Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/err.h>
4#include <linux/slab.h>
5#include <linux/spinlock.h>
6#include "messages.h"
7#include "ctree.h"
8#include "volumes.h"
9#include "extent_map.h"
10#include "compression.h"
11#include "btrfs_inode.h"
12
13
14static struct kmem_cache *extent_map_cache;
15
16int __init extent_map_init(void)
17{
18 extent_map_cache = kmem_cache_create("btrfs_extent_map",
19 sizeof(struct extent_map), 0,
20 SLAB_MEM_SPREAD, NULL);
21 if (!extent_map_cache)
22 return -ENOMEM;
23 return 0;
24}
25
26void __cold extent_map_exit(void)
27{
28 kmem_cache_destroy(extent_map_cache);
29}
30
31/*
32 * Initialize the extent tree @tree. Should be called for each new inode or
33 * other user of the extent_map interface.
34 */
35void extent_map_tree_init(struct extent_map_tree *tree)
36{
37 tree->map = RB_ROOT_CACHED;
38 INIT_LIST_HEAD(&tree->modified_extents);
39 rwlock_init(&tree->lock);
40}
41
42/*
43 * Allocate a new extent_map structure. The new structure is returned with a
44 * reference count of one and needs to be freed using free_extent_map()
45 */
46struct extent_map *alloc_extent_map(void)
47{
48 struct extent_map *em;
49 em = kmem_cache_zalloc(extent_map_cache, GFP_NOFS);
50 if (!em)
51 return NULL;
52 RB_CLEAR_NODE(&em->rb_node);
53 em->compress_type = BTRFS_COMPRESS_NONE;
54 refcount_set(&em->refs, 1);
55 INIT_LIST_HEAD(&em->list);
56 return em;
57}
58
59/*
60 * Drop the reference out on @em by one and free the structure if the reference
61 * count hits zero.
62 */
63void free_extent_map(struct extent_map *em)
64{
65 if (!em)
66 return;
67 if (refcount_dec_and_test(&em->refs)) {
68 WARN_ON(extent_map_in_tree(em));
69 WARN_ON(!list_empty(&em->list));
70 if (test_bit(EXTENT_FLAG_FS_MAPPING, &em->flags))
71 kfree(em->map_lookup);
72 kmem_cache_free(extent_map_cache, em);
73 }
74}
75
76/* Do the math around the end of an extent, handling wrapping. */
77static u64 range_end(u64 start, u64 len)
78{
79 if (start + len < start)
80 return (u64)-1;
81 return start + len;
82}
83
84static int tree_insert(struct rb_root_cached *root, struct extent_map *em)
85{
86 struct rb_node **p = &root->rb_root.rb_node;
87 struct rb_node *parent = NULL;
88 struct extent_map *entry = NULL;
89 struct rb_node *orig_parent = NULL;
90 u64 end = range_end(em->start, em->len);
91 bool leftmost = true;
92
93 while (*p) {
94 parent = *p;
95 entry = rb_entry(parent, struct extent_map, rb_node);
96
97 if (em->start < entry->start) {
98 p = &(*p)->rb_left;
99 } else if (em->start >= extent_map_end(entry)) {
100 p = &(*p)->rb_right;
101 leftmost = false;
102 } else {
103 return -EEXIST;
104 }
105 }
106
107 orig_parent = parent;
108 while (parent && em->start >= extent_map_end(entry)) {
109 parent = rb_next(parent);
110 entry = rb_entry(parent, struct extent_map, rb_node);
111 }
112 if (parent)
113 if (end > entry->start && em->start < extent_map_end(entry))
114 return -EEXIST;
115
116 parent = orig_parent;
117 entry = rb_entry(parent, struct extent_map, rb_node);
118 while (parent && em->start < entry->start) {
119 parent = rb_prev(parent);
120 entry = rb_entry(parent, struct extent_map, rb_node);
121 }
122 if (parent)
123 if (end > entry->start && em->start < extent_map_end(entry))
124 return -EEXIST;
125
126 rb_link_node(&em->rb_node, orig_parent, p);
127 rb_insert_color_cached(&em->rb_node, root, leftmost);
128 return 0;
129}
130
131/*
132 * Search through the tree for an extent_map with a given offset. If it can't
133 * be found, try to find some neighboring extents
134 */
135static struct rb_node *__tree_search(struct rb_root *root, u64 offset,
136 struct rb_node **prev_or_next_ret)
137{
138 struct rb_node *n = root->rb_node;
139 struct rb_node *prev = NULL;
140 struct rb_node *orig_prev = NULL;
141 struct extent_map *entry;
142 struct extent_map *prev_entry = NULL;
143
144 ASSERT(prev_or_next_ret);
145
146 while (n) {
147 entry = rb_entry(n, struct extent_map, rb_node);
148 prev = n;
149 prev_entry = entry;
150
151 if (offset < entry->start)
152 n = n->rb_left;
153 else if (offset >= extent_map_end(entry))
154 n = n->rb_right;
155 else
156 return n;
157 }
158
159 orig_prev = prev;
160 while (prev && offset >= extent_map_end(prev_entry)) {
161 prev = rb_next(prev);
162 prev_entry = rb_entry(prev, struct extent_map, rb_node);
163 }
164
165 /*
166 * Previous extent map found, return as in this case the caller does not
167 * care about the next one.
168 */
169 if (prev) {
170 *prev_or_next_ret = prev;
171 return NULL;
172 }
173
174 prev = orig_prev;
175 prev_entry = rb_entry(prev, struct extent_map, rb_node);
176 while (prev && offset < prev_entry->start) {
177 prev = rb_prev(prev);
178 prev_entry = rb_entry(prev, struct extent_map, rb_node);
179 }
180 *prev_or_next_ret = prev;
181
182 return NULL;
183}
184
185/* Check to see if two extent_map structs are adjacent and safe to merge. */
186static int mergable_maps(struct extent_map *prev, struct extent_map *next)
187{
188 if (test_bit(EXTENT_FLAG_PINNED, &prev->flags))
189 return 0;
190
191 /*
192 * don't merge compressed extents, we need to know their
193 * actual size
194 */
195 if (test_bit(EXTENT_FLAG_COMPRESSED, &prev->flags))
196 return 0;
197
198 if (test_bit(EXTENT_FLAG_LOGGING, &prev->flags) ||
199 test_bit(EXTENT_FLAG_LOGGING, &next->flags))
200 return 0;
201
202 /*
203 * We don't want to merge stuff that hasn't been written to the log yet
204 * since it may not reflect exactly what is on disk, and that would be
205 * bad.
206 */
207 if (!list_empty(&prev->list) || !list_empty(&next->list))
208 return 0;
209
210 ASSERT(next->block_start != EXTENT_MAP_DELALLOC &&
211 prev->block_start != EXTENT_MAP_DELALLOC);
212
213 if (prev->map_lookup || next->map_lookup)
214 ASSERT(test_bit(EXTENT_FLAG_FS_MAPPING, &prev->flags) &&
215 test_bit(EXTENT_FLAG_FS_MAPPING, &next->flags));
216
217 if (extent_map_end(prev) == next->start &&
218 prev->flags == next->flags &&
219 prev->map_lookup == next->map_lookup &&
220 ((next->block_start == EXTENT_MAP_HOLE &&
221 prev->block_start == EXTENT_MAP_HOLE) ||
222 (next->block_start == EXTENT_MAP_INLINE &&
223 prev->block_start == EXTENT_MAP_INLINE) ||
224 (next->block_start < EXTENT_MAP_LAST_BYTE - 1 &&
225 next->block_start == extent_map_block_end(prev)))) {
226 return 1;
227 }
228 return 0;
229}
230
231static void try_merge_map(struct extent_map_tree *tree, struct extent_map *em)
232{
233 struct extent_map *merge = NULL;
234 struct rb_node *rb;
235
236 /*
237 * We can't modify an extent map that is in the tree and that is being
238 * used by another task, as it can cause that other task to see it in
239 * inconsistent state during the merging. We always have 1 reference for
240 * the tree and 1 for this task (which is unpinning the extent map or
241 * clearing the logging flag), so anything > 2 means it's being used by
242 * other tasks too.
243 */
244 if (refcount_read(&em->refs) > 2)
245 return;
246
247 if (em->start != 0) {
248 rb = rb_prev(&em->rb_node);
249 if (rb)
250 merge = rb_entry(rb, struct extent_map, rb_node);
251 if (rb && mergable_maps(merge, em)) {
252 em->start = merge->start;
253 em->orig_start = merge->orig_start;
254 em->len += merge->len;
255 em->block_len += merge->block_len;
256 em->block_start = merge->block_start;
257 em->mod_len = (em->mod_len + em->mod_start) - merge->mod_start;
258 em->mod_start = merge->mod_start;
259 em->generation = max(em->generation, merge->generation);
260 set_bit(EXTENT_FLAG_MERGED, &em->flags);
261
262 rb_erase_cached(&merge->rb_node, &tree->map);
263 RB_CLEAR_NODE(&merge->rb_node);
264 free_extent_map(merge);
265 }
266 }
267
268 rb = rb_next(&em->rb_node);
269 if (rb)
270 merge = rb_entry(rb, struct extent_map, rb_node);
271 if (rb && mergable_maps(em, merge)) {
272 em->len += merge->len;
273 em->block_len += merge->block_len;
274 rb_erase_cached(&merge->rb_node, &tree->map);
275 RB_CLEAR_NODE(&merge->rb_node);
276 em->mod_len = (merge->mod_start + merge->mod_len) - em->mod_start;
277 em->generation = max(em->generation, merge->generation);
278 set_bit(EXTENT_FLAG_MERGED, &em->flags);
279 free_extent_map(merge);
280 }
281}
282
283/*
284 * Unpin an extent from the cache.
285 *
286 * @tree: tree to unpin the extent in
287 * @start: logical offset in the file
288 * @len: length of the extent
289 * @gen: generation that this extent has been modified in
290 *
291 * Called after an extent has been written to disk properly. Set the generation
292 * to the generation that actually added the file item to the inode so we know
293 * we need to sync this extent when we call fsync().
294 */
295int unpin_extent_cache(struct extent_map_tree *tree, u64 start, u64 len,
296 u64 gen)
297{
298 int ret = 0;
299 struct extent_map *em;
300 bool prealloc = false;
301
302 write_lock(&tree->lock);
303 em = lookup_extent_mapping(tree, start, len);
304
305 WARN_ON(!em || em->start != start);
306
307 if (!em)
308 goto out;
309
310 em->generation = gen;
311 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
312 em->mod_start = em->start;
313 em->mod_len = em->len;
314
315 if (test_bit(EXTENT_FLAG_FILLING, &em->flags)) {
316 prealloc = true;
317 clear_bit(EXTENT_FLAG_FILLING, &em->flags);
318 }
319
320 try_merge_map(tree, em);
321
322 if (prealloc) {
323 em->mod_start = em->start;
324 em->mod_len = em->len;
325 }
326
327 free_extent_map(em);
328out:
329 write_unlock(&tree->lock);
330 return ret;
331
332}
333
334void clear_em_logging(struct extent_map_tree *tree, struct extent_map *em)
335{
336 lockdep_assert_held_write(&tree->lock);
337
338 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
339 if (extent_map_in_tree(em))
340 try_merge_map(tree, em);
341}
342
343static inline void setup_extent_mapping(struct extent_map_tree *tree,
344 struct extent_map *em,
345 int modified)
346{
347 refcount_inc(&em->refs);
348 em->mod_start = em->start;
349 em->mod_len = em->len;
350
351 if (modified)
352 list_move(&em->list, &tree->modified_extents);
353 else
354 try_merge_map(tree, em);
355}
356
357static void extent_map_device_set_bits(struct extent_map *em, unsigned bits)
358{
359 struct map_lookup *map = em->map_lookup;
360 u64 stripe_size = em->orig_block_len;
361 int i;
362
363 for (i = 0; i < map->num_stripes; i++) {
364 struct btrfs_io_stripe *stripe = &map->stripes[i];
365 struct btrfs_device *device = stripe->dev;
366
367 set_extent_bits_nowait(&device->alloc_state, stripe->physical,
368 stripe->physical + stripe_size - 1, bits);
369 }
370}
371
372static void extent_map_device_clear_bits(struct extent_map *em, unsigned bits)
373{
374 struct map_lookup *map = em->map_lookup;
375 u64 stripe_size = em->orig_block_len;
376 int i;
377
378 for (i = 0; i < map->num_stripes; i++) {
379 struct btrfs_io_stripe *stripe = &map->stripes[i];
380 struct btrfs_device *device = stripe->dev;
381
382 __clear_extent_bit(&device->alloc_state, stripe->physical,
383 stripe->physical + stripe_size - 1, bits,
384 NULL, GFP_NOWAIT, NULL);
385 }
386}
387
388/*
389 * Add new extent map to the extent tree
390 *
391 * @tree: tree to insert new map in
392 * @em: map to insert
393 * @modified: indicate whether the given @em should be added to the
394 * modified list, which indicates the extent needs to be logged
395 *
396 * Insert @em into @tree or perform a simple forward/backward merge with
397 * existing mappings. The extent_map struct passed in will be inserted
398 * into the tree directly, with an additional reference taken, or a
399 * reference dropped if the merge attempt was successful.
400 */
401int add_extent_mapping(struct extent_map_tree *tree,
402 struct extent_map *em, int modified)
403{
404 int ret = 0;
405
406 lockdep_assert_held_write(&tree->lock);
407
408 ret = tree_insert(&tree->map, em);
409 if (ret)
410 goto out;
411
412 setup_extent_mapping(tree, em, modified);
413 if (test_bit(EXTENT_FLAG_FS_MAPPING, &em->flags)) {
414 extent_map_device_set_bits(em, CHUNK_ALLOCATED);
415 extent_map_device_clear_bits(em, CHUNK_TRIMMED);
416 }
417out:
418 return ret;
419}
420
421static struct extent_map *
422__lookup_extent_mapping(struct extent_map_tree *tree,
423 u64 start, u64 len, int strict)
424{
425 struct extent_map *em;
426 struct rb_node *rb_node;
427 struct rb_node *prev_or_next = NULL;
428 u64 end = range_end(start, len);
429
430 rb_node = __tree_search(&tree->map.rb_root, start, &prev_or_next);
431 if (!rb_node) {
432 if (prev_or_next)
433 rb_node = prev_or_next;
434 else
435 return NULL;
436 }
437
438 em = rb_entry(rb_node, struct extent_map, rb_node);
439
440 if (strict && !(end > em->start && start < extent_map_end(em)))
441 return NULL;
442
443 refcount_inc(&em->refs);
444 return em;
445}
446
447/*
448 * Lookup extent_map that intersects @start + @len range.
449 *
450 * @tree: tree to lookup in
451 * @start: byte offset to start the search
452 * @len: length of the lookup range
453 *
454 * Find and return the first extent_map struct in @tree that intersects the
455 * [start, len] range. There may be additional objects in the tree that
456 * intersect, so check the object returned carefully to make sure that no
457 * additional lookups are needed.
458 */
459struct extent_map *lookup_extent_mapping(struct extent_map_tree *tree,
460 u64 start, u64 len)
461{
462 return __lookup_extent_mapping(tree, start, len, 1);
463}
464
465/*
466 * Find a nearby extent map intersecting @start + @len (not an exact search).
467 *
468 * @tree: tree to lookup in
469 * @start: byte offset to start the search
470 * @len: length of the lookup range
471 *
472 * Find and return the first extent_map struct in @tree that intersects the
473 * [start, len] range.
474 *
475 * If one can't be found, any nearby extent may be returned
476 */
477struct extent_map *search_extent_mapping(struct extent_map_tree *tree,
478 u64 start, u64 len)
479{
480 return __lookup_extent_mapping(tree, start, len, 0);
481}
482
483/*
484 * Remove an extent_map from the extent tree.
485 *
486 * @tree: extent tree to remove from
487 * @em: extent map being removed
488 *
489 * Remove @em from @tree. No reference counts are dropped, and no checks
490 * are done to see if the range is in use.
491 */
492void remove_extent_mapping(struct extent_map_tree *tree, struct extent_map *em)
493{
494 lockdep_assert_held_write(&tree->lock);
495
496 WARN_ON(test_bit(EXTENT_FLAG_PINNED, &em->flags));
497 rb_erase_cached(&em->rb_node, &tree->map);
498 if (!test_bit(EXTENT_FLAG_LOGGING, &em->flags))
499 list_del_init(&em->list);
500 if (test_bit(EXTENT_FLAG_FS_MAPPING, &em->flags))
501 extent_map_device_clear_bits(em, CHUNK_ALLOCATED);
502 RB_CLEAR_NODE(&em->rb_node);
503}
504
505void replace_extent_mapping(struct extent_map_tree *tree,
506 struct extent_map *cur,
507 struct extent_map *new,
508 int modified)
509{
510 lockdep_assert_held_write(&tree->lock);
511
512 WARN_ON(test_bit(EXTENT_FLAG_PINNED, &cur->flags));
513 ASSERT(extent_map_in_tree(cur));
514 if (!test_bit(EXTENT_FLAG_LOGGING, &cur->flags))
515 list_del_init(&cur->list);
516 rb_replace_node_cached(&cur->rb_node, &new->rb_node, &tree->map);
517 RB_CLEAR_NODE(&cur->rb_node);
518
519 setup_extent_mapping(tree, new, modified);
520}
521
522static struct extent_map *next_extent_map(const struct extent_map *em)
523{
524 struct rb_node *next;
525
526 next = rb_next(&em->rb_node);
527 if (!next)
528 return NULL;
529 return container_of(next, struct extent_map, rb_node);
530}
531
532static struct extent_map *prev_extent_map(struct extent_map *em)
533{
534 struct rb_node *prev;
535
536 prev = rb_prev(&em->rb_node);
537 if (!prev)
538 return NULL;
539 return container_of(prev, struct extent_map, rb_node);
540}
541
542/*
543 * Helper for btrfs_get_extent. Given an existing extent in the tree,
544 * the existing extent is the nearest extent to map_start,
545 * and an extent that you want to insert, deal with overlap and insert
546 * the best fitted new extent into the tree.
547 */
548static noinline int merge_extent_mapping(struct extent_map_tree *em_tree,
549 struct extent_map *existing,
550 struct extent_map *em,
551 u64 map_start)
552{
553 struct extent_map *prev;
554 struct extent_map *next;
555 u64 start;
556 u64 end;
557 u64 start_diff;
558
559 BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
560
561 if (existing->start > map_start) {
562 next = existing;
563 prev = prev_extent_map(next);
564 } else {
565 prev = existing;
566 next = next_extent_map(prev);
567 }
568
569 start = prev ? extent_map_end(prev) : em->start;
570 start = max_t(u64, start, em->start);
571 end = next ? next->start : extent_map_end(em);
572 end = min_t(u64, end, extent_map_end(em));
573 start_diff = start - em->start;
574 em->start = start;
575 em->len = end - start;
576 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
577 !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
578 em->block_start += start_diff;
579 em->block_len = em->len;
580 }
581 return add_extent_mapping(em_tree, em, 0);
582}
583
584/*
585 * Add extent mapping into em_tree.
586 *
587 * @fs_info: the filesystem
588 * @em_tree: extent tree into which we want to insert the extent mapping
589 * @em_in: extent we are inserting
590 * @start: start of the logical range btrfs_get_extent() is requesting
591 * @len: length of the logical range btrfs_get_extent() is requesting
592 *
593 * Note that @em_in's range may be different from [start, start+len),
594 * but they must be overlapped.
595 *
596 * Insert @em_in into @em_tree. In case there is an overlapping range, handle
597 * the -EEXIST by either:
598 * a) Returning the existing extent in @em_in if @start is within the
599 * existing em.
600 * b) Merge the existing extent with @em_in passed in.
601 *
602 * Return 0 on success, otherwise -EEXIST.
603 *
604 */
605int btrfs_add_extent_mapping(struct btrfs_fs_info *fs_info,
606 struct extent_map_tree *em_tree,
607 struct extent_map **em_in, u64 start, u64 len)
608{
609 int ret;
610 struct extent_map *em = *em_in;
611
612 /*
613 * Tree-checker should have rejected any inline extent with non-zero
614 * file offset. Here just do a sanity check.
615 */
616 if (em->block_start == EXTENT_MAP_INLINE)
617 ASSERT(em->start == 0);
618
619 ret = add_extent_mapping(em_tree, em, 0);
620 /* it is possible that someone inserted the extent into the tree
621 * while we had the lock dropped. It is also possible that
622 * an overlapping map exists in the tree
623 */
624 if (ret == -EEXIST) {
625 struct extent_map *existing;
626
627 ret = 0;
628
629 existing = search_extent_mapping(em_tree, start, len);
630
631 trace_btrfs_handle_em_exist(fs_info, existing, em, start, len);
632
633 /*
634 * existing will always be non-NULL, since there must be
635 * extent causing the -EEXIST.
636 */
637 if (start >= existing->start &&
638 start < extent_map_end(existing)) {
639 free_extent_map(em);
640 *em_in = existing;
641 ret = 0;
642 } else {
643 u64 orig_start = em->start;
644 u64 orig_len = em->len;
645
646 /*
647 * The existing extent map is the one nearest to
648 * the [start, start + len) range which overlaps
649 */
650 ret = merge_extent_mapping(em_tree, existing,
651 em, start);
652 if (ret) {
653 free_extent_map(em);
654 *em_in = NULL;
655 WARN_ONCE(ret,
656"unexpected error %d: merge existing(start %llu len %llu) with em(start %llu len %llu)\n",
657 ret, existing->start, existing->len,
658 orig_start, orig_len);
659 }
660 free_extent_map(existing);
661 }
662 }
663
664 ASSERT(ret == 0 || ret == -EEXIST);
665 return ret;
666}
667
668/*
669 * Drop all extent maps from a tree in the fastest possible way, rescheduling
670 * if needed. This avoids searching the tree, from the root down to the first
671 * extent map, before each deletion.
672 */
673static void drop_all_extent_maps_fast(struct extent_map_tree *tree)
674{
675 write_lock(&tree->lock);
676 while (!RB_EMPTY_ROOT(&tree->map.rb_root)) {
677 struct extent_map *em;
678 struct rb_node *node;
679
680 node = rb_first_cached(&tree->map);
681 em = rb_entry(node, struct extent_map, rb_node);
682 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
683 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
684 remove_extent_mapping(tree, em);
685 free_extent_map(em);
686 cond_resched_rwlock_write(&tree->lock);
687 }
688 write_unlock(&tree->lock);
689}
690
691/*
692 * Drop all extent maps in a given range.
693 *
694 * @inode: The target inode.
695 * @start: Start offset of the range.
696 * @end: End offset of the range (inclusive value).
697 * @skip_pinned: Indicate if pinned extent maps should be ignored or not.
698 *
699 * This drops all the extent maps that intersect the given range [@start, @end].
700 * Extent maps that partially overlap the range and extend behind or beyond it,
701 * are split.
702 * The caller should have locked an appropriate file range in the inode's io
703 * tree before calling this function.
704 */
705void btrfs_drop_extent_map_range(struct btrfs_inode *inode, u64 start, u64 end,
706 bool skip_pinned)
707{
708 struct extent_map *split;
709 struct extent_map *split2;
710 struct extent_map *em;
711 struct extent_map_tree *em_tree = &inode->extent_tree;
712 u64 len = end - start + 1;
713
714 WARN_ON(end < start);
715 if (end == (u64)-1) {
716 if (start == 0 && !skip_pinned) {
717 drop_all_extent_maps_fast(em_tree);
718 return;
719 }
720 len = (u64)-1;
721 } else {
722 /* Make end offset exclusive for use in the loop below. */
723 end++;
724 }
725
726 /*
727 * It's ok if we fail to allocate the extent maps, see the comment near
728 * the bottom of the loop below. We only need two spare extent maps in
729 * the worst case, where the first extent map that intersects our range
730 * starts before the range and the last extent map that intersects our
731 * range ends after our range (and they might be the same extent map),
732 * because we need to split those two extent maps at the boundaries.
733 */
734 split = alloc_extent_map();
735 split2 = alloc_extent_map();
736
737 write_lock(&em_tree->lock);
738 em = lookup_extent_mapping(em_tree, start, len);
739
740 while (em) {
741 /* extent_map_end() returns exclusive value (last byte + 1). */
742 const u64 em_end = extent_map_end(em);
743 struct extent_map *next_em = NULL;
744 u64 gen;
745 unsigned long flags;
746 bool modified;
747 bool compressed;
748
749 if (em_end < end) {
750 next_em = next_extent_map(em);
751 if (next_em) {
752 if (next_em->start < end)
753 refcount_inc(&next_em->refs);
754 else
755 next_em = NULL;
756 }
757 }
758
759 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
760 start = em_end;
761 if (end != (u64)-1)
762 len = start + len - em_end;
763 goto next;
764 }
765
766 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
767 clear_bit(EXTENT_FLAG_LOGGING, &flags);
768 modified = !list_empty(&em->list);
769
770 /*
771 * The extent map does not cross our target range, so no need to
772 * split it, we can remove it directly.
773 */
774 if (em->start >= start && em_end <= end)
775 goto remove_em;
776
777 flags = em->flags;
778 gen = em->generation;
779 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
780
781 if (em->start < start) {
782 if (!split) {
783 split = split2;
784 split2 = NULL;
785 if (!split)
786 goto remove_em;
787 }
788 split->start = em->start;
789 split->len = start - em->start;
790
791 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
792 split->orig_start = em->orig_start;
793 split->block_start = em->block_start;
794
795 if (compressed)
796 split->block_len = em->block_len;
797 else
798 split->block_len = split->len;
799 split->orig_block_len = max(split->block_len,
800 em->orig_block_len);
801 split->ram_bytes = em->ram_bytes;
802 } else {
803 split->orig_start = split->start;
804 split->block_len = 0;
805 split->block_start = em->block_start;
806 split->orig_block_len = 0;
807 split->ram_bytes = split->len;
808 }
809
810 split->generation = gen;
811 split->flags = flags;
812 split->compress_type = em->compress_type;
813 replace_extent_mapping(em_tree, em, split, modified);
814 free_extent_map(split);
815 split = split2;
816 split2 = NULL;
817 }
818 if (em_end > end) {
819 if (!split) {
820 split = split2;
821 split2 = NULL;
822 if (!split)
823 goto remove_em;
824 }
825 split->start = start + len;
826 split->len = em_end - (start + len);
827 split->block_start = em->block_start;
828 split->flags = flags;
829 split->compress_type = em->compress_type;
830 split->generation = gen;
831
832 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
833 split->orig_block_len = max(em->block_len,
834 em->orig_block_len);
835
836 split->ram_bytes = em->ram_bytes;
837 if (compressed) {
838 split->block_len = em->block_len;
839 split->orig_start = em->orig_start;
840 } else {
841 const u64 diff = start + len - em->start;
842
843 split->block_len = split->len;
844 split->block_start += diff;
845 split->orig_start = em->orig_start;
846 }
847 } else {
848 split->ram_bytes = split->len;
849 split->orig_start = split->start;
850 split->block_len = 0;
851 split->orig_block_len = 0;
852 }
853
854 if (extent_map_in_tree(em)) {
855 replace_extent_mapping(em_tree, em, split,
856 modified);
857 } else {
858 int ret;
859
860 ret = add_extent_mapping(em_tree, split,
861 modified);
862 /* Logic error, shouldn't happen. */
863 ASSERT(ret == 0);
864 if (WARN_ON(ret != 0) && modified)
865 btrfs_set_inode_full_sync(inode);
866 }
867 free_extent_map(split);
868 split = NULL;
869 }
870remove_em:
871 if (extent_map_in_tree(em)) {
872 /*
873 * If the extent map is still in the tree it means that
874 * either of the following is true:
875 *
876 * 1) It fits entirely in our range (doesn't end beyond
877 * it or starts before it);
878 *
879 * 2) It starts before our range and/or ends after our
880 * range, and we were not able to allocate the extent
881 * maps for split operations, @split and @split2.
882 *
883 * If we are at case 2) then we just remove the entire
884 * extent map - this is fine since if anyone needs it to
885 * access the subranges outside our range, will just
886 * load it again from the subvolume tree's file extent
887 * item. However if the extent map was in the list of
888 * modified extents, then we must mark the inode for a
889 * full fsync, otherwise a fast fsync will miss this
890 * extent if it's new and needs to be logged.
891 */
892 if ((em->start < start || em_end > end) && modified) {
893 ASSERT(!split);
894 btrfs_set_inode_full_sync(inode);
895 }
896 remove_extent_mapping(em_tree, em);
897 }
898
899 /*
900 * Once for the tree reference (we replaced or removed the
901 * extent map from the tree).
902 */
903 free_extent_map(em);
904next:
905 /* Once for us (for our lookup reference). */
906 free_extent_map(em);
907
908 em = next_em;
909 }
910
911 write_unlock(&em_tree->lock);
912
913 free_extent_map(split);
914 free_extent_map(split2);
915}
916
917/*
918 * Replace a range in the inode's extent map tree with a new extent map.
919 *
920 * @inode: The target inode.
921 * @new_em: The new extent map to add to the inode's extent map tree.
922 * @modified: Indicate if the new extent map should be added to the list of
923 * modified extents (for fast fsync tracking).
924 *
925 * Drops all the extent maps in the inode's extent map tree that intersect the
926 * range of the new extent map and adds the new extent map to the tree.
927 * The caller should have locked an appropriate file range in the inode's io
928 * tree before calling this function.
929 */
930int btrfs_replace_extent_map_range(struct btrfs_inode *inode,
931 struct extent_map *new_em,
932 bool modified)
933{
934 const u64 end = new_em->start + new_em->len - 1;
935 struct extent_map_tree *tree = &inode->extent_tree;
936 int ret;
937
938 ASSERT(!extent_map_in_tree(new_em));
939
940 /*
941 * The caller has locked an appropriate file range in the inode's io
942 * tree, but getting -EEXIST when adding the new extent map can still
943 * happen in case there are extents that partially cover the range, and
944 * this is due to two tasks operating on different parts of the extent.
945 * See commit 18e83ac75bfe67 ("Btrfs: fix unexpected EEXIST from
946 * btrfs_get_extent") for an example and details.
947 */
948 do {
949 btrfs_drop_extent_map_range(inode, new_em->start, end, false);
950 write_lock(&tree->lock);
951 ret = add_extent_mapping(tree, new_em, modified);
952 write_unlock(&tree->lock);
953 } while (ret == -EEXIST);
954
955 return ret;
956}
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/err.h>
4#include <linux/slab.h>
5#include <linux/spinlock.h>
6#include "messages.h"
7#include "ctree.h"
8#include "extent_map.h"
9#include "compression.h"
10#include "btrfs_inode.h"
11#include "disk-io.h"
12
13
14static struct kmem_cache *extent_map_cache;
15
16int __init extent_map_init(void)
17{
18 extent_map_cache = kmem_cache_create("btrfs_extent_map",
19 sizeof(struct extent_map), 0, 0, NULL);
20 if (!extent_map_cache)
21 return -ENOMEM;
22 return 0;
23}
24
25void __cold extent_map_exit(void)
26{
27 kmem_cache_destroy(extent_map_cache);
28}
29
30/*
31 * Initialize the extent tree @tree. Should be called for each new inode or
32 * other user of the extent_map interface.
33 */
34void extent_map_tree_init(struct extent_map_tree *tree)
35{
36 tree->root = RB_ROOT;
37 INIT_LIST_HEAD(&tree->modified_extents);
38 rwlock_init(&tree->lock);
39}
40
41/*
42 * Allocate a new extent_map structure. The new structure is returned with a
43 * reference count of one and needs to be freed using free_extent_map()
44 */
45struct extent_map *alloc_extent_map(void)
46{
47 struct extent_map *em;
48 em = kmem_cache_zalloc(extent_map_cache, GFP_NOFS);
49 if (!em)
50 return NULL;
51 RB_CLEAR_NODE(&em->rb_node);
52 refcount_set(&em->refs, 1);
53 INIT_LIST_HEAD(&em->list);
54 return em;
55}
56
57/*
58 * Drop the reference out on @em by one and free the structure if the reference
59 * count hits zero.
60 */
61void free_extent_map(struct extent_map *em)
62{
63 if (!em)
64 return;
65 if (refcount_dec_and_test(&em->refs)) {
66 WARN_ON(extent_map_in_tree(em));
67 WARN_ON(!list_empty(&em->list));
68 kmem_cache_free(extent_map_cache, em);
69 }
70}
71
72/* Do the math around the end of an extent, handling wrapping. */
73static u64 range_end(u64 start, u64 len)
74{
75 if (start + len < start)
76 return (u64)-1;
77 return start + len;
78}
79
80static void remove_em(struct btrfs_inode *inode, struct extent_map *em)
81{
82 struct btrfs_fs_info *fs_info = inode->root->fs_info;
83
84 rb_erase(&em->rb_node, &inode->extent_tree.root);
85 RB_CLEAR_NODE(&em->rb_node);
86
87 if (!btrfs_is_testing(fs_info) && is_fstree(btrfs_root_id(inode->root)))
88 percpu_counter_dec(&fs_info->evictable_extent_maps);
89}
90
91static int tree_insert(struct rb_root *root, struct extent_map *em)
92{
93 struct rb_node **p = &root->rb_node;
94 struct rb_node *parent = NULL;
95 struct extent_map *entry = NULL;
96 struct rb_node *orig_parent = NULL;
97 u64 end = range_end(em->start, em->len);
98
99 while (*p) {
100 parent = *p;
101 entry = rb_entry(parent, struct extent_map, rb_node);
102
103 if (em->start < entry->start)
104 p = &(*p)->rb_left;
105 else if (em->start >= extent_map_end(entry))
106 p = &(*p)->rb_right;
107 else
108 return -EEXIST;
109 }
110
111 orig_parent = parent;
112 while (parent && em->start >= extent_map_end(entry)) {
113 parent = rb_next(parent);
114 entry = rb_entry(parent, struct extent_map, rb_node);
115 }
116 if (parent)
117 if (end > entry->start && em->start < extent_map_end(entry))
118 return -EEXIST;
119
120 parent = orig_parent;
121 entry = rb_entry(parent, struct extent_map, rb_node);
122 while (parent && em->start < entry->start) {
123 parent = rb_prev(parent);
124 entry = rb_entry(parent, struct extent_map, rb_node);
125 }
126 if (parent)
127 if (end > entry->start && em->start < extent_map_end(entry))
128 return -EEXIST;
129
130 rb_link_node(&em->rb_node, orig_parent, p);
131 rb_insert_color(&em->rb_node, root);
132 return 0;
133}
134
135/*
136 * Search through the tree for an extent_map with a given offset. If it can't
137 * be found, try to find some neighboring extents
138 */
139static struct rb_node *__tree_search(struct rb_root *root, u64 offset,
140 struct rb_node **prev_or_next_ret)
141{
142 struct rb_node *n = root->rb_node;
143 struct rb_node *prev = NULL;
144 struct rb_node *orig_prev = NULL;
145 struct extent_map *entry;
146 struct extent_map *prev_entry = NULL;
147
148 ASSERT(prev_or_next_ret);
149
150 while (n) {
151 entry = rb_entry(n, struct extent_map, rb_node);
152 prev = n;
153 prev_entry = entry;
154
155 if (offset < entry->start)
156 n = n->rb_left;
157 else if (offset >= extent_map_end(entry))
158 n = n->rb_right;
159 else
160 return n;
161 }
162
163 orig_prev = prev;
164 while (prev && offset >= extent_map_end(prev_entry)) {
165 prev = rb_next(prev);
166 prev_entry = rb_entry(prev, struct extent_map, rb_node);
167 }
168
169 /*
170 * Previous extent map found, return as in this case the caller does not
171 * care about the next one.
172 */
173 if (prev) {
174 *prev_or_next_ret = prev;
175 return NULL;
176 }
177
178 prev = orig_prev;
179 prev_entry = rb_entry(prev, struct extent_map, rb_node);
180 while (prev && offset < prev_entry->start) {
181 prev = rb_prev(prev);
182 prev_entry = rb_entry(prev, struct extent_map, rb_node);
183 }
184 *prev_or_next_ret = prev;
185
186 return NULL;
187}
188
189static inline u64 extent_map_block_len(const struct extent_map *em)
190{
191 if (extent_map_is_compressed(em))
192 return em->disk_num_bytes;
193 return em->len;
194}
195
196static inline u64 extent_map_block_end(const struct extent_map *em)
197{
198 const u64 block_start = extent_map_block_start(em);
199 const u64 block_end = block_start + extent_map_block_len(em);
200
201 if (block_end < block_start)
202 return (u64)-1;
203
204 return block_end;
205}
206
207static bool can_merge_extent_map(const struct extent_map *em)
208{
209 if (em->flags & EXTENT_FLAG_PINNED)
210 return false;
211
212 /* Don't merge compressed extents, we need to know their actual size. */
213 if (extent_map_is_compressed(em))
214 return false;
215
216 if (em->flags & EXTENT_FLAG_LOGGING)
217 return false;
218
219 /*
220 * We don't want to merge stuff that hasn't been written to the log yet
221 * since it may not reflect exactly what is on disk, and that would be
222 * bad.
223 */
224 if (!list_empty(&em->list))
225 return false;
226
227 return true;
228}
229
230/* Check to see if two extent_map structs are adjacent and safe to merge. */
231static bool mergeable_maps(const struct extent_map *prev, const struct extent_map *next)
232{
233 if (extent_map_end(prev) != next->start)
234 return false;
235
236 /*
237 * The merged flag is not an on-disk flag, it just indicates we had the
238 * extent maps of 2 (or more) adjacent extents merged, so factor it out.
239 */
240 if ((prev->flags & ~EXTENT_FLAG_MERGED) !=
241 (next->flags & ~EXTENT_FLAG_MERGED))
242 return false;
243
244 if (next->disk_bytenr < EXTENT_MAP_LAST_BYTE - 1)
245 return extent_map_block_start(next) == extent_map_block_end(prev);
246
247 /* HOLES and INLINE extents. */
248 return next->disk_bytenr == prev->disk_bytenr;
249}
250
251/*
252 * Handle the on-disk data extents merge for @prev and @next.
253 *
254 * @prev: left extent to merge
255 * @next: right extent to merge
256 * @merged: the extent we will not discard after the merge; updated with new values
257 *
258 * After this, one of the two extents is the new merged extent and the other is
259 * removed from the tree and likely freed. Note that @merged is one of @prev/@next
260 * so there is const/non-const aliasing occurring here.
261 *
262 * Only touches disk_bytenr/disk_num_bytes/offset/ram_bytes.
263 * For now only uncompressed regular extent can be merged.
264 */
265static void merge_ondisk_extents(const struct extent_map *prev, const struct extent_map *next,
266 struct extent_map *merged)
267{
268 u64 new_disk_bytenr;
269 u64 new_disk_num_bytes;
270 u64 new_offset;
271
272 /* @prev and @next should not be compressed. */
273 ASSERT(!extent_map_is_compressed(prev));
274 ASSERT(!extent_map_is_compressed(next));
275
276 /*
277 * There are two different cases where @prev and @next can be merged.
278 *
279 * 1) They are referring to the same data extent:
280 *
281 * |<----- data extent A ----->|
282 * |<- prev ->|<- next ->|
283 *
284 * 2) They are referring to different data extents but still adjacent:
285 *
286 * |<-- data extent A -->|<-- data extent B -->|
287 * |<- prev ->|<- next ->|
288 *
289 * The calculation here always merges the data extents first, then updates
290 * @offset using the new data extents.
291 *
292 * For case 1), the merged data extent would be the same.
293 * For case 2), we just merge the two data extents into one.
294 */
295 new_disk_bytenr = min(prev->disk_bytenr, next->disk_bytenr);
296 new_disk_num_bytes = max(prev->disk_bytenr + prev->disk_num_bytes,
297 next->disk_bytenr + next->disk_num_bytes) -
298 new_disk_bytenr;
299 new_offset = prev->disk_bytenr + prev->offset - new_disk_bytenr;
300
301 merged->disk_bytenr = new_disk_bytenr;
302 merged->disk_num_bytes = new_disk_num_bytes;
303 merged->ram_bytes = new_disk_num_bytes;
304 merged->offset = new_offset;
305}
306
307static void dump_extent_map(struct btrfs_fs_info *fs_info, const char *prefix,
308 struct extent_map *em)
309{
310 if (!IS_ENABLED(CONFIG_BTRFS_DEBUG))
311 return;
312 btrfs_crit(fs_info,
313"%s, start=%llu len=%llu disk_bytenr=%llu disk_num_bytes=%llu ram_bytes=%llu offset=%llu flags=0x%x",
314 prefix, em->start, em->len, em->disk_bytenr, em->disk_num_bytes,
315 em->ram_bytes, em->offset, em->flags);
316 ASSERT(0);
317}
318
319/* Internal sanity checks for btrfs debug builds. */
320static void validate_extent_map(struct btrfs_fs_info *fs_info, struct extent_map *em)
321{
322 if (!IS_ENABLED(CONFIG_BTRFS_DEBUG))
323 return;
324 if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE) {
325 if (em->disk_num_bytes == 0)
326 dump_extent_map(fs_info, "zero disk_num_bytes", em);
327 if (em->offset + em->len > em->ram_bytes)
328 dump_extent_map(fs_info, "ram_bytes too small", em);
329 if (em->offset + em->len > em->disk_num_bytes &&
330 !extent_map_is_compressed(em))
331 dump_extent_map(fs_info, "disk_num_bytes too small", em);
332 if (!extent_map_is_compressed(em) &&
333 em->ram_bytes != em->disk_num_bytes)
334 dump_extent_map(fs_info,
335 "ram_bytes mismatch with disk_num_bytes for non-compressed em",
336 em);
337 } else if (em->offset) {
338 dump_extent_map(fs_info, "non-zero offset for hole/inline", em);
339 }
340}
341
342static void try_merge_map(struct btrfs_inode *inode, struct extent_map *em)
343{
344 struct btrfs_fs_info *fs_info = inode->root->fs_info;
345 struct extent_map *merge = NULL;
346 struct rb_node *rb;
347
348 /*
349 * We can't modify an extent map that is in the tree and that is being
350 * used by another task, as it can cause that other task to see it in
351 * inconsistent state during the merging. We always have 1 reference for
352 * the tree and 1 for this task (which is unpinning the extent map or
353 * clearing the logging flag), so anything > 2 means it's being used by
354 * other tasks too.
355 */
356 if (refcount_read(&em->refs) > 2)
357 return;
358
359 if (!can_merge_extent_map(em))
360 return;
361
362 if (em->start != 0) {
363 rb = rb_prev(&em->rb_node);
364 if (rb)
365 merge = rb_entry(rb, struct extent_map, rb_node);
366 if (rb && can_merge_extent_map(merge) && mergeable_maps(merge, em)) {
367 em->start = merge->start;
368 em->len += merge->len;
369 em->generation = max(em->generation, merge->generation);
370
371 if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE)
372 merge_ondisk_extents(merge, em, em);
373 em->flags |= EXTENT_FLAG_MERGED;
374
375 validate_extent_map(fs_info, em);
376 remove_em(inode, merge);
377 free_extent_map(merge);
378 }
379 }
380
381 rb = rb_next(&em->rb_node);
382 if (rb)
383 merge = rb_entry(rb, struct extent_map, rb_node);
384 if (rb && can_merge_extent_map(merge) && mergeable_maps(em, merge)) {
385 em->len += merge->len;
386 if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE)
387 merge_ondisk_extents(em, merge, em);
388 validate_extent_map(fs_info, em);
389 em->generation = max(em->generation, merge->generation);
390 em->flags |= EXTENT_FLAG_MERGED;
391 remove_em(inode, merge);
392 free_extent_map(merge);
393 }
394}
395
396/*
397 * Unpin an extent from the cache.
398 *
399 * @inode: the inode from which we are unpinning an extent range
400 * @start: logical offset in the file
401 * @len: length of the extent
402 * @gen: generation that this extent has been modified in
403 *
404 * Called after an extent has been written to disk properly. Set the generation
405 * to the generation that actually added the file item to the inode so we know
406 * we need to sync this extent when we call fsync().
407 *
408 * Returns: 0 on success
409 * -ENOENT when the extent is not found in the tree
410 * -EUCLEAN if the found extent does not match the expected start
411 */
412int unpin_extent_cache(struct btrfs_inode *inode, u64 start, u64 len, u64 gen)
413{
414 struct btrfs_fs_info *fs_info = inode->root->fs_info;
415 struct extent_map_tree *tree = &inode->extent_tree;
416 int ret = 0;
417 struct extent_map *em;
418
419 write_lock(&tree->lock);
420 em = lookup_extent_mapping(tree, start, len);
421
422 if (WARN_ON(!em)) {
423 btrfs_warn(fs_info,
424"no extent map found for inode %llu (root %lld) when unpinning extent range [%llu, %llu), generation %llu",
425 btrfs_ino(inode), btrfs_root_id(inode->root),
426 start, start + len, gen);
427 ret = -ENOENT;
428 goto out;
429 }
430
431 if (WARN_ON(em->start != start)) {
432 btrfs_warn(fs_info,
433"found extent map for inode %llu (root %lld) with unexpected start offset %llu when unpinning extent range [%llu, %llu), generation %llu",
434 btrfs_ino(inode), btrfs_root_id(inode->root),
435 em->start, start, start + len, gen);
436 ret = -EUCLEAN;
437 goto out;
438 }
439
440 em->generation = gen;
441 em->flags &= ~EXTENT_FLAG_PINNED;
442
443 try_merge_map(inode, em);
444
445out:
446 write_unlock(&tree->lock);
447 free_extent_map(em);
448 return ret;
449
450}
451
452void clear_em_logging(struct btrfs_inode *inode, struct extent_map *em)
453{
454 lockdep_assert_held_write(&inode->extent_tree.lock);
455
456 em->flags &= ~EXTENT_FLAG_LOGGING;
457 if (extent_map_in_tree(em))
458 try_merge_map(inode, em);
459}
460
461static inline void setup_extent_mapping(struct btrfs_inode *inode,
462 struct extent_map *em,
463 int modified)
464{
465 refcount_inc(&em->refs);
466
467 ASSERT(list_empty(&em->list));
468
469 if (modified)
470 list_add(&em->list, &inode->extent_tree.modified_extents);
471 else
472 try_merge_map(inode, em);
473}
474
475/*
476 * Add a new extent map to an inode's extent map tree.
477 *
478 * @inode: the target inode
479 * @em: map to insert
480 * @modified: indicate whether the given @em should be added to the
481 * modified list, which indicates the extent needs to be logged
482 *
483 * Insert @em into the @inode's extent map tree or perform a simple
484 * forward/backward merge with existing mappings. The extent_map struct passed
485 * in will be inserted into the tree directly, with an additional reference
486 * taken, or a reference dropped if the merge attempt was successful.
487 */
488static int add_extent_mapping(struct btrfs_inode *inode,
489 struct extent_map *em, int modified)
490{
491 struct extent_map_tree *tree = &inode->extent_tree;
492 struct btrfs_root *root = inode->root;
493 struct btrfs_fs_info *fs_info = root->fs_info;
494 int ret;
495
496 lockdep_assert_held_write(&tree->lock);
497
498 validate_extent_map(fs_info, em);
499 ret = tree_insert(&tree->root, em);
500 if (ret)
501 return ret;
502
503 setup_extent_mapping(inode, em, modified);
504
505 if (!btrfs_is_testing(fs_info) && is_fstree(btrfs_root_id(root)))
506 percpu_counter_inc(&fs_info->evictable_extent_maps);
507
508 return 0;
509}
510
511static struct extent_map *
512__lookup_extent_mapping(struct extent_map_tree *tree,
513 u64 start, u64 len, int strict)
514{
515 struct extent_map *em;
516 struct rb_node *rb_node;
517 struct rb_node *prev_or_next = NULL;
518 u64 end = range_end(start, len);
519
520 rb_node = __tree_search(&tree->root, start, &prev_or_next);
521 if (!rb_node) {
522 if (prev_or_next)
523 rb_node = prev_or_next;
524 else
525 return NULL;
526 }
527
528 em = rb_entry(rb_node, struct extent_map, rb_node);
529
530 if (strict && !(end > em->start && start < extent_map_end(em)))
531 return NULL;
532
533 refcount_inc(&em->refs);
534 return em;
535}
536
537/*
538 * Lookup extent_map that intersects @start + @len range.
539 *
540 * @tree: tree to lookup in
541 * @start: byte offset to start the search
542 * @len: length of the lookup range
543 *
544 * Find and return the first extent_map struct in @tree that intersects the
545 * [start, len] range. There may be additional objects in the tree that
546 * intersect, so check the object returned carefully to make sure that no
547 * additional lookups are needed.
548 */
549struct extent_map *lookup_extent_mapping(struct extent_map_tree *tree,
550 u64 start, u64 len)
551{
552 return __lookup_extent_mapping(tree, start, len, 1);
553}
554
555/*
556 * Find a nearby extent map intersecting @start + @len (not an exact search).
557 *
558 * @tree: tree to lookup in
559 * @start: byte offset to start the search
560 * @len: length of the lookup range
561 *
562 * Find and return the first extent_map struct in @tree that intersects the
563 * [start, len] range.
564 *
565 * If one can't be found, any nearby extent may be returned
566 */
567struct extent_map *search_extent_mapping(struct extent_map_tree *tree,
568 u64 start, u64 len)
569{
570 return __lookup_extent_mapping(tree, start, len, 0);
571}
572
573/*
574 * Remove an extent_map from its inode's extent tree.
575 *
576 * @inode: the inode the extent map belongs to
577 * @em: extent map being removed
578 *
579 * Remove @em from the extent tree of @inode. No reference counts are dropped,
580 * and no checks are done to see if the range is in use.
581 */
582void remove_extent_mapping(struct btrfs_inode *inode, struct extent_map *em)
583{
584 struct extent_map_tree *tree = &inode->extent_tree;
585
586 lockdep_assert_held_write(&tree->lock);
587
588 WARN_ON(em->flags & EXTENT_FLAG_PINNED);
589 if (!(em->flags & EXTENT_FLAG_LOGGING))
590 list_del_init(&em->list);
591
592 remove_em(inode, em);
593}
594
595static void replace_extent_mapping(struct btrfs_inode *inode,
596 struct extent_map *cur,
597 struct extent_map *new,
598 int modified)
599{
600 struct btrfs_fs_info *fs_info = inode->root->fs_info;
601 struct extent_map_tree *tree = &inode->extent_tree;
602
603 lockdep_assert_held_write(&tree->lock);
604
605 validate_extent_map(fs_info, new);
606
607 WARN_ON(cur->flags & EXTENT_FLAG_PINNED);
608 ASSERT(extent_map_in_tree(cur));
609 if (!(cur->flags & EXTENT_FLAG_LOGGING))
610 list_del_init(&cur->list);
611 rb_replace_node(&cur->rb_node, &new->rb_node, &tree->root);
612 RB_CLEAR_NODE(&cur->rb_node);
613
614 setup_extent_mapping(inode, new, modified);
615}
616
617static struct extent_map *next_extent_map(const struct extent_map *em)
618{
619 struct rb_node *next;
620
621 next = rb_next(&em->rb_node);
622 if (!next)
623 return NULL;
624 return container_of(next, struct extent_map, rb_node);
625}
626
627static struct extent_map *prev_extent_map(struct extent_map *em)
628{
629 struct rb_node *prev;
630
631 prev = rb_prev(&em->rb_node);
632 if (!prev)
633 return NULL;
634 return container_of(prev, struct extent_map, rb_node);
635}
636
637/*
638 * Helper for btrfs_get_extent. Given an existing extent in the tree,
639 * the existing extent is the nearest extent to map_start,
640 * and an extent that you want to insert, deal with overlap and insert
641 * the best fitted new extent into the tree.
642 */
643static noinline int merge_extent_mapping(struct btrfs_inode *inode,
644 struct extent_map *existing,
645 struct extent_map *em,
646 u64 map_start)
647{
648 struct extent_map *prev;
649 struct extent_map *next;
650 u64 start;
651 u64 end;
652 u64 start_diff;
653
654 if (map_start < em->start || map_start >= extent_map_end(em))
655 return -EINVAL;
656
657 if (existing->start > map_start) {
658 next = existing;
659 prev = prev_extent_map(next);
660 } else {
661 prev = existing;
662 next = next_extent_map(prev);
663 }
664
665 start = prev ? extent_map_end(prev) : em->start;
666 start = max_t(u64, start, em->start);
667 end = next ? next->start : extent_map_end(em);
668 end = min_t(u64, end, extent_map_end(em));
669 start_diff = start - em->start;
670 em->start = start;
671 em->len = end - start;
672 if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE)
673 em->offset += start_diff;
674 return add_extent_mapping(inode, em, 0);
675}
676
677/*
678 * Add extent mapping into an inode's extent map tree.
679 *
680 * @inode: target inode
681 * @em_in: extent we are inserting
682 * @start: start of the logical range btrfs_get_extent() is requesting
683 * @len: length of the logical range btrfs_get_extent() is requesting
684 *
685 * Note that @em_in's range may be different from [start, start+len),
686 * but they must be overlapped.
687 *
688 * Insert @em_in into the inode's extent map tree. In case there is an
689 * overlapping range, handle the -EEXIST by either:
690 * a) Returning the existing extent in @em_in if @start is within the
691 * existing em.
692 * b) Merge the existing extent with @em_in passed in.
693 *
694 * Return 0 on success, otherwise -EEXIST.
695 *
696 */
697int btrfs_add_extent_mapping(struct btrfs_inode *inode,
698 struct extent_map **em_in, u64 start, u64 len)
699{
700 int ret;
701 struct extent_map *em = *em_in;
702 struct btrfs_fs_info *fs_info = inode->root->fs_info;
703
704 /*
705 * Tree-checker should have rejected any inline extent with non-zero
706 * file offset. Here just do a sanity check.
707 */
708 if (em->disk_bytenr == EXTENT_MAP_INLINE)
709 ASSERT(em->start == 0);
710
711 ret = add_extent_mapping(inode, em, 0);
712 /* it is possible that someone inserted the extent into the tree
713 * while we had the lock dropped. It is also possible that
714 * an overlapping map exists in the tree
715 */
716 if (ret == -EEXIST) {
717 struct extent_map *existing;
718
719 existing = search_extent_mapping(&inode->extent_tree, start, len);
720
721 trace_btrfs_handle_em_exist(fs_info, existing, em, start, len);
722
723 /*
724 * existing will always be non-NULL, since there must be
725 * extent causing the -EEXIST.
726 */
727 if (start >= existing->start &&
728 start < extent_map_end(existing)) {
729 free_extent_map(em);
730 *em_in = existing;
731 ret = 0;
732 } else {
733 u64 orig_start = em->start;
734 u64 orig_len = em->len;
735
736 /*
737 * The existing extent map is the one nearest to
738 * the [start, start + len) range which overlaps
739 */
740 ret = merge_extent_mapping(inode, existing, em, start);
741 if (WARN_ON(ret)) {
742 free_extent_map(em);
743 *em_in = NULL;
744 btrfs_warn(fs_info,
745"extent map merge error existing [%llu, %llu) with em [%llu, %llu) start %llu",
746 existing->start, extent_map_end(existing),
747 orig_start, orig_start + orig_len, start);
748 }
749 free_extent_map(existing);
750 }
751 }
752
753 ASSERT(ret == 0 || ret == -EEXIST);
754 return ret;
755}
756
757/*
758 * Drop all extent maps from a tree in the fastest possible way, rescheduling
759 * if needed. This avoids searching the tree, from the root down to the first
760 * extent map, before each deletion.
761 */
762static void drop_all_extent_maps_fast(struct btrfs_inode *inode)
763{
764 struct extent_map_tree *tree = &inode->extent_tree;
765 struct rb_node *node;
766
767 write_lock(&tree->lock);
768 node = rb_first(&tree->root);
769 while (node) {
770 struct extent_map *em;
771 struct rb_node *next = rb_next(node);
772
773 em = rb_entry(node, struct extent_map, rb_node);
774 em->flags &= ~(EXTENT_FLAG_PINNED | EXTENT_FLAG_LOGGING);
775 remove_extent_mapping(inode, em);
776 free_extent_map(em);
777
778 if (cond_resched_rwlock_write(&tree->lock))
779 node = rb_first(&tree->root);
780 else
781 node = next;
782 }
783 write_unlock(&tree->lock);
784}
785
786/*
787 * Drop all extent maps in a given range.
788 *
789 * @inode: The target inode.
790 * @start: Start offset of the range.
791 * @end: End offset of the range (inclusive value).
792 * @skip_pinned: Indicate if pinned extent maps should be ignored or not.
793 *
794 * This drops all the extent maps that intersect the given range [@start, @end].
795 * Extent maps that partially overlap the range and extend behind or beyond it,
796 * are split.
797 * The caller should have locked an appropriate file range in the inode's io
798 * tree before calling this function.
799 */
800void btrfs_drop_extent_map_range(struct btrfs_inode *inode, u64 start, u64 end,
801 bool skip_pinned)
802{
803 struct extent_map *split;
804 struct extent_map *split2;
805 struct extent_map *em;
806 struct extent_map_tree *em_tree = &inode->extent_tree;
807 u64 len = end - start + 1;
808
809 WARN_ON(end < start);
810 if (end == (u64)-1) {
811 if (start == 0 && !skip_pinned) {
812 drop_all_extent_maps_fast(inode);
813 return;
814 }
815 len = (u64)-1;
816 } else {
817 /* Make end offset exclusive for use in the loop below. */
818 end++;
819 }
820
821 /*
822 * It's ok if we fail to allocate the extent maps, see the comment near
823 * the bottom of the loop below. We only need two spare extent maps in
824 * the worst case, where the first extent map that intersects our range
825 * starts before the range and the last extent map that intersects our
826 * range ends after our range (and they might be the same extent map),
827 * because we need to split those two extent maps at the boundaries.
828 */
829 split = alloc_extent_map();
830 split2 = alloc_extent_map();
831
832 write_lock(&em_tree->lock);
833 em = lookup_extent_mapping(em_tree, start, len);
834
835 while (em) {
836 /* extent_map_end() returns exclusive value (last byte + 1). */
837 const u64 em_end = extent_map_end(em);
838 struct extent_map *next_em = NULL;
839 u64 gen;
840 unsigned long flags;
841 bool modified;
842
843 if (em_end < end) {
844 next_em = next_extent_map(em);
845 if (next_em) {
846 if (next_em->start < end)
847 refcount_inc(&next_em->refs);
848 else
849 next_em = NULL;
850 }
851 }
852
853 if (skip_pinned && (em->flags & EXTENT_FLAG_PINNED)) {
854 start = em_end;
855 goto next;
856 }
857
858 flags = em->flags;
859 /*
860 * In case we split the extent map, we want to preserve the
861 * EXTENT_FLAG_LOGGING flag on our extent map, but we don't want
862 * it on the new extent maps.
863 */
864 em->flags &= ~(EXTENT_FLAG_PINNED | EXTENT_FLAG_LOGGING);
865 modified = !list_empty(&em->list);
866
867 /*
868 * The extent map does not cross our target range, so no need to
869 * split it, we can remove it directly.
870 */
871 if (em->start >= start && em_end <= end)
872 goto remove_em;
873
874 gen = em->generation;
875
876 if (em->start < start) {
877 if (!split) {
878 split = split2;
879 split2 = NULL;
880 if (!split)
881 goto remove_em;
882 }
883 split->start = em->start;
884 split->len = start - em->start;
885
886 if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE) {
887 split->disk_bytenr = em->disk_bytenr;
888 split->disk_num_bytes = em->disk_num_bytes;
889 split->offset = em->offset;
890 split->ram_bytes = em->ram_bytes;
891 } else {
892 split->disk_bytenr = em->disk_bytenr;
893 split->disk_num_bytes = 0;
894 split->offset = 0;
895 split->ram_bytes = split->len;
896 }
897
898 split->generation = gen;
899 split->flags = flags;
900 replace_extent_mapping(inode, em, split, modified);
901 free_extent_map(split);
902 split = split2;
903 split2 = NULL;
904 }
905 if (em_end > end) {
906 if (!split) {
907 split = split2;
908 split2 = NULL;
909 if (!split)
910 goto remove_em;
911 }
912 split->start = end;
913 split->len = em_end - end;
914 split->disk_bytenr = em->disk_bytenr;
915 split->flags = flags;
916 split->generation = gen;
917
918 if (em->disk_bytenr < EXTENT_MAP_LAST_BYTE) {
919 split->disk_num_bytes = em->disk_num_bytes;
920 split->offset = em->offset + end - em->start;
921 split->ram_bytes = em->ram_bytes;
922 } else {
923 split->disk_num_bytes = 0;
924 split->offset = 0;
925 split->ram_bytes = split->len;
926 }
927
928 if (extent_map_in_tree(em)) {
929 replace_extent_mapping(inode, em, split, modified);
930 } else {
931 int ret;
932
933 ret = add_extent_mapping(inode, split, modified);
934 /* Logic error, shouldn't happen. */
935 ASSERT(ret == 0);
936 if (WARN_ON(ret != 0) && modified)
937 btrfs_set_inode_full_sync(inode);
938 }
939 free_extent_map(split);
940 split = NULL;
941 }
942remove_em:
943 if (extent_map_in_tree(em)) {
944 /*
945 * If the extent map is still in the tree it means that
946 * either of the following is true:
947 *
948 * 1) It fits entirely in our range (doesn't end beyond
949 * it or starts before it);
950 *
951 * 2) It starts before our range and/or ends after our
952 * range, and we were not able to allocate the extent
953 * maps for split operations, @split and @split2.
954 *
955 * If we are at case 2) then we just remove the entire
956 * extent map - this is fine since if anyone needs it to
957 * access the subranges outside our range, will just
958 * load it again from the subvolume tree's file extent
959 * item. However if the extent map was in the list of
960 * modified extents, then we must mark the inode for a
961 * full fsync, otherwise a fast fsync will miss this
962 * extent if it's new and needs to be logged.
963 */
964 if ((em->start < start || em_end > end) && modified) {
965 ASSERT(!split);
966 btrfs_set_inode_full_sync(inode);
967 }
968 remove_extent_mapping(inode, em);
969 }
970
971 /*
972 * Once for the tree reference (we replaced or removed the
973 * extent map from the tree).
974 */
975 free_extent_map(em);
976next:
977 /* Once for us (for our lookup reference). */
978 free_extent_map(em);
979
980 em = next_em;
981 }
982
983 write_unlock(&em_tree->lock);
984
985 free_extent_map(split);
986 free_extent_map(split2);
987}
988
989/*
990 * Replace a range in the inode's extent map tree with a new extent map.
991 *
992 * @inode: The target inode.
993 * @new_em: The new extent map to add to the inode's extent map tree.
994 * @modified: Indicate if the new extent map should be added to the list of
995 * modified extents (for fast fsync tracking).
996 *
997 * Drops all the extent maps in the inode's extent map tree that intersect the
998 * range of the new extent map and adds the new extent map to the tree.
999 * The caller should have locked an appropriate file range in the inode's io
1000 * tree before calling this function.
1001 */
1002int btrfs_replace_extent_map_range(struct btrfs_inode *inode,
1003 struct extent_map *new_em,
1004 bool modified)
1005{
1006 const u64 end = new_em->start + new_em->len - 1;
1007 struct extent_map_tree *tree = &inode->extent_tree;
1008 int ret;
1009
1010 ASSERT(!extent_map_in_tree(new_em));
1011
1012 /*
1013 * The caller has locked an appropriate file range in the inode's io
1014 * tree, but getting -EEXIST when adding the new extent map can still
1015 * happen in case there are extents that partially cover the range, and
1016 * this is due to two tasks operating on different parts of the extent.
1017 * See commit 18e83ac75bfe67 ("Btrfs: fix unexpected EEXIST from
1018 * btrfs_get_extent") for an example and details.
1019 */
1020 do {
1021 btrfs_drop_extent_map_range(inode, new_em->start, end, false);
1022 write_lock(&tree->lock);
1023 ret = add_extent_mapping(inode, new_em, modified);
1024 write_unlock(&tree->lock);
1025 } while (ret == -EEXIST);
1026
1027 return ret;
1028}
1029
1030/*
1031 * Split off the first pre bytes from the extent_map at [start, start + len],
1032 * and set the block_start for it to new_logical.
1033 *
1034 * This function is used when an ordered_extent needs to be split.
1035 */
1036int split_extent_map(struct btrfs_inode *inode, u64 start, u64 len, u64 pre,
1037 u64 new_logical)
1038{
1039 struct extent_map_tree *em_tree = &inode->extent_tree;
1040 struct extent_map *em;
1041 struct extent_map *split_pre = NULL;
1042 struct extent_map *split_mid = NULL;
1043 int ret = 0;
1044 unsigned long flags;
1045
1046 ASSERT(pre != 0);
1047 ASSERT(pre < len);
1048
1049 split_pre = alloc_extent_map();
1050 if (!split_pre)
1051 return -ENOMEM;
1052 split_mid = alloc_extent_map();
1053 if (!split_mid) {
1054 ret = -ENOMEM;
1055 goto out_free_pre;
1056 }
1057
1058 lock_extent(&inode->io_tree, start, start + len - 1, NULL);
1059 write_lock(&em_tree->lock);
1060 em = lookup_extent_mapping(em_tree, start, len);
1061 if (!em) {
1062 ret = -EIO;
1063 goto out_unlock;
1064 }
1065
1066 ASSERT(em->len == len);
1067 ASSERT(!extent_map_is_compressed(em));
1068 ASSERT(em->disk_bytenr < EXTENT_MAP_LAST_BYTE);
1069 ASSERT(em->flags & EXTENT_FLAG_PINNED);
1070 ASSERT(!(em->flags & EXTENT_FLAG_LOGGING));
1071 ASSERT(!list_empty(&em->list));
1072
1073 flags = em->flags;
1074 em->flags &= ~EXTENT_FLAG_PINNED;
1075
1076 /* First, replace the em with a new extent_map starting from * em->start */
1077 split_pre->start = em->start;
1078 split_pre->len = pre;
1079 split_pre->disk_bytenr = new_logical;
1080 split_pre->disk_num_bytes = split_pre->len;
1081 split_pre->offset = 0;
1082 split_pre->ram_bytes = split_pre->len;
1083 split_pre->flags = flags;
1084 split_pre->generation = em->generation;
1085
1086 replace_extent_mapping(inode, em, split_pre, 1);
1087
1088 /*
1089 * Now we only have an extent_map at:
1090 * [em->start, em->start + pre]
1091 */
1092
1093 /* Insert the middle extent_map. */
1094 split_mid->start = em->start + pre;
1095 split_mid->len = em->len - pre;
1096 split_mid->disk_bytenr = extent_map_block_start(em) + pre;
1097 split_mid->disk_num_bytes = split_mid->len;
1098 split_mid->offset = 0;
1099 split_mid->ram_bytes = split_mid->len;
1100 split_mid->flags = flags;
1101 split_mid->generation = em->generation;
1102 add_extent_mapping(inode, split_mid, 1);
1103
1104 /* Once for us */
1105 free_extent_map(em);
1106 /* Once for the tree */
1107 free_extent_map(em);
1108
1109out_unlock:
1110 write_unlock(&em_tree->lock);
1111 unlock_extent(&inode->io_tree, start, start + len - 1, NULL);
1112 free_extent_map(split_mid);
1113out_free_pre:
1114 free_extent_map(split_pre);
1115 return ret;
1116}
1117
1118struct btrfs_em_shrink_ctx {
1119 long nr_to_scan;
1120 long scanned;
1121};
1122
1123static long btrfs_scan_inode(struct btrfs_inode *inode, struct btrfs_em_shrink_ctx *ctx)
1124{
1125 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1126 const u64 cur_fs_gen = btrfs_get_fs_generation(fs_info);
1127 struct extent_map_tree *tree = &inode->extent_tree;
1128 long nr_dropped = 0;
1129 struct rb_node *node;
1130
1131 lockdep_assert_held_write(&tree->lock);
1132
1133 /*
1134 * Take the mmap lock so that we serialize with the inode logging phase
1135 * of fsync because we may need to set the full sync flag on the inode,
1136 * in case we have to remove extent maps in the tree's list of modified
1137 * extents. If we set the full sync flag in the inode while an fsync is
1138 * in progress, we may risk missing new extents because before the flag
1139 * is set, fsync decides to only wait for writeback to complete and then
1140 * during inode logging it sees the flag set and uses the subvolume tree
1141 * to find new extents, which may not be there yet because ordered
1142 * extents haven't completed yet.
1143 *
1144 * We also do a try lock because we don't want to block for too long and
1145 * we are holding the extent map tree's lock in write mode.
1146 */
1147 if (!down_read_trylock(&inode->i_mmap_lock))
1148 return 0;
1149
1150 node = rb_first(&tree->root);
1151 while (node) {
1152 struct rb_node *next = rb_next(node);
1153 struct extent_map *em;
1154
1155 em = rb_entry(node, struct extent_map, rb_node);
1156 ctx->scanned++;
1157
1158 if (em->flags & EXTENT_FLAG_PINNED)
1159 goto next;
1160
1161 /*
1162 * If the inode is in the list of modified extents (new) and its
1163 * generation is the same (or is greater than) the current fs
1164 * generation, it means it was not yet persisted so we have to
1165 * set the full sync flag so that the next fsync will not miss
1166 * it.
1167 */
1168 if (!list_empty(&em->list) && em->generation >= cur_fs_gen)
1169 btrfs_set_inode_full_sync(inode);
1170
1171 remove_extent_mapping(inode, em);
1172 trace_btrfs_extent_map_shrinker_remove_em(inode, em);
1173 /* Drop the reference for the tree. */
1174 free_extent_map(em);
1175 nr_dropped++;
1176next:
1177 if (ctx->scanned >= ctx->nr_to_scan)
1178 break;
1179
1180 /*
1181 * Stop if we need to reschedule or there's contention on the
1182 * lock. This is to avoid slowing other tasks trying to take the
1183 * lock.
1184 */
1185 if (need_resched() || rwlock_needbreak(&tree->lock) ||
1186 btrfs_fs_closing(fs_info))
1187 break;
1188 node = next;
1189 }
1190 up_read(&inode->i_mmap_lock);
1191
1192 return nr_dropped;
1193}
1194
1195static struct btrfs_inode *find_first_inode_to_shrink(struct btrfs_root *root,
1196 u64 min_ino)
1197{
1198 struct btrfs_inode *inode;
1199 unsigned long from = min_ino;
1200
1201 xa_lock(&root->inodes);
1202 while (true) {
1203 struct extent_map_tree *tree;
1204
1205 inode = xa_find(&root->inodes, &from, ULONG_MAX, XA_PRESENT);
1206 if (!inode)
1207 break;
1208
1209 tree = &inode->extent_tree;
1210
1211 /*
1212 * We want to be fast so if the lock is busy we don't want to
1213 * spend time waiting for it (some task is about to do IO for
1214 * the inode).
1215 */
1216 if (!write_trylock(&tree->lock))
1217 goto next;
1218
1219 /*
1220 * Skip inode if it doesn't have loaded extent maps, so we avoid
1221 * getting a reference and doing an iput later. This includes
1222 * cases like files that were opened for things like stat(2), or
1223 * files with all extent maps previously released through the
1224 * release folio callback (btrfs_release_folio()) or released in
1225 * a previous run, or directories which never have extent maps.
1226 */
1227 if (RB_EMPTY_ROOT(&tree->root)) {
1228 write_unlock(&tree->lock);
1229 goto next;
1230 }
1231
1232 if (igrab(&inode->vfs_inode))
1233 break;
1234
1235 write_unlock(&tree->lock);
1236next:
1237 from = btrfs_ino(inode) + 1;
1238 cond_resched_lock(&root->inodes.xa_lock);
1239 }
1240 xa_unlock(&root->inodes);
1241
1242 return inode;
1243}
1244
1245static long btrfs_scan_root(struct btrfs_root *root, struct btrfs_em_shrink_ctx *ctx)
1246{
1247 struct btrfs_fs_info *fs_info = root->fs_info;
1248 struct btrfs_inode *inode;
1249 long nr_dropped = 0;
1250 u64 min_ino = fs_info->em_shrinker_last_ino + 1;
1251
1252 inode = find_first_inode_to_shrink(root, min_ino);
1253 while (inode) {
1254 nr_dropped += btrfs_scan_inode(inode, ctx);
1255 write_unlock(&inode->extent_tree.lock);
1256
1257 min_ino = btrfs_ino(inode) + 1;
1258 fs_info->em_shrinker_last_ino = btrfs_ino(inode);
1259 iput(&inode->vfs_inode);
1260
1261 if (ctx->scanned >= ctx->nr_to_scan || btrfs_fs_closing(fs_info))
1262 break;
1263
1264 cond_resched();
1265
1266 inode = find_first_inode_to_shrink(root, min_ino);
1267 }
1268
1269 if (inode) {
1270 /*
1271 * There are still inodes in this root or we happened to process
1272 * the last one and reached the scan limit. In either case set
1273 * the current root to this one, so we'll resume from the next
1274 * inode if there is one or we will find out this was the last
1275 * one and move to the next root.
1276 */
1277 fs_info->em_shrinker_last_root = btrfs_root_id(root);
1278 } else {
1279 /*
1280 * No more inodes in this root, set extent_map_shrinker_last_ino to 0 so
1281 * that when processing the next root we start from its first inode.
1282 */
1283 fs_info->em_shrinker_last_ino = 0;
1284 fs_info->em_shrinker_last_root = btrfs_root_id(root) + 1;
1285 }
1286
1287 return nr_dropped;
1288}
1289
1290static void btrfs_extent_map_shrinker_worker(struct work_struct *work)
1291{
1292 struct btrfs_fs_info *fs_info;
1293 struct btrfs_em_shrink_ctx ctx;
1294 u64 start_root_id;
1295 u64 next_root_id;
1296 bool cycled = false;
1297 long nr_dropped = 0;
1298
1299 fs_info = container_of(work, struct btrfs_fs_info, em_shrinker_work);
1300
1301 ctx.scanned = 0;
1302 ctx.nr_to_scan = atomic64_read(&fs_info->em_shrinker_nr_to_scan);
1303
1304 start_root_id = fs_info->em_shrinker_last_root;
1305 next_root_id = fs_info->em_shrinker_last_root;
1306
1307 if (trace_btrfs_extent_map_shrinker_scan_enter_enabled()) {
1308 s64 nr = percpu_counter_sum_positive(&fs_info->evictable_extent_maps);
1309
1310 trace_btrfs_extent_map_shrinker_scan_enter(fs_info, nr);
1311 }
1312
1313 while (ctx.scanned < ctx.nr_to_scan && !btrfs_fs_closing(fs_info)) {
1314 struct btrfs_root *root;
1315 unsigned long count;
1316
1317 cond_resched();
1318
1319 spin_lock(&fs_info->fs_roots_radix_lock);
1320 count = radix_tree_gang_lookup(&fs_info->fs_roots_radix,
1321 (void **)&root,
1322 (unsigned long)next_root_id, 1);
1323 if (count == 0) {
1324 spin_unlock(&fs_info->fs_roots_radix_lock);
1325 if (start_root_id > 0 && !cycled) {
1326 next_root_id = 0;
1327 fs_info->em_shrinker_last_root = 0;
1328 fs_info->em_shrinker_last_ino = 0;
1329 cycled = true;
1330 continue;
1331 }
1332 break;
1333 }
1334 next_root_id = btrfs_root_id(root) + 1;
1335 root = btrfs_grab_root(root);
1336 spin_unlock(&fs_info->fs_roots_radix_lock);
1337
1338 if (!root)
1339 continue;
1340
1341 if (is_fstree(btrfs_root_id(root)))
1342 nr_dropped += btrfs_scan_root(root, &ctx);
1343
1344 btrfs_put_root(root);
1345 }
1346
1347 if (trace_btrfs_extent_map_shrinker_scan_exit_enabled()) {
1348 s64 nr = percpu_counter_sum_positive(&fs_info->evictable_extent_maps);
1349
1350 trace_btrfs_extent_map_shrinker_scan_exit(fs_info, nr_dropped, nr);
1351 }
1352
1353 atomic64_set(&fs_info->em_shrinker_nr_to_scan, 0);
1354}
1355
1356void btrfs_free_extent_maps(struct btrfs_fs_info *fs_info, long nr_to_scan)
1357{
1358 /*
1359 * Do nothing if the shrinker is already running. In case of high memory
1360 * pressure we can have a lot of tasks calling us and all passing the
1361 * same nr_to_scan value, but in reality we may need only to free
1362 * nr_to_scan extent maps (or less). In case we need to free more than
1363 * that, we will be called again by the fs shrinker, so no worries about
1364 * not doing enough work to reclaim memory from extent maps.
1365 * We can also be repeatedly called with the same nr_to_scan value
1366 * simply because the shrinker runs asynchronously and multiple calls
1367 * to this function are made before the shrinker does enough progress.
1368 *
1369 * That's why we set the atomic counter to nr_to_scan only if its
1370 * current value is zero, instead of incrementing the counter by
1371 * nr_to_scan.
1372 */
1373 if (atomic64_cmpxchg(&fs_info->em_shrinker_nr_to_scan, 0, nr_to_scan) != 0)
1374 return;
1375
1376 queue_work(system_unbound_wq, &fs_info->em_shrinker_work);
1377}
1378
1379void btrfs_init_extent_map_shrinker_work(struct btrfs_fs_info *fs_info)
1380{
1381 atomic64_set(&fs_info->em_shrinker_nr_to_scan, 0);
1382 INIT_WORK(&fs_info->em_shrinker_work, btrfs_extent_map_shrinker_worker);
1383}