EXT4_I(inode)->i_es_lock);
713 es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
714 if (!es || es->es_lblk > end)
715 __es_insert_extent(inode, &newes);
716 write_unlock(&EXT4_I(inode)->i_es_lock);
720 * ext4_es_lookup_extent() looks up an extent in extent status tree.
722 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
724 * Return: 1 on found, 0 on not
726 int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
727 struct extent_status *es)
729 struct ext4_es_tree *tree;
730 struct extent_status *es1 = NULL;
731 struct rb_node *node;
734 trace_ext4_es_lookup_extent_enter(inode, lblk);
735 es_debug("lookup extent in block %u\n", lblk);
737 tree = &EXT4_I(inode)->i_es_tree;
738 read_lock(&EXT4_I(inode)->i_es_lock);
740 /* find extent in cache firstly */
741 es->es_lblk = es->es_len = es->es_pblk = 0;
742 if (tree->cache_es) {
743 es1 = tree->cache_es;
744 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
745 es_debug("%u cached by [%u/%u)\n",
746 lblk, es1->es_lblk, es1->es_len);
752 node = tree->root.rb_node;
754 es1 = rb_entry(node, struct extent_status, rb_node);
755 if (lblk < es1->es_lblk)
756 node = node->rb_left;
757 else if (lblk > ext4_es_end(es1))
758 node = node->rb_right;
768 es->es_lblk = es1->es_lblk;
769 es->es_len = es1->es_len;
770 es->es_pblk = es1->es_pblk;
773 read_unlock(&EXT4_I(inode)->i_es_lock);
775 trace_ext4_es_lookup_extent_exit(inode, es, found);
779 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
782 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
783 struct rb_node *node;
784 struct extent_status *es;
785 struct extent_status orig_es;
786 ext4_lblk_t len1, len2;
792 es = __es_tree_search(&tree->root, lblk);
795 if (es->es_lblk > end)
798 /* Simply invalidate cache_es. */
799 tree->cache_es = NULL;
801 orig_es.es_lblk = es->es_lblk;
802 orig_es.es_len = es->es_len;
803 orig_es.es_pblk = es->es_pblk;
805 len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
806 len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
811 struct extent_status newes;
813 newes.es_lblk = end + 1;
815 if (ext4_es_is_written(&orig_es) ||
816 ext4_es_is_unwritten(&orig_es)) {
817 block = ext4_es_pblock(&orig_es) +
818 orig_es.es_len - len2;
819 ext4_es_store_pblock(&newes, block);
821 ext4_es_store_status(&newes, ext4_es_status(&orig_es));
822 err = __es_insert_extent(inode, &newes);
824 es->es_lblk = orig_es.es_lblk;
825 es->es_len = orig_es.es_len;
826 if ((err == -ENOMEM) &&
827 __ext4_es_shrink(EXT4_SB(inode->i_sb), 1,
833 es->es_lblk = end + 1;
835 if (ext4_es_is_written(es) ||
836 ext4_es_is_unwritten(es)) {
837 block = orig_es.es_pblk + orig_es.es_len - len2;
838 ext4_es_store_pblock(es, block);
845 node = rb_next(&es->rb_node);
847 es = rb_entry(node, struct extent_status, rb_node);
852 while (es && ext4_es_end(es) <= end) {
853 node = rb_next(&es->rb_node);
854 rb_erase(&es->rb_node, &tree->root);
855 ext4_es_free_extent(inode, es);
860 es = rb_entry(node, struct extent_status, rb_node);
863 if (es && es->es_lblk < end + 1) {
864 ext4_lblk_t orig_len = es->es_len;
866 len1 = ext4_es_end(es) - end;
867 es->es_lblk = end + 1;
869 if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
870 block = es->es_pblk + orig_len - len1;
871 ext4_es_store_pblock(es, block);
880 * ext4_es_remove_extent() removes a space from a extent status tree.
882 * Return 0 on success, error code on failure.
884 int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
890 trace_ext4_es_remove_extent(inode, lblk, len);
891 es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
892 lblk, len, inode->i_ino);
897 end = lblk + len - 1;
900 write_lock(&EXT4_I(inode)->i_es_lock);
901 err = __es_remove_extent(inode, lblk, end);
902 write_unlock(&EXT4_I(inode)->i_es_lock);
903 ext4_es_print_tree(inode);
907 static int ext4_inode_touch_time_cmp(void *priv, struct list_head *a,
910 struct ext4_inode_info *eia, *eib;
911 eia = list_entry(a, struct ext4_inode_info, i_es_lru);
912 eib = list_entry(b, struct ext4_inode_info, i_es_lru);
914 if (ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
915 !ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
917 if (!ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
918 ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
920 if (eia->i_touch_when == eib->i_touch_when)
922 if (time_after(eia->i_touch_when, eib->i_touch_when))
928 static int __ext4_es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
929 struct ext4_inode_info *locked_ei)
931 struct ext4_inode_info *ei;
932 struct list_head *cur, *tmp;
935 int retried = 0, skip_precached = 1, nr_skipped = 0;
937 spin_lock(&sbi->s_es_lru_lock);
940 list_for_each_safe(cur, tmp, &sbi->s_es_lru) {
944 * If we have already reclaimed all extents from extent
945 * status tree, just stop the loop immediately.
947 if (percpu_counter_read_positive(&sbi->s_extent_cache_cnt) == 0)
950 ei = list_entry(cur, struct ext4_inode_info, i_es_lru);
953 * Skip the inode that is newer than the last_sorted
954 * time. Normally we try hard to avoid shrinking
955 * precached inodes, but we will as a last resort.
957 if ((sbi->s_es_last_sorted < ei->i_touch_when) ||
958 (skip_precached && ext4_test_inode_state(&ei->vfs_inode,
959 EXT4_STATE_EXT_PRECACHED))) {
961 list_move_tail(cur, &skipped);
965 if (ei->i_es_lru_nr == 0 || ei == locked_ei)
968 write_lock(&ei->i_es_lock);
969 shrunk = __es_try_to_reclaim_extents(ei, nr_to_scan);
970 if (ei->i_es_lru_nr == 0)
971 list_del_init(&ei->i_es_lru);
972 write_unlock(&ei->i_es_lock);
975 nr_to_scan -= shrunk;
980 /* Move the newer inodes into the tail of the LRU list. */
981 list_splice_tail(&skipped, &sbi->s_es_lru);
982 INIT_LIST_HEAD(&skipped);
985 * If we skipped any inodes, and we weren't able to make any
986 * forward progress, sort the list and try again.
988 if ((nr_shrunk == 0) && nr_skipped && !retried) {
990 list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
991 sbi->s_es_last_sorted = jiffies;
992 ei = list_first_entry(&sbi->s_es_lru, struct ext4_inode_info,
995 * If there are no non-precached inodes left on the
996 * list, start releasing precached extents.
998 if (ext4_test_inode_state(&ei->vfs_inode,
999 EXT4_STATE_EXT_PRECACHED))
1004 spin_unlock(&sbi->s_es_lru_lock);
1006 if (locked_ei && nr_shrunk == 0)
1007 nr_shrunk = __es_try_to_reclaim_extents(locked_ei, nr_to_scan);
1012 static unsigned long ext4_es_count(struct shrinker *shrink,
1013 struct shrink_control *sc)
1016 struct ext4_sb_info *sbi;
1018 sbi = container_of(shrink, struct ext4_sb_info, s_es_shrinker);
1019 nr = percpu_counter_read_positive(&sbi->s_extent_cache_cnt);
1020 trace_ext4_es_shrink_enter(sbi->s_sb, sc->nr_to_scan, nr);
1024 static unsigned long ext4_es_scan(struct shrinker *shrink,
1025 struct shrink_control *sc)
1027 struct ext4_sb_info *sbi = container_of(shrink,
1028 struct ext4_sb_info, s_es_shrinker);
1029 int nr_to_scan = sc->nr_to_scan;
1032 ret = percpu_counter_read_positive(&sbi->s_extent_cache_cnt);
1033 trace_ext4_es_shrink_enter(sbi->s_sb, nr_to_scan, ret);
1038 nr_shrunk = __ext4_es_shrink(sbi, nr_to_scan, NULL);
1040 trace_ext4_es_shrink_exit(sbi->s_sb, nr_shrunk, ret);
1044 void ext4_es_register_shrinker(struct ext4_sb_info *sbi)
1046 INIT_LIST_HEAD(&sbi->s_es_lru);
1047 spin_lock_init(&sbi->s_es_lru_lock);
1048 sbi->s_es_last_sorted = 0;
1049 sbi->s_es_shrinker.scan_objects = ext4_es_scan;
1050 sbi->s_es_shrinker.count_objects = ext4_es_count;
1051 sbi->s_es_shrinker.seeks = DEFAULT_SEEKS;
1052 register_shrinker(&sbi->s_es_shrinker);
1055 void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi)
1057 unregister_shrinker(&sbi->s_es_shrinker);
1060 void ext4_es_lru_add(struct inode *inode)
1062 struct ext4_inode_info *ei = EXT4_I(inode);
1063 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1065 ei->i_touch_when = jiffies;
1067 if (!list_empty(&ei->i_es_lru))
1070 spin_lock(&sbi->s_es_lru_lock);
1071 if (list_empty(&ei->i_es_lru))
1072 list_add_tail(&ei->i_es_lru, &sbi->s_es_lru);
1073 spin_unlock(&sbi->s_es_lru_lock);
1076 void ext4_es_lru_del(struct inode *inode)
1078 struct ext4_inode_info *ei = EXT4_I(inode);
1079 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1081 spin_lock(&sbi->s_es_lru_lock);
1082 if (!list_empty(&ei->i_es_lru))
1083 list_del_init(&ei->i_es_lru);
1084 spin_unlock(&sbi->s_es_lru_lock);
1087 static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
1090 struct inode *inode = &ei->vfs_inode;
1091 struct ext4_es_tree *tree = &ei->i_es_tree;
1092 struct rb_node *node;
1093 struct extent_status *es;
1094 unsigned long nr_shrunk = 0;
1095 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1096 DEFAULT_RATELIMIT_BURST);
1098 if (ei->i_es_lru_nr == 0)
1101 if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
1103 ext4_warning(inode->i_sb, "forced shrink of precached extents");
1105 node = rb_first(&tree->root);
1106 while (node != NULL) {
1107 es = rb_entry(node, struct extent_status, rb_node);
1108 node = rb_next(&es->rb_node);
1110 * We can't reclaim delayed extent from status tree because
1111 * fiemap, bigallic, and seek_data/hole need to use it.
1113 if (!ext4_es_is_delayed(es)) {
1114 rb_erase(&es->rb_node, &tree->root);
1115 ext4_es_free_extent(inode, es);
1117 if (--nr_to_scan == 0)
1121 tree->cache_es = NULL;