2 * f2fs extent cache support
4 * Copyright (c) 2015 Motorola Mobility
5 * Copyright (c) 2015 Samsung Electronics
6 * Authors: Jaegeuk Kim <jaegeuk@kernel.org>
7 * Chao Yu <chao2.yu@samsung.com>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
15 #include <linux/f2fs_fs.h>
19 #include <trace/events/f2fs.h>
21 static struct kmem_cache *extent_tree_slab;
22 static struct kmem_cache *extent_node_slab;
24 static struct extent_node *__attach_extent_node(struct f2fs_sb_info *sbi,
25 struct extent_tree *et, struct extent_info *ei,
26 struct rb_node *parent, struct rb_node **p)
28 struct extent_node *en;
30 en = kmem_cache_alloc(extent_node_slab, GFP_ATOMIC);
35 INIT_LIST_HEAD(&en->list);
37 rb_link_node(&en->rb_node, parent, p);
38 rb_insert_color(&en->rb_node, &et->root);
40 atomic_inc(&sbi->total_ext_node);
44 static void __detach_extent_node(struct f2fs_sb_info *sbi,
45 struct extent_tree *et, struct extent_node *en)
47 rb_erase(&en->rb_node, &et->root);
49 atomic_dec(&sbi->total_ext_node);
51 if (et->cached_en == en)
55 static struct extent_tree *__grab_extent_tree(struct inode *inode)
57 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
58 struct extent_tree *et;
59 nid_t ino = inode->i_ino;
61 down_write(&sbi->extent_tree_lock);
62 et = radix_tree_lookup(&sbi->extent_tree_root, ino);
64 et = f2fs_kmem_cache_alloc(extent_tree_slab, GFP_NOFS);
65 f2fs_radix_tree_insert(&sbi->extent_tree_root, ino, et);
66 memset(et, 0, sizeof(struct extent_tree));
70 rwlock_init(&et->lock);
71 atomic_set(&et->refcount, 0);
73 atomic_inc(&sbi->total_ext_tree);
75 atomic_inc(&et->refcount);
76 up_write(&sbi->extent_tree_lock);
78 /* never died until evict_inode */
79 F2FS_I(inode)->extent_tree = et;
84 static struct extent_node *__lookup_extent_tree(struct f2fs_sb_info *sbi,
85 struct extent_tree *et, unsigned int fofs)
87 struct rb_node *node = et->root.rb_node;
88 struct extent_node *en = et->cached_en;
91 struct extent_info *cei = &en->ei;
93 if (cei->fofs <= fofs && cei->fofs + cei->len > fofs) {
94 stat_inc_cached_node_hit(sbi);
100 en = rb_entry(node, struct extent_node, rb_node);
102 if (fofs < en->ei.fofs) {
103 node = node->rb_left;
104 } else if (fofs >= en->ei.fofs + en->ei.len) {
105 node = node->rb_right;
107 stat_inc_rbtree_node_hit(sbi);
114 static struct extent_node *__init_extent_tree(struct f2fs_sb_info *sbi,
115 struct extent_tree *et, struct extent_info *ei)
117 struct rb_node **p = &et->root.rb_node;
118 struct extent_node *en;
120 en = __attach_extent_node(sbi, et, ei, NULL, p);
124 et->largest = en->ei;
129 static unsigned int __free_extent_tree(struct f2fs_sb_info *sbi,
130 struct extent_tree *et, bool free_all)
132 struct rb_node *node, *next;
133 struct extent_node *en;
134 unsigned int count = et->count;
136 node = rb_first(&et->root);
138 next = rb_next(node);
139 en = rb_entry(node, struct extent_node, rb_node);
142 spin_lock(&sbi->extent_lock);
143 if (!list_empty(&en->list))
144 list_del_init(&en->list);
145 spin_unlock(&sbi->extent_lock);
148 if (free_all || list_empty(&en->list)) {
149 __detach_extent_node(sbi, et, en);
150 kmem_cache_free(extent_node_slab, en);
155 return count - et->count;
158 static void __drop_largest_extent(struct inode *inode,
159 pgoff_t fofs, unsigned int len)
161 struct extent_info *largest = &F2FS_I(inode)->extent_tree->largest;
163 if (fofs < largest->fofs + largest->len && fofs + len > largest->fofs)
167 void f2fs_init_extent_tree(struct inode *inode, struct f2fs_extent *i_ext)
169 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
170 struct extent_tree *et;
171 struct extent_node *en;
172 struct extent_info ei;
174 if (!f2fs_may_extent_tree(inode))
177 et = __grab_extent_tree(inode);
179 if (!i_ext || le32_to_cpu(i_ext->len) < F2FS_MIN_EXTENT_LEN)
182 set_extent_info(&ei, le32_to_cpu(i_ext->fofs),
183 le32_to_cpu(i_ext->blk), le32_to_cpu(i_ext->len));
185 write_lock(&et->lock);
189 en = __init_extent_tree(sbi, et, &ei);
191 spin_lock(&sbi->extent_lock);
192 list_add_tail(&en->list, &sbi->extent_list);
193 spin_unlock(&sbi->extent_lock);
196 write_unlock(&et->lock);
199 static bool f2fs_lookup_extent_tree(struct inode *inode, pgoff_t pgofs,
200 struct extent_info *ei)
202 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
203 struct extent_tree *et = F2FS_I(inode)->extent_tree;
204 struct extent_node *en;
207 f2fs_bug_on(sbi, !et);
209 trace_f2fs_lookup_extent_tree_start(inode, pgofs);
211 read_lock(&et->lock);
213 if (et->largest.fofs <= pgofs &&
214 et->largest.fofs + et->largest.len > pgofs) {
217 stat_inc_largest_node_hit(sbi);
221 en = __lookup_extent_tree(sbi, et, pgofs);
224 spin_lock(&sbi->extent_lock);
225 if (!list_empty(&en->list))
226 list_move_tail(&en->list, &sbi->extent_list);
228 spin_unlock(&sbi->extent_lock);
232 stat_inc_total_hit(sbi);
233 read_unlock(&et->lock);
235 trace_f2fs_lookup_extent_tree_end(inode, pgofs, ei);
241 * lookup extent at @fofs, if hit, return the extent
242 * if not, return NULL and
243 * @prev_ex: extent before fofs
244 * @next_ex: extent after fofs
245 * @insert_p: insert point for new extent at fofs
246 * in order to simpfy the insertion after.
247 * tree must stay unchanged between lookup and insertion.
249 static struct extent_node *__lookup_extent_tree_ret(struct extent_tree *et,
251 struct extent_node **prev_ex,
252 struct extent_node **next_ex,
253 struct rb_node ***insert_p,
254 struct rb_node **insert_parent)
256 struct rb_node **pnode = &et->root.rb_node;
257 struct rb_node *parent = NULL, *tmp_node;
258 struct extent_node *en = et->cached_en;
261 *insert_parent = NULL;
265 if (RB_EMPTY_ROOT(&et->root))
269 struct extent_info *cei = &en->ei;
271 if (cei->fofs <= fofs && cei->fofs + cei->len > fofs)
272 goto lookup_neighbors;
277 en = rb_entry(*pnode, struct extent_node, rb_node);
279 if (fofs < en->ei.fofs)
280 pnode = &(*pnode)->rb_left;
281 else if (fofs >= en->ei.fofs + en->ei.len)
282 pnode = &(*pnode)->rb_right;
284 goto lookup_neighbors;
288 *insert_parent = parent;
290 en = rb_entry(parent, struct extent_node, rb_node);
292 if (parent && fofs > en->ei.fofs)
293 tmp_node = rb_next(parent);
294 *next_ex = tmp_node ?
295 rb_entry(tmp_node, struct extent_node, rb_node) : NULL;
298 if (parent && fofs < en->ei.fofs)
299 tmp_node = rb_prev(parent);
300 *prev_ex = tmp_node ?
301 rb_entry(tmp_node, struct extent_node, rb_node) : NULL;
305 if (fofs == en->ei.fofs) {
306 /* lookup prev node for merging backward later */
307 tmp_node = rb_prev(&en->rb_node);
308 *prev_ex = tmp_node ?
309 rb_entry(tmp_node, struct extent_node, rb_node) : NULL;
311 if (fofs == en->ei.fofs + en->ei.len - 1) {
312 /* lookup next node for merging frontward later */
313 tmp_node = rb_next(&en->rb_node);
314 *next_ex = tmp_node ?
315 rb_entry(tmp_node, struct extent_node, rb_node) : NULL;
320 static struct extent_node *__try_merge_extent_node(struct f2fs_sb_info *sbi,
321 struct extent_tree *et, struct extent_info *ei,
322 struct extent_node **den,
323 struct extent_node *prev_ex,
324 struct extent_node *next_ex)
326 struct extent_node *en = NULL;
328 if (prev_ex && __is_back_mergeable(ei, &prev_ex->ei)) {
329 prev_ex->ei.len += ei->len;
334 if (next_ex && __is_front_mergeable(ei, &next_ex->ei)) {
336 __detach_extent_node(sbi, et, prev_ex);
339 next_ex->ei.fofs = ei->fofs;
340 next_ex->ei.blk = ei->blk;
341 next_ex->ei.len += ei->len;
346 __try_update_largest_extent(et, en);
352 static struct extent_node *__insert_extent_tree(struct f2fs_sb_info *sbi,
353 struct extent_tree *et, struct extent_info *ei,
354 struct rb_node **insert_p,
355 struct rb_node *insert_parent)
357 struct rb_node **p = &et->root.rb_node;
358 struct rb_node *parent = NULL;
359 struct extent_node *en = NULL;
361 if (insert_p && insert_parent) {
362 parent = insert_parent;
369 en = rb_entry(parent, struct extent_node, rb_node);
371 if (ei->fofs < en->ei.fofs)
373 else if (ei->fofs >= en->ei.fofs + en->ei.len)
379 en = __attach_extent_node(sbi, et, ei, parent, p);
383 __try_update_largest_extent(et, en);
388 static unsigned int f2fs_update_extent_tree_range(struct inode *inode,
389 pgoff_t fofs, block_t blkaddr, unsigned int len)
391 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
392 struct extent_tree *et = F2FS_I(inode)->extent_tree;
393 struct extent_node *en = NULL, *en1 = NULL;
394 struct extent_node *prev_en = NULL, *next_en = NULL;
395 struct extent_info ei, dei, prev;
396 struct rb_node **insert_p = NULL, *insert_parent = NULL;
397 unsigned int end = fofs + len;
398 unsigned int pos = (unsigned int)fofs;
403 trace_f2fs_update_extent_tree_range(inode, fofs, blkaddr, len);
405 write_lock(&et->lock);
407 if (is_inode_flag_set(F2FS_I(inode), FI_NO_EXTENT)) {
408 write_unlock(&et->lock);
416 * drop largest extent before lookup, in case it's already
417 * been shrunk from extent tree
419 __drop_largest_extent(inode, fofs, len);
421 /* 1. lookup first extent node in range [fofs, fofs + len - 1] */
422 en = __lookup_extent_tree_ret(et, fofs, &prev_en, &next_en,
423 &insert_p, &insert_parent);
427 /* 2. invlidate all extent nodes in range [fofs, fofs + len - 1] */
428 while (en && en->ei.fofs < end) {
429 unsigned int org_end;
430 int parts = 0; /* # of parts current extent split into */
432 next_en = en1 = NULL;
435 org_end = dei.fofs + dei.len;
436 f2fs_bug_on(sbi, pos >= org_end);
438 if (pos > dei.fofs && pos - dei.fofs >= F2FS_MIN_EXTENT_LEN) {
439 en->ei.len = pos - en->ei.fofs;
444 if (end < org_end && org_end - end >= F2FS_MIN_EXTENT_LEN) {
446 set_extent_info(&ei, end,
447 end - dei.fofs + dei.blk,
449 en1 = __insert_extent_tree(sbi, et, &ei,
454 en->ei.blk += end - dei.fofs;
455 en->ei.len -= end - dei.fofs;
462 struct rb_node *node = rb_next(&en->rb_node);
465 rb_entry(node, struct extent_node, rb_node)
470 __try_update_largest_extent(et, en);
472 __detach_extent_node(sbi, et, en);
475 * if original extent is split into zero or two parts, extent
476 * tree has been altered by deletion or insertion, therefore
477 * invalidate pointers regard to tree.
481 insert_parent = NULL;
484 /* update in global extent list */
485 spin_lock(&sbi->extent_lock);
486 if (!parts && !list_empty(&en->list))
489 list_add_tail(&en1->list, &sbi->extent_list);
490 spin_unlock(&sbi->extent_lock);
492 /* release extent node */
494 kmem_cache_free(extent_node_slab, en);
499 /* 3. update extent in extent cache */
501 struct extent_node *den = NULL;
503 set_extent_info(&ei, fofs, blkaddr, len);
504 en1 = __try_merge_extent_node(sbi, et, &ei, &den,
507 en1 = __insert_extent_tree(sbi, et, &ei,
508 insert_p, insert_parent);
510 /* give up extent_cache, if split and small updates happen */
512 prev.len < F2FS_MIN_EXTENT_LEN &&
513 et->largest.len < F2FS_MIN_EXTENT_LEN) {
515 set_inode_flag(F2FS_I(inode), FI_NO_EXTENT);
518 spin_lock(&sbi->extent_lock);
520 if (list_empty(&en1->list))
521 list_add_tail(&en1->list, &sbi->extent_list);
523 list_move_tail(&en1->list, &sbi->extent_list);
525 if (den && !list_empty(&den->list))
526 list_del(&den->list);
527 spin_unlock(&sbi->extent_lock);
530 kmem_cache_free(extent_node_slab, den);
533 if (is_inode_flag_set(F2FS_I(inode), FI_NO_EXTENT))
534 __free_extent_tree(sbi, et, true);
536 write_unlock(&et->lock);
538 return !__is_extent_same(&prev, &et->largest);
541 unsigned int f2fs_shrink_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink)
543 struct extent_tree *treevec[EXT_TREE_VEC_SIZE];
544 struct extent_node *en, *tmp;
545 unsigned long ino = F2FS_ROOT_INO(sbi);
546 struct radix_tree_root *root = &sbi->extent_tree_root;
548 unsigned int node_cnt = 0, tree_cnt = 0;
551 if (!test_opt(sbi, EXTENT_CACHE))
554 if (!down_write_trylock(&sbi->extent_tree_lock))
557 /* 1. remove unreferenced extent tree */
558 while ((found = radix_tree_gang_lookup(root,
559 (void **)treevec, ino, EXT_TREE_VEC_SIZE))) {
562 ino = treevec[found - 1]->ino + 1;
563 for (i = 0; i < found; i++) {
564 struct extent_tree *et = treevec[i];
566 if (!atomic_read(&et->refcount)) {
567 write_lock(&et->lock);
568 node_cnt += __free_extent_tree(sbi, et, true);
569 write_unlock(&et->lock);
571 radix_tree_delete(root, et->ino);
572 kmem_cache_free(extent_tree_slab, et);
573 atomic_dec(&sbi->total_ext_tree);
576 if (node_cnt + tree_cnt >= nr_shrink)
581 up_write(&sbi->extent_tree_lock);
583 /* 2. remove LRU extent entries */
584 if (!down_write_trylock(&sbi->extent_tree_lock))
587 remained = nr_shrink - (node_cnt + tree_cnt);
589 spin_lock(&sbi->extent_lock);
590 list_for_each_entry_safe(en, tmp, &sbi->extent_list, list) {
593 list_del_init(&en->list);
595 spin_unlock(&sbi->extent_lock);
598 * reset ino for searching victims from beginning of global extent tree.
600 ino = F2FS_ROOT_INO(sbi);
602 while ((found = radix_tree_gang_lookup(root,
603 (void **)treevec, ino, EXT_TREE_VEC_SIZE))) {
606 ino = treevec[found - 1]->ino + 1;
607 for (i = 0; i < found; i++) {
608 struct extent_tree *et = treevec[i];
610 if (write_trylock(&et->lock)) {
611 node_cnt += __free_extent_tree(sbi, et, false);
612 write_unlock(&et->lock);
615 if (node_cnt + tree_cnt >= nr_shrink)
620 up_write(&sbi->extent_tree_lock);
622 trace_f2fs_shrink_extent_tree(sbi, node_cnt, tree_cnt);
624 return node_cnt + tree_cnt;
627 unsigned int f2fs_destroy_extent_node(struct inode *inode)
629 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
630 struct extent_tree *et = F2FS_I(inode)->extent_tree;
631 unsigned int node_cnt = 0;
636 write_lock(&et->lock);
637 node_cnt = __free_extent_tree(sbi, et, true);
638 write_unlock(&et->lock);
643 void f2fs_destroy_extent_tree(struct inode *inode)
645 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
646 struct extent_tree *et = F2FS_I(inode)->extent_tree;
647 unsigned int node_cnt = 0;
652 if (inode->i_nlink && !is_bad_inode(inode) && et->count) {
653 atomic_dec(&et->refcount);
657 /* free all extent info belong to this extent tree */
658 node_cnt = f2fs_destroy_extent_node(inode);
660 /* delete extent tree entry in radix tree */
661 down_write(&sbi->extent_tree_lock);
662 atomic_dec(&et->refcount);
663 f2fs_bug_on(sbi, atomic_read(&et->refcount) || et->count);
664 radix_tree_delete(&sbi->extent_tree_root, inode->i_ino);
665 kmem_cache_free(extent_tree_slab, et);
666 atomic_dec(&sbi->total_ext_tree);
667 up_write(&sbi->extent_tree_lock);
669 F2FS_I(inode)->extent_tree = NULL;
671 trace_f2fs_destroy_extent_tree(inode, node_cnt);
674 bool f2fs_lookup_extent_cache(struct inode *inode, pgoff_t pgofs,
675 struct extent_info *ei)
677 if (!f2fs_may_extent_tree(inode))
680 return f2fs_lookup_extent_tree(inode, pgofs, ei);
683 void f2fs_update_extent_cache(struct dnode_of_data *dn)
685 struct f2fs_inode_info *fi = F2FS_I(dn->inode);
688 if (!f2fs_may_extent_tree(dn->inode))
691 f2fs_bug_on(F2FS_I_SB(dn->inode), dn->data_blkaddr == NEW_ADDR);
694 fofs = start_bidx_of_node(ofs_of_node(dn->node_page), fi) +
697 if (f2fs_update_extent_tree_range(dn->inode, fofs, dn->data_blkaddr, 1))
701 void f2fs_update_extent_cache_range(struct dnode_of_data *dn,
702 pgoff_t fofs, block_t blkaddr, unsigned int len)
705 if (!f2fs_may_extent_tree(dn->inode))
708 if (f2fs_update_extent_tree_range(dn->inode, fofs, blkaddr, len))
712 void init_extent_cache_info(struct f2fs_sb_info *sbi)
714 INIT_RADIX_TREE(&sbi->extent_tree_root, GFP_NOIO);
715 init_rwsem(&sbi->extent_tree_lock);
716 INIT_LIST_HEAD(&sbi->extent_list);
717 spin_lock_init(&sbi->extent_lock);
718 atomic_set(&sbi->total_ext_tree, 0);
719 atomic_set(&sbi->total_ext_node, 0);
722 int __init create_extent_cache(void)
724 extent_tree_slab = f2fs_kmem_cache_create("f2fs_extent_tree",
725 sizeof(struct extent_tree));
726 if (!extent_tree_slab)
728 extent_node_slab = f2fs_kmem_cache_create("f2fs_extent_node",
729 sizeof(struct extent_node));
730 if (!extent_node_slab) {
731 kmem_cache_destroy(extent_tree_slab);
737 void destroy_extent_cache(void)
739 kmem_cache_destroy(extent_node_slab);
740 kmem_cache_destroy(extent_tree_slab);