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,
149 int __init ext4_init_es(void)
151 ext4_es_cachep = KMEM_CACHE(extent_status, SLAB_RECLAIM_ACCOUNT);
152 if (ext4_es_cachep == NULL)
157 void ext4_exit_es(void)
160 kmem_cache_destroy(ext4_es_cachep);
163 void ext4_es_init_tree(struct ext4_es_tree *tree)
165 tree->root = RB_ROOT;
166 tree->cache_es = NULL;
170 static void ext4_es_print_tree(struct inode *inode)
172 struct ext4_es_tree *tree;
173 struct rb_node *node;
175 printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
176 tree = &EXT4_I(inode)->i_es_tree;
177 node = rb_first(&tree->root);
179 struct extent_status *es;
180 es = rb_entry(node, struct extent_status, rb_node);
181 printk(KERN_DEBUG " [%u/%u) %llu %llx",
182 es->es_lblk, es->es_len,
183 ext4_es_pblock(es), ext4_es_status(es));
184 node = rb_next(node);
186 printk(KERN_DEBUG "\n");
189 #define ext4_es_print_tree(inode)
192 static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
194 BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
195 return es->es_lblk + es->es_len - 1;
199 * search through the tree for an delayed extent with a given offset. If
200 * it can't be found, try to find next extent.
202 static struct extent_status *__es_tree_search(struct rb_root *root,
205 struct rb_node *node = root->rb_node;
206 struct extent_status *es = NULL;
209 es = rb_entry(node, struct extent_status, rb_node);
210 if (lblk < es->es_lblk)
211 node = node->rb_left;
212 else if (lblk > ext4_es_end(es))
213 node = node->rb_right;
218 if (es && lblk < es->es_lblk)
221 if (es && lblk > ext4_es_end(es)) {
222 node = rb_next(&es->rb_node);
223 return node ? rb_entry(node, struct extent_status, rb_node) :
231 * ext4_es_find_delayed_extent: find the 1st delayed extent covering @es->lblk
232 * if it exists, otherwise, the next extent after @es->lblk.
234 * @inode: the inode which owns delayed extents
235 * @lblk: the offset where we start to search
236 * @es: delayed extent that we found
238 void ext4_es_find_delayed_extent(struct inode *inode, ext4_lblk_t lblk,
239 struct extent_status *es)
241 struct ext4_es_tree *tree = NULL;
242 struct extent_status *es1 = NULL;
243 struct rb_node *node;
246 trace_ext4_es_find_delayed_extent_enter(inode, lblk);
248 read_lock(&EXT4_I(inode)->i_es_lock);
249 tree = &EXT4_I(inode)->i_es_tree;
251 /* find extent in cache firstly */
252 es->es_lblk = es->es_len = es->es_pblk = 0;
253 if (tree->cache_es) {
254 es1 = tree->cache_es;
255 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
256 es_debug("%u cached by [%u/%u) %llu %llx\n",
257 lblk, es1->es_lblk, es1->es_len,
258 ext4_es_pblock(es1), ext4_es_status(es1));
263 es1 = __es_tree_search(&tree->root, lblk);
266 if (es1 && !ext4_es_is_delayed(es1)) {
267 while ((node = rb_next(&es1->rb_node)) != NULL) {
268 es1 = rb_entry(node, struct extent_status, rb_node);
269 if (ext4_es_is_delayed(es1))
274 if (es1 && ext4_es_is_delayed(es1)) {
275 tree->cache_es = es1;
276 es->es_lblk = es1->es_lblk;
277 es->es_len = es1->es_len;
278 es->es_pblk = es1->es_pblk;
281 read_unlock(&EXT4_I(inode)->i_es_lock);
283 trace_ext4_es_find_delayed_extent_exit(inode, es);
286 static struct extent_status *
287 ext4_es_alloc_extent(struct inode *inode, ext4_lblk_t lblk, ext4_lblk_t len,
290 struct extent_status *es;
291 es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
300 static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
302 kmem_cache_free(ext4_es_cachep, es);
306 * Check whether or not two extents can be merged
308 * - logical block number is contiguous
309 * - physical block number is contiguous
312 static int ext4_es_can_be_merged(struct extent_status *es1,
313 struct extent_status *es2)
315 if (es1->es_lblk + es1->es_len != es2->es_lblk)
318 if (ext4_es_status(es1) != ext4_es_status(es2))
321 if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
322 (ext4_es_pblock(es1) + es1->es_len != ext4_es_pblock(es2)))
328 static struct extent_status *
329 ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
331 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
332 struct extent_status *es1;
333 struct rb_node *node;
335 node = rb_prev(&es->rb_node);
339 es1 = rb_entry(node, struct extent_status, rb_node);
340 if (ext4_es_can_be_merged(es1, es)) {
341 es1->es_len += es->es_len;
342 rb_erase(&es->rb_node, &tree->root);
343 ext4_es_free_extent(inode, es);
350 static struct extent_status *
351 ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
353 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
354 struct extent_status *es1;
355 struct rb_node *node;
357 node = rb_next(&es->rb_node);
361 es1 = rb_entry(node, struct extent_status, rb_node);
362 if (ext4_es_can_be_merged(es, es1)) {
363 es->es_len += es1->es_len;
364 rb_erase(node, &tree->root);
365 ext4_es_free_extent(inode, es1);
371 static int __es_insert_extent(struct inode *inode, struct extent_status *newes)
373 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
374 struct rb_node **p = &tree->root.rb_node;
375 struct rb_node *parent = NULL;
376 struct extent_status *es;
380 es = rb_entry(parent, struct extent_status, rb_node);
382 if (newes->es_lblk < es->es_lblk) {
383 if (ext4_es_can_be_merged(newes, es)) {
385 * Here we can modify es_lblk directly
386 * because it isn't overlapped.
388 es->es_lblk = newes->es_lblk;
389 es->es_len += newes->es_len;
390 if (ext4_es_is_written(es) ||
391 ext4_es_is_unwritten(es))
392 ext4_es_store_pblock(es,
394 es = ext4_es_try_to_merge_left(inode, es);
398 } else if (newes->es_lblk > ext4_es_end(es)) {
399 if (ext4_es_can_be_merged(es, newes)) {
400 es->es_len += newes->es_len;
401 es = ext4_es_try_to_merge_right(inode, es);
411 es = ext4_es_alloc_extent(inode, newes->es_lblk, newes->es_len,
415 rb_link_node(&es->rb_node, parent, p);
416 rb_insert_color(&es->rb_node, &tree->root);
424 * ext4_es_insert_extent() adds a space to a extent status tree.
426 * ext4_es_insert_extent is called by ext4_da_write_begin and
427 * ext4_es_remove_extent.
429 * Return 0 on success, error code on failure.
431 int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk,
432 ext4_lblk_t len, ext4_fsblk_t pblk,
433 unsigned long long status)
435 struct extent_status newes;
436 ext4_lblk_t end = lblk + len - 1;
439 es_debug("add [%u/%u) %llu %llx to extent status tree of inode %lu\n",
440 lblk, len, pblk, status, inode->i_ino);
444 newes.es_lblk = lblk;
446 ext4_es_store_pblock(&newes, pblk);
447 ext4_es_store_status(&newes, status);
448 trace_ext4_es_insert_extent(inode, &newes);
450 write_lock(&EXT4_I(inode)->i_es_lock);
451 err = __es_remove_extent(inode, lblk, end);
454 err = __es_insert_extent(inode, &newes);
457 write_unlock(&EXT4_I(inode)->i_es_lock);
459 ext4_es_print_tree(inode);
465 * ext4_es_lookup_extent() looks up an extent in extent status tree.
467 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
469 * Return: 1 on found, 0 on not
471 int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
472 struct extent_status *es)
474 struct ext4_es_tree *tree;
475 struct extent_status *es1 = NULL;
476 struct rb_node *node;
479 trace_ext4_es_lookup_extent_enter(inode, lblk);
480 es_debug("lookup extent in block %u\n", lblk);
482 tree = &EXT4_I(inode)->i_es_tree;
483 read_lock(&EXT4_I(inode)->i_es_lock);
485 /* find extent in cache firstly */
486 es->es_lblk = es->es_len = es->es_pblk = 0;
487 if (tree->cache_es) {
488 es1 = tree->cache_es;
489 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
490 es_debug("%u cached by [%u/%u)\n",
491 lblk, es1->es_lblk, es1->es_len);
497 node = tree->root.rb_node;
499 es1 = rb_entry(node, struct extent_status, rb_node);
500 if (lblk < es1->es_lblk)
501 node = node->rb_left;
502 else if (lblk > ext4_es_end(es1))
503 node = node->rb_right;
513 es->es_lblk = es1->es_lblk;
514 es->es_len = es1->es_len;
515 es->es_pblk = es1->es_pblk;
518 read_unlock(&EXT4_I(inode)->i_es_lock);
520 trace_ext4_es_lookup_extent_exit(inode, es, found);
524 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
527 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
528 struct rb_node *node;
529 struct extent_status *es;
530 struct extent_status orig_es;
531 ext4_lblk_t len1, len2;
535 es = __es_tree_search(&tree->root, lblk);
538 if (es->es_lblk > end)
541 /* Simply invalidate cache_es. */
542 tree->cache_es = NULL;
544 orig_es.es_lblk = es->es_lblk;
545 orig_es.es_len = es->es_len;
546 orig_es.es_pblk = es->es_pblk;
548 len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
549 len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
554 struct extent_status newes;
556 newes.es_lblk = end + 1;
558 if (ext4_es_is_written(&orig_es) ||
559 ext4_es_is_unwritten(&orig_es)) {
560 block = ext4_es_pblock(&orig_es) +
561 orig_es.es_len - len2;
562 ext4_es_store_pblock(&newes, block);
564 ext4_es_store_status(&newes, ext4_es_status(&orig_es));
565 err = __es_insert_extent(inode, &newes);
567 es->es_lblk = orig_es.es_lblk;
568 es->es_len = orig_es.es_len;
572 es->es_lblk = end + 1;
574 if (ext4_es_is_written(es) ||
575 ext4_es_is_unwritten(es)) {
576 block = orig_es.es_pblk + orig_es.es_len - len2;
577 ext4_es_store_pblock(es, block);
584 node = rb_next(&es->rb_node);
586 es = rb_entry(node, struct extent_status, rb_node);
591 while (es && ext4_es_end(es) <= end) {
592 node = rb_next(&es->rb_node);
593 rb_erase(&es->rb_node, &tree->root);
594 ext4_es_free_extent(inode, es);
599 es = rb_entry(node, struct extent_status, rb_node);
602 if (es && es->es_lblk < end + 1) {
603 ext4_lblk_t orig_len = es->es_len;
605 len1 = ext4_es_end(es) - end;
606 es->es_lblk = end + 1;
608 if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
609 block = es->es_pblk + orig_len - len1;
610 ext4_es_store_pblock(es, block);
619 * ext4_es_remove_extent() removes a space from a extent status tree.
621 * Return 0 on success, error code on failure.
623 int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
629 trace_ext4_es_remove_extent(inode, lblk, len);
630 es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
631 lblk, len, inode->i_ino);
633 end = lblk + len - 1;
636 write_lock(&EXT4_I(inode)->i_es_lock);
637 err = __es_remove_extent(inode, lblk, end);
638 write_unlock(&EXT4_I(inode)->i_es_lock);
639 ext4_es_print_tree(inode);