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 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;
112 static struct extent_node *__try_back_merge(struct f2fs_sb_info *sbi,
113 struct extent_tree *et, struct extent_node *en)
115 struct extent_node *prev;
116 struct rb_node *node;
118 node = rb_prev(&en->rb_node);
122 prev = rb_entry(node, struct extent_node, rb_node);
123 if (__is_back_mergeable(&en->ei, &prev->ei)) {
124 en->ei.fofs = prev->ei.fofs;
125 en->ei.blk = prev->ei.blk;
126 en->ei.len += prev->ei.len;
127 __detach_extent_node(sbi, et, prev);
133 static struct extent_node *__try_front_merge(struct f2fs_sb_info *sbi,
134 struct extent_tree *et, struct extent_node *en)
136 struct extent_node *next;
137 struct rb_node *node;
139 node = rb_next(&en->rb_node);
143 next = rb_entry(node, struct extent_node, rb_node);
144 if (__is_front_mergeable(&en->ei, &next->ei)) {
145 en->ei.len += next->ei.len;
146 __detach_extent_node(sbi, et, next);
152 static struct extent_node *__insert_extent_tree(struct f2fs_sb_info *sbi,
153 struct extent_tree *et, struct extent_info *ei,
154 struct extent_node **den)
156 struct rb_node **p = &et->root.rb_node;
157 struct rb_node *parent = NULL;
158 struct extent_node *en;
162 en = rb_entry(parent, struct extent_node, rb_node);
164 if (ei->fofs < en->ei.fofs) {
165 if (__is_front_mergeable(ei, &en->ei)) {
166 f2fs_bug_on(sbi, !den);
167 en->ei.fofs = ei->fofs;
168 en->ei.blk = ei->blk;
169 en->ei.len += ei->len;
170 *den = __try_back_merge(sbi, et, en);
174 } else if (ei->fofs >= en->ei.fofs + en->ei.len) {
175 if (__is_back_mergeable(ei, &en->ei)) {
176 f2fs_bug_on(sbi, !den);
177 en->ei.len += ei->len;
178 *den = __try_front_merge(sbi, et, en);
187 en = __attach_extent_node(sbi, et, ei, parent, p);
191 if (en->ei.len > et->largest.len)
192 et->largest = en->ei;
197 static unsigned int __free_extent_tree(struct f2fs_sb_info *sbi,
198 struct extent_tree *et, bool free_all)
200 struct rb_node *node, *next;
201 struct extent_node *en;
202 unsigned int count = et->count;
204 node = rb_first(&et->root);
206 next = rb_next(node);
207 en = rb_entry(node, struct extent_node, rb_node);
210 spin_lock(&sbi->extent_lock);
211 if (!list_empty(&en->list))
212 list_del_init(&en->list);
213 spin_unlock(&sbi->extent_lock);
216 if (free_all || list_empty(&en->list)) {
217 __detach_extent_node(sbi, et, en);
218 kmem_cache_free(extent_node_slab, en);
223 return count - et->count;
226 void f2fs_drop_largest_extent(struct inode *inode, pgoff_t fofs)
228 struct extent_info *largest = &F2FS_I(inode)->extent_tree->largest;
230 if (largest->fofs <= fofs && largest->fofs + largest->len > fofs)
234 void f2fs_init_extent_tree(struct inode *inode, struct f2fs_extent *i_ext)
236 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
237 struct extent_tree *et;
238 struct extent_node *en;
239 struct extent_info ei;
241 if (!f2fs_may_extent_tree(inode))
244 et = __grab_extent_tree(inode);
246 if (!i_ext || le32_to_cpu(i_ext->len) < F2FS_MIN_EXTENT_LEN)
249 set_extent_info(&ei, le32_to_cpu(i_ext->fofs),
250 le32_to_cpu(i_ext->blk), le32_to_cpu(i_ext->len));
252 write_lock(&et->lock);
256 en = __insert_extent_tree(sbi, et, &ei, NULL);
258 spin_lock(&sbi->extent_lock);
259 list_add_tail(&en->list, &sbi->extent_list);
260 spin_unlock(&sbi->extent_lock);
263 write_unlock(&et->lock);
266 static bool f2fs_lookup_extent_tree(struct inode *inode, pgoff_t pgofs,
267 struct extent_info *ei)
269 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
270 struct extent_tree *et = F2FS_I(inode)->extent_tree;
271 struct extent_node *en;
274 f2fs_bug_on(sbi, !et);
276 trace_f2fs_lookup_extent_tree_start(inode, pgofs);
278 read_lock(&et->lock);
280 if (et->largest.fofs <= pgofs &&
281 et->largest.fofs + et->largest.len > pgofs) {
284 stat_inc_read_hit(sbi);
285 stat_inc_largest_node_hit(sbi);
289 en = __lookup_extent_tree(sbi, et, pgofs);
292 spin_lock(&sbi->extent_lock);
293 if (!list_empty(&en->list))
294 list_move_tail(&en->list, &sbi->extent_list);
296 spin_unlock(&sbi->extent_lock);
298 stat_inc_read_hit(sbi);
301 stat_inc_total_hit(sbi);
302 read_unlock(&et->lock);
304 trace_f2fs_lookup_extent_tree_end(inode, pgofs, ei);
310 * lookup extent at @fofs, if hit, return the extent
311 * if not, return NULL and
312 * @prev_ex: extent before fofs
313 * @next_ex: extent after fofs
314 * @insert_p: insert point for new extent at fofs
315 * in order to simpfy the insertion after.
316 * tree must stay unchanged between lookup and insertion.
318 static struct extent_node *__lookup_extent_tree_ret(struct extent_tree *et,
320 struct extent_node **prev_ex,
321 struct extent_node **next_ex,
322 struct rb_node ***insert_p,
323 struct rb_node **insert_parent)
325 struct rb_node **pnode = &et->root.rb_node;
326 struct rb_node *parent = NULL, *tmp_node;
327 struct extent_node *en;
330 struct extent_info *cei = &et->cached_en->ei;
332 if (cei->fofs <= fofs && cei->fofs + cei->len > fofs)
333 return et->cached_en;
338 en = rb_entry(*pnode, struct extent_node, rb_node);
340 if (fofs < en->ei.fofs)
341 pnode = &(*pnode)->rb_left;
342 else if (fofs >= en->ei.fofs + en->ei.len)
343 pnode = &(*pnode)->rb_right;
349 *insert_parent = parent;
351 en = rb_entry(parent, struct extent_node, rb_node);
353 if (parent && fofs > en->ei.fofs)
354 tmp_node = rb_next(parent);
355 *next_ex = tmp_node ?
356 rb_entry(tmp_node, struct extent_node, rb_node) : NULL;
359 if (parent && fofs < en->ei.fofs)
360 tmp_node = rb_prev(parent);
361 *prev_ex = tmp_node ?
362 rb_entry(tmp_node, struct extent_node, rb_node) : NULL;
367 static struct extent_node *__insert_extent_tree_ret(struct f2fs_sb_info *sbi,
368 struct extent_tree *et, struct extent_info *ei,
369 struct extent_node **den,
370 struct extent_node *prev_ex,
371 struct extent_node *next_ex,
372 struct rb_node **insert_p,
373 struct rb_node *insert_parent)
375 struct rb_node **p = &et->root.rb_node;
376 struct rb_node *parent = NULL;
377 struct extent_node *en = NULL;
380 if (prev_ex && __is_back_mergeable(ei, &prev_ex->ei)) {
381 f2fs_bug_on(sbi, !den);
383 prev_ex->ei.len += ei->len;
387 if (next_ex && __is_front_mergeable(ei, &next_ex->ei)) {
388 f2fs_bug_on(sbi, !den);
390 __detach_extent_node(sbi, et, prev_ex);
393 next_ex->ei.fofs = ei->fofs;
394 next_ex->ei.blk = ei->blk;
395 next_ex->ei.len += ei->len;
401 if (insert_p && insert_parent) {
402 parent = insert_parent;
409 en = rb_entry(parent, struct extent_node, rb_node);
411 if (ei->fofs < en->ei.fofs)
413 else if (ei->fofs >= en->ei.fofs + en->ei.len)
419 en = __attach_extent_node(sbi, et, ei, parent, p);
423 if (en->ei.len > et->largest.len)
424 et->largest = en->ei;
429 /* return true, if on-disk extent should be updated */
430 static bool f2fs_update_extent_tree(struct inode *inode, pgoff_t fofs,
433 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
434 struct extent_tree *et = F2FS_I(inode)->extent_tree;
435 struct extent_node *en = NULL, *en1 = NULL, *en2 = NULL, *en3 = NULL;
436 struct extent_node *den = NULL, *prev_ex = NULL, *next_ex = NULL;
437 struct extent_info ei, dei, prev;
438 struct rb_node **insert_p = NULL, *insert_parent = NULL;
444 trace_f2fs_update_extent_tree(inode, fofs, blkaddr);
446 write_lock(&et->lock);
448 if (is_inode_flag_set(F2FS_I(inode), FI_NO_EXTENT)) {
449 write_unlock(&et->lock);
456 /* we do not guarantee that the largest extent is cached all the time */
457 f2fs_drop_largest_extent(inode, fofs);
459 /* 1. lookup and remove existing extent info in cache */
460 en = __lookup_extent_tree_ret(et, fofs, &prev_ex, &next_ex,
461 &insert_p, &insert_parent);
466 __detach_extent_node(sbi, et, en);
468 /* 2. if extent can be split, try to split it */
469 if (dei.len > F2FS_MIN_EXTENT_LEN) {
470 /* insert left part of split extent into cache */
471 if (fofs - dei.fofs >= F2FS_MIN_EXTENT_LEN) {
472 set_extent_info(&ei, dei.fofs, dei.blk,
474 en1 = __insert_extent_tree_ret(sbi, et, &ei, NULL,
475 NULL, NULL, NULL, NULL);
478 /* insert right part of split extent into cache */
479 endofs = dei.fofs + dei.len - 1;
480 if (endofs - fofs >= F2FS_MIN_EXTENT_LEN) {
481 set_extent_info(&ei, fofs + 1,
482 fofs - dei.fofs + dei.blk + 1, endofs - fofs);
483 en2 = __insert_extent_tree_ret(sbi, et, &ei, NULL,
484 NULL, NULL, NULL, NULL);
489 /* 3. update extent in extent cache */
491 set_extent_info(&ei, fofs, blkaddr, 1);
492 en3 = __insert_extent_tree_ret(sbi, et, &ei, &den,
493 prev_ex, next_ex, insert_p, insert_parent);
495 /* give up extent_cache, if split and small updates happen */
497 prev.len < F2FS_MIN_EXTENT_LEN &&
498 et->largest.len < F2FS_MIN_EXTENT_LEN) {
500 set_inode_flag(F2FS_I(inode), FI_NO_EXTENT);
504 /* 4. update in global extent list */
505 spin_lock(&sbi->extent_lock);
506 if (en && !list_empty(&en->list))
509 * en1 and en2 split from en, they will become more and more smaller
510 * fragments after splitting several times. So if the length is smaller
511 * than F2FS_MIN_EXTENT_LEN, we will not add them into extent tree.
514 list_add_tail(&en1->list, &sbi->extent_list);
516 list_add_tail(&en2->list, &sbi->extent_list);
518 if (list_empty(&en3->list))
519 list_add_tail(&en3->list, &sbi->extent_list);
521 list_move_tail(&en3->list, &sbi->extent_list);
523 if (den && !list_empty(&den->list))
524 list_del(&den->list);
525 spin_unlock(&sbi->extent_lock);
527 /* 5. release extent node */
529 kmem_cache_free(extent_node_slab, en);
531 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 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);
597 while ((found = radix_tree_gang_lookup(root,
598 (void **)treevec, ino, EXT_TREE_VEC_SIZE))) {
601 ino = treevec[found - 1]->ino + 1;
602 for (i = 0; i < found; i++) {
603 struct extent_tree *et = treevec[i];
605 write_lock(&et->lock);
606 node_cnt += __free_extent_tree(sbi, et, false);
607 write_unlock(&et->lock);
609 if (node_cnt + tree_cnt >= nr_shrink)
614 up_write(&sbi->extent_tree_lock);
616 trace_f2fs_shrink_extent_tree(sbi, node_cnt, tree_cnt);
618 return node_cnt + tree_cnt;
621 unsigned int f2fs_destroy_extent_node(struct inode *inode)
623 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
624 struct extent_tree *et = F2FS_I(inode)->extent_tree;
625 unsigned int node_cnt = 0;
630 write_lock(&et->lock);
631 node_cnt = __free_extent_tree(sbi, et, true);
632 write_unlock(&et->lock);
637 void f2fs_destroy_extent_tree(struct inode *inode)
639 struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
640 struct extent_tree *et = F2FS_I(inode)->extent_tree;
641 unsigned int node_cnt = 0;
646 if (inode->i_nlink && !is_bad_inode(inode) && et->count) {
647 atomic_dec(&et->refcount);
651 /* free all extent info belong to this extent tree */
652 node_cnt = f2fs_destroy_extent_node(inode);
654 /* delete extent tree entry in radix tree */
655 down_write(&sbi->extent_tree_lock);
656 atomic_dec(&et->refcount);
657 f2fs_bug_on(sbi, atomic_read(&et->refcount) || et->count);
658 radix_tree_delete(&sbi->extent_tree_root, inode->i_ino);
659 kmem_cache_free(extent_tree_slab, et);
660 sbi->total_ext_tree--;
661 up_write(&sbi->extent_tree_lock);
663 F2FS_I(inode)->extent_tree = NULL;
665 trace_f2fs_destroy_extent_tree(inode, node_cnt);
668 bool f2fs_lookup_extent_cache(struct inode *inode, pgoff_t pgofs,
669 struct extent_info *ei)
671 if (!f2fs_may_extent_tree(inode))
674 return f2fs_lookup_extent_tree(inode, pgofs, ei);
677 void f2fs_update_extent_cache(struct dnode_of_data *dn)
679 struct f2fs_inode_info *fi = F2FS_I(dn->inode);
682 if (!f2fs_may_extent_tree(dn->inode))
685 f2fs_bug_on(F2FS_I_SB(dn->inode), dn->data_blkaddr == NEW_ADDR);
687 fofs = start_bidx_of_node(ofs_of_node(dn->node_page), fi) +
690 if (f2fs_update_extent_tree(dn->inode, fofs, dn->data_blkaddr))
694 void init_extent_cache_info(struct f2fs_sb_info *sbi)
696 INIT_RADIX_TREE(&sbi->extent_tree_root, GFP_NOIO);
697 init_rwsem(&sbi->extent_tree_lock);
698 INIT_LIST_HEAD(&sbi->extent_list);
699 spin_lock_init(&sbi->extent_lock);
700 sbi->total_ext_tree = 0;
701 atomic_set(&sbi->total_ext_node, 0);
704 int __init create_extent_cache(void)
706 extent_tree_slab = f2fs_kmem_cache_create("f2fs_extent_tree",
707 sizeof(struct extent_tree));
708 if (!extent_tree_slab)
710 extent_node_slab = f2fs_kmem_cache_create("f2fs_extent_node",
711 sizeof(struct extent_node));
712 if (!extent_node_slab) {
713 kmem_cache_destroy(extent_tree_slab);
719 void destroy_extent_cache(void)
721 kmem_cache_destroy(extent_node_slab);
722 kmem_cache_destroy(extent_tree_slab);