?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf bpf: Move perf_event_output() from stdio.h to bpf.h
[linux/fpc-iii.git] / fs / btrfs / extent-tree.c
bloba1febf155747e05c21212b65bf2d3ff758f37868
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6 #include <linux/sched.h>
7 #include <linux/sched/signal.h>
8 #include <linux/pagemap.h>
9 #include <linux/writeback.h>
10 #include <linux/blkdev.h>
11 #include <linux/sort.h>
12 #include <linux/rcupdate.h>
13 #include <linux/kthread.h>
14 #include <linux/slab.h>
15 #include <linux/ratelimit.h>
16 #include <linux/percpu_counter.h>
17 #include <linux/lockdep.h>
18 #include <linux/crc32c.h>
19 #include "tree-log.h"
20 #include "disk-io.h"
21 #include "print-tree.h"
22 #include "volumes.h"
23 #include "raid56.h"
24 #include "locking.h"
25 #include "free-space-cache.h"
26 #include "free-space-tree.h"
27 #include "math.h"
28 #include "sysfs.h"
29 #include "qgroup.h"
30 #include "ref-verify.h"
31
32 #undef SCRAMBLE_DELAYED_REFS
33
34 /*
35 * control flags for do_chunk_alloc's force field
36 * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
37 * if we really need one.
38 *
39 * CHUNK_ALLOC_LIMITED means to only try and allocate one
40 * if we have very few chunks already allocated. This is
41 * used as part of the clustering code to help make sure
42 * we have a good pool of storage to cluster in, without
43 * filling the FS with empty chunks
44 *
45 * CHUNK_ALLOC_FORCE means it must try to allocate one
46 *
47 */
48 enum {
49 CHUNK_ALLOC_NO_FORCE = 0,
50 CHUNK_ALLOC_LIMITED = 1,
51 CHUNK_ALLOC_FORCE = 2,
52 };
53
54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
55 struct btrfs_delayed_ref_node *node, u64 parent,
56 u64 root_objectid, u64 owner_objectid,
57 u64 owner_offset, int refs_to_drop,
58 struct btrfs_delayed_extent_op *extra_op);
59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
60 struct extent_buffer *leaf,
61 struct btrfs_extent_item *ei);
62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
63 u64 parent, u64 root_objectid,
64 u64 flags, u64 owner, u64 offset,
65 struct btrfs_key *ins, int ref_mod);
66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
67 struct btrfs_delayed_ref_node *node,
68 struct btrfs_delayed_extent_op *extent_op);
69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
70 int force);
71 static int find_next_key(struct btrfs_path *path, int level,
72 struct btrfs_key *key);
73 static void dump_space_info(struct btrfs_fs_info *fs_info,
74 struct btrfs_space_info *info, u64 bytes,
75 int dump_block_groups);
76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
77 u64 num_bytes);
78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
79 struct btrfs_space_info *space_info,
80 u64 num_bytes);
81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
82 struct btrfs_space_info *space_info,
83 u64 num_bytes);
84
85 static noinline int
86 block_group_cache_done(struct btrfs_block_group_cache *cache)
87 {
88 smp_mb();
89 return cache->cached == BTRFS_CACHE_FINISHED ||
90 cache->cached == BTRFS_CACHE_ERROR;
91 }
92
93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
94 {
95 return (cache->flags & bits) == bits;
96 }
97
98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
99 {
100 atomic_inc(&cache->count);
101 }
102
103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
104 {
105 if (atomic_dec_and_test(&cache->count)) {
106 WARN_ON(cache->pinned > 0);
107 WARN_ON(cache->reserved > 0);
108
109 /*
110 * If not empty, someone is still holding mutex of
111 * full_stripe_lock, which can only be released by caller.
112 * And it will definitely cause use-after-free when caller
113 * tries to release full stripe lock.
114 *
115 * No better way to resolve, but only to warn.
116 */
117 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
118 kfree(cache->free_space_ctl);
119 kfree(cache);
120 }
121 }
122
123 /*
124 * this adds the block group to the fs_info rb tree for the block group
125 * cache
126 */
127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
128 struct btrfs_block_group_cache *block_group)
129 {
130 struct rb_node **p;
131 struct rb_node *parent = NULL;
132 struct btrfs_block_group_cache *cache;
133
134 spin_lock(&info->block_group_cache_lock);
135 p = &info->block_group_cache_tree.rb_node;
136
137 while (*p) {
138 parent = *p;
139 cache = rb_entry(parent, struct btrfs_block_group_cache,
140 cache_node);
141 if (block_group->key.objectid < cache->key.objectid) {
142 p = &(*p)->rb_left;
143 } else if (block_group->key.objectid > cache->key.objectid) {
144 p = &(*p)->rb_right;
145 } else {
146 spin_unlock(&info->block_group_cache_lock);
147 return -EEXIST;
148 }
149 }
150
151 rb_link_node(&block_group->cache_node, parent, p);
152 rb_insert_color(&block_group->cache_node,
153 &info->block_group_cache_tree);
154
155 if (info->first_logical_byte > block_group->key.objectid)
156 info->first_logical_byte = block_group->key.objectid;
157
158 spin_unlock(&info->block_group_cache_lock);
159
160 return 0;
161 }
162
163 /*
164 * This will return the block group at or after bytenr if contains is 0, else
165 * it will return the block group that contains the bytenr
166 */
167 static struct btrfs_block_group_cache *
168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
169 int contains)
170 {
171 struct btrfs_block_group_cache *cache, *ret = NULL;
172 struct rb_node *n;
173 u64 end, start;
174
175 spin_lock(&info->block_group_cache_lock);
176 n = info->block_group_cache_tree.rb_node;
177
178 while (n) {
179 cache = rb_entry(n, struct btrfs_block_group_cache,
180 cache_node);
181 end = cache->key.objectid + cache->key.offset - 1;
182 start = cache->key.objectid;
183
184 if (bytenr < start) {
185 if (!contains && (!ret || start < ret->key.objectid))
186 ret = cache;
187 n = n->rb_left;
188 } else if (bytenr > start) {
189 if (contains && bytenr <= end) {
190 ret = cache;
191 break;
192 }
193 n = n->rb_right;
194 } else {
195 ret = cache;
196 break;
197 }
198 }
199 if (ret) {
200 btrfs_get_block_group(ret);
201 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
202 info->first_logical_byte = ret->key.objectid;
203 }
204 spin_unlock(&info->block_group_cache_lock);
205
206 return ret;
207 }
208
209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
210 u64 start, u64 num_bytes)
211 {
212 u64 end = start + num_bytes - 1;
213 set_extent_bits(&fs_info->freed_extents[0],
214 start, end, EXTENT_UPTODATE);
215 set_extent_bits(&fs_info->freed_extents[1],
216 start, end, EXTENT_UPTODATE);
217 return 0;
218 }
219
220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
221 {
222 struct btrfs_fs_info *fs_info = cache->fs_info;
223 u64 start, end;
224
225 start = cache->key.objectid;
226 end = start + cache->key.offset - 1;
227
228 clear_extent_bits(&fs_info->freed_extents[0],
229 start, end, EXTENT_UPTODATE);
230 clear_extent_bits(&fs_info->freed_extents[1],
231 start, end, EXTENT_UPTODATE);
232 }
233
234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
235 {
236 struct btrfs_fs_info *fs_info = cache->fs_info;
237 u64 bytenr;
238 u64 *logical;
239 int stripe_len;
240 int i, nr, ret;
241
242 if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
243 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
244 cache->bytes_super += stripe_len;
245 ret = add_excluded_extent(fs_info, cache->key.objectid,
246 stripe_len);
247 if (ret)
248 return ret;
249 }
250
251 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252 bytenr = btrfs_sb_offset(i);
253 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
254 bytenr, &logical, &nr, &stripe_len);
255 if (ret)
256 return ret;
257
258 while (nr--) {
259 u64 start, len;
260
261 if (logical[nr] > cache->key.objectid +
262 cache->key.offset)
263 continue;
264
265 if (logical[nr] + stripe_len <= cache->key.objectid)
266 continue;
267
268 start = logical[nr];
269 if (start < cache->key.objectid) {
270 start = cache->key.objectid;
271 len = (logical[nr] + stripe_len) - start;
272 } else {
273 len = min_t(u64, stripe_len,
274 cache->key.objectid +
275 cache->key.offset - start);
276 }
277
278 cache->bytes_super += len;
279 ret = add_excluded_extent(fs_info, start, len);
280 if (ret) {
281 kfree(logical);
282 return ret;
283 }
284 }
285
286 kfree(logical);
287 }
288 return 0;
289 }
290
291 static struct btrfs_caching_control *
292 get_caching_control(struct btrfs_block_group_cache *cache)
293 {
294 struct btrfs_caching_control *ctl;
295
296 spin_lock(&cache->lock);
297 if (!cache->caching_ctl) {
298 spin_unlock(&cache->lock);
299 return NULL;
300 }
301
302 ctl = cache->caching_ctl;
303 refcount_inc(&ctl->count);
304 spin_unlock(&cache->lock);
305 return ctl;
306 }
307
308 static void put_caching_control(struct btrfs_caching_control *ctl)
309 {
310 if (refcount_dec_and_test(&ctl->count))
311 kfree(ctl);
312 }
313
314 #ifdef CONFIG_BTRFS_DEBUG
315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
316 {
317 struct btrfs_fs_info *fs_info = block_group->fs_info;
318 u64 start = block_group->key.objectid;
319 u64 len = block_group->key.offset;
320 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
321 fs_info->nodesize : fs_info->sectorsize;
322 u64 step = chunk << 1;
323
324 while (len > chunk) {
325 btrfs_remove_free_space(block_group, start, chunk);
326 start += step;
327 if (len < step)
328 len = 0;
329 else
330 len -= step;
331 }
332 }
333 #endif
334
335 /*
336 * this is only called by cache_block_group, since we could have freed extents
337 * we need to check the pinned_extents for any extents that can't be used yet
338 * since their free space will be released as soon as the transaction commits.
339 */
340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
341 u64 start, u64 end)
342 {
343 struct btrfs_fs_info *info = block_group->fs_info;
344 u64 extent_start, extent_end, size, total_added = 0;
345 int ret;
346
347 while (start < end) {
348 ret = find_first_extent_bit(info->pinned_extents, start,
349 &extent_start, &extent_end,
350 EXTENT_DIRTY | EXTENT_UPTODATE,
351 NULL);
352 if (ret)
353 break;
354
355 if (extent_start <= start) {
356 start = extent_end + 1;
357 } else if (extent_start > start && extent_start < end) {
358 size = extent_start - start;
359 total_added += size;
360 ret = btrfs_add_free_space(block_group, start,
361 size);
362 BUG_ON(ret); /* -ENOMEM or logic error */
363 start = extent_end + 1;
364 } else {
365 break;
366 }
367 }
368
369 if (start < end) {
370 size = end - start;
371 total_added += size;
372 ret = btrfs_add_free_space(block_group, start, size);
373 BUG_ON(ret); /* -ENOMEM or logic error */
374 }
375
376 return total_added;
377 }
378
379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
380 {
381 struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
382 struct btrfs_fs_info *fs_info = block_group->fs_info;
383 struct btrfs_root *extent_root = fs_info->extent_root;
384 struct btrfs_path *path;
385 struct extent_buffer *leaf;
386 struct btrfs_key key;
387 u64 total_found = 0;
388 u64 last = 0;
389 u32 nritems;
390 int ret;
391 bool wakeup = true;
392
393 path = btrfs_alloc_path();
394 if (!path)
395 return -ENOMEM;
396
397 last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
398
399 #ifdef CONFIG_BTRFS_DEBUG
400 /*
401 * If we're fragmenting we don't want to make anybody think we can
402 * allocate from this block group until we've had a chance to fragment
403 * the free space.
404 */
405 if (btrfs_should_fragment_free_space(block_group))
406 wakeup = false;
407 #endif
408 /*
409 * We don't want to deadlock with somebody trying to allocate a new
410 * extent for the extent root while also trying to search the extent
411 * root to add free space. So we skip locking and search the commit
412 * root, since its read-only
413 */
414 path->skip_locking = 1;
415 path->search_commit_root = 1;
416 path->reada = READA_FORWARD;
417
418 key.objectid = last;
419 key.offset = 0;
420 key.type = BTRFS_EXTENT_ITEM_KEY;
421
422 next:
423 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
424 if (ret < 0)
425 goto out;
426
427 leaf = path->nodes[0];
428 nritems = btrfs_header_nritems(leaf);
429
430 while (1) {
431 if (btrfs_fs_closing(fs_info) > 1) {
432 last = (u64)-1;
433 break;
434 }
435
436 if (path->slots[0] < nritems) {
437 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
438 } else {
439 ret = find_next_key(path, 0, &key);
440 if (ret)
441 break;
442
443 if (need_resched() ||
444 rwsem_is_contended(&fs_info->commit_root_sem)) {
445 if (wakeup)
446 caching_ctl->progress = last;
447 btrfs_release_path(path);
448 up_read(&fs_info->commit_root_sem);
449 mutex_unlock(&caching_ctl->mutex);
450 cond_resched();
451 mutex_lock(&caching_ctl->mutex);
452 down_read(&fs_info->commit_root_sem);
453 goto next;
454 }
455
456 ret = btrfs_next_leaf(extent_root, path);
457 if (ret < 0)
458 goto out;
459 if (ret)
460 break;
461 leaf = path->nodes[0];
462 nritems = btrfs_header_nritems(leaf);
463 continue;
464 }
465
466 if (key.objectid < last) {
467 key.objectid = last;
468 key.offset = 0;
469 key.type = BTRFS_EXTENT_ITEM_KEY;
470
471 if (wakeup)
472 caching_ctl->progress = last;
473 btrfs_release_path(path);
474 goto next;
475 }
476
477 if (key.objectid < block_group->key.objectid) {
478 path->slots[0]++;
479 continue;
480 }
481
482 if (key.objectid >= block_group->key.objectid +
483 block_group->key.offset)
484 break;
485
486 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
487 key.type == BTRFS_METADATA_ITEM_KEY) {
488 total_found += add_new_free_space(block_group, last,
489 key.objectid);
490 if (key.type == BTRFS_METADATA_ITEM_KEY)
491 last = key.objectid +
492 fs_info->nodesize;
493 else
494 last = key.objectid + key.offset;
495
496 if (total_found > CACHING_CTL_WAKE_UP) {
497 total_found = 0;
498 if (wakeup)
499 wake_up(&caching_ctl->wait);
500 }
501 }
502 path->slots[0]++;
503 }
504 ret = 0;
505
506 total_found += add_new_free_space(block_group, last,
507 block_group->key.objectid +
508 block_group->key.offset);
509 caching_ctl->progress = (u64)-1;
510
511 out:
512 btrfs_free_path(path);
513 return ret;
514 }
515
516 static noinline void caching_thread(struct btrfs_work *work)
517 {
518 struct btrfs_block_group_cache *block_group;
519 struct btrfs_fs_info *fs_info;
520 struct btrfs_caching_control *caching_ctl;
521 int ret;
522
523 caching_ctl = container_of(work, struct btrfs_caching_control, work);
524 block_group = caching_ctl->block_group;
525 fs_info = block_group->fs_info;
526
527 mutex_lock(&caching_ctl->mutex);
528 down_read(&fs_info->commit_root_sem);
529
530 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
531 ret = load_free_space_tree(caching_ctl);
532 else
533 ret = load_extent_tree_free(caching_ctl);
534
535 spin_lock(&block_group->lock);
536 block_group->caching_ctl = NULL;
537 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
538 spin_unlock(&block_group->lock);
539
540 #ifdef CONFIG_BTRFS_DEBUG
541 if (btrfs_should_fragment_free_space(block_group)) {
542 u64 bytes_used;
543
544 spin_lock(&block_group->space_info->lock);
545 spin_lock(&block_group->lock);
546 bytes_used = block_group->key.offset -
547 btrfs_block_group_used(&block_group->item);
548 block_group->space_info->bytes_used += bytes_used >> 1;
549 spin_unlock(&block_group->lock);
550 spin_unlock(&block_group->space_info->lock);
551 fragment_free_space(block_group);
552 }
553 #endif
554
555 caching_ctl->progress = (u64)-1;
556
557 up_read(&fs_info->commit_root_sem);
558 free_excluded_extents(block_group);
559 mutex_unlock(&caching_ctl->mutex);
560
561 wake_up(&caching_ctl->wait);
562
563 put_caching_control(caching_ctl);
564 btrfs_put_block_group(block_group);
565 }
566
567 static int cache_block_group(struct btrfs_block_group_cache *cache,
568 int load_cache_only)
569 {
570 DEFINE_WAIT(wait);
571 struct btrfs_fs_info *fs_info = cache->fs_info;
572 struct btrfs_caching_control *caching_ctl;
573 int ret = 0;
574
575 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
576 if (!caching_ctl)
577 return -ENOMEM;
578
579 INIT_LIST_HEAD(&caching_ctl->list);
580 mutex_init(&caching_ctl->mutex);
581 init_waitqueue_head(&caching_ctl->wait);
582 caching_ctl->block_group = cache;
583 caching_ctl->progress = cache->key.objectid;
584 refcount_set(&caching_ctl->count, 1);
585 btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
586 caching_thread, NULL, NULL);
587
588 spin_lock(&cache->lock);
589 /*
590 * This should be a rare occasion, but this could happen I think in the
591 * case where one thread starts to load the space cache info, and then
592 * some other thread starts a transaction commit which tries to do an
593 * allocation while the other thread is still loading the space cache
594 * info. The previous loop should have kept us from choosing this block
595 * group, but if we've moved to the state where we will wait on caching
596 * block groups we need to first check if we're doing a fast load here,
597 * so we can wait for it to finish, otherwise we could end up allocating
598 * from a block group who's cache gets evicted for one reason or
599 * another.
600 */
601 while (cache->cached == BTRFS_CACHE_FAST) {
602 struct btrfs_caching_control *ctl;
603
604 ctl = cache->caching_ctl;
605 refcount_inc(&ctl->count);
606 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
607 spin_unlock(&cache->lock);
608
609 schedule();
610
611 finish_wait(&ctl->wait, &wait);
612 put_caching_control(ctl);
613 spin_lock(&cache->lock);
614 }
615
616 if (cache->cached != BTRFS_CACHE_NO) {
617 spin_unlock(&cache->lock);
618 kfree(caching_ctl);
619 return 0;
620 }
621 WARN_ON(cache->caching_ctl);
622 cache->caching_ctl = caching_ctl;
623 cache->cached = BTRFS_CACHE_FAST;
624 spin_unlock(&cache->lock);
625
626 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
627 mutex_lock(&caching_ctl->mutex);
628 ret = load_free_space_cache(fs_info, cache);
629
630 spin_lock(&cache->lock);
631 if (ret == 1) {
632 cache->caching_ctl = NULL;
633 cache->cached = BTRFS_CACHE_FINISHED;
634 cache->last_byte_to_unpin = (u64)-1;
635 caching_ctl->progress = (u64)-1;
636 } else {
637 if (load_cache_only) {
638 cache->caching_ctl = NULL;
639 cache->cached = BTRFS_CACHE_NO;
640 } else {
641 cache->cached = BTRFS_CACHE_STARTED;
642 cache->has_caching_ctl = 1;
643 }
644 }
645 spin_unlock(&cache->lock);
646 #ifdef CONFIG_BTRFS_DEBUG
647 if (ret == 1 &&
648 btrfs_should_fragment_free_space(cache)) {
649 u64 bytes_used;
650
651 spin_lock(&cache->space_info->lock);
652 spin_lock(&cache->lock);
653 bytes_used = cache->key.offset -
654 btrfs_block_group_used(&cache->item);
655 cache->space_info->bytes_used += bytes_used >> 1;
656 spin_unlock(&cache->lock);
657 spin_unlock(&cache->space_info->lock);
658 fragment_free_space(cache);
659 }
660 #endif
661 mutex_unlock(&caching_ctl->mutex);
662
663 wake_up(&caching_ctl->wait);
664 if (ret == 1) {
665 put_caching_control(caching_ctl);
666 free_excluded_extents(cache);
667 return 0;
668 }
669 } else {
670 /*
671 * We're either using the free space tree or no caching at all.
672 * Set cached to the appropriate value and wakeup any waiters.
673 */
674 spin_lock(&cache->lock);
675 if (load_cache_only) {
676 cache->caching_ctl = NULL;
677 cache->cached = BTRFS_CACHE_NO;
678 } else {
679 cache->cached = BTRFS_CACHE_STARTED;
680 cache->has_caching_ctl = 1;
681 }
682 spin_unlock(&cache->lock);
683 wake_up(&caching_ctl->wait);
684 }
685
686 if (load_cache_only) {
687 put_caching_control(caching_ctl);
688 return 0;
689 }
690
691 down_write(&fs_info->commit_root_sem);
692 refcount_inc(&caching_ctl->count);
693 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
694 up_write(&fs_info->commit_root_sem);
695
696 btrfs_get_block_group(cache);
697
698 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
699
700 return ret;
701 }
702
703 /*
704 * return the block group that starts at or after bytenr
705 */
706 static struct btrfs_block_group_cache *
707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
708 {
709 return block_group_cache_tree_search(info, bytenr, 0);
710 }
711
712 /*
713 * return the block group that contains the given bytenr
714 */
715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
716 struct btrfs_fs_info *info,
717 u64 bytenr)
718 {
719 return block_group_cache_tree_search(info, bytenr, 1);
720 }
721
722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
723 u64 flags)
724 {
725 struct list_head *head = &info->space_info;
726 struct btrfs_space_info *found;
727
728 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
729
730 rcu_read_lock();
731 list_for_each_entry_rcu(found, head, list) {
732 if (found->flags & flags) {
733 rcu_read_unlock();
734 return found;
735 }
736 }
737 rcu_read_unlock();
738 return NULL;
739 }
740
741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
742 bool metadata, u64 root_objectid)
743 {
744 struct btrfs_space_info *space_info;
745 u64 flags;
746
747 if (metadata) {
748 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
749 flags = BTRFS_BLOCK_GROUP_SYSTEM;
750 else
751 flags = BTRFS_BLOCK_GROUP_METADATA;
752 } else {
753 flags = BTRFS_BLOCK_GROUP_DATA;
754 }
755
756 space_info = __find_space_info(fs_info, flags);
757 ASSERT(space_info);
758 percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
759 BTRFS_TOTAL_BYTES_PINNED_BATCH);
760 }
761
762 /*
763 * after adding space to the filesystem, we need to clear the full flags
764 * on all the space infos.
765 */
766 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
767 {
768 struct list_head *head = &info->space_info;
769 struct btrfs_space_info *found;
770
771 rcu_read_lock();
772 list_for_each_entry_rcu(found, head, list)
773 found->full = 0;
774 rcu_read_unlock();
775 }
776
777 /* simple helper to search for an existing data extent at a given offset */
778 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
779 {
780 int ret;
781 struct btrfs_key key;
782 struct btrfs_path *path;
783
784 path = btrfs_alloc_path();
785 if (!path)
786 return -ENOMEM;
787
788 key.objectid = start;
789 key.offset = len;
790 key.type = BTRFS_EXTENT_ITEM_KEY;
791 ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
792 btrfs_free_path(path);
793 return ret;
794 }
795
796 /*
797 * helper function to lookup reference count and flags of a tree block.
798 *
799 * the head node for delayed ref is used to store the sum of all the
800 * reference count modifications queued up in the rbtree. the head
801 * node may also store the extent flags to set. This way you can check
802 * to see what the reference count and extent flags would be if all of
803 * the delayed refs are not processed.
804 */
805 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
806 struct btrfs_fs_info *fs_info, u64 bytenr,
807 u64 offset, int metadata, u64 *refs, u64 *flags)
808 {
809 struct btrfs_delayed_ref_head *head;
810 struct btrfs_delayed_ref_root *delayed_refs;
811 struct btrfs_path *path;
812 struct btrfs_extent_item *ei;
813 struct extent_buffer *leaf;
814 struct btrfs_key key;
815 u32 item_size;
816 u64 num_refs;
817 u64 extent_flags;
818 int ret;
819
820 /*
821 * If we don't have skinny metadata, don't bother doing anything
822 * different
823 */
824 if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
825 offset = fs_info->nodesize;
826 metadata = 0;
827 }
828
829 path = btrfs_alloc_path();
830 if (!path)
831 return -ENOMEM;
832
833 if (!trans) {
834 path->skip_locking = 1;
835 path->search_commit_root = 1;
836 }
837
838 search_again:
839 key.objectid = bytenr;
840 key.offset = offset;
841 if (metadata)
842 key.type = BTRFS_METADATA_ITEM_KEY;
843 else
844 key.type = BTRFS_EXTENT_ITEM_KEY;
845
846 ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
847 if (ret < 0)
848 goto out_free;
849
850 if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
851 if (path->slots[0]) {
852 path->slots[0]--;
853 btrfs_item_key_to_cpu(path->nodes[0], &key,
854 path->slots[0]);
855 if (key.objectid == bytenr &&
856 key.type == BTRFS_EXTENT_ITEM_KEY &&
857 key.offset == fs_info->nodesize)
858 ret = 0;
859 }
860 }
861
862 if (ret == 0) {
863 leaf = path->nodes[0];
864 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
865 if (item_size >= sizeof(*ei)) {
866 ei = btrfs_item_ptr(leaf, path->slots[0],
867 struct btrfs_extent_item);
868 num_refs = btrfs_extent_refs(leaf, ei);
869 extent_flags = btrfs_extent_flags(leaf, ei);
870 } else {
871 ret = -EINVAL;
872 btrfs_print_v0_err(fs_info);
873 if (trans)
874 btrfs_abort_transaction(trans, ret);
875 else
876 btrfs_handle_fs_error(fs_info, ret, NULL);
877
878 goto out_free;
879 }
880
881 BUG_ON(num_refs == 0);
882 } else {
883 num_refs = 0;
884 extent_flags = 0;
885 ret = 0;
886 }
887
888 if (!trans)
889 goto out;
890
891 delayed_refs = &trans->transaction->delayed_refs;
892 spin_lock(&delayed_refs->lock);
893 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
894 if (head) {
895 if (!mutex_trylock(&head->mutex)) {
896 refcount_inc(&head->refs);
897 spin_unlock(&delayed_refs->lock);
898
899 btrfs_release_path(path);
900
901 /*
902 * Mutex was contended, block until it's released and try
903 * again
904 */
905 mutex_lock(&head->mutex);
906 mutex_unlock(&head->mutex);
907 btrfs_put_delayed_ref_head(head);
908 goto search_again;
909 }
910 spin_lock(&head->lock);
911 if (head->extent_op && head->extent_op->update_flags)
912 extent_flags |= head->extent_op->flags_to_set;
913 else
914 BUG_ON(num_refs == 0);
915
916 num_refs += head->ref_mod;
917 spin_unlock(&head->lock);
918 mutex_unlock(&head->mutex);
919 }
920 spin_unlock(&delayed_refs->lock);
921 out:
922 WARN_ON(num_refs == 0);
923 if (refs)
924 *refs = num_refs;
925 if (flags)
926 *flags = extent_flags;
927 out_free:
928 btrfs_free_path(path);
929 return ret;
930 }
931
932 /*
933 * Back reference rules. Back refs have three main goals:
934 *
935 * 1) differentiate between all holders of references to an extent so that
936 * when a reference is dropped we can make sure it was a valid reference
937 * before freeing the extent.
938 *
939 * 2) Provide enough information to quickly find the holders of an extent
940 * if we notice a given block is corrupted or bad.
941 *
942 * 3) Make it easy to migrate blocks for FS shrinking or storage pool
943 * maintenance. This is actually the same as #2, but with a slightly
944 * different use case.
945 *
946 * There are two kinds of back refs. The implicit back refs is optimized
947 * for pointers in non-shared tree blocks. For a given pointer in a block,
948 * back refs of this kind provide information about the block's owner tree
949 * and the pointer's key. These information allow us to find the block by
950 * b-tree searching. The full back refs is for pointers in tree blocks not
951 * referenced by their owner trees. The location of tree block is recorded
952 * in the back refs. Actually the full back refs is generic, and can be
953 * used in all cases the implicit back refs is used. The major shortcoming
954 * of the full back refs is its overhead. Every time a tree block gets
955 * COWed, we have to update back refs entry for all pointers in it.
956 *
957 * For a newly allocated tree block, we use implicit back refs for
958 * pointers in it. This means most tree related operations only involve
959 * implicit back refs. For a tree block created in old transaction, the
960 * only way to drop a reference to it is COW it. So we can detect the
961 * event that tree block loses its owner tree's reference and do the
962 * back refs conversion.
963 *
964 * When a tree block is COWed through a tree, there are four cases:
965 *
966 * The reference count of the block is one and the tree is the block's
967 * owner tree. Nothing to do in this case.
968 *
969 * The reference count of the block is one and the tree is not the
970 * block's owner tree. In this case, full back refs is used for pointers
971 * in the block. Remove these full back refs, add implicit back refs for
972 * every pointers in the new block.
973 *
974 * The reference count of the block is greater than one and the tree is
975 * the block's owner tree. In this case, implicit back refs is used for
976 * pointers in the block. Add full back refs for every pointers in the
977 * block, increase lower level extents' reference counts. The original
978 * implicit back refs are entailed to the new block.
979 *
980 * The reference count of the block is greater than one and the tree is
981 * not the block's owner tree. Add implicit back refs for every pointer in
982 * the new block, increase lower level extents' reference count.
983 *
984 * Back Reference Key composing:
985 *
986 * The key objectid corresponds to the first byte in the extent,
987 * The key type is used to differentiate between types of back refs.
988 * There are different meanings of the key offset for different types
989 * of back refs.
990 *
991 * File extents can be referenced by:
992 *
993 * - multiple snapshots, subvolumes, or different generations in one subvol
994 * - different files inside a single subvolume
995 * - different offsets inside a file (bookend extents in file.c)
996 *
997 * The extent ref structure for the implicit back refs has fields for:
998 *
999 * - Objectid of the subvolume root
1000 * - objectid of the file holding the reference
1001 * - original offset in the file
1002 * - how many bookend extents
1003 *
1004 * The key offset for the implicit back refs is hash of the first
1005 * three fields.
1006 *
1007 * The extent ref structure for the full back refs has field for:
1008 *
1009 * - number of pointers in the tree leaf
1010 *
1011 * The key offset for the implicit back refs is the first byte of
1012 * the tree leaf
1013 *
1014 * When a file extent is allocated, The implicit back refs is used.
1015 * the fields are filled in:
1016 *
1017 * (root_key.objectid, inode objectid, offset in file, 1)
1018 *
1019 * When a file extent is removed file truncation, we find the
1020 * corresponding implicit back refs and check the following fields:
1021 *
1022 * (btrfs_header_owner(leaf), inode objectid, offset in file)
1023 *
1024 * Btree extents can be referenced by:
1025 *
1026 * - Different subvolumes
1027 *
1028 * Both the implicit back refs and the full back refs for tree blocks
1029 * only consist of key. The key offset for the implicit back refs is
1030 * objectid of block's owner tree. The key offset for the full back refs
1031 * is the first byte of parent block.
1032 *
1033 * When implicit back refs is used, information about the lowest key and
1034 * level of the tree block are required. These information are stored in
1035 * tree block info structure.
1036 */
1037
1038 /*
1039 * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1040 * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1041 * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1042 */
1043 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1044 struct btrfs_extent_inline_ref *iref,
1045 enum btrfs_inline_ref_type is_data)
1046 {
1047 int type = btrfs_extent_inline_ref_type(eb, iref);
1048 u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1049
1050 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1051 type == BTRFS_SHARED_BLOCK_REF_KEY ||
1052 type == BTRFS_SHARED_DATA_REF_KEY ||
1053 type == BTRFS_EXTENT_DATA_REF_KEY) {
1054 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1055 if (type == BTRFS_TREE_BLOCK_REF_KEY)
1056 return type;
1057 if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1058 ASSERT(eb->fs_info);
1059 /*
1060 * Every shared one has parent tree
1061 * block, which must be aligned to
1062 * nodesize.
1063 */
1064 if (offset &&
1065 IS_ALIGNED(offset, eb->fs_info->nodesize))
1066 return type;
1067 }
1068 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1069 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1070 return type;
1071 if (type == BTRFS_SHARED_DATA_REF_KEY) {
1072 ASSERT(eb->fs_info);
1073 /*
1074 * Every shared one has parent tree
1075 * block, which must be aligned to
1076 * nodesize.
1077 */
1078 if (offset &&
1079 IS_ALIGNED(offset, eb->fs_info->nodesize))
1080 return type;
1081 }
1082 } else {
1083 ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1084 return type;
1085 }
1086 }
1087
1088 btrfs_print_leaf((struct extent_buffer *)eb);
1089 btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1090 eb->start, type);
1091 WARN_ON(1);
1092
1093 return BTRFS_REF_TYPE_INVALID;
1094 }
1095
1096 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1097 {
1098 u32 high_crc = ~(u32)0;
1099 u32 low_crc = ~(u32)0;
1100 __le64 lenum;
1101
1102 lenum = cpu_to_le64(root_objectid);
1103 high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1104 lenum = cpu_to_le64(owner);
1105 low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1106 lenum = cpu_to_le64(offset);
1107 low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1108
1109 return ((u64)high_crc << 31) ^ (u64)low_crc;
1110 }
1111
1112 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1113 struct btrfs_extent_data_ref *ref)
1114 {
1115 return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1116 btrfs_extent_data_ref_objectid(leaf, ref),
1117 btrfs_extent_data_ref_offset(leaf, ref));
1118 }
1119
1120 static int match_extent_data_ref(struct extent_buffer *leaf,
1121 struct btrfs_extent_data_ref *ref,
1122 u64 root_objectid, u64 owner, u64 offset)
1123 {
1124 if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1125 btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1126 btrfs_extent_data_ref_offset(leaf, ref) != offset)
1127 return 0;
1128 return 1;
1129 }
1130
1131 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1132 struct btrfs_path *path,
1133 u64 bytenr, u64 parent,
1134 u64 root_objectid,
1135 u64 owner, u64 offset)
1136 {
1137 struct btrfs_root *root = trans->fs_info->extent_root;
1138 struct btrfs_key key;
1139 struct btrfs_extent_data_ref *ref;
1140 struct extent_buffer *leaf;
1141 u32 nritems;
1142 int ret;
1143 int recow;
1144 int err = -ENOENT;
1145
1146 key.objectid = bytenr;
1147 if (parent) {
1148 key.type = BTRFS_SHARED_DATA_REF_KEY;
1149 key.offset = parent;
1150 } else {
1151 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1152 key.offset = hash_extent_data_ref(root_objectid,
1153 owner, offset);
1154 }
1155 again:
1156 recow = 0;
1157 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1158 if (ret < 0) {
1159 err = ret;
1160 goto fail;
1161 }
1162
1163 if (parent) {
1164 if (!ret)
1165 return 0;
1166 goto fail;
1167 }
1168
1169 leaf = path->nodes[0];
1170 nritems = btrfs_header_nritems(leaf);
1171 while (1) {
1172 if (path->slots[0] >= nritems) {
1173 ret = btrfs_next_leaf(root, path);
1174 if (ret < 0)
1175 err = ret;
1176 if (ret)
1177 goto fail;
1178
1179 leaf = path->nodes[0];
1180 nritems = btrfs_header_nritems(leaf);
1181 recow = 1;
1182 }
1183
1184 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1185 if (key.objectid != bytenr ||
1186 key.type != BTRFS_EXTENT_DATA_REF_KEY)
1187 goto fail;
1188
1189 ref = btrfs_item_ptr(leaf, path->slots[0],
1190 struct btrfs_extent_data_ref);
1191
1192 if (match_extent_data_ref(leaf, ref, root_objectid,
1193 owner, offset)) {
1194 if (recow) {
1195 btrfs_release_path(path);
1196 goto again;
1197 }
1198 err = 0;
1199 break;
1200 }
1201 path->slots[0]++;
1202 }
1203 fail:
1204 return err;
1205 }
1206
1207 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1208 struct btrfs_path *path,
1209 u64 bytenr, u64 parent,
1210 u64 root_objectid, u64 owner,
1211 u64 offset, int refs_to_add)
1212 {
1213 struct btrfs_root *root = trans->fs_info->extent_root;
1214 struct btrfs_key key;
1215 struct extent_buffer *leaf;
1216 u32 size;
1217 u32 num_refs;
1218 int ret;
1219
1220 key.objectid = bytenr;
1221 if (parent) {
1222 key.type = BTRFS_SHARED_DATA_REF_KEY;
1223 key.offset = parent;
1224 size = sizeof(struct btrfs_shared_data_ref);
1225 } else {
1226 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1227 key.offset = hash_extent_data_ref(root_objectid,
1228 owner, offset);
1229 size = sizeof(struct btrfs_extent_data_ref);
1230 }
1231
1232 ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1233 if (ret && ret != -EEXIST)
1234 goto fail;
1235
1236 leaf = path->nodes[0];
1237 if (parent) {
1238 struct btrfs_shared_data_ref *ref;
1239 ref = btrfs_item_ptr(leaf, path->slots[0],
1240 struct btrfs_shared_data_ref);
1241 if (ret == 0) {
1242 btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1243 } else {
1244 num_refs = btrfs_shared_data_ref_count(leaf, ref);
1245 num_refs += refs_to_add;
1246 btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1247 }
1248 } else {
1249 struct btrfs_extent_data_ref *ref;
1250 while (ret == -EEXIST) {
1251 ref = btrfs_item_ptr(leaf, path->slots[0],
1252 struct btrfs_extent_data_ref);
1253 if (match_extent_data_ref(leaf, ref, root_objectid,
1254 owner, offset))
1255 break;
1256 btrfs_release_path(path);
1257 key.offset++;
1258 ret = btrfs_insert_empty_item(trans, root, path, &key,
1259 size);
1260 if (ret && ret != -EEXIST)
1261 goto fail;
1262
1263 leaf = path->nodes[0];
1264 }
1265 ref = btrfs_item_ptr(leaf, path->slots[0],
1266 struct btrfs_extent_data_ref);
1267 if (ret == 0) {
1268 btrfs_set_extent_data_ref_root(leaf, ref,
1269 root_objectid);
1270 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1271 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1272 btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1273 } else {
1274 num_refs = btrfs_extent_data_ref_count(leaf, ref);
1275 num_refs += refs_to_add;
1276 btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1277 }
1278 }
1279 btrfs_mark_buffer_dirty(leaf);
1280 ret = 0;
1281 fail:
1282 btrfs_release_path(path);
1283 return ret;
1284 }
1285
1286 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1287 struct btrfs_path *path,
1288 int refs_to_drop, int *last_ref)
1289 {
1290 struct btrfs_key key;
1291 struct btrfs_extent_data_ref *ref1 = NULL;
1292 struct btrfs_shared_data_ref *ref2 = NULL;
1293 struct extent_buffer *leaf;
1294 u32 num_refs = 0;
1295 int ret = 0;
1296
1297 leaf = path->nodes[0];
1298 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1299
1300 if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1301 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1302 struct btrfs_extent_data_ref);
1303 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1304 } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1305 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1306 struct btrfs_shared_data_ref);
1307 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1308 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1309 btrfs_print_v0_err(trans->fs_info);
1310 btrfs_abort_transaction(trans, -EINVAL);
1311 return -EINVAL;
1312 } else {
1313 BUG();
1314 }
1315
1316 BUG_ON(num_refs < refs_to_drop);
1317 num_refs -= refs_to_drop;
1318
1319 if (num_refs == 0) {
1320 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1321 *last_ref = 1;
1322 } else {
1323 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1324 btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1325 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1326 btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1327 btrfs_mark_buffer_dirty(leaf);
1328 }
1329 return ret;
1330 }
1331
1332 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1333 struct btrfs_extent_inline_ref *iref)
1334 {
1335 struct btrfs_key key;
1336 struct extent_buffer *leaf;
1337 struct btrfs_extent_data_ref *ref1;
1338 struct btrfs_shared_data_ref *ref2;
1339 u32 num_refs = 0;
1340 int type;
1341
1342 leaf = path->nodes[0];
1343 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1344
1345 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1346 if (iref) {
1347 /*
1348 * If type is invalid, we should have bailed out earlier than
1349 * this call.
1350 */
1351 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1352 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1353 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1354 ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1355 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1356 } else {
1357 ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1358 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1359 }
1360 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1361 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1362 struct btrfs_extent_data_ref);
1363 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1364 } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1365 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1366 struct btrfs_shared_data_ref);
1367 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1368 } else {
1369 WARN_ON(1);
1370 }
1371 return num_refs;
1372 }
1373
1374 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1375 struct btrfs_path *path,
1376 u64 bytenr, u64 parent,
1377 u64 root_objectid)
1378 {
1379 struct btrfs_root *root = trans->fs_info->extent_root;
1380 struct btrfs_key key;
1381 int ret;
1382
1383 key.objectid = bytenr;
1384 if (parent) {
1385 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1386 key.offset = parent;
1387 } else {
1388 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1389 key.offset = root_objectid;
1390 }
1391
1392 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1393 if (ret > 0)
1394 ret = -ENOENT;
1395 return ret;
1396 }
1397
1398 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1399 struct btrfs_path *path,
1400 u64 bytenr, u64 parent,
1401 u64 root_objectid)
1402 {
1403 struct btrfs_key key;
1404 int ret;
1405
1406 key.objectid = bytenr;
1407 if (parent) {
1408 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1409 key.offset = parent;
1410 } else {
1411 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1412 key.offset = root_objectid;
1413 }
1414
1415 ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1416 path, &key, 0);
1417 btrfs_release_path(path);
1418 return ret;
1419 }
1420
1421 static inline int extent_ref_type(u64 parent, u64 owner)
1422 {
1423 int type;
1424 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1425 if (parent > 0)
1426 type = BTRFS_SHARED_BLOCK_REF_KEY;
1427 else
1428 type = BTRFS_TREE_BLOCK_REF_KEY;
1429 } else {
1430 if (parent > 0)
1431 type = BTRFS_SHARED_DATA_REF_KEY;
1432 else
1433 type = BTRFS_EXTENT_DATA_REF_KEY;
1434 }
1435 return type;
1436 }
1437
1438 static int find_next_key(struct btrfs_path *path, int level,
1439 struct btrfs_key *key)
1440
1441 {
1442 for (; level < BTRFS_MAX_LEVEL; level++) {
1443 if (!path->nodes[level])
1444 break;
1445 if (path->slots[level] + 1 >=
1446 btrfs_header_nritems(path->nodes[level]))
1447 continue;
1448 if (level == 0)
1449 btrfs_item_key_to_cpu(path->nodes[level], key,
1450 path->slots[level] + 1);
1451 else
1452 btrfs_node_key_to_cpu(path->nodes[level], key,
1453 path->slots[level] + 1);
1454 return 0;
1455 }
1456 return 1;
1457 }
1458
1459 /*
1460 * look for inline back ref. if back ref is found, *ref_ret is set
1461 * to the address of inline back ref, and 0 is returned.
1462 *
1463 * if back ref isn't found, *ref_ret is set to the address where it
1464 * should be inserted, and -ENOENT is returned.
1465 *
1466 * if insert is true and there are too many inline back refs, the path
1467 * points to the extent item, and -EAGAIN is returned.
1468 *
1469 * NOTE: inline back refs are ordered in the same way that back ref
1470 * items in the tree are ordered.
1471 */
1472 static noinline_for_stack
1473 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1474 struct btrfs_path *path,
1475 struct btrfs_extent_inline_ref **ref_ret,
1476 u64 bytenr, u64 num_bytes,
1477 u64 parent, u64 root_objectid,
1478 u64 owner, u64 offset, int insert)
1479 {
1480 struct btrfs_fs_info *fs_info = trans->fs_info;
1481 struct btrfs_root *root = fs_info->extent_root;
1482 struct btrfs_key key;
1483 struct extent_buffer *leaf;
1484 struct btrfs_extent_item *ei;
1485 struct btrfs_extent_inline_ref *iref;
1486 u64 flags;
1487 u64 item_size;
1488 unsigned long ptr;
1489 unsigned long end;
1490 int extra_size;
1491 int type;
1492 int want;
1493 int ret;
1494 int err = 0;
1495 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1496 int needed;
1497
1498 key.objectid = bytenr;
1499 key.type = BTRFS_EXTENT_ITEM_KEY;
1500 key.offset = num_bytes;
1501
1502 want = extent_ref_type(parent, owner);
1503 if (insert) {
1504 extra_size = btrfs_extent_inline_ref_size(want);
1505 path->keep_locks = 1;
1506 } else
1507 extra_size = -1;
1508
1509 /*
1510 * Owner is our level, so we can just add one to get the level for the
1511 * block we are interested in.
1512 */
1513 if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1514 key.type = BTRFS_METADATA_ITEM_KEY;
1515 key.offset = owner;
1516 }
1517
1518 again:
1519 ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1520 if (ret < 0) {
1521 err = ret;
1522 goto out;
1523 }
1524
1525 /*
1526 * We may be a newly converted file system which still has the old fat
1527 * extent entries for metadata, so try and see if we have one of those.
1528 */
1529 if (ret > 0 && skinny_metadata) {
1530 skinny_metadata = false;
1531 if (path->slots[0]) {
1532 path->slots[0]--;
1533 btrfs_item_key_to_cpu(path->nodes[0], &key,
1534 path->slots[0]);
1535 if (key.objectid == bytenr &&
1536 key.type == BTRFS_EXTENT_ITEM_KEY &&
1537 key.offset == num_bytes)
1538 ret = 0;
1539 }
1540 if (ret) {
1541 key.objectid = bytenr;
1542 key.type = BTRFS_EXTENT_ITEM_KEY;
1543 key.offset = num_bytes;
1544 btrfs_release_path(path);
1545 goto again;
1546 }
1547 }
1548
1549 if (ret && !insert) {
1550 err = -ENOENT;
1551 goto out;
1552 } else if (WARN_ON(ret)) {
1553 err = -EIO;
1554 goto out;
1555 }
1556
1557 leaf = path->nodes[0];
1558 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1559 if (unlikely(item_size < sizeof(*ei))) {
1560 err = -EINVAL;
1561 btrfs_print_v0_err(fs_info);
1562 btrfs_abort_transaction(trans, err);
1563 goto out;
1564 }
1565
1566 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1567 flags = btrfs_extent_flags(leaf, ei);
1568
1569 ptr = (unsigned long)(ei + 1);
1570 end = (unsigned long)ei + item_size;
1571
1572 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1573 ptr += sizeof(struct btrfs_tree_block_info);
1574 BUG_ON(ptr > end);
1575 }
1576
1577 if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1578 needed = BTRFS_REF_TYPE_DATA;
1579 else
1580 needed = BTRFS_REF_TYPE_BLOCK;
1581
1582 err = -ENOENT;
1583 while (1) {
1584 if (ptr >= end) {
1585 WARN_ON(ptr > end);
1586 break;
1587 }
1588 iref = (struct btrfs_extent_inline_ref *)ptr;
1589 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1590 if (type == BTRFS_REF_TYPE_INVALID) {
1591 err = -EUCLEAN;
1592 goto out;
1593 }
1594
1595 if (want < type)
1596 break;
1597 if (want > type) {
1598 ptr += btrfs_extent_inline_ref_size(type);
1599 continue;
1600 }
1601
1602 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1603 struct btrfs_extent_data_ref *dref;
1604 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1605 if (match_extent_data_ref(leaf, dref, root_objectid,
1606 owner, offset)) {
1607 err = 0;
1608 break;
1609 }
1610 if (hash_extent_data_ref_item(leaf, dref) <
1611 hash_extent_data_ref(root_objectid, owner, offset))
1612 break;
1613 } else {
1614 u64 ref_offset;
1615 ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1616 if (parent > 0) {
1617 if (parent == ref_offset) {
1618 err = 0;
1619 break;
1620 }
1621 if (ref_offset < parent)
1622 break;
1623 } else {
1624 if (root_objectid == ref_offset) {
1625 err = 0;
1626 break;
1627 }
1628 if (ref_offset < root_objectid)
1629 break;
1630 }
1631 }
1632 ptr += btrfs_extent_inline_ref_size(type);
1633 }
1634 if (err == -ENOENT && insert) {
1635 if (item_size + extra_size >=
1636 BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1637 err = -EAGAIN;
1638 goto out;
1639 }
1640 /*
1641 * To add new inline back ref, we have to make sure
1642 * there is no corresponding back ref item.
1643 * For simplicity, we just do not add new inline back
1644 * ref if there is any kind of item for this block
1645 */
1646 if (find_next_key(path, 0, &key) == 0 &&
1647 key.objectid == bytenr &&
1648 key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1649 err = -EAGAIN;
1650 goto out;
1651 }
1652 }
1653 *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1654 out:
1655 if (insert) {
1656 path->keep_locks = 0;
1657 btrfs_unlock_up_safe(path, 1);
1658 }
1659 return err;
1660 }
1661
1662 /*
1663 * helper to add new inline back ref
1664 */
1665 static noinline_for_stack
1666 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1667 struct btrfs_path *path,
1668 struct btrfs_extent_inline_ref *iref,
1669 u64 parent, u64 root_objectid,
1670 u64 owner, u64 offset, int refs_to_add,
1671 struct btrfs_delayed_extent_op *extent_op)
1672 {
1673 struct extent_buffer *leaf;
1674 struct btrfs_extent_item *ei;
1675 unsigned long ptr;
1676 unsigned long end;
1677 unsigned long item_offset;
1678 u64 refs;
1679 int size;
1680 int type;
1681
1682 leaf = path->nodes[0];
1683 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1684 item_offset = (unsigned long)iref - (unsigned long)ei;
1685
1686 type = extent_ref_type(parent, owner);
1687 size = btrfs_extent_inline_ref_size(type);
1688
1689 btrfs_extend_item(fs_info, path, size);
1690
1691 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1692 refs = btrfs_extent_refs(leaf, ei);
1693 refs += refs_to_add;
1694 btrfs_set_extent_refs(leaf, ei, refs);
1695 if (extent_op)
1696 __run_delayed_extent_op(extent_op, leaf, ei);
1697
1698 ptr = (unsigned long)ei + item_offset;
1699 end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1700 if (ptr < end - size)
1701 memmove_extent_buffer(leaf, ptr + size, ptr,
1702 end - size - ptr);
1703
1704 iref = (struct btrfs_extent_inline_ref *)ptr;
1705 btrfs_set_extent_inline_ref_type(leaf, iref, type);
1706 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1707 struct btrfs_extent_data_ref *dref;
1708 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1709 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1710 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1711 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1712 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1713 } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1714 struct btrfs_shared_data_ref *sref;
1715 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1716 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1717 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1718 } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1719 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1720 } else {
1721 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1722 }
1723 btrfs_mark_buffer_dirty(leaf);
1724 }
1725
1726 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1727 struct btrfs_path *path,
1728 struct btrfs_extent_inline_ref **ref_ret,
1729 u64 bytenr, u64 num_bytes, u64 parent,
1730 u64 root_objectid, u64 owner, u64 offset)
1731 {
1732 int ret;
1733
1734 ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1735 num_bytes, parent, root_objectid,
1736 owner, offset, 0);
1737 if (ret != -ENOENT)
1738 return ret;
1739
1740 btrfs_release_path(path);
1741 *ref_ret = NULL;
1742
1743 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1744 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1745 root_objectid);
1746 } else {
1747 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1748 root_objectid, owner, offset);
1749 }
1750 return ret;
1751 }
1752
1753 /*
1754 * helper to update/remove inline back ref
1755 */
1756 static noinline_for_stack
1757 void update_inline_extent_backref(struct btrfs_path *path,
1758 struct btrfs_extent_inline_ref *iref,
1759 int refs_to_mod,
1760 struct btrfs_delayed_extent_op *extent_op,
1761 int *last_ref)
1762 {
1763 struct extent_buffer *leaf = path->nodes[0];
1764 struct btrfs_fs_info *fs_info = leaf->fs_info;
1765 struct btrfs_extent_item *ei;
1766 struct btrfs_extent_data_ref *dref = NULL;
1767 struct btrfs_shared_data_ref *sref = NULL;
1768 unsigned long ptr;
1769 unsigned long end;
1770 u32 item_size;
1771 int size;
1772 int type;
1773 u64 refs;
1774
1775 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1776 refs = btrfs_extent_refs(leaf, ei);
1777 WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1778 refs += refs_to_mod;
1779 btrfs_set_extent_refs(leaf, ei, refs);
1780 if (extent_op)
1781 __run_delayed_extent_op(extent_op, leaf, ei);
1782
1783 /*
1784 * If type is invalid, we should have bailed out after
1785 * lookup_inline_extent_backref().
1786 */
1787 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1788 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1789
1790 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1791 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1792 refs = btrfs_extent_data_ref_count(leaf, dref);
1793 } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1794 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1795 refs = btrfs_shared_data_ref_count(leaf, sref);
1796 } else {
1797 refs = 1;
1798 BUG_ON(refs_to_mod != -1);
1799 }
1800
1801 BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1802 refs += refs_to_mod;
1803
1804 if (refs > 0) {
1805 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1806 btrfs_set_extent_data_ref_count(leaf, dref, refs);
1807 else
1808 btrfs_set_shared_data_ref_count(leaf, sref, refs);
1809 } else {
1810 *last_ref = 1;
1811 size = btrfs_extent_inline_ref_size(type);
1812 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1813 ptr = (unsigned long)iref;
1814 end = (unsigned long)ei + item_size;
1815 if (ptr + size < end)
1816 memmove_extent_buffer(leaf, ptr, ptr + size,
1817 end - ptr - size);
1818 item_size -= size;
1819 btrfs_truncate_item(fs_info, path, item_size, 1);
1820 }
1821 btrfs_mark_buffer_dirty(leaf);
1822 }
1823
1824 static noinline_for_stack
1825 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1826 struct btrfs_path *path,
1827 u64 bytenr, u64 num_bytes, u64 parent,
1828 u64 root_objectid, u64 owner,
1829 u64 offset, int refs_to_add,
1830 struct btrfs_delayed_extent_op *extent_op)
1831 {
1832 struct btrfs_extent_inline_ref *iref;
1833 int ret;
1834
1835 ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1836 num_bytes, parent, root_objectid,
1837 owner, offset, 1);
1838 if (ret == 0) {
1839 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1840 update_inline_extent_backref(path, iref, refs_to_add,
1841 extent_op, NULL);
1842 } else if (ret == -ENOENT) {
1843 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1844 root_objectid, owner, offset,
1845 refs_to_add, extent_op);
1846 ret = 0;
1847 }
1848 return ret;
1849 }
1850
1851 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1852 struct btrfs_path *path,
1853 u64 bytenr, u64 parent, u64 root_objectid,
1854 u64 owner, u64 offset, int refs_to_add)
1855 {
1856 int ret;
1857 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1858 BUG_ON(refs_to_add != 1);
1859 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1860 root_objectid);
1861 } else {
1862 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1863 root_objectid, owner, offset,
1864 refs_to_add);
1865 }
1866 return ret;
1867 }
1868
1869 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1870 struct btrfs_path *path,
1871 struct btrfs_extent_inline_ref *iref,
1872 int refs_to_drop, int is_data, int *last_ref)
1873 {
1874 int ret = 0;
1875
1876 BUG_ON(!is_data && refs_to_drop != 1);
1877 if (iref) {
1878 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1879 last_ref);
1880 } else if (is_data) {
1881 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1882 last_ref);
1883 } else {
1884 *last_ref = 1;
1885 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1886 }
1887 return ret;
1888 }
1889
1890 #define in_range(b, first, len) ((b) >= (first) && (b) < (first) + (len))
1891 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1892 u64 *discarded_bytes)
1893 {
1894 int j, ret = 0;
1895 u64 bytes_left, end;
1896 u64 aligned_start = ALIGN(start, 1 << 9);
1897
1898 if (WARN_ON(start != aligned_start)) {
1899 len -= aligned_start - start;
1900 len = round_down(len, 1 << 9);
1901 start = aligned_start;
1902 }
1903
1904 *discarded_bytes = 0;
1905
1906 if (!len)
1907 return 0;
1908
1909 end = start + len;
1910 bytes_left = len;
1911
1912 /* Skip any superblocks on this device. */
1913 for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1914 u64 sb_start = btrfs_sb_offset(j);
1915 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1916 u64 size = sb_start - start;
1917
1918 if (!in_range(sb_start, start, bytes_left) &&
1919 !in_range(sb_end, start, bytes_left) &&
1920 !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1921 continue;
1922
1923 /*
1924 * Superblock spans beginning of range. Adjust start and
1925 * try again.
1926 */
1927 if (sb_start <= start) {
1928 start += sb_end - start;
1929 if (start > end) {
1930 bytes_left = 0;
1931 break;
1932 }
1933 bytes_left = end - start;
1934 continue;
1935 }
1936
1937 if (size) {
1938 ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1939 GFP_NOFS, 0);
1940 if (!ret)
1941 *discarded_bytes += size;
1942 else if (ret != -EOPNOTSUPP)
1943 return ret;
1944 }
1945
1946 start = sb_end;
1947 if (start > end) {
1948 bytes_left = 0;
1949 break;
1950 }
1951 bytes_left = end - start;
1952 }
1953
1954 if (bytes_left) {
1955 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1956 GFP_NOFS, 0);
1957 if (!ret)
1958 *discarded_bytes += bytes_left;
1959 }
1960 return ret;
1961 }
1962
1963 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1964 u64 num_bytes, u64 *actual_bytes)
1965 {
1966 int ret;
1967 u64 discarded_bytes = 0;
1968 struct btrfs_bio *bbio = NULL;
1969
1970
1971 /*
1972 * Avoid races with device replace and make sure our bbio has devices
1973 * associated to its stripes that don't go away while we are discarding.
1974 */
1975 btrfs_bio_counter_inc_blocked(fs_info);
1976 /* Tell the block device(s) that the sectors can be discarded */
1977 ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1978 &bbio, 0);
1979 /* Error condition is -ENOMEM */
1980 if (!ret) {
1981 struct btrfs_bio_stripe *stripe = bbio->stripes;
1982 int i;
1983
1984
1985 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1986 u64 bytes;
1987 struct request_queue *req_q;
1988
1989 if (!stripe->dev->bdev) {
1990 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1991 continue;
1992 }
1993 req_q = bdev_get_queue(stripe->dev->bdev);
1994 if (!blk_queue_discard(req_q))
1995 continue;
1996
1997 ret = btrfs_issue_discard(stripe->dev->bdev,
1998 stripe->physical,
1999 stripe->length,
2000 &bytes);
2001 if (!ret)
2002 discarded_bytes += bytes;
2003 else if (ret != -EOPNOTSUPP)
2004 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2005
2006 /*
2007 * Just in case we get back EOPNOTSUPP for some reason,
2008 * just ignore the return value so we don't screw up
2009 * people calling discard_extent.
2010 */
2011 ret = 0;
2012 }
2013 btrfs_put_bbio(bbio);
2014 }
2015 btrfs_bio_counter_dec(fs_info);
2016
2017 if (actual_bytes)
2018 *actual_bytes = discarded_bytes;
2019
2020
2021 if (ret == -EOPNOTSUPP)
2022 ret = 0;
2023 return ret;
2024 }
2025
2026 /* Can return -ENOMEM */
2027 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2028 struct btrfs_root *root,
2029 u64 bytenr, u64 num_bytes, u64 parent,
2030 u64 root_objectid, u64 owner, u64 offset)
2031 {
2032 struct btrfs_fs_info *fs_info = root->fs_info;
2033 int old_ref_mod, new_ref_mod;
2034 int ret;
2035
2036 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2037 root_objectid == BTRFS_TREE_LOG_OBJECTID);
2038
2039 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2040 owner, offset, BTRFS_ADD_DELAYED_REF);
2041
2042 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2043 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2044 num_bytes, parent,
2045 root_objectid, (int)owner,
2046 BTRFS_ADD_DELAYED_REF, NULL,
2047 &old_ref_mod, &new_ref_mod);
2048 } else {
2049 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2050 num_bytes, parent,
2051 root_objectid, owner, offset,
2052 0, BTRFS_ADD_DELAYED_REF,
2053 &old_ref_mod, &new_ref_mod);
2054 }
2055
2056 if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2057 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2058
2059 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2060 }
2061
2062 return ret;
2063 }
2064
2065 /*
2066 * __btrfs_inc_extent_ref - insert backreference for a given extent
2067 *
2068 * @trans: Handle of transaction
2069 *
2070 * @node: The delayed ref node used to get the bytenr/length for
2071 * extent whose references are incremented.
2072 *
2073 * @parent: If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2074 * BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2075 * bytenr of the parent block. Since new extents are always
2076 * created with indirect references, this will only be the case
2077 * when relocating a shared extent. In that case, root_objectid
2078 * will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2079 * be 0
2080 *
2081 * @root_objectid: The id of the root where this modification has originated,
2082 * this can be either one of the well-known metadata trees or
2083 * the subvolume id which references this extent.
2084 *
2085 * @owner: For data extents it is the inode number of the owning file.
2086 * For metadata extents this parameter holds the level in the
2087 * tree of the extent.
2088 *
2089 * @offset: For metadata extents the offset is ignored and is currently
2090 * always passed as 0. For data extents it is the fileoffset
2091 * this extent belongs to.
2092 *
2093 * @refs_to_add Number of references to add
2094 *
2095 * @extent_op Pointer to a structure, holding information necessary when
2096 * updating a tree block's flags
2097 *
2098 */
2099 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2100 struct btrfs_delayed_ref_node *node,
2101 u64 parent, u64 root_objectid,
2102 u64 owner, u64 offset, int refs_to_add,
2103 struct btrfs_delayed_extent_op *extent_op)
2104 {
2105 struct btrfs_path *path;
2106 struct extent_buffer *leaf;
2107 struct btrfs_extent_item *item;
2108 struct btrfs_key key;
2109 u64 bytenr = node->bytenr;
2110 u64 num_bytes = node->num_bytes;
2111 u64 refs;
2112 int ret;
2113
2114 path = btrfs_alloc_path();
2115 if (!path)
2116 return -ENOMEM;
2117
2118 path->reada = READA_FORWARD;
2119 path->leave_spinning = 1;
2120 /* this will setup the path even if it fails to insert the back ref */
2121 ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2122 parent, root_objectid, owner,
2123 offset, refs_to_add, extent_op);
2124 if ((ret < 0 && ret != -EAGAIN) || !ret)
2125 goto out;
2126
2127 /*
2128 * Ok we had -EAGAIN which means we didn't have space to insert and
2129 * inline extent ref, so just update the reference count and add a
2130 * normal backref.
2131 */
2132 leaf = path->nodes[0];
2133 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2134 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2135 refs = btrfs_extent_refs(leaf, item);
2136 btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2137 if (extent_op)
2138 __run_delayed_extent_op(extent_op, leaf, item);
2139
2140 btrfs_mark_buffer_dirty(leaf);
2141 btrfs_release_path(path);
2142
2143 path->reada = READA_FORWARD;
2144 path->leave_spinning = 1;
2145 /* now insert the actual backref */
2146 ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2147 owner, offset, refs_to_add);
2148 if (ret)
2149 btrfs_abort_transaction(trans, ret);
2150 out:
2151 btrfs_free_path(path);
2152 return ret;
2153 }
2154
2155 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2156 struct btrfs_delayed_ref_node *node,
2157 struct btrfs_delayed_extent_op *extent_op,
2158 int insert_reserved)
2159 {
2160 int ret = 0;
2161 struct btrfs_delayed_data_ref *ref;
2162 struct btrfs_key ins;
2163 u64 parent = 0;
2164 u64 ref_root = 0;
2165 u64 flags = 0;
2166
2167 ins.objectid = node->bytenr;
2168 ins.offset = node->num_bytes;
2169 ins.type = BTRFS_EXTENT_ITEM_KEY;
2170
2171 ref = btrfs_delayed_node_to_data_ref(node);
2172 trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2173
2174 if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2175 parent = ref->parent;
2176 ref_root = ref->root;
2177
2178 if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2179 if (extent_op)
2180 flags |= extent_op->flags_to_set;
2181 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2182 flags, ref->objectid,
2183 ref->offset, &ins,
2184 node->ref_mod);
2185 } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2186 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2187 ref->objectid, ref->offset,
2188 node->ref_mod, extent_op);
2189 } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2190 ret = __btrfs_free_extent(trans, node, parent,
2191 ref_root, ref->objectid,
2192 ref->offset, node->ref_mod,
2193 extent_op);
2194 } else {
2195 BUG();
2196 }
2197 return ret;
2198 }
2199
2200 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2201 struct extent_buffer *leaf,
2202 struct btrfs_extent_item *ei)
2203 {
2204 u64 flags = btrfs_extent_flags(leaf, ei);
2205 if (extent_op->update_flags) {
2206 flags |= extent_op->flags_to_set;
2207 btrfs_set_extent_flags(leaf, ei, flags);
2208 }
2209
2210 if (extent_op->update_key) {
2211 struct btrfs_tree_block_info *bi;
2212 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2213 bi = (struct btrfs_tree_block_info *)(ei + 1);
2214 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2215 }
2216 }
2217
2218 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2219 struct btrfs_delayed_ref_head *head,
2220 struct btrfs_delayed_extent_op *extent_op)
2221 {
2222 struct btrfs_fs_info *fs_info = trans->fs_info;
2223 struct btrfs_key key;
2224 struct btrfs_path *path;
2225 struct btrfs_extent_item *ei;
2226 struct extent_buffer *leaf;
2227 u32 item_size;
2228 int ret;
2229 int err = 0;
2230 int metadata = !extent_op->is_data;
2231
2232 if (trans->aborted)
2233 return 0;
2234
2235 if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2236 metadata = 0;
2237
2238 path = btrfs_alloc_path();
2239 if (!path)
2240 return -ENOMEM;
2241
2242 key.objectid = head->bytenr;
2243
2244 if (metadata) {
2245 key.type = BTRFS_METADATA_ITEM_KEY;
2246 key.offset = extent_op->level;
2247 } else {
2248 key.type = BTRFS_EXTENT_ITEM_KEY;
2249 key.offset = head->num_bytes;
2250 }
2251
2252 again:
2253 path->reada = READA_FORWARD;
2254 path->leave_spinning = 1;
2255 ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2256 if (ret < 0) {
2257 err = ret;
2258 goto out;
2259 }
2260 if (ret > 0) {
2261 if (metadata) {
2262 if (path->slots[0] > 0) {
2263 path->slots[0]--;
2264 btrfs_item_key_to_cpu(path->nodes[0], &key,
2265 path->slots[0]);
2266 if (key.objectid == head->bytenr &&
2267 key.type == BTRFS_EXTENT_ITEM_KEY &&
2268 key.offset == head->num_bytes)
2269 ret = 0;
2270 }
2271 if (ret > 0) {
2272 btrfs_release_path(path);
2273 metadata = 0;
2274
2275 key.objectid = head->bytenr;
2276 key.offset = head->num_bytes;
2277 key.type = BTRFS_EXTENT_ITEM_KEY;
2278 goto again;
2279 }
2280 } else {
2281 err = -EIO;
2282 goto out;
2283 }
2284 }
2285
2286 leaf = path->nodes[0];
2287 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2288
2289 if (unlikely(item_size < sizeof(*ei))) {
2290 err = -EINVAL;
2291 btrfs_print_v0_err(fs_info);
2292 btrfs_abort_transaction(trans, err);
2293 goto out;
2294 }
2295
2296 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2297 __run_delayed_extent_op(extent_op, leaf, ei);
2298
2299 btrfs_mark_buffer_dirty(leaf);
2300 out:
2301 btrfs_free_path(path);
2302 return err;
2303 }
2304
2305 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2306 struct btrfs_delayed_ref_node *node,
2307 struct btrfs_delayed_extent_op *extent_op,
2308 int insert_reserved)
2309 {
2310 int ret = 0;
2311 struct btrfs_delayed_tree_ref *ref;
2312 u64 parent = 0;
2313 u64 ref_root = 0;
2314
2315 ref = btrfs_delayed_node_to_tree_ref(node);
2316 trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2317
2318 if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2319 parent = ref->parent;
2320 ref_root = ref->root;
2321
2322 if (node->ref_mod != 1) {
2323 btrfs_err(trans->fs_info,
2324 "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2325 node->bytenr, node->ref_mod, node->action, ref_root,
2326 parent);
2327 return -EIO;
2328 }
2329 if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2330 BUG_ON(!extent_op || !extent_op->update_flags);
2331 ret = alloc_reserved_tree_block(trans, node, extent_op);
2332 } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2333 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2334 ref->level, 0, 1, extent_op);
2335 } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2336 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2337 ref->level, 0, 1, extent_op);
2338 } else {
2339 BUG();
2340 }
2341 return ret;
2342 }
2343
2344 /* helper function to actually process a single delayed ref entry */
2345 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2346 struct btrfs_delayed_ref_node *node,
2347 struct btrfs_delayed_extent_op *extent_op,
2348 int insert_reserved)
2349 {
2350 int ret = 0;
2351
2352 if (trans->aborted) {
2353 if (insert_reserved)
2354 btrfs_pin_extent(trans->fs_info, node->bytenr,
2355 node->num_bytes, 1);
2356 return 0;
2357 }
2358
2359 if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2360 node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2361 ret = run_delayed_tree_ref(trans, node, extent_op,
2362 insert_reserved);
2363 else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2364 node->type == BTRFS_SHARED_DATA_REF_KEY)
2365 ret = run_delayed_data_ref(trans, node, extent_op,
2366 insert_reserved);
2367 else
2368 BUG();
2369 if (ret && insert_reserved)
2370 btrfs_pin_extent(trans->fs_info, node->bytenr,
2371 node->num_bytes, 1);
2372 return ret;
2373 }
2374
2375 static inline struct btrfs_delayed_ref_node *
2376 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2377 {
2378 struct btrfs_delayed_ref_node *ref;
2379
2380 if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2381 return NULL;
2382
2383 /*
2384 * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2385 * This is to prevent a ref count from going down to zero, which deletes
2386 * the extent item from the extent tree, when there still are references
2387 * to add, which would fail because they would not find the extent item.
2388 */
2389 if (!list_empty(&head->ref_add_list))
2390 return list_first_entry(&head->ref_add_list,
2391 struct btrfs_delayed_ref_node, add_list);
2392
2393 ref = rb_entry(rb_first_cached(&head->ref_tree),
2394 struct btrfs_delayed_ref_node, ref_node);
2395 ASSERT(list_empty(&ref->add_list));
2396 return ref;
2397 }
2398
2399 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2400 struct btrfs_delayed_ref_head *head)
2401 {
2402 spin_lock(&delayed_refs->lock);
2403 head->processing = 0;
2404 delayed_refs->num_heads_ready++;
2405 spin_unlock(&delayed_refs->lock);
2406 btrfs_delayed_ref_unlock(head);
2407 }
2408
2409 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2410 struct btrfs_delayed_ref_head *head)
2411 {
2412 struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2413 int ret;
2414
2415 if (!extent_op)
2416 return 0;
2417 head->extent_op = NULL;
2418 if (head->must_insert_reserved) {
2419 btrfs_free_delayed_extent_op(extent_op);
2420 return 0;
2421 }
2422 spin_unlock(&head->lock);
2423 ret = run_delayed_extent_op(trans, head, extent_op);
2424 btrfs_free_delayed_extent_op(extent_op);
2425 return ret ? ret : 1;
2426 }
2427
2428 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2429 struct btrfs_delayed_ref_head *head)
2430 {
2431
2432 struct btrfs_fs_info *fs_info = trans->fs_info;
2433 struct btrfs_delayed_ref_root *delayed_refs;
2434 int ret;
2435
2436 delayed_refs = &trans->transaction->delayed_refs;
2437
2438 ret = cleanup_extent_op(trans, head);
2439 if (ret < 0) {
2440 unselect_delayed_ref_head(delayed_refs, head);
2441 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2442 return ret;
2443 } else if (ret) {
2444 return ret;
2445 }
2446
2447 /*
2448 * Need to drop our head ref lock and re-acquire the delayed ref lock
2449 * and then re-check to make sure nobody got added.
2450 */
2451 spin_unlock(&head->lock);
2452 spin_lock(&delayed_refs->lock);
2453 spin_lock(&head->lock);
2454 if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2455 spin_unlock(&head->lock);
2456 spin_unlock(&delayed_refs->lock);
2457 return 1;
2458 }
2459 delayed_refs->num_heads--;
2460 rb_erase_cached(&head->href_node, &delayed_refs->href_root);
2461 RB_CLEAR_NODE(&head->href_node);
2462 spin_unlock(&head->lock);
2463 spin_unlock(&delayed_refs->lock);
2464 atomic_dec(&delayed_refs->num_entries);
2465
2466 trace_run_delayed_ref_head(fs_info, head, 0);
2467
2468 if (head->total_ref_mod < 0) {
2469 struct btrfs_space_info *space_info;
2470 u64 flags;
2471
2472 if (head->is_data)
2473 flags = BTRFS_BLOCK_GROUP_DATA;
2474 else if (head->is_system)
2475 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2476 else
2477 flags = BTRFS_BLOCK_GROUP_METADATA;
2478 space_info = __find_space_info(fs_info, flags);
2479 ASSERT(space_info);
2480 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2481 -head->num_bytes,
2482 BTRFS_TOTAL_BYTES_PINNED_BATCH);
2483
2484 if (head->is_data) {
2485 spin_lock(&delayed_refs->lock);
2486 delayed_refs->pending_csums -= head->num_bytes;
2487 spin_unlock(&delayed_refs->lock);
2488 }
2489 }
2490
2491 if (head->must_insert_reserved) {
2492 btrfs_pin_extent(fs_info, head->bytenr,
2493 head->num_bytes, 1);
2494 if (head->is_data) {
2495 ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2496 head->num_bytes);
2497 }
2498 }
2499
2500 /* Also free its reserved qgroup space */
2501 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2502 head->qgroup_reserved);
2503 btrfs_delayed_ref_unlock(head);
2504 btrfs_put_delayed_ref_head(head);
2505 return 0;
2506 }
2507
2508 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2509 struct btrfs_trans_handle *trans)
2510 {
2511 struct btrfs_delayed_ref_root *delayed_refs =
2512 &trans->transaction->delayed_refs;
2513 struct btrfs_delayed_ref_head *head = NULL;
2514 int ret;
2515
2516 spin_lock(&delayed_refs->lock);
2517 head = btrfs_select_ref_head(delayed_refs);
2518 if (!head) {
2519 spin_unlock(&delayed_refs->lock);
2520 return head;
2521 }
2522
2523 /*
2524 * Grab the lock that says we are going to process all the refs for
2525 * this head
2526 */
2527 ret = btrfs_delayed_ref_lock(delayed_refs, head);
2528 spin_unlock(&delayed_refs->lock);
2529
2530 /*
2531 * We may have dropped the spin lock to get the head mutex lock, and
2532 * that might have given someone else time to free the head. If that's
2533 * true, it has been removed from our list and we can move on.
2534 */
2535 if (ret == -EAGAIN)
2536 head = ERR_PTR(-EAGAIN);
2537
2538 return head;
2539 }
2540
2541 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2542 struct btrfs_delayed_ref_head *locked_ref,
2543 unsigned long *run_refs)
2544 {
2545 struct btrfs_fs_info *fs_info = trans->fs_info;
2546 struct btrfs_delayed_ref_root *delayed_refs;
2547 struct btrfs_delayed_extent_op *extent_op;
2548 struct btrfs_delayed_ref_node *ref;
2549 int must_insert_reserved = 0;
2550 int ret;
2551
2552 delayed_refs = &trans->transaction->delayed_refs;
2553
2554 lockdep_assert_held(&locked_ref->mutex);
2555 lockdep_assert_held(&locked_ref->lock);
2556
2557 while ((ref = select_delayed_ref(locked_ref))) {
2558 if (ref->seq &&
2559 btrfs_check_delayed_seq(fs_info, ref->seq)) {
2560 spin_unlock(&locked_ref->lock);
2561 unselect_delayed_ref_head(delayed_refs, locked_ref);
2562 return -EAGAIN;
2563 }
2564
2565 (*run_refs)++;
2566 ref->in_tree = 0;
2567 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2568 RB_CLEAR_NODE(&ref->ref_node);
2569 if (!list_empty(&ref->add_list))
2570 list_del(&ref->add_list);
2571 /*
2572 * When we play the delayed ref, also correct the ref_mod on
2573 * head
2574 */
2575 switch (ref->action) {
2576 case BTRFS_ADD_DELAYED_REF:
2577 case BTRFS_ADD_DELAYED_EXTENT:
2578 locked_ref->ref_mod -= ref->ref_mod;
2579 break;
2580 case BTRFS_DROP_DELAYED_REF:
2581 locked_ref->ref_mod += ref->ref_mod;
2582 break;
2583 default:
2584 WARN_ON(1);
2585 }
2586 atomic_dec(&delayed_refs->num_entries);
2587
2588 /*
2589 * Record the must_insert_reserved flag before we drop the
2590 * spin lock.
2591 */
2592 must_insert_reserved = locked_ref->must_insert_reserved;
2593 locked_ref->must_insert_reserved = 0;
2594
2595 extent_op = locked_ref->extent_op;
2596 locked_ref->extent_op = NULL;
2597 spin_unlock(&locked_ref->lock);
2598
2599 ret = run_one_delayed_ref(trans, ref, extent_op,
2600 must_insert_reserved);
2601
2602 btrfs_free_delayed_extent_op(extent_op);
2603 if (ret) {
2604 unselect_delayed_ref_head(delayed_refs, locked_ref);
2605 btrfs_put_delayed_ref(ref);
2606 btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2607 ret);
2608 return ret;
2609 }
2610
2611 btrfs_put_delayed_ref(ref);
2612 cond_resched();
2613
2614 spin_lock(&locked_ref->lock);
2615 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2616 }
2617
2618 return 0;
2619 }
2620
2621 /*
2622 * Returns 0 on success or if called with an already aborted transaction.
2623 * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2624 */
2625 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2626 unsigned long nr)
2627 {
2628 struct btrfs_fs_info *fs_info = trans->fs_info;
2629 struct btrfs_delayed_ref_root *delayed_refs;
2630 struct btrfs_delayed_ref_head *locked_ref = NULL;
2631 ktime_t start = ktime_get();
2632 int ret;
2633 unsigned long count = 0;
2634 unsigned long actual_count = 0;
2635
2636 delayed_refs = &trans->transaction->delayed_refs;
2637 do {
2638 if (!locked_ref) {
2639 locked_ref = btrfs_obtain_ref_head(trans);
2640 if (IS_ERR_OR_NULL(locked_ref)) {
2641 if (PTR_ERR(locked_ref) == -EAGAIN) {
2642 continue;
2643 } else {
2644 break;
2645 }
2646 }
2647 count++;
2648 }
2649 /*
2650 * We need to try and merge add/drops of the same ref since we
2651 * can run into issues with relocate dropping the implicit ref
2652 * and then it being added back again before the drop can
2653 * finish. If we merged anything we need to re-loop so we can
2654 * get a good ref.
2655 * Or we can get node references of the same type that weren't
2656 * merged when created due to bumps in the tree mod seq, and
2657 * we need to merge them to prevent adding an inline extent
2658 * backref before dropping it (triggering a BUG_ON at
2659 * insert_inline_extent_backref()).
2660 */
2661 spin_lock(&locked_ref->lock);
2662 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2663
2664 ret = btrfs_run_delayed_refs_for_head(trans, locked_ref,
2665 &actual_count);
2666 if (ret < 0 && ret != -EAGAIN) {
2667 /*
2668 * Error, btrfs_run_delayed_refs_for_head already
2669 * unlocked everything so just bail out
2670 */
2671 return ret;
2672 } else if (!ret) {
2673 /*
2674 * Success, perform the usual cleanup of a processed
2675 * head
2676 */
2677 ret = cleanup_ref_head(trans, locked_ref);
2678 if (ret > 0 ) {
2679 /* We dropped our lock, we need to loop. */
2680 ret = 0;
2681 continue;
2682 } else if (ret) {
2683 return ret;
2684 }
2685 }
2686
2687 /*
2688 * Either success case or btrfs_run_delayed_refs_for_head
2689 * returned -EAGAIN, meaning we need to select another head
2690 */
2691
2692 locked_ref = NULL;
2693 cond_resched();
2694 } while ((nr != -1 && count < nr) || locked_ref);
2695
2696 /*
2697 * We don't want to include ref heads since we can have empty ref heads
2698 * and those will drastically skew our runtime down since we just do
2699 * accounting, no actual extent tree updates.
2700 */
2701 if (actual_count > 0) {
2702 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2703 u64 avg;
2704
2705 /*
2706 * We weigh the current average higher than our current runtime
2707 * to avoid large swings in the average.
2708 */
2709 spin_lock(&delayed_refs->lock);
2710 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2711 fs_info->avg_delayed_ref_runtime = avg >> 2; /* div by 4 */
2712 spin_unlock(&delayed_refs->lock);
2713 }
2714 return 0;
2715 }
2716
2717 #ifdef SCRAMBLE_DELAYED_REFS
2718 /*
2719 * Normally delayed refs get processed in ascending bytenr order. This
2720 * correlates in most cases to the order added. To expose dependencies on this
2721 * order, we start to process the tree in the middle instead of the beginning
2722 */
2723 static u64 find_middle(struct rb_root *root)
2724 {
2725 struct rb_node *n = root->rb_node;
2726 struct btrfs_delayed_ref_node *entry;
2727 int alt = 1;
2728 u64 middle;
2729 u64 first = 0, last = 0;
2730
2731 n = rb_first(root);
2732 if (n) {
2733 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2734 first = entry->bytenr;
2735 }
2736 n = rb_last(root);
2737 if (n) {
2738 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2739 last = entry->bytenr;
2740 }
2741 n = root->rb_node;
2742
2743 while (n) {
2744 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2745 WARN_ON(!entry->in_tree);
2746
2747 middle = entry->bytenr;
2748
2749 if (alt)
2750 n = n->rb_left;
2751 else
2752 n = n->rb_right;
2753
2754 alt = 1 - alt;
2755 }
2756 return middle;
2757 }
2758 #endif
2759
2760 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2761 {
2762 u64 num_bytes;
2763
2764 num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2765 sizeof(struct btrfs_extent_inline_ref));
2766 if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2767 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2768
2769 /*
2770 * We don't ever fill up leaves all the way so multiply by 2 just to be
2771 * closer to what we're really going to want to use.
2772 */
2773 return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2774 }
2775
2776 /*
2777 * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2778 * would require to store the csums for that many bytes.
2779 */
2780 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2781 {
2782 u64 csum_size;
2783 u64 num_csums_per_leaf;
2784 u64 num_csums;
2785
2786 csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2787 num_csums_per_leaf = div64_u64(csum_size,
2788 (u64)btrfs_super_csum_size(fs_info->super_copy));
2789 num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2790 num_csums += num_csums_per_leaf - 1;
2791 num_csums = div64_u64(num_csums, num_csums_per_leaf);
2792 return num_csums;
2793 }
2794
2795 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans)
2796 {
2797 struct btrfs_fs_info *fs_info = trans->fs_info;
2798 struct btrfs_block_rsv *global_rsv;
2799 u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2800 u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2801 unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2802 u64 num_bytes, num_dirty_bgs_bytes;
2803 int ret = 0;
2804
2805 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2806 num_heads = heads_to_leaves(fs_info, num_heads);
2807 if (num_heads > 1)
2808 num_bytes += (num_heads - 1) * fs_info->nodesize;
2809 num_bytes <<= 1;
2810 num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2811 fs_info->nodesize;
2812 num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2813 num_dirty_bgs);
2814 global_rsv = &fs_info->global_block_rsv;
2815
2816 /*
2817 * If we can't allocate any more chunks lets make sure we have _lots_ of
2818 * wiggle room since running delayed refs can create more delayed refs.
2819 */
2820 if (global_rsv->space_info->full) {
2821 num_dirty_bgs_bytes <<= 1;
2822 num_bytes <<= 1;
2823 }
2824
2825 spin_lock(&global_rsv->lock);
2826 if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2827 ret = 1;
2828 spin_unlock(&global_rsv->lock);
2829 return ret;
2830 }
2831
2832 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
2833 {
2834 u64 num_entries =
2835 atomic_read(&trans->transaction->delayed_refs.num_entries);
2836 u64 avg_runtime;
2837 u64 val;
2838
2839 smp_mb();
2840 avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
2841 val = num_entries * avg_runtime;
2842 if (val >= NSEC_PER_SEC)
2843 return 1;
2844 if (val >= NSEC_PER_SEC / 2)
2845 return 2;
2846
2847 return btrfs_check_space_for_delayed_refs(trans);
2848 }
2849
2850 struct async_delayed_refs {
2851 struct btrfs_root *root;
2852 u64 transid;
2853 int count;
2854 int error;
2855 int sync;
2856 struct completion wait;
2857 struct btrfs_work work;
2858 };
2859
2860 static inline struct async_delayed_refs *
2861 to_async_delayed_refs(struct btrfs_work *work)
2862 {
2863 return container_of(work, struct async_delayed_refs, work);
2864 }
2865
2866 static void delayed_ref_async_start(struct btrfs_work *work)
2867 {
2868 struct async_delayed_refs *async = to_async_delayed_refs(work);
2869 struct btrfs_trans_handle *trans;
2870 struct btrfs_fs_info *fs_info = async->root->fs_info;
2871 int ret;
2872
2873 /* if the commit is already started, we don't need to wait here */
2874 if (btrfs_transaction_blocked(fs_info))
2875 goto done;
2876
2877 trans = btrfs_join_transaction(async->root);
2878 if (IS_ERR(trans)) {
2879 async->error = PTR_ERR(trans);
2880 goto done;
2881 }
2882
2883 /*
2884 * trans->sync means that when we call end_transaction, we won't
2885 * wait on delayed refs
2886 */
2887 trans->sync = true;
2888
2889 /* Don't bother flushing if we got into a different transaction */
2890 if (trans->transid > async->transid)
2891 goto end;
2892
2893 ret = btrfs_run_delayed_refs(trans, async->count);
2894 if (ret)
2895 async->error = ret;
2896 end:
2897 ret = btrfs_end_transaction(trans);
2898 if (ret && !async->error)
2899 async->error = ret;
2900 done:
2901 if (async->sync)
2902 complete(&async->wait);
2903 else
2904 kfree(async);
2905 }
2906
2907 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2908 unsigned long count, u64 transid, int wait)
2909 {
2910 struct async_delayed_refs *async;
2911 int ret;
2912
2913 async = kmalloc(sizeof(*async), GFP_NOFS);
2914 if (!async)
2915 return -ENOMEM;
2916
2917 async->root = fs_info->tree_root;
2918 async->count = count;
2919 async->error = 0;
2920 async->transid = transid;
2921 if (wait)
2922 async->sync = 1;
2923 else
2924 async->sync = 0;
2925 init_completion(&async->wait);
2926
2927 btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2928 delayed_ref_async_start, NULL, NULL);
2929
2930 btrfs_queue_work(fs_info->extent_workers, &async->work);
2931
2932 if (wait) {
2933 wait_for_completion(&async->wait);
2934 ret = async->error;
2935 kfree(async);
2936 return ret;
2937 }
2938 return 0;
2939 }
2940
2941 /*
2942 * this starts processing the delayed reference count updates and
2943 * extent insertions we have queued up so far. count can be
2944 * 0, which means to process everything in the tree at the start
2945 * of the run (but not newly added entries), or it can be some target
2946 * number you'd like to process.
2947 *
2948 * Returns 0 on success or if called with an aborted transaction
2949 * Returns <0 on error and aborts the transaction
2950 */
2951 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2952 unsigned long count)
2953 {
2954 struct btrfs_fs_info *fs_info = trans->fs_info;
2955 struct rb_node *node;
2956 struct btrfs_delayed_ref_root *delayed_refs;
2957 struct btrfs_delayed_ref_head *head;
2958 int ret;
2959 int run_all = count == (unsigned long)-1;
2960
2961 /* We'll clean this up in btrfs_cleanup_transaction */
2962 if (trans->aborted)
2963 return 0;
2964
2965 if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2966 return 0;
2967
2968 delayed_refs = &trans->transaction->delayed_refs;
2969 if (count == 0)
2970 count = atomic_read(&delayed_refs->num_entries) * 2;
2971
2972 again:
2973 #ifdef SCRAMBLE_DELAYED_REFS
2974 delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2975 #endif
2976 ret = __btrfs_run_delayed_refs(trans, count);
2977 if (ret < 0) {
2978 btrfs_abort_transaction(trans, ret);
2979 return ret;
2980 }
2981
2982 if (run_all) {
2983 if (!list_empty(&trans->new_bgs))
2984 btrfs_create_pending_block_groups(trans);
2985
2986 spin_lock(&delayed_refs->lock);
2987 node = rb_first_cached(&delayed_refs->href_root);
2988 if (!node) {
2989 spin_unlock(&delayed_refs->lock);
2990 goto out;
2991 }
2992 head = rb_entry(node, struct btrfs_delayed_ref_head,
2993 href_node);
2994 refcount_inc(&head->refs);
2995 spin_unlock(&delayed_refs->lock);
2996
2997 /* Mutex was contended, block until it's released and retry. */
2998 mutex_lock(&head->mutex);
2999 mutex_unlock(&head->mutex);
3000
3001 btrfs_put_delayed_ref_head(head);
3002 cond_resched();
3003 goto again;
3004 }
3005 out:
3006 return 0;
3007 }
3008
3009 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3010 struct btrfs_fs_info *fs_info,
3011 u64 bytenr, u64 num_bytes, u64 flags,
3012 int level, int is_data)
3013 {
3014 struct btrfs_delayed_extent_op *extent_op;
3015 int ret;
3016
3017 extent_op = btrfs_alloc_delayed_extent_op();
3018 if (!extent_op)
3019 return -ENOMEM;
3020
3021 extent_op->flags_to_set = flags;
3022 extent_op->update_flags = true;
3023 extent_op->update_key = false;
3024 extent_op->is_data = is_data ? true : false;
3025 extent_op->level = level;
3026
3027 ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3028 num_bytes, extent_op);
3029 if (ret)
3030 btrfs_free_delayed_extent_op(extent_op);
3031 return ret;
3032 }
3033
3034 static noinline int check_delayed_ref(struct btrfs_root *root,
3035 struct btrfs_path *path,
3036 u64 objectid, u64 offset, u64 bytenr)
3037 {
3038 struct btrfs_delayed_ref_head *head;
3039 struct btrfs_delayed_ref_node *ref;
3040 struct btrfs_delayed_data_ref *data_ref;
3041 struct btrfs_delayed_ref_root *delayed_refs;
3042 struct btrfs_transaction *cur_trans;
3043 struct rb_node *node;
3044 int ret = 0;
3045
3046 spin_lock(&root->fs_info->trans_lock);
3047 cur_trans = root->fs_info->running_transaction;
3048 if (cur_trans)
3049 refcount_inc(&cur_trans->use_count);
3050 spin_unlock(&root->fs_info->trans_lock);
3051 if (!cur_trans)
3052 return 0;
3053
3054 delayed_refs = &cur_trans->delayed_refs;
3055 spin_lock(&delayed_refs->lock);
3056 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3057 if (!head) {
3058 spin_unlock(&delayed_refs->lock);
3059 btrfs_put_transaction(cur_trans);
3060 return 0;
3061 }
3062
3063 if (!mutex_trylock(&head->mutex)) {
3064 refcount_inc(&head->refs);
3065 spin_unlock(&delayed_refs->lock);
3066
3067 btrfs_release_path(path);
3068
3069 /*
3070 * Mutex was contended, block until it's released and let
3071 * caller try again
3072 */
3073 mutex_lock(&head->mutex);
3074 mutex_unlock(&head->mutex);
3075 btrfs_put_delayed_ref_head(head);
3076 btrfs_put_transaction(cur_trans);
3077 return -EAGAIN;
3078 }
3079 spin_unlock(&delayed_refs->lock);
3080
3081 spin_lock(&head->lock);
3082 /*
3083 * XXX: We should replace this with a proper search function in the
3084 * future.
3085 */
3086 for (node = rb_first_cached(&head->ref_tree); node;
3087 node = rb_next(node)) {
3088 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3089 /* If it's a shared ref we know a cross reference exists */
3090 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3091 ret = 1;
3092 break;
3093 }
3094
3095 data_ref = btrfs_delayed_node_to_data_ref(ref);
3096
3097 /*
3098 * If our ref doesn't match the one we're currently looking at
3099 * then we have a cross reference.
3100 */
3101 if (data_ref->root != root->root_key.objectid ||
3102 data_ref->objectid != objectid ||
3103 data_ref->offset != offset) {
3104 ret = 1;
3105 break;
3106 }
3107 }
3108 spin_unlock(&head->lock);
3109 mutex_unlock(&head->mutex);
3110 btrfs_put_transaction(cur_trans);
3111 return ret;
3112 }
3113
3114 static noinline int check_committed_ref(struct btrfs_root *root,
3115 struct btrfs_path *path,
3116 u64 objectid, u64 offset, u64 bytenr)
3117 {
3118 struct btrfs_fs_info *fs_info = root->fs_info;
3119 struct btrfs_root *extent_root = fs_info->extent_root;
3120 struct extent_buffer *leaf;
3121 struct btrfs_extent_data_ref *ref;
3122 struct btrfs_extent_inline_ref *iref;
3123 struct btrfs_extent_item *ei;
3124 struct btrfs_key key;
3125 u32 item_size;
3126 int type;
3127 int ret;
3128
3129 key.objectid = bytenr;
3130 key.offset = (u64)-1;
3131 key.type = BTRFS_EXTENT_ITEM_KEY;
3132
3133 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3134 if (ret < 0)
3135 goto out;
3136 BUG_ON(ret == 0); /* Corruption */
3137
3138 ret = -ENOENT;
3139 if (path->slots[0] == 0)
3140 goto out;
3141
3142 path->slots[0]--;
3143 leaf = path->nodes[0];
3144 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3145
3146 if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3147 goto out;
3148
3149 ret = 1;
3150 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3151 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3152
3153 if (item_size != sizeof(*ei) +
3154 btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3155 goto out;
3156
3157 if (btrfs_extent_generation(leaf, ei) <=
3158 btrfs_root_last_snapshot(&root->root_item))
3159 goto out;
3160
3161 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3162
3163 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3164 if (type != BTRFS_EXTENT_DATA_REF_KEY)
3165 goto out;
3166
3167 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3168 if (btrfs_extent_refs(leaf, ei) !=
3169 btrfs_extent_data_ref_count(leaf, ref) ||
3170 btrfs_extent_data_ref_root(leaf, ref) !=
3171 root->root_key.objectid ||
3172 btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3173 btrfs_extent_data_ref_offset(leaf, ref) != offset)
3174 goto out;
3175
3176 ret = 0;
3177 out:
3178 return ret;
3179 }
3180
3181 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3182 u64 bytenr)
3183 {
3184 struct btrfs_path *path;
3185 int ret;
3186
3187 path = btrfs_alloc_path();
3188 if (!path)
3189 return -ENOMEM;
3190
3191 do {
3192 ret = check_committed_ref(root, path, objectid,
3193 offset, bytenr);
3194 if (ret && ret != -ENOENT)
3195 goto out;
3196
3197 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3198 } while (ret == -EAGAIN);
3199
3200 out:
3201 btrfs_free_path(path);
3202 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3203 WARN_ON(ret > 0);
3204 return ret;
3205 }
3206
3207 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3208 struct btrfs_root *root,
3209 struct extent_buffer *buf,
3210 int full_backref, int inc)
3211 {
3212 struct btrfs_fs_info *fs_info = root->fs_info;
3213 u64 bytenr;
3214 u64 num_bytes;
3215 u64 parent;
3216 u64 ref_root;
3217 u32 nritems;
3218 struct btrfs_key key;
3219 struct btrfs_file_extent_item *fi;
3220 int i;
3221 int level;
3222 int ret = 0;
3223 int (*process_func)(struct btrfs_trans_handle *,
3224 struct btrfs_root *,
3225 u64, u64, u64, u64, u64, u64);
3226
3227
3228 if (btrfs_is_testing(fs_info))
3229 return 0;
3230
3231 ref_root = btrfs_header_owner(buf);
3232 nritems = btrfs_header_nritems(buf);
3233 level = btrfs_header_level(buf);
3234
3235 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3236 return 0;
3237
3238 if (inc)
3239 process_func = btrfs_inc_extent_ref;
3240 else
3241 process_func = btrfs_free_extent;
3242
3243 if (full_backref)
3244 parent = buf->start;
3245 else
3246 parent = 0;
3247
3248 for (i = 0; i < nritems; i++) {
3249 if (level == 0) {
3250 btrfs_item_key_to_cpu(buf, &key, i);
3251 if (key.type != BTRFS_EXTENT_DATA_KEY)
3252 continue;
3253 fi = btrfs_item_ptr(buf, i,
3254 struct btrfs_file_extent_item);
3255 if (btrfs_file_extent_type(buf, fi) ==
3256 BTRFS_FILE_EXTENT_INLINE)
3257 continue;
3258 bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3259 if (bytenr == 0)
3260 continue;
3261
3262 num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3263 key.offset -= btrfs_file_extent_offset(buf, fi);
3264 ret = process_func(trans, root, bytenr, num_bytes,
3265 parent, ref_root, key.objectid,
3266 key.offset);
3267 if (ret)
3268 goto fail;
3269 } else {
3270 bytenr = btrfs_node_blockptr(buf, i);
3271 num_bytes = fs_info->nodesize;
3272 ret = process_func(trans, root, bytenr, num_bytes,
3273 parent, ref_root, level - 1, 0);
3274 if (ret)
3275 goto fail;
3276 }
3277 }
3278 return 0;
3279 fail:
3280 return ret;
3281 }
3282
3283 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3284 struct extent_buffer *buf, int full_backref)
3285 {
3286 return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3287 }
3288
3289 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3290 struct extent_buffer *buf, int full_backref)
3291 {
3292 return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3293 }
3294
3295 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3296 struct btrfs_fs_info *fs_info,
3297 struct btrfs_path *path,
3298 struct btrfs_block_group_cache *cache)
3299 {
3300 int ret;
3301 struct btrfs_root *extent_root = fs_info->extent_root;
3302 unsigned long bi;
3303 struct extent_buffer *leaf;
3304
3305 ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3306 if (ret) {
3307 if (ret > 0)
3308 ret = -ENOENT;
3309 goto fail;
3310 }
3311
3312 leaf = path->nodes[0];
3313 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3314 write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3315 btrfs_mark_buffer_dirty(leaf);
3316 fail:
3317 btrfs_release_path(path);
3318 return ret;
3319
3320 }
3321
3322 static struct btrfs_block_group_cache *
3323 next_block_group(struct btrfs_fs_info *fs_info,
3324 struct btrfs_block_group_cache *cache)
3325 {
3326 struct rb_node *node;
3327
3328 spin_lock(&fs_info->block_group_cache_lock);
3329
3330 /* If our block group was removed, we need a full search. */
3331 if (RB_EMPTY_NODE(&cache->cache_node)) {
3332 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3333
3334 spin_unlock(&fs_info->block_group_cache_lock);
3335 btrfs_put_block_group(cache);
3336 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3337 }
3338 node = rb_next(&cache->cache_node);
3339 btrfs_put_block_group(cache);
3340 if (node) {
3341 cache = rb_entry(node, struct btrfs_block_group_cache,
3342 cache_node);
3343 btrfs_get_block_group(cache);
3344 } else
3345 cache = NULL;
3346 spin_unlock(&fs_info->block_group_cache_lock);
3347 return cache;
3348 }
3349
3350 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3351 struct btrfs_trans_handle *trans,
3352 struct btrfs_path *path)
3353 {
3354 struct btrfs_fs_info *fs_info = block_group->fs_info;
3355 struct btrfs_root *root = fs_info->tree_root;
3356 struct inode *inode = NULL;
3357 struct extent_changeset *data_reserved = NULL;
3358 u64 alloc_hint = 0;
3359 int dcs = BTRFS_DC_ERROR;
3360 u64 num_pages = 0;
3361 int retries = 0;
3362 int ret = 0;
3363
3364 /*
3365 * If this block group is smaller than 100 megs don't bother caching the
3366 * block group.
3367 */
3368 if (block_group->key.offset < (100 * SZ_1M)) {
3369 spin_lock(&block_group->lock);
3370 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3371 spin_unlock(&block_group->lock);
3372 return 0;
3373 }
3374
3375 if (trans->aborted)
3376 return 0;
3377 again:
3378 inode = lookup_free_space_inode(fs_info, block_group, path);
3379 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3380 ret = PTR_ERR(inode);
3381 btrfs_release_path(path);
3382 goto out;
3383 }
3384
3385 if (IS_ERR(inode)) {
3386 BUG_ON(retries);
3387 retries++;
3388
3389 if (block_group->ro)
3390 goto out_free;
3391
3392 ret = create_free_space_inode(fs_info, trans, block_group,
3393 path);
3394 if (ret)
3395 goto out_free;
3396 goto again;
3397 }
3398
3399 /*
3400 * We want to set the generation to 0, that way if anything goes wrong
3401 * from here on out we know not to trust this cache when we load up next
3402 * time.
3403 */
3404 BTRFS_I(inode)->generation = 0;
3405 ret = btrfs_update_inode(trans, root, inode);
3406 if (ret) {
3407 /*
3408 * So theoretically we could recover from this, simply set the
3409 * super cache generation to 0 so we know to invalidate the
3410 * cache, but then we'd have to keep track of the block groups
3411 * that fail this way so we know we _have_ to reset this cache
3412 * before the next commit or risk reading stale cache. So to
3413 * limit our exposure to horrible edge cases lets just abort the
3414 * transaction, this only happens in really bad situations
3415 * anyway.
3416 */
3417 btrfs_abort_transaction(trans, ret);
3418 goto out_put;
3419 }
3420 WARN_ON(ret);
3421
3422 /* We've already setup this transaction, go ahead and exit */
3423 if (block_group->cache_generation == trans->transid &&
3424 i_size_read(inode)) {
3425 dcs = BTRFS_DC_SETUP;
3426 goto out_put;
3427 }
3428
3429 if (i_size_read(inode) > 0) {
3430 ret = btrfs_check_trunc_cache_free_space(fs_info,
3431 &fs_info->global_block_rsv);
3432 if (ret)
3433 goto out_put;
3434
3435 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3436 if (ret)
3437 goto out_put;
3438 }
3439
3440 spin_lock(&block_group->lock);
3441 if (block_group->cached != BTRFS_CACHE_FINISHED ||
3442 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3443 /*
3444 * don't bother trying to write stuff out _if_
3445 * a) we're not cached,
3446 * b) we're with nospace_cache mount option,
3447 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3448 */
3449 dcs = BTRFS_DC_WRITTEN;
3450 spin_unlock(&block_group->lock);
3451 goto out_put;
3452 }
3453 spin_unlock(&block_group->lock);
3454
3455 /*
3456 * We hit an ENOSPC when setting up the cache in this transaction, just
3457 * skip doing the setup, we've already cleared the cache so we're safe.
3458 */
3459 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3460 ret = -ENOSPC;
3461 goto out_put;
3462 }
3463
3464 /*
3465 * Try to preallocate enough space based on how big the block group is.
3466 * Keep in mind this has to include any pinned space which could end up
3467 * taking up quite a bit since it's not folded into the other space
3468 * cache.
3469 */
3470 num_pages = div_u64(block_group->key.offset, SZ_256M);
3471 if (!num_pages)
3472 num_pages = 1;
3473
3474 num_pages *= 16;
3475 num_pages *= PAGE_SIZE;
3476
3477 ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3478 if (ret)
3479 goto out_put;
3480
3481 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3482 num_pages, num_pages,
3483 &alloc_hint);
3484 /*
3485 * Our cache requires contiguous chunks so that we don't modify a bunch
3486 * of metadata or split extents when writing the cache out, which means
3487 * we can enospc if we are heavily fragmented in addition to just normal
3488 * out of space conditions. So if we hit this just skip setting up any
3489 * other block groups for this transaction, maybe we'll unpin enough
3490 * space the next time around.
3491 */
3492 if (!ret)
3493 dcs = BTRFS_DC_SETUP;
3494 else if (ret == -ENOSPC)
3495 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3496
3497 out_put:
3498 iput(inode);
3499 out_free:
3500 btrfs_release_path(path);
3501 out:
3502 spin_lock(&block_group->lock);
3503 if (!ret && dcs == BTRFS_DC_SETUP)
3504 block_group->cache_generation = trans->transid;
3505 block_group->disk_cache_state = dcs;
3506 spin_unlock(&block_group->lock);
3507
3508 extent_changeset_free(data_reserved);
3509 return ret;
3510 }
3511
3512 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3513 struct btrfs_fs_info *fs_info)
3514 {
3515 struct btrfs_block_group_cache *cache, *tmp;
3516 struct btrfs_transaction *cur_trans = trans->transaction;
3517 struct btrfs_path *path;
3518
3519 if (list_empty(&cur_trans->dirty_bgs) ||
3520 !btrfs_test_opt(fs_info, SPACE_CACHE))
3521 return 0;
3522
3523 path = btrfs_alloc_path();
3524 if (!path)
3525 return -ENOMEM;
3526
3527 /* Could add new block groups, use _safe just in case */
3528 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3529 dirty_list) {
3530 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3531 cache_save_setup(cache, trans, path);
3532 }
3533
3534 btrfs_free_path(path);
3535 return 0;
3536 }
3537
3538 /*
3539 * transaction commit does final block group cache writeback during a
3540 * critical section where nothing is allowed to change the FS. This is
3541 * required in order for the cache to actually match the block group,
3542 * but can introduce a lot of latency into the commit.
3543 *
3544 * So, btrfs_start_dirty_block_groups is here to kick off block group
3545 * cache IO. There's a chance we'll have to redo some of it if the
3546 * block group changes again during the commit, but it greatly reduces
3547 * the commit latency by getting rid of the easy block groups while
3548 * we're still allowing others to join the commit.
3549 */
3550 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3551 {
3552 struct btrfs_fs_info *fs_info = trans->fs_info;
3553 struct btrfs_block_group_cache *cache;
3554 struct btrfs_transaction *cur_trans = trans->transaction;
3555 int ret = 0;
3556 int should_put;
3557 struct btrfs_path *path = NULL;
3558 LIST_HEAD(dirty);
3559 struct list_head *io = &cur_trans->io_bgs;
3560 int num_started = 0;
3561 int loops = 0;
3562
3563 spin_lock(&cur_trans->dirty_bgs_lock);
3564 if (list_empty(&cur_trans->dirty_bgs)) {
3565 spin_unlock(&cur_trans->dirty_bgs_lock);
3566 return 0;
3567 }
3568 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3569 spin_unlock(&cur_trans->dirty_bgs_lock);
3570
3571 again:
3572 /*
3573 * make sure all the block groups on our dirty list actually
3574 * exist
3575 */
3576 btrfs_create_pending_block_groups(trans);
3577
3578 if (!path) {
3579 path = btrfs_alloc_path();
3580 if (!path)
3581 return -ENOMEM;
3582 }
3583
3584 /*
3585 * cache_write_mutex is here only to save us from balance or automatic
3586 * removal of empty block groups deleting this block group while we are
3587 * writing out the cache
3588 */
3589 mutex_lock(&trans->transaction->cache_write_mutex);
3590 while (!list_empty(&dirty)) {
3591 cache = list_first_entry(&dirty,
3592 struct btrfs_block_group_cache,
3593 dirty_list);
3594 /*
3595 * this can happen if something re-dirties a block
3596 * group that is already under IO. Just wait for it to
3597 * finish and then do it all again
3598 */
3599 if (!list_empty(&cache->io_list)) {
3600 list_del_init(&cache->io_list);
3601 btrfs_wait_cache_io(trans, cache, path);
3602 btrfs_put_block_group(cache);
3603 }
3604
3605
3606 /*
3607 * btrfs_wait_cache_io uses the cache->dirty_list to decide
3608 * if it should update the cache_state. Don't delete
3609 * until after we wait.
3610 *
3611 * Since we're not running in the commit critical section
3612 * we need the dirty_bgs_lock to protect from update_block_group
3613 */
3614 spin_lock(&cur_trans->dirty_bgs_lock);
3615 list_del_init(&cache->dirty_list);
3616 spin_unlock(&cur_trans->dirty_bgs_lock);
3617
3618 should_put = 1;
3619
3620 cache_save_setup(cache, trans, path);
3621
3622 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3623 cache->io_ctl.inode = NULL;
3624 ret = btrfs_write_out_cache(fs_info, trans,
3625 cache, path);
3626 if (ret == 0 && cache->io_ctl.inode) {
3627 num_started++;
3628 should_put = 0;
3629
3630 /*
3631 * The cache_write_mutex is protecting the
3632 * io_list, also refer to the definition of
3633 * btrfs_transaction::io_bgs for more details
3634 */
3635 list_add_tail(&cache->io_list, io);
3636 } else {
3637 /*
3638 * if we failed to write the cache, the
3639 * generation will be bad and life goes on
3640 */
3641 ret = 0;
3642 }
3643 }
3644 if (!ret) {
3645 ret = write_one_cache_group(trans, fs_info,
3646 path, cache);
3647 /*
3648 * Our block group might still be attached to the list
3649 * of new block groups in the transaction handle of some
3650 * other task (struct btrfs_trans_handle->new_bgs). This
3651 * means its block group item isn't yet in the extent
3652 * tree. If this happens ignore the error, as we will
3653 * try again later in the critical section of the
3654 * transaction commit.
3655 */
3656 if (ret == -ENOENT) {
3657 ret = 0;
3658 spin_lock(&cur_trans->dirty_bgs_lock);
3659 if (list_empty(&cache->dirty_list)) {
3660 list_add_tail(&cache->dirty_list,
3661 &cur_trans->dirty_bgs);
3662 btrfs_get_block_group(cache);
3663 }
3664 spin_unlock(&cur_trans->dirty_bgs_lock);
3665 } else if (ret) {
3666 btrfs_abort_transaction(trans, ret);
3667 }
3668 }
3669
3670 /* if its not on the io list, we need to put the block group */
3671 if (should_put)
3672 btrfs_put_block_group(cache);
3673
3674 if (ret)
3675 break;
3676
3677 /*
3678 * Avoid blocking other tasks for too long. It might even save
3679 * us from writing caches for block groups that are going to be
3680 * removed.
3681 */
3682 mutex_unlock(&trans->transaction->cache_write_mutex);
3683 mutex_lock(&trans->transaction->cache_write_mutex);
3684 }
3685 mutex_unlock(&trans->transaction->cache_write_mutex);
3686
3687 /*
3688 * go through delayed refs for all the stuff we've just kicked off
3689 * and then loop back (just once)
3690 */
3691 ret = btrfs_run_delayed_refs(trans, 0);
3692 if (!ret && loops == 0) {
3693 loops++;
3694 spin_lock(&cur_trans->dirty_bgs_lock);
3695 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3696 /*
3697 * dirty_bgs_lock protects us from concurrent block group
3698 * deletes too (not just cache_write_mutex).
3699 */
3700 if (!list_empty(&dirty)) {
3701 spin_unlock(&cur_trans->dirty_bgs_lock);
3702 goto again;
3703 }
3704 spin_unlock(&cur_trans->dirty_bgs_lock);
3705 } else if (ret < 0) {
3706 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3707 }
3708
3709 btrfs_free_path(path);
3710 return ret;
3711 }
3712
3713 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3714 struct btrfs_fs_info *fs_info)
3715 {
3716 struct btrfs_block_group_cache *cache;
3717 struct btrfs_transaction *cur_trans = trans->transaction;
3718 int ret = 0;
3719 int should_put;
3720 struct btrfs_path *path;
3721 struct list_head *io = &cur_trans->io_bgs;
3722 int num_started = 0;
3723
3724 path = btrfs_alloc_path();
3725 if (!path)
3726 return -ENOMEM;
3727
3728 /*
3729 * Even though we are in the critical section of the transaction commit,
3730 * we can still have concurrent tasks adding elements to this
3731 * transaction's list of dirty block groups. These tasks correspond to
3732 * endio free space workers started when writeback finishes for a
3733 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3734 * allocate new block groups as a result of COWing nodes of the root
3735 * tree when updating the free space inode. The writeback for the space
3736 * caches is triggered by an earlier call to
3737 * btrfs_start_dirty_block_groups() and iterations of the following
3738 * loop.
3739 * Also we want to do the cache_save_setup first and then run the
3740 * delayed refs to make sure we have the best chance at doing this all
3741 * in one shot.
3742 */
3743 spin_lock(&cur_trans->dirty_bgs_lock);
3744 while (!list_empty(&cur_trans->dirty_bgs)) {
3745 cache = list_first_entry(&cur_trans->dirty_bgs,
3746 struct btrfs_block_group_cache,
3747 dirty_list);
3748
3749 /*
3750 * this can happen if cache_save_setup re-dirties a block
3751 * group that is already under IO. Just wait for it to
3752 * finish and then do it all again
3753 */
3754 if (!list_empty(&cache->io_list)) {
3755 spin_unlock(&cur_trans->dirty_bgs_lock);
3756 list_del_init(&cache->io_list);
3757 btrfs_wait_cache_io(trans, cache, path);
3758 btrfs_put_block_group(cache);
3759 spin_lock(&cur_trans->dirty_bgs_lock);
3760 }
3761
3762 /*
3763 * don't remove from the dirty list until after we've waited
3764 * on any pending IO
3765 */
3766 list_del_init(&cache->dirty_list);
3767 spin_unlock(&cur_trans->dirty_bgs_lock);
3768 should_put = 1;
3769
3770 cache_save_setup(cache, trans, path);
3771
3772 if (!ret)
3773 ret = btrfs_run_delayed_refs(trans,
3774 (unsigned long) -1);
3775
3776 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3777 cache->io_ctl.inode = NULL;
3778 ret = btrfs_write_out_cache(fs_info, trans,
3779 cache, path);
3780 if (ret == 0 && cache->io_ctl.inode) {
3781 num_started++;
3782 should_put = 0;
3783 list_add_tail(&cache->io_list, io);
3784 } else {
3785 /*
3786 * if we failed to write the cache, the
3787 * generation will be bad and life goes on
3788 */
3789 ret = 0;
3790 }
3791 }
3792 if (!ret) {
3793 ret = write_one_cache_group(trans, fs_info,
3794 path, cache);
3795 /*
3796 * One of the free space endio workers might have
3797 * created a new block group while updating a free space
3798 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3799 * and hasn't released its transaction handle yet, in
3800 * which case the new block group is still attached to
3801 * its transaction handle and its creation has not
3802 * finished yet (no block group item in the extent tree
3803 * yet, etc). If this is the case, wait for all free
3804 * space endio workers to finish and retry. This is a
3805 * a very rare case so no need for a more efficient and
3806 * complex approach.
3807 */
3808 if (ret == -ENOENT) {
3809 wait_event(cur_trans->writer_wait,
3810 atomic_read(&cur_trans->num_writers) == 1);
3811 ret = write_one_cache_group(trans, fs_info,
3812 path, cache);
3813 }
3814 if (ret)
3815 btrfs_abort_transaction(trans, ret);
3816 }
3817
3818 /* if its not on the io list, we need to put the block group */
3819 if (should_put)
3820 btrfs_put_block_group(cache);
3821 spin_lock(&cur_trans->dirty_bgs_lock);
3822 }
3823 spin_unlock(&cur_trans->dirty_bgs_lock);
3824
3825 /*
3826 * Refer to the definition of io_bgs member for details why it's safe
3827 * to use it without any locking
3828 */
3829 while (!list_empty(io)) {
3830 cache = list_first_entry(io, struct btrfs_block_group_cache,
3831 io_list);
3832 list_del_init(&cache->io_list);
3833 btrfs_wait_cache_io(trans, cache, path);
3834 btrfs_put_block_group(cache);
3835 }
3836
3837 btrfs_free_path(path);
3838 return ret;
3839 }
3840
3841 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3842 {
3843 struct btrfs_block_group_cache *block_group;
3844 int readonly = 0;
3845
3846 block_group = btrfs_lookup_block_group(fs_info, bytenr);
3847 if (!block_group || block_group->ro)
3848 readonly = 1;
3849 if (block_group)
3850 btrfs_put_block_group(block_group);
3851 return readonly;
3852 }
3853
3854 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3855 {
3856 struct btrfs_block_group_cache *bg;
3857 bool ret = true;
3858
3859 bg = btrfs_lookup_block_group(fs_info, bytenr);
3860 if (!bg)
3861 return false;
3862
3863 spin_lock(&bg->lock);
3864 if (bg->ro)
3865 ret = false;
3866 else
3867 atomic_inc(&bg->nocow_writers);
3868 spin_unlock(&bg->lock);
3869
3870 /* no put on block group, done by btrfs_dec_nocow_writers */
3871 if (!ret)
3872 btrfs_put_block_group(bg);
3873
3874 return ret;
3875
3876 }
3877
3878 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3879 {
3880 struct btrfs_block_group_cache *bg;
3881
3882 bg = btrfs_lookup_block_group(fs_info, bytenr);
3883 ASSERT(bg);
3884 if (atomic_dec_and_test(&bg->nocow_writers))
3885 wake_up_var(&bg->nocow_writers);
3886 /*
3887 * Once for our lookup and once for the lookup done by a previous call
3888 * to btrfs_inc_nocow_writers()
3889 */
3890 btrfs_put_block_group(bg);
3891 btrfs_put_block_group(bg);
3892 }
3893
3894 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3895 {
3896 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3897 }
3898
3899 static const char *alloc_name(u64 flags)
3900 {
3901 switch (flags) {
3902 case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3903 return "mixed";
3904 case BTRFS_BLOCK_GROUP_METADATA:
3905 return "metadata";
3906 case BTRFS_BLOCK_GROUP_DATA:
3907 return "data";
3908 case BTRFS_BLOCK_GROUP_SYSTEM:
3909 return "system";
3910 default:
3911 WARN_ON(1);
3912 return "invalid-combination";
3913 };
3914 }
3915
3916 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3917 {
3918
3919 struct btrfs_space_info *space_info;
3920 int i;
3921 int ret;
3922
3923 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3924 if (!space_info)
3925 return -ENOMEM;
3926
3927 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3928 GFP_KERNEL);
3929 if (ret) {
3930 kfree(space_info);
3931 return ret;
3932 }
3933
3934 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3935 INIT_LIST_HEAD(&space_info->block_groups[i]);
3936 init_rwsem(&space_info->groups_sem);
3937 spin_lock_init(&space_info->lock);
3938 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3939 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3940 init_waitqueue_head(&space_info->wait);
3941 INIT_LIST_HEAD(&space_info->ro_bgs);
3942 INIT_LIST_HEAD(&space_info->tickets);
3943 INIT_LIST_HEAD(&space_info->priority_tickets);
3944
3945 ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3946 info->space_info_kobj, "%s",
3947 alloc_name(space_info->flags));
3948 if (ret) {
3949 percpu_counter_destroy(&space_info->total_bytes_pinned);
3950 kfree(space_info);
3951 return ret;
3952 }
3953
3954 list_add_rcu(&space_info->list, &info->space_info);
3955 if (flags & BTRFS_BLOCK_GROUP_DATA)
3956 info->data_sinfo = space_info;
3957
3958 return ret;
3959 }
3960
3961 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
3962 u64 total_bytes, u64 bytes_used,
3963 u64 bytes_readonly,
3964 struct btrfs_space_info **space_info)
3965 {
3966 struct btrfs_space_info *found;
3967 int factor;
3968
3969 factor = btrfs_bg_type_to_factor(flags);
3970
3971 found = __find_space_info(info, flags);
3972 ASSERT(found);
3973 spin_lock(&found->lock);
3974 found->total_bytes += total_bytes;
3975 found->disk_total += total_bytes * factor;
3976 found->bytes_used += bytes_used;
3977 found->disk_used += bytes_used * factor;
3978 found->bytes_readonly += bytes_readonly;
3979 if (total_bytes > 0)
3980 found->full = 0;
3981 space_info_add_new_bytes(info, found, total_bytes -
3982 bytes_used - bytes_readonly);
3983 spin_unlock(&found->lock);
3984 *space_info = found;
3985 }
3986
3987 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3988 {
3989 u64 extra_flags = chunk_to_extended(flags) &
3990 BTRFS_EXTENDED_PROFILE_MASK;
3991
3992 write_seqlock(&fs_info->profiles_lock);
3993 if (flags & BTRFS_BLOCK_GROUP_DATA)
3994 fs_info->avail_data_alloc_bits |= extra_flags;
3995 if (flags & BTRFS_BLOCK_GROUP_METADATA)
3996 fs_info->avail_metadata_alloc_bits |= extra_flags;
3997 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3998 fs_info->avail_system_alloc_bits |= extra_flags;
3999 write_sequnlock(&fs_info->profiles_lock);
4000 }
4001
4002 /*
4003 * returns target flags in extended format or 0 if restripe for this
4004 * chunk_type is not in progress
4005 *
4006 * should be called with balance_lock held
4007 */
4008 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4009 {
4010 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4011 u64 target = 0;
4012
4013 if (!bctl)
4014 return 0;
4015
4016 if (flags & BTRFS_BLOCK_GROUP_DATA &&
4017 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4018 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4019 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4020 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4021 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4022 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4023 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4024 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4025 }
4026
4027 return target;
4028 }
4029
4030 /*
4031 * @flags: available profiles in extended format (see ctree.h)
4032 *
4033 * Returns reduced profile in chunk format. If profile changing is in
4034 * progress (either running or paused) picks the target profile (if it's
4035 * already available), otherwise falls back to plain reducing.
4036 */
4037 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4038 {
4039 u64 num_devices = fs_info->fs_devices->rw_devices;
4040 u64 target;
4041 u64 raid_type;
4042 u64 allowed = 0;
4043
4044 /*
4045 * see if restripe for this chunk_type is in progress, if so
4046 * try to reduce to the target profile
4047 */
4048 spin_lock(&fs_info->balance_lock);
4049 target = get_restripe_target(fs_info, flags);
4050 if (target) {
4051 /* pick target profile only if it's already available */
4052 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4053 spin_unlock(&fs_info->balance_lock);
4054 return extended_to_chunk(target);
4055 }
4056 }
4057 spin_unlock(&fs_info->balance_lock);
4058
4059 /* First, mask out the RAID levels which aren't possible */
4060 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4061 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4062 allowed |= btrfs_raid_array[raid_type].bg_flag;
4063 }
4064 allowed &= flags;
4065
4066 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4067 allowed = BTRFS_BLOCK_GROUP_RAID6;
4068 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4069 allowed = BTRFS_BLOCK_GROUP_RAID5;
4070 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4071 allowed = BTRFS_BLOCK_GROUP_RAID10;
4072 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4073 allowed = BTRFS_BLOCK_GROUP_RAID1;
4074 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4075 allowed = BTRFS_BLOCK_GROUP_RAID0;
4076
4077 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4078
4079 return extended_to_chunk(flags | allowed);
4080 }
4081
4082 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4083 {
4084 unsigned seq;
4085 u64 flags;
4086
4087 do {
4088 flags = orig_flags;
4089 seq = read_seqbegin(&fs_info->profiles_lock);
4090
4091 if (flags & BTRFS_BLOCK_GROUP_DATA)
4092 flags |= fs_info->avail_data_alloc_bits;
4093 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4094 flags |= fs_info->avail_system_alloc_bits;
4095 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4096 flags |= fs_info->avail_metadata_alloc_bits;
4097 } while (read_seqretry(&fs_info->profiles_lock, seq));
4098
4099 return btrfs_reduce_alloc_profile(fs_info, flags);
4100 }
4101
4102 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4103 {
4104 struct btrfs_fs_info *fs_info = root->fs_info;
4105 u64 flags;
4106 u64 ret;
4107
4108 if (data)
4109 flags = BTRFS_BLOCK_GROUP_DATA;
4110 else if (root == fs_info->chunk_root)
4111 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4112 else
4113 flags = BTRFS_BLOCK_GROUP_METADATA;
4114
4115 ret = get_alloc_profile(fs_info, flags);
4116 return ret;
4117 }
4118
4119 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4120 {
4121 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4122 }
4123
4124 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4125 {
4126 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4127 }
4128
4129 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4130 {
4131 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4132 }
4133
4134 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4135 bool may_use_included)
4136 {
4137 ASSERT(s_info);
4138 return s_info->bytes_used + s_info->bytes_reserved +
4139 s_info->bytes_pinned + s_info->bytes_readonly +
4140 (may_use_included ? s_info->bytes_may_use : 0);
4141 }
4142
4143 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4144 {
4145 struct btrfs_root *root = inode->root;
4146 struct btrfs_fs_info *fs_info = root->fs_info;
4147 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4148 u64 used;
4149 int ret = 0;
4150 int need_commit = 2;
4151 int have_pinned_space;
4152
4153 /* make sure bytes are sectorsize aligned */
4154 bytes = ALIGN(bytes, fs_info->sectorsize);
4155
4156 if (btrfs_is_free_space_inode(inode)) {
4157 need_commit = 0;
4158 ASSERT(current->journal_info);
4159 }
4160
4161 again:
4162 /* make sure we have enough space to handle the data first */
4163 spin_lock(&data_sinfo->lock);
4164 used = btrfs_space_info_used(data_sinfo, true);
4165
4166 if (used + bytes > data_sinfo->total_bytes) {
4167 struct btrfs_trans_handle *trans;
4168
4169 /*
4170 * if we don't have enough free bytes in this space then we need
4171 * to alloc a new chunk.
4172 */
4173 if (!data_sinfo->full) {
4174 u64 alloc_target;
4175
4176 data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4177 spin_unlock(&data_sinfo->lock);
4178
4179 alloc_target = btrfs_data_alloc_profile(fs_info);
4180 /*
4181 * It is ugly that we don't call nolock join
4182 * transaction for the free space inode case here.
4183 * But it is safe because we only do the data space
4184 * reservation for the free space cache in the
4185 * transaction context, the common join transaction
4186 * just increase the counter of the current transaction
4187 * handler, doesn't try to acquire the trans_lock of
4188 * the fs.
4189 */
4190 trans = btrfs_join_transaction(root);
4191 if (IS_ERR(trans))
4192 return PTR_ERR(trans);
4193
4194 ret = do_chunk_alloc(trans, alloc_target,
4195 CHUNK_ALLOC_NO_FORCE);
4196 btrfs_end_transaction(trans);
4197 if (ret < 0) {
4198 if (ret != -ENOSPC)
4199 return ret;
4200 else {
4201 have_pinned_space = 1;
4202 goto commit_trans;
4203 }
4204 }
4205
4206 goto again;
4207 }
4208
4209 /*
4210 * If we don't have enough pinned space to deal with this
4211 * allocation, and no removed chunk in current transaction,
4212 * don't bother committing the transaction.
4213 */
4214 have_pinned_space = __percpu_counter_compare(
4215 &data_sinfo->total_bytes_pinned,
4216 used + bytes - data_sinfo->total_bytes,
4217 BTRFS_TOTAL_BYTES_PINNED_BATCH);
4218 spin_unlock(&data_sinfo->lock);
4219
4220 /* commit the current transaction and try again */
4221 commit_trans:
4222 if (need_commit) {
4223 need_commit--;
4224
4225 if (need_commit > 0) {
4226 btrfs_start_delalloc_roots(fs_info, -1);
4227 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4228 (u64)-1);
4229 }
4230
4231 trans = btrfs_join_transaction(root);
4232 if (IS_ERR(trans))
4233 return PTR_ERR(trans);
4234 if (have_pinned_space >= 0 ||
4235 test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4236 &trans->transaction->flags) ||
4237 need_commit > 0) {
4238 ret = btrfs_commit_transaction(trans);
4239 if (ret)
4240 return ret;
4241 /*
4242 * The cleaner kthread might still be doing iput
4243 * operations. Wait for it to finish so that
4244 * more space is released.
4245 */
4246 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4247 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4248 goto again;
4249 } else {
4250 btrfs_end_transaction(trans);
4251 }
4252 }
4253
4254 trace_btrfs_space_reservation(fs_info,
4255 "space_info:enospc",
4256 data_sinfo->flags, bytes, 1);
4257 return -ENOSPC;
4258 }
4259 data_sinfo->bytes_may_use += bytes;
4260 trace_btrfs_space_reservation(fs_info, "space_info",
4261 data_sinfo->flags, bytes, 1);
4262 spin_unlock(&data_sinfo->lock);
4263
4264 return 0;
4265 }
4266
4267 int btrfs_check_data_free_space(struct inode *inode,
4268 struct extent_changeset **reserved, u64 start, u64 len)
4269 {
4270 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4271 int ret;
4272
4273 /* align the range */
4274 len = round_up(start + len, fs_info->sectorsize) -
4275 round_down(start, fs_info->sectorsize);
4276 start = round_down(start, fs_info->sectorsize);
4277
4278 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4279 if (ret < 0)
4280 return ret;
4281
4282 /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4283 ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4284 if (ret < 0)
4285 btrfs_free_reserved_data_space_noquota(inode, start, len);
4286 else
4287 ret = 0;
4288 return ret;
4289 }
4290
4291 /*
4292 * Called if we need to clear a data reservation for this inode
4293 * Normally in a error case.
4294 *
4295 * This one will *NOT* use accurate qgroup reserved space API, just for case
4296 * which we can't sleep and is sure it won't affect qgroup reserved space.
4297 * Like clear_bit_hook().
4298 */
4299 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4300 u64 len)
4301 {
4302 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4303 struct btrfs_space_info *data_sinfo;
4304
4305 /* Make sure the range is aligned to sectorsize */
4306 len = round_up(start + len, fs_info->sectorsize) -
4307 round_down(start, fs_info->sectorsize);
4308 start = round_down(start, fs_info->sectorsize);
4309
4310 data_sinfo = fs_info->data_sinfo;
4311 spin_lock(&data_sinfo->lock);
4312 if (WARN_ON(data_sinfo->bytes_may_use < len))
4313 data_sinfo->bytes_may_use = 0;
4314 else
4315 data_sinfo->bytes_may_use -= len;
4316 trace_btrfs_space_reservation(fs_info, "space_info",
4317 data_sinfo->flags, len, 0);
4318 spin_unlock(&data_sinfo->lock);
4319 }
4320
4321 /*
4322 * Called if we need to clear a data reservation for this inode
4323 * Normally in a error case.
4324 *
4325 * This one will handle the per-inode data rsv map for accurate reserved
4326 * space framework.
4327 */
4328 void btrfs_free_reserved_data_space(struct inode *inode,
4329 struct extent_changeset *reserved, u64 start, u64 len)
4330 {
4331 struct btrfs_root *root = BTRFS_I(inode)->root;
4332
4333 /* Make sure the range is aligned to sectorsize */
4334 len = round_up(start + len, root->fs_info->sectorsize) -
4335 round_down(start, root->fs_info->sectorsize);
4336 start = round_down(start, root->fs_info->sectorsize);
4337
4338 btrfs_free_reserved_data_space_noquota(inode, start, len);
4339 btrfs_qgroup_free_data(inode, reserved, start, len);
4340 }
4341
4342 static void force_metadata_allocation(struct btrfs_fs_info *info)
4343 {
4344 struct list_head *head = &info->space_info;
4345 struct btrfs_space_info *found;
4346
4347 rcu_read_lock();
4348 list_for_each_entry_rcu(found, head, list) {
4349 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4350 found->force_alloc = CHUNK_ALLOC_FORCE;
4351 }
4352 rcu_read_unlock();
4353 }
4354
4355 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4356 {
4357 return (global->size << 1);
4358 }
4359
4360 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4361 struct btrfs_space_info *sinfo, int force)
4362 {
4363 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4364 u64 bytes_used = btrfs_space_info_used(sinfo, false);
4365 u64 thresh;
4366
4367 if (force == CHUNK_ALLOC_FORCE)
4368 return 1;
4369
4370 /*
4371 * We need to take into account the global rsv because for all intents
4372 * and purposes it's used space. Don't worry about locking the
4373 * global_rsv, it doesn't change except when the transaction commits.
4374 */
4375 if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4376 bytes_used += calc_global_rsv_need_space(global_rsv);
4377
4378 /*
4379 * in limited mode, we want to have some free space up to
4380 * about 1% of the FS size.
4381 */
4382 if (force == CHUNK_ALLOC_LIMITED) {
4383 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4384 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4385
4386 if (sinfo->total_bytes - bytes_used < thresh)
4387 return 1;
4388 }
4389
4390 if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4391 return 0;
4392 return 1;
4393 }
4394
4395 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4396 {
4397 u64 num_dev;
4398
4399 if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4400 BTRFS_BLOCK_GROUP_RAID0 |
4401 BTRFS_BLOCK_GROUP_RAID5 |
4402 BTRFS_BLOCK_GROUP_RAID6))
4403 num_dev = fs_info->fs_devices->rw_devices;
4404 else if (type & BTRFS_BLOCK_GROUP_RAID1)
4405 num_dev = 2;
4406 else
4407 num_dev = 1; /* DUP or single */
4408
4409 return num_dev;
4410 }
4411
4412 /*
4413 * If @is_allocation is true, reserve space in the system space info necessary
4414 * for allocating a chunk, otherwise if it's false, reserve space necessary for
4415 * removing a chunk.
4416 */
4417 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4418 {
4419 struct btrfs_fs_info *fs_info = trans->fs_info;
4420 struct btrfs_space_info *info;
4421 u64 left;
4422 u64 thresh;
4423 int ret = 0;
4424 u64 num_devs;
4425
4426 /*
4427 * Needed because we can end up allocating a system chunk and for an
4428 * atomic and race free space reservation in the chunk block reserve.
4429 */
4430 lockdep_assert_held(&fs_info->chunk_mutex);
4431
4432 info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4433 spin_lock(&info->lock);
4434 left = info->total_bytes - btrfs_space_info_used(info, true);
4435 spin_unlock(&info->lock);
4436
4437 num_devs = get_profile_num_devs(fs_info, type);
4438
4439 /* num_devs device items to update and 1 chunk item to add or remove */
4440 thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4441 btrfs_calc_trans_metadata_size(fs_info, 1);
4442
4443 if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4444 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4445 left, thresh, type);
4446 dump_space_info(fs_info, info, 0, 0);
4447 }
4448
4449 if (left < thresh) {
4450 u64 flags = btrfs_system_alloc_profile(fs_info);
4451
4452 /*
4453 * Ignore failure to create system chunk. We might end up not
4454 * needing it, as we might not need to COW all nodes/leafs from
4455 * the paths we visit in the chunk tree (they were already COWed
4456 * or created in the current transaction for example).
4457 */
4458 ret = btrfs_alloc_chunk(trans, flags);
4459 }
4460
4461 if (!ret) {
4462 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4463 &fs_info->chunk_block_rsv,
4464 thresh, BTRFS_RESERVE_NO_FLUSH);
4465 if (!ret)
4466 trans->chunk_bytes_reserved += thresh;
4467 }
4468 }
4469
4470 /*
4471 * If force is CHUNK_ALLOC_FORCE:
4472 * - return 1 if it successfully allocates a chunk,
4473 * - return errors including -ENOSPC otherwise.
4474 * If force is NOT CHUNK_ALLOC_FORCE:
4475 * - return 0 if it doesn't need to allocate a new chunk,
4476 * - return 1 if it successfully allocates a chunk,
4477 * - return errors including -ENOSPC otherwise.
4478 */
4479 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4480 int force)
4481 {
4482 struct btrfs_fs_info *fs_info = trans->fs_info;
4483 struct btrfs_space_info *space_info;
4484 bool wait_for_alloc = false;
4485 bool should_alloc = false;
4486 int ret = 0;
4487
4488 /* Don't re-enter if we're already allocating a chunk */
4489 if (trans->allocating_chunk)
4490 return -ENOSPC;
4491
4492 space_info = __find_space_info(fs_info, flags);
4493 ASSERT(space_info);
4494
4495 do {
4496 spin_lock(&space_info->lock);
4497 if (force < space_info->force_alloc)
4498 force = space_info->force_alloc;
4499 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4500 if (space_info->full) {
4501 /* No more free physical space */
4502 if (should_alloc)
4503 ret = -ENOSPC;
4504 else
4505 ret = 0;
4506 spin_unlock(&space_info->lock);
4507 return ret;
4508 } else if (!should_alloc) {
4509 spin_unlock(&space_info->lock);
4510 return 0;
4511 } else if (space_info->chunk_alloc) {
4512 /*
4513 * Someone is already allocating, so we need to block
4514 * until this someone is finished and then loop to
4515 * recheck if we should continue with our allocation
4516 * attempt.
4517 */
4518 wait_for_alloc = true;
4519 spin_unlock(&space_info->lock);
4520 mutex_lock(&fs_info->chunk_mutex);
4521 mutex_unlock(&fs_info->chunk_mutex);
4522 } else {
4523 /* Proceed with allocation */
4524 space_info->chunk_alloc = 1;
4525 wait_for_alloc = false;
4526 spin_unlock(&space_info->lock);
4527 }
4528
4529 cond_resched();
4530 } while (wait_for_alloc);
4531
4532 mutex_lock(&fs_info->chunk_mutex);
4533 trans->allocating_chunk = true;
4534
4535 /*
4536 * If we have mixed data/metadata chunks we want to make sure we keep
4537 * allocating mixed chunks instead of individual chunks.
4538 */
4539 if (btrfs_mixed_space_info(space_info))
4540 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4541
4542 /*
4543 * if we're doing a data chunk, go ahead and make sure that
4544 * we keep a reasonable number of metadata chunks allocated in the
4545 * FS as well.
4546 */
4547 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4548 fs_info->data_chunk_allocations++;
4549 if (!(fs_info->data_chunk_allocations %
4550 fs_info->metadata_ratio))
4551 force_metadata_allocation(fs_info);
4552 }
4553
4554 /*
4555 * Check if we have enough space in SYSTEM chunk because we may need
4556 * to update devices.
4557 */
4558 check_system_chunk(trans, flags);
4559
4560 ret = btrfs_alloc_chunk(trans, flags);
4561 trans->allocating_chunk = false;
4562
4563 spin_lock(&space_info->lock);
4564 if (ret < 0) {
4565 if (ret == -ENOSPC)
4566 space_info->full = 1;
4567 else
4568 goto out;
4569 } else {
4570 ret = 1;
4571 space_info->max_extent_size = 0;
4572 }
4573
4574 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4575 out:
4576 space_info->chunk_alloc = 0;
4577 spin_unlock(&space_info->lock);
4578 mutex_unlock(&fs_info->chunk_mutex);
4579 /*
4580 * When we allocate a new chunk we reserve space in the chunk block
4581 * reserve to make sure we can COW nodes/leafs in the chunk tree or
4582 * add new nodes/leafs to it if we end up needing to do it when
4583 * inserting the chunk item and updating device items as part of the
4584 * second phase of chunk allocation, performed by
4585 * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4586 * large number of new block groups to create in our transaction
4587 * handle's new_bgs list to avoid exhausting the chunk block reserve
4588 * in extreme cases - like having a single transaction create many new
4589 * block groups when starting to write out the free space caches of all
4590 * the block groups that were made dirty during the lifetime of the
4591 * transaction.
4592 */
4593 if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
4594 btrfs_create_pending_block_groups(trans);
4595
4596 return ret;
4597 }
4598
4599 static int can_overcommit(struct btrfs_fs_info *fs_info,
4600 struct btrfs_space_info *space_info, u64 bytes,
4601 enum btrfs_reserve_flush_enum flush,
4602 bool system_chunk)
4603 {
4604 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4605 u64 profile;
4606 u64 space_size;
4607 u64 avail;
4608 u64 used;
4609 int factor;
4610
4611 /* Don't overcommit when in mixed mode. */
4612 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4613 return 0;
4614
4615 if (system_chunk)
4616 profile = btrfs_system_alloc_profile(fs_info);
4617 else
4618 profile = btrfs_metadata_alloc_profile(fs_info);
4619
4620 used = btrfs_space_info_used(space_info, false);
4621
4622 /*
4623 * We only want to allow over committing if we have lots of actual space
4624 * free, but if we don't have enough space to handle the global reserve
4625 * space then we could end up having a real enospc problem when trying
4626 * to allocate a chunk or some other such important allocation.
4627 */
4628 spin_lock(&global_rsv->lock);
4629 space_size = calc_global_rsv_need_space(global_rsv);
4630 spin_unlock(&global_rsv->lock);
4631 if (used + space_size >= space_info->total_bytes)
4632 return 0;
4633
4634 used += space_info->bytes_may_use;
4635
4636 avail = atomic64_read(&fs_info->free_chunk_space);
4637
4638 /*
4639 * If we have dup, raid1 or raid10 then only half of the free
4640 * space is actually useable. For raid56, the space info used
4641 * doesn't include the parity drive, so we don't have to
4642 * change the math
4643 */
4644 factor = btrfs_bg_type_to_factor(profile);
4645 avail = div_u64(avail, factor);
4646
4647 /*
4648 * If we aren't flushing all things, let us overcommit up to
4649 * 1/2th of the space. If we can flush, don't let us overcommit
4650 * too much, let it overcommit up to 1/8 of the space.
4651 */
4652 if (flush == BTRFS_RESERVE_FLUSH_ALL)
4653 avail >>= 3;
4654 else
4655 avail >>= 1;
4656
4657 if (used + bytes < space_info->total_bytes + avail)
4658 return 1;
4659 return 0;
4660 }
4661
4662 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4663 unsigned long nr_pages, int nr_items)
4664 {
4665 struct super_block *sb = fs_info->sb;
4666
4667 if (down_read_trylock(&sb->s_umount)) {
4668 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4669 up_read(&sb->s_umount);
4670 } else {
4671 /*
4672 * We needn't worry the filesystem going from r/w to r/o though
4673 * we don't acquire ->s_umount mutex, because the filesystem
4674 * should guarantee the delalloc inodes list be empty after
4675 * the filesystem is readonly(all dirty pages are written to
4676 * the disk).
4677 */
4678 btrfs_start_delalloc_roots(fs_info, nr_items);
4679 if (!current->journal_info)
4680 btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4681 }
4682 }
4683
4684 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4685 u64 to_reclaim)
4686 {
4687 u64 bytes;
4688 u64 nr;
4689
4690 bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4691 nr = div64_u64(to_reclaim, bytes);
4692 if (!nr)
4693 nr = 1;
4694 return nr;
4695 }
4696
4697 #define EXTENT_SIZE_PER_ITEM SZ_256K
4698
4699 /*
4700 * shrink metadata reservation for delalloc
4701 */
4702 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4703 u64 orig, bool wait_ordered)
4704 {
4705 struct btrfs_space_info *space_info;
4706 struct btrfs_trans_handle *trans;
4707 u64 delalloc_bytes;
4708 u64 max_reclaim;
4709 u64 items;
4710 long time_left;
4711 unsigned long nr_pages;
4712 int loops;
4713
4714 /* Calc the number of the pages we need flush for space reservation */
4715 items = calc_reclaim_items_nr(fs_info, to_reclaim);
4716 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4717
4718 trans = (struct btrfs_trans_handle *)current->journal_info;
4719 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4720
4721 delalloc_bytes = percpu_counter_sum_positive(
4722 &fs_info->delalloc_bytes);
4723 if (delalloc_bytes == 0) {
4724 if (trans)
4725 return;
4726 if (wait_ordered)
4727 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4728 return;
4729 }
4730
4731 loops = 0;
4732 while (delalloc_bytes && loops < 3) {
4733 max_reclaim = min(delalloc_bytes, to_reclaim);
4734 nr_pages = max_reclaim >> PAGE_SHIFT;
4735 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4736 /*
4737 * We need to wait for the async pages to actually start before
4738 * we do anything.
4739 */
4740 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4741 if (!max_reclaim)
4742 goto skip_async;
4743
4744 if (max_reclaim <= nr_pages)
4745 max_reclaim = 0;
4746 else
4747 max_reclaim -= nr_pages;
4748
4749 wait_event(fs_info->async_submit_wait,
4750 atomic_read(&fs_info->async_delalloc_pages) <=
4751 (int)max_reclaim);
4752 skip_async:
4753 spin_lock(&space_info->lock);
4754 if (list_empty(&space_info->tickets) &&
4755 list_empty(&space_info->priority_tickets)) {
4756 spin_unlock(&space_info->lock);
4757 break;
4758 }
4759 spin_unlock(&space_info->lock);
4760
4761 loops++;
4762 if (wait_ordered && !trans) {
4763 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4764 } else {
4765 time_left = schedule_timeout_killable(1);
4766 if (time_left)
4767 break;
4768 }
4769 delalloc_bytes = percpu_counter_sum_positive(
4770 &fs_info->delalloc_bytes);
4771 }
4772 }
4773
4774 struct reserve_ticket {
4775 u64 bytes;
4776 int error;
4777 struct list_head list;
4778 wait_queue_head_t wait;
4779 };
4780
4781 /**
4782 * maybe_commit_transaction - possibly commit the transaction if its ok to
4783 * @root - the root we're allocating for
4784 * @bytes - the number of bytes we want to reserve
4785 * @force - force the commit
4786 *
4787 * This will check to make sure that committing the transaction will actually
4788 * get us somewhere and then commit the transaction if it does. Otherwise it
4789 * will return -ENOSPC.
4790 */
4791 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4792 struct btrfs_space_info *space_info)
4793 {
4794 struct reserve_ticket *ticket = NULL;
4795 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4796 struct btrfs_trans_handle *trans;
4797 u64 bytes;
4798
4799 trans = (struct btrfs_trans_handle *)current->journal_info;
4800 if (trans)
4801 return -EAGAIN;
4802
4803 spin_lock(&space_info->lock);
4804 if (!list_empty(&space_info->priority_tickets))
4805 ticket = list_first_entry(&space_info->priority_tickets,
4806 struct reserve_ticket, list);
4807 else if (!list_empty(&space_info->tickets))
4808 ticket = list_first_entry(&space_info->tickets,
4809 struct reserve_ticket, list);
4810 bytes = (ticket) ? ticket->bytes : 0;
4811 spin_unlock(&space_info->lock);
4812
4813 if (!bytes)
4814 return 0;
4815
4816 /* See if there is enough pinned space to make this reservation */
4817 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4818 bytes,
4819 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4820 goto commit;
4821
4822 /*
4823 * See if there is some space in the delayed insertion reservation for
4824 * this reservation.
4825 */
4826 if (space_info != delayed_rsv->space_info)
4827 return -ENOSPC;
4828
4829 spin_lock(&delayed_rsv->lock);
4830 if (delayed_rsv->size > bytes)
4831 bytes = 0;
4832 else
4833 bytes -= delayed_rsv->size;
4834 spin_unlock(&delayed_rsv->lock);
4835
4836 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4837 bytes,
4838 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) {
4839 return -ENOSPC;
4840 }
4841
4842 commit:
4843 trans = btrfs_join_transaction(fs_info->extent_root);
4844 if (IS_ERR(trans))
4845 return -ENOSPC;
4846
4847 return btrfs_commit_transaction(trans);
4848 }
4849
4850 /*
4851 * Try to flush some data based on policy set by @state. This is only advisory
4852 * and may fail for various reasons. The caller is supposed to examine the
4853 * state of @space_info to detect the outcome.
4854 */
4855 static void flush_space(struct btrfs_fs_info *fs_info,
4856 struct btrfs_space_info *space_info, u64 num_bytes,
4857 int state)
4858 {
4859 struct btrfs_root *root = fs_info->extent_root;
4860 struct btrfs_trans_handle *trans;
4861 int nr;
4862 int ret = 0;
4863
4864 switch (state) {
4865 case FLUSH_DELAYED_ITEMS_NR:
4866 case FLUSH_DELAYED_ITEMS:
4867 if (state == FLUSH_DELAYED_ITEMS_NR)
4868 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4869 else
4870 nr = -1;
4871
4872 trans = btrfs_join_transaction(root);
4873 if (IS_ERR(trans)) {
4874 ret = PTR_ERR(trans);
4875 break;
4876 }
4877 ret = btrfs_run_delayed_items_nr(trans, nr);
4878 btrfs_end_transaction(trans);
4879 break;
4880 case FLUSH_DELALLOC:
4881 case FLUSH_DELALLOC_WAIT:
4882 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4883 state == FLUSH_DELALLOC_WAIT);
4884 break;
4885 case ALLOC_CHUNK:
4886 trans = btrfs_join_transaction(root);
4887 if (IS_ERR(trans)) {
4888 ret = PTR_ERR(trans);
4889 break;
4890 }
4891 ret = do_chunk_alloc(trans,
4892 btrfs_metadata_alloc_profile(fs_info),
4893 CHUNK_ALLOC_NO_FORCE);
4894 btrfs_end_transaction(trans);
4895 if (ret > 0 || ret == -ENOSPC)
4896 ret = 0;
4897 break;
4898 case COMMIT_TRANS:
4899 ret = may_commit_transaction(fs_info, space_info);
4900 break;
4901 default:
4902 ret = -ENOSPC;
4903 break;
4904 }
4905
4906 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4907 ret);
4908 return;
4909 }
4910
4911 static inline u64
4912 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4913 struct btrfs_space_info *space_info,
4914 bool system_chunk)
4915 {
4916 struct reserve_ticket *ticket;
4917 u64 used;
4918 u64 expected;
4919 u64 to_reclaim = 0;
4920
4921 list_for_each_entry(ticket, &space_info->tickets, list)
4922 to_reclaim += ticket->bytes;
4923 list_for_each_entry(ticket, &space_info->priority_tickets, list)
4924 to_reclaim += ticket->bytes;
4925 if (to_reclaim)
4926 return to_reclaim;
4927
4928 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4929 if (can_overcommit(fs_info, space_info, to_reclaim,
4930 BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4931 return 0;
4932
4933 used = btrfs_space_info_used(space_info, true);
4934
4935 if (can_overcommit(fs_info, space_info, SZ_1M,
4936 BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4937 expected = div_factor_fine(space_info->total_bytes, 95);
4938 else
4939 expected = div_factor_fine(space_info->total_bytes, 90);
4940
4941 if (used > expected)
4942 to_reclaim = used - expected;
4943 else
4944 to_reclaim = 0;
4945 to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4946 space_info->bytes_reserved);
4947 return to_reclaim;
4948 }
4949
4950 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
4951 struct btrfs_space_info *space_info,
4952 u64 used, bool system_chunk)
4953 {
4954 u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4955
4956 /* If we're just plain full then async reclaim just slows us down. */
4957 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4958 return 0;
4959
4960 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4961 system_chunk))
4962 return 0;
4963
4964 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4965 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4966 }
4967
4968 static void wake_all_tickets(struct list_head *head)
4969 {
4970 struct reserve_ticket *ticket;
4971
4972 while (!list_empty(head)) {
4973 ticket = list_first_entry(head, struct reserve_ticket, list);
4974 list_del_init(&ticket->list);
4975 ticket->error = -ENOSPC;
4976 wake_up(&ticket->wait);
4977 }
4978 }
4979
4980 /*
4981 * This is for normal flushers, we can wait all goddamned day if we want to. We
4982 * will loop and continuously try to flush as long as we are making progress.
4983 * We count progress as clearing off tickets each time we have to loop.
4984 */
4985 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4986 {
4987 struct btrfs_fs_info *fs_info;
4988 struct btrfs_space_info *space_info;
4989 u64 to_reclaim;
4990 int flush_state;
4991 int commit_cycles = 0;
4992 u64 last_tickets_id;
4993
4994 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4995 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4996
4997 spin_lock(&space_info->lock);
4998 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
4999 false);
5000 if (!to_reclaim) {
5001 space_info->flush = 0;
5002 spin_unlock(&space_info->lock);
5003 return;
5004 }
5005 last_tickets_id = space_info->tickets_id;
5006 spin_unlock(&space_info->lock);
5007
5008 flush_state = FLUSH_DELAYED_ITEMS_NR;
5009 do {
5010 flush_space(fs_info, space_info, to_reclaim, flush_state);
5011 spin_lock(&space_info->lock);
5012 if (list_empty(&space_info->tickets)) {
5013 space_info->flush = 0;
5014 spin_unlock(&space_info->lock);
5015 return;
5016 }
5017 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5018 space_info,
5019 false);
5020 if (last_tickets_id == space_info->tickets_id) {
5021 flush_state++;
5022 } else {
5023 last_tickets_id = space_info->tickets_id;
5024 flush_state = FLUSH_DELAYED_ITEMS_NR;
5025 if (commit_cycles)
5026 commit_cycles--;
5027 }
5028
5029 if (flush_state > COMMIT_TRANS) {
5030 commit_cycles++;
5031 if (commit_cycles > 2) {
5032 wake_all_tickets(&space_info->tickets);
5033 space_info->flush = 0;
5034 } else {
5035 flush_state = FLUSH_DELAYED_ITEMS_NR;
5036 }
5037 }
5038 spin_unlock(&space_info->lock);
5039 } while (flush_state <= COMMIT_TRANS);
5040 }
5041
5042 void btrfs_init_async_reclaim_work(struct work_struct *work)
5043 {
5044 INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5045 }
5046
5047 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5048 struct btrfs_space_info *space_info,
5049 struct reserve_ticket *ticket)
5050 {
5051 u64 to_reclaim;
5052 int flush_state = FLUSH_DELAYED_ITEMS_NR;
5053
5054 spin_lock(&space_info->lock);
5055 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5056 false);
5057 if (!to_reclaim) {
5058 spin_unlock(&space_info->lock);
5059 return;
5060 }
5061 spin_unlock(&space_info->lock);
5062
5063 do {
5064 flush_space(fs_info, space_info, to_reclaim, flush_state);
5065 flush_state++;
5066 spin_lock(&space_info->lock);
5067 if (ticket->bytes == 0) {
5068 spin_unlock(&space_info->lock);
5069 return;
5070 }
5071 spin_unlock(&space_info->lock);
5072
5073 /*
5074 * Priority flushers can't wait on delalloc without
5075 * deadlocking.
5076 */
5077 if (flush_state == FLUSH_DELALLOC ||
5078 flush_state == FLUSH_DELALLOC_WAIT)
5079 flush_state = ALLOC_CHUNK;
5080 } while (flush_state < COMMIT_TRANS);
5081 }
5082
5083 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5084 struct btrfs_space_info *space_info,
5085 struct reserve_ticket *ticket, u64 orig_bytes)
5086
5087 {
5088 DEFINE_WAIT(wait);
5089 int ret = 0;
5090
5091 spin_lock(&space_info->lock);
5092 while (ticket->bytes > 0 && ticket->error == 0) {
5093 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5094 if (ret) {
5095 ret = -EINTR;
5096 break;
5097 }
5098 spin_unlock(&space_info->lock);
5099
5100 schedule();
5101
5102 finish_wait(&ticket->wait, &wait);
5103 spin_lock(&space_info->lock);
5104 }
5105 if (!ret)
5106 ret = ticket->error;
5107 if (!list_empty(&ticket->list))
5108 list_del_init(&ticket->list);
5109 if (ticket->bytes && ticket->bytes < orig_bytes) {
5110 u64 num_bytes = orig_bytes - ticket->bytes;
5111 space_info->bytes_may_use -= num_bytes;
5112 trace_btrfs_space_reservation(fs_info, "space_info",
5113 space_info->flags, num_bytes, 0);
5114 }
5115 spin_unlock(&space_info->lock);
5116
5117 return ret;
5118 }
5119
5120 /**
5121 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5122 * @root - the root we're allocating for
5123 * @space_info - the space info we want to allocate from
5124 * @orig_bytes - the number of bytes we want
5125 * @flush - whether or not we can flush to make our reservation
5126 *
5127 * This will reserve orig_bytes number of bytes from the space info associated
5128 * with the block_rsv. If there is not enough space it will make an attempt to
5129 * flush out space to make room. It will do this by flushing delalloc if
5130 * possible or committing the transaction. If flush is 0 then no attempts to
5131 * regain reservations will be made and this will fail if there is not enough
5132 * space already.
5133 */
5134 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5135 struct btrfs_space_info *space_info,
5136 u64 orig_bytes,
5137 enum btrfs_reserve_flush_enum flush,
5138 bool system_chunk)
5139 {
5140 struct reserve_ticket ticket;
5141 u64 used;
5142 int ret = 0;
5143
5144 ASSERT(orig_bytes);
5145 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5146
5147 spin_lock(&space_info->lock);
5148 ret = -ENOSPC;
5149 used = btrfs_space_info_used(space_info, true);
5150
5151 /*
5152 * If we have enough space then hooray, make our reservation and carry
5153 * on. If not see if we can overcommit, and if we can, hooray carry on.
5154 * If not things get more complicated.
5155 */
5156 if (used + orig_bytes <= space_info->total_bytes) {
5157 space_info->bytes_may_use += orig_bytes;
5158 trace_btrfs_space_reservation(fs_info, "space_info",
5159 space_info->flags, orig_bytes, 1);
5160 ret = 0;
5161 } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5162 system_chunk)) {
5163 space_info->bytes_may_use += orig_bytes;
5164 trace_btrfs_space_reservation(fs_info, "space_info",
5165 space_info->flags, orig_bytes, 1);
5166 ret = 0;
5167 }
5168
5169 /*
5170 * If we couldn't make a reservation then setup our reservation ticket
5171 * and kick the async worker if it's not already running.
5172 *
5173 * If we are a priority flusher then we just need to add our ticket to
5174 * the list and we will do our own flushing further down.
5175 */
5176 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5177 ticket.bytes = orig_bytes;
5178 ticket.error = 0;
5179 init_waitqueue_head(&ticket.wait);
5180 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5181 list_add_tail(&ticket.list, &space_info->tickets);
5182 if (!space_info->flush) {
5183 space_info->flush = 1;
5184 trace_btrfs_trigger_flush(fs_info,
5185 space_info->flags,
5186 orig_bytes, flush,
5187 "enospc");
5188 queue_work(system_unbound_wq,
5189 &fs_info->async_reclaim_work);
5190 }
5191 } else {
5192 list_add_tail(&ticket.list,
5193 &space_info->priority_tickets);
5194 }
5195 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5196 used += orig_bytes;
5197 /*
5198 * We will do the space reservation dance during log replay,
5199 * which means we won't have fs_info->fs_root set, so don't do
5200 * the async reclaim as we will panic.
5201 */
5202 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5203 need_do_async_reclaim(fs_info, space_info,
5204 used, system_chunk) &&
5205 !work_busy(&fs_info->async_reclaim_work)) {
5206 trace_btrfs_trigger_flush(fs_info, space_info->flags,
5207 orig_bytes, flush, "preempt");
5208 queue_work(system_unbound_wq,
5209 &fs_info->async_reclaim_work);
5210 }
5211 }
5212 spin_unlock(&space_info->lock);
5213 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5214 return ret;
5215
5216 if (flush == BTRFS_RESERVE_FLUSH_ALL)
5217 return wait_reserve_ticket(fs_info, space_info, &ticket,
5218 orig_bytes);
5219
5220 ret = 0;
5221 priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5222 spin_lock(&space_info->lock);
5223 if (ticket.bytes) {
5224 if (ticket.bytes < orig_bytes) {
5225 u64 num_bytes = orig_bytes - ticket.bytes;
5226 space_info->bytes_may_use -= num_bytes;
5227 trace_btrfs_space_reservation(fs_info, "space_info",
5228 space_info->flags,
5229 num_bytes, 0);
5230
5231 }
5232 list_del_init(&ticket.list);
5233 ret = -ENOSPC;
5234 }
5235 spin_unlock(&space_info->lock);
5236 ASSERT(list_empty(&ticket.list));
5237 return ret;
5238 }
5239
5240 /**
5241 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5242 * @root - the root we're allocating for
5243 * @block_rsv - the block_rsv we're allocating for
5244 * @orig_bytes - the number of bytes we want
5245 * @flush - whether or not we can flush to make our reservation
5246 *
5247 * This will reserve orgi_bytes number of bytes from the space info associated
5248 * with the block_rsv. If there is not enough space it will make an attempt to
5249 * flush out space to make room. It will do this by flushing delalloc if
5250 * possible or committing the transaction. If flush is 0 then no attempts to
5251 * regain reservations will be made and this will fail if there is not enough
5252 * space already.
5253 */
5254 static int reserve_metadata_bytes(struct btrfs_root *root,
5255 struct btrfs_block_rsv *block_rsv,
5256 u64 orig_bytes,
5257 enum btrfs_reserve_flush_enum flush)
5258 {
5259 struct btrfs_fs_info *fs_info = root->fs_info;
5260 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5261 int ret;
5262 bool system_chunk = (root == fs_info->chunk_root);
5263
5264 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5265 orig_bytes, flush, system_chunk);
5266 if (ret == -ENOSPC &&
5267 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5268 if (block_rsv != global_rsv &&
5269 !block_rsv_use_bytes(global_rsv, orig_bytes))
5270 ret = 0;
5271 }
5272 if (ret == -ENOSPC) {
5273 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5274 block_rsv->space_info->flags,
5275 orig_bytes, 1);
5276
5277 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5278 dump_space_info(fs_info, block_rsv->space_info,
5279 orig_bytes, 0);
5280 }
5281 return ret;
5282 }
5283
5284 static struct btrfs_block_rsv *get_block_rsv(
5285 const struct btrfs_trans_handle *trans,
5286 const struct btrfs_root *root)
5287 {
5288 struct btrfs_fs_info *fs_info = root->fs_info;
5289 struct btrfs_block_rsv *block_rsv = NULL;
5290
5291 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5292 (root == fs_info->csum_root && trans->adding_csums) ||
5293 (root == fs_info->uuid_root))
5294 block_rsv = trans->block_rsv;
5295
5296 if (!block_rsv)
5297 block_rsv = root->block_rsv;
5298
5299 if (!block_rsv)
5300 block_rsv = &fs_info->empty_block_rsv;
5301
5302 return block_rsv;
5303 }
5304
5305 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5306 u64 num_bytes)
5307 {
5308 int ret = -ENOSPC;
5309 spin_lock(&block_rsv->lock);
5310 if (block_rsv->reserved >= num_bytes) {
5311 block_rsv->reserved -= num_bytes;
5312 if (block_rsv->reserved < block_rsv->size)
5313 block_rsv->full = 0;
5314 ret = 0;
5315 }
5316 spin_unlock(&block_rsv->lock);
5317 return ret;
5318 }
5319
5320 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5321 u64 num_bytes, bool update_size)
5322 {
5323 spin_lock(&block_rsv->lock);
5324 block_rsv->reserved += num_bytes;
5325 if (update_size)
5326 block_rsv->size += num_bytes;
5327 else if (block_rsv->reserved >= block_rsv->size)
5328 block_rsv->full = 1;
5329 spin_unlock(&block_rsv->lock);
5330 }
5331
5332 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5333 struct btrfs_block_rsv *dest, u64 num_bytes,
5334 int min_factor)
5335 {
5336 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5337 u64 min_bytes;
5338
5339 if (global_rsv->space_info != dest->space_info)
5340 return -ENOSPC;
5341
5342 spin_lock(&global_rsv->lock);
5343 min_bytes = div_factor(global_rsv->size, min_factor);
5344 if (global_rsv->reserved < min_bytes + num_bytes) {
5345 spin_unlock(&global_rsv->lock);
5346 return -ENOSPC;
5347 }
5348 global_rsv->reserved -= num_bytes;
5349 if (global_rsv->reserved < global_rsv->size)
5350 global_rsv->full = 0;
5351 spin_unlock(&global_rsv->lock);
5352
5353 block_rsv_add_bytes(dest, num_bytes, true);
5354 return 0;
5355 }
5356
5357 /*
5358 * This is for space we already have accounted in space_info->bytes_may_use, so
5359 * basically when we're returning space from block_rsv's.
5360 */
5361 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5362 struct btrfs_space_info *space_info,
5363 u64 num_bytes)
5364 {
5365 struct reserve_ticket *ticket;
5366 struct list_head *head;
5367 u64 used;
5368 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5369 bool check_overcommit = false;
5370
5371 spin_lock(&space_info->lock);
5372 head = &space_info->priority_tickets;
5373
5374 /*
5375 * If we are over our limit then we need to check and see if we can
5376 * overcommit, and if we can't then we just need to free up our space
5377 * and not satisfy any requests.
5378 */
5379 used = btrfs_space_info_used(space_info, true);
5380 if (used - num_bytes >= space_info->total_bytes)
5381 check_overcommit = true;
5382 again:
5383 while (!list_empty(head) && num_bytes) {
5384 ticket = list_first_entry(head, struct reserve_ticket,
5385 list);
5386 /*
5387 * We use 0 bytes because this space is already reserved, so
5388 * adding the ticket space would be a double count.
5389 */
5390 if (check_overcommit &&
5391 !can_overcommit(fs_info, space_info, 0, flush, false))
5392 break;
5393 if (num_bytes >= ticket->bytes) {
5394 list_del_init(&ticket->list);
5395 num_bytes -= ticket->bytes;
5396 ticket->bytes = 0;
5397 space_info->tickets_id++;
5398 wake_up(&ticket->wait);
5399 } else {
5400 ticket->bytes -= num_bytes;
5401 num_bytes = 0;
5402 }
5403 }
5404
5405 if (num_bytes && head == &space_info->priority_tickets) {
5406 head = &space_info->tickets;
5407 flush = BTRFS_RESERVE_FLUSH_ALL;
5408 goto again;
5409 }
5410 space_info->bytes_may_use -= num_bytes;
5411 trace_btrfs_space_reservation(fs_info, "space_info",
5412 space_info->flags, num_bytes, 0);
5413 spin_unlock(&space_info->lock);
5414 }
5415
5416 /*
5417 * This is for newly allocated space that isn't accounted in
5418 * space_info->bytes_may_use yet. So if we allocate a chunk or unpin an extent
5419 * we use this helper.
5420 */
5421 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5422 struct btrfs_space_info *space_info,
5423 u64 num_bytes)
5424 {
5425 struct reserve_ticket *ticket;
5426 struct list_head *head = &space_info->priority_tickets;
5427
5428 again:
5429 while (!list_empty(head) && num_bytes) {
5430 ticket = list_first_entry(head, struct reserve_ticket,
5431 list);
5432 if (num_bytes >= ticket->bytes) {
5433 trace_btrfs_space_reservation(fs_info, "space_info",
5434 space_info->flags,
5435 ticket->bytes, 1);
5436 list_del_init(&ticket->list);
5437 num_bytes -= ticket->bytes;
5438 space_info->bytes_may_use += ticket->bytes;
5439 ticket->bytes = 0;
5440 space_info->tickets_id++;
5441 wake_up(&ticket->wait);
5442 } else {
5443 trace_btrfs_space_reservation(fs_info, "space_info",
5444 space_info->flags,
5445 num_bytes, 1);
5446 space_info->bytes_may_use += num_bytes;
5447 ticket->bytes -= num_bytes;
5448 num_bytes = 0;
5449 }
5450 }
5451
5452 if (num_bytes && head == &space_info->priority_tickets) {
5453 head = &space_info->tickets;
5454 goto again;
5455 }
5456 }
5457
5458 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5459 struct btrfs_block_rsv *block_rsv,
5460 struct btrfs_block_rsv *dest, u64 num_bytes,
5461 u64 *qgroup_to_release_ret)
5462 {
5463 struct btrfs_space_info *space_info = block_rsv->space_info;
5464 u64 qgroup_to_release = 0;
5465 u64 ret;
5466
5467 spin_lock(&block_rsv->lock);
5468 if (num_bytes == (u64)-1) {
5469 num_bytes = block_rsv->size;
5470 qgroup_to_release = block_rsv->qgroup_rsv_size;
5471 }
5472 block_rsv->size -= num_bytes;
5473 if (block_rsv->reserved >= block_rsv->size) {
5474 num_bytes = block_rsv->reserved - block_rsv->size;
5475 block_rsv->reserved = block_rsv->size;
5476 block_rsv->full = 1;
5477 } else {
5478 num_bytes = 0;
5479 }
5480 if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5481 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5482 block_rsv->qgroup_rsv_size;
5483 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5484 } else {
5485 qgroup_to_release = 0;
5486 }
5487 spin_unlock(&block_rsv->lock);
5488
5489 ret = num_bytes;
5490 if (num_bytes > 0) {
5491 if (dest) {
5492 spin_lock(&dest->lock);
5493 if (!dest->full) {
5494 u64 bytes_to_add;
5495
5496 bytes_to_add = dest->size - dest->reserved;
5497 bytes_to_add = min(num_bytes, bytes_to_add);
5498 dest->reserved += bytes_to_add;
5499 if (dest->reserved >= dest->size)
5500 dest->full = 1;
5501 num_bytes -= bytes_to_add;
5502 }
5503 spin_unlock(&dest->lock);
5504 }
5505 if (num_bytes)
5506 space_info_add_old_bytes(fs_info, space_info,
5507 num_bytes);
5508 }
5509 if (qgroup_to_release_ret)
5510 *qgroup_to_release_ret = qgroup_to_release;
5511 return ret;
5512 }
5513
5514 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5515 struct btrfs_block_rsv *dst, u64 num_bytes,
5516 bool update_size)
5517 {
5518 int ret;
5519
5520 ret = block_rsv_use_bytes(src, num_bytes);
5521 if (ret)
5522 return ret;
5523
5524 block_rsv_add_bytes(dst, num_bytes, update_size);
5525 return 0;
5526 }
5527
5528 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5529 {
5530 memset(rsv, 0, sizeof(*rsv));
5531 spin_lock_init(&rsv->lock);
5532 rsv->type = type;
5533 }
5534
5535 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5536 struct btrfs_block_rsv *rsv,
5537 unsigned short type)
5538 {
5539 btrfs_init_block_rsv(rsv, type);
5540 rsv->space_info = __find_space_info(fs_info,
5541 BTRFS_BLOCK_GROUP_METADATA);
5542 }
5543
5544 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5545 unsigned short type)
5546 {
5547 struct btrfs_block_rsv *block_rsv;
5548
5549 block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5550 if (!block_rsv)
5551 return NULL;
5552
5553 btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5554 return block_rsv;
5555 }
5556
5557 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5558 struct btrfs_block_rsv *rsv)
5559 {
5560 if (!rsv)
5561 return;
5562 btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5563 kfree(rsv);
5564 }
5565
5566 int btrfs_block_rsv_add(struct btrfs_root *root,
5567 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5568 enum btrfs_reserve_flush_enum flush)
5569 {
5570 int ret;
5571
5572 if (num_bytes == 0)
5573 return 0;
5574
5575 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5576 if (!ret)
5577 block_rsv_add_bytes(block_rsv, num_bytes, true);
5578
5579 return ret;
5580 }
5581
5582 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5583 {
5584 u64 num_bytes = 0;
5585 int ret = -ENOSPC;
5586
5587 if (!block_rsv)
5588 return 0;
5589
5590 spin_lock(&block_rsv->lock);
5591 num_bytes = div_factor(block_rsv->size, min_factor);
5592 if (block_rsv->reserved >= num_bytes)
5593 ret = 0;
5594 spin_unlock(&block_rsv->lock);
5595
5596 return ret;
5597 }
5598
5599 int btrfs_block_rsv_refill(struct btrfs_root *root,
5600 struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5601 enum btrfs_reserve_flush_enum flush)
5602 {
5603 u64 num_bytes = 0;
5604 int ret = -ENOSPC;
5605
5606 if (!block_rsv)
5607 return 0;
5608
5609 spin_lock(&block_rsv->lock);
5610 num_bytes = min_reserved;
5611 if (block_rsv->reserved >= num_bytes)
5612 ret = 0;
5613 else
5614 num_bytes -= block_rsv->reserved;
5615 spin_unlock(&block_rsv->lock);
5616
5617 if (!ret)
5618 return 0;
5619
5620 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5621 if (!ret) {
5622 block_rsv_add_bytes(block_rsv, num_bytes, false);
5623 return 0;
5624 }
5625
5626 return ret;
5627 }
5628
5629 /**
5630 * btrfs_inode_rsv_refill - refill the inode block rsv.
5631 * @inode - the inode we are refilling.
5632 * @flush - the flusing restriction.
5633 *
5634 * Essentially the same as btrfs_block_rsv_refill, except it uses the
5635 * block_rsv->size as the minimum size. We'll either refill the missing amount
5636 * or return if we already have enough space. This will also handle the resreve
5637 * tracepoint for the reserved amount.
5638 */
5639 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5640 enum btrfs_reserve_flush_enum flush)
5641 {
5642 struct btrfs_root *root = inode->root;
5643 struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5644 u64 num_bytes = 0;
5645 u64 qgroup_num_bytes = 0;
5646 int ret = -ENOSPC;
5647
5648 spin_lock(&block_rsv->lock);
5649 if (block_rsv->reserved < block_rsv->size)
5650 num_bytes = block_rsv->size - block_rsv->reserved;
5651 if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5652 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5653 block_rsv->qgroup_rsv_reserved;
5654 spin_unlock(&block_rsv->lock);
5655
5656 if (num_bytes == 0)
5657 return 0;
5658
5659 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5660 if (ret)
5661 return ret;
5662 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5663 if (!ret) {
5664 block_rsv_add_bytes(block_rsv, num_bytes, false);
5665 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5666 btrfs_ino(inode), num_bytes, 1);
5667
5668 /* Don't forget to increase qgroup_rsv_reserved */
5669 spin_lock(&block_rsv->lock);
5670 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5671 spin_unlock(&block_rsv->lock);
5672 } else
5673 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5674 return ret;
5675 }
5676
5677 /**
5678 * btrfs_inode_rsv_release - release any excessive reservation.
5679 * @inode - the inode we need to release from.
5680 * @qgroup_free - free or convert qgroup meta.
5681 * Unlike normal operation, qgroup meta reservation needs to know if we are
5682 * freeing qgroup reservation or just converting it into per-trans. Normally
5683 * @qgroup_free is true for error handling, and false for normal release.
5684 *
5685 * This is the same as btrfs_block_rsv_release, except that it handles the
5686 * tracepoint for the reservation.
5687 */
5688 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5689 {
5690 struct btrfs_fs_info *fs_info = inode->root->fs_info;
5691 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5692 struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5693 u64 released = 0;
5694 u64 qgroup_to_release = 0;
5695
5696 /*
5697 * Since we statically set the block_rsv->size we just want to say we
5698 * are releasing 0 bytes, and then we'll just get the reservation over
5699 * the size free'd.
5700 */
5701 released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5702 &qgroup_to_release);
5703 if (released > 0)
5704 trace_btrfs_space_reservation(fs_info, "delalloc",
5705 btrfs_ino(inode), released, 0);
5706 if (qgroup_free)
5707 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5708 else
5709 btrfs_qgroup_convert_reserved_meta(inode->root,
5710 qgroup_to_release);
5711 }
5712
5713 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5714 struct btrfs_block_rsv *block_rsv,
5715 u64 num_bytes)
5716 {
5717 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5718
5719 if (global_rsv == block_rsv ||
5720 block_rsv->space_info != global_rsv->space_info)
5721 global_rsv = NULL;
5722 block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5723 }
5724
5725 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5726 {
5727 struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5728 struct btrfs_space_info *sinfo = block_rsv->space_info;
5729 u64 num_bytes;
5730
5731 /*
5732 * The global block rsv is based on the size of the extent tree, the
5733 * checksum tree and the root tree. If the fs is empty we want to set
5734 * it to a minimal amount for safety.
5735 */
5736 num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5737 btrfs_root_used(&fs_info->csum_root->root_item) +
5738 btrfs_root_used(&fs_info->tree_root->root_item);
5739 num_bytes = max_t(u64, num_bytes, SZ_16M);
5740
5741 spin_lock(&sinfo->lock);
5742 spin_lock(&block_rsv->lock);
5743
5744 block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5745
5746 if (block_rsv->reserved < block_rsv->size) {
5747 num_bytes = btrfs_space_info_used(sinfo, true);
5748 if (sinfo->total_bytes > num_bytes) {
5749 num_bytes = sinfo->total_bytes - num_bytes;
5750 num_bytes = min(num_bytes,
5751 block_rsv->size - block_rsv->reserved);
5752 block_rsv->reserved += num_bytes;
5753 sinfo->bytes_may_use += num_bytes;
5754 trace_btrfs_space_reservation(fs_info, "space_info",
5755 sinfo->flags, num_bytes,
5756 1);
5757 }
5758 } else if (block_rsv->reserved > block_rsv->size) {
5759 num_bytes = block_rsv->reserved - block_rsv->size;
5760 sinfo->bytes_may_use -= num_bytes;
5761 trace_btrfs_space_reservation(fs_info, "space_info",
5762 sinfo->flags, num_bytes, 0);
5763 block_rsv->reserved = block_rsv->size;
5764 }
5765
5766 if (block_rsv->reserved == block_rsv->size)
5767 block_rsv->full = 1;
5768 else
5769 block_rsv->full = 0;
5770
5771 spin_unlock(&block_rsv->lock);
5772 spin_unlock(&sinfo->lock);
5773 }
5774
5775 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5776 {
5777 struct btrfs_space_info *space_info;
5778
5779 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5780 fs_info->chunk_block_rsv.space_info = space_info;
5781
5782 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5783 fs_info->global_block_rsv.space_info = space_info;
5784 fs_info->trans_block_rsv.space_info = space_info;
5785 fs_info->empty_block_rsv.space_info = space_info;
5786 fs_info->delayed_block_rsv.space_info = space_info;
5787
5788 fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5789 fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5790 fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5791 fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5792 if (fs_info->quota_root)
5793 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5794 fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5795
5796 update_global_block_rsv(fs_info);
5797 }
5798
5799 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5800 {
5801 block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5802 (u64)-1, NULL);
5803 WARN_ON(fs_info->trans_block_rsv.size > 0);
5804 WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5805 WARN_ON(fs_info->chunk_block_rsv.size > 0);
5806 WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5807 WARN_ON(fs_info->delayed_block_rsv.size > 0);
5808 WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5809 }
5810
5811
5812 /*
5813 * To be called after all the new block groups attached to the transaction
5814 * handle have been created (btrfs_create_pending_block_groups()).
5815 */
5816 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5817 {
5818 struct btrfs_fs_info *fs_info = trans->fs_info;
5819
5820 if (!trans->chunk_bytes_reserved)
5821 return;
5822
5823 WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5824
5825 block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5826 trans->chunk_bytes_reserved, NULL);
5827 trans->chunk_bytes_reserved = 0;
5828 }
5829
5830 /*
5831 * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5832 * root: the root of the parent directory
5833 * rsv: block reservation
5834 * items: the number of items that we need do reservation
5835 * use_global_rsv: allow fallback to the global block reservation
5836 *
5837 * This function is used to reserve the space for snapshot/subvolume
5838 * creation and deletion. Those operations are different with the
5839 * common file/directory operations, they change two fs/file trees
5840 * and root tree, the number of items that the qgroup reserves is
5841 * different with the free space reservation. So we can not use
5842 * the space reservation mechanism in start_transaction().
5843 */
5844 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5845 struct btrfs_block_rsv *rsv, int items,
5846 bool use_global_rsv)
5847 {
5848 u64 qgroup_num_bytes = 0;
5849 u64 num_bytes;
5850 int ret;
5851 struct btrfs_fs_info *fs_info = root->fs_info;
5852 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5853
5854 if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5855 /* One for parent inode, two for dir entries */
5856 qgroup_num_bytes = 3 * fs_info->nodesize;
5857 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5858 qgroup_num_bytes, true);
5859 if (ret)
5860 return ret;
5861 }
5862
5863 num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5864 rsv->space_info = __find_space_info(fs_info,
5865 BTRFS_BLOCK_GROUP_METADATA);
5866 ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5867 BTRFS_RESERVE_FLUSH_ALL);
5868
5869 if (ret == -ENOSPC && use_global_rsv)
5870 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
5871
5872 if (ret && qgroup_num_bytes)
5873 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5874
5875 return ret;
5876 }
5877
5878 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5879 struct btrfs_block_rsv *rsv)
5880 {
5881 btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5882 }
5883
5884 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5885 struct btrfs_inode *inode)
5886 {
5887 struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5888 u64 reserve_size = 0;
5889 u64 qgroup_rsv_size = 0;
5890 u64 csum_leaves;
5891 unsigned outstanding_extents;
5892
5893 lockdep_assert_held(&inode->lock);
5894 outstanding_extents = inode->outstanding_extents;
5895 if (outstanding_extents)
5896 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5897 outstanding_extents + 1);
5898 csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
5899 inode->csum_bytes);
5900 reserve_size += btrfs_calc_trans_metadata_size(fs_info,
5901 csum_leaves);
5902 /*
5903 * For qgroup rsv, the calculation is very simple:
5904 * account one nodesize for each outstanding extent
5905 *
5906 * This is overestimating in most cases.
5907 */
5908 qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
5909
5910 spin_lock(&block_rsv->lock);
5911 block_rsv->size = reserve_size;
5912 block_rsv->qgroup_rsv_size = qgroup_rsv_size;
5913 spin_unlock(&block_rsv->lock);
5914 }
5915
5916 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5917 {
5918 struct btrfs_fs_info *fs_info = inode->root->fs_info;
5919 unsigned nr_extents;
5920 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5921 int ret = 0;
5922 bool delalloc_lock = true;
5923
5924 /* If we are a free space inode we need to not flush since we will be in
5925 * the middle of a transaction commit. We also don't need the delalloc
5926 * mutex since we won't race with anybody. We need this mostly to make
5927 * lockdep shut its filthy mouth.
5928 *
5929 * If we have a transaction open (can happen if we call truncate_block
5930 * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5931 */
5932 if (btrfs_is_free_space_inode(inode)) {
5933 flush = BTRFS_RESERVE_NO_FLUSH;
5934 delalloc_lock = false;
5935 } else {
5936 if (current->journal_info)
5937 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5938
5939 if (btrfs_transaction_in_commit(fs_info))
5940 schedule_timeout(1);
5941 }
5942
5943 if (delalloc_lock)
5944 mutex_lock(&inode->delalloc_mutex);
5945
5946 num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5947
5948 /* Add our new extents and calculate the new rsv size. */
5949 spin_lock(&inode->lock);
5950 nr_extents = count_max_extents(num_bytes);
5951 btrfs_mod_outstanding_extents(inode, nr_extents);
5952 inode->csum_bytes += num_bytes;
5953 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5954 spin_unlock(&inode->lock);
5955
5956 ret = btrfs_inode_rsv_refill(inode, flush);
5957 if (unlikely(ret))
5958 goto out_fail;
5959
5960 if (delalloc_lock)
5961 mutex_unlock(&inode->delalloc_mutex);
5962 return 0;
5963
5964 out_fail:
5965 spin_lock(&inode->lock);
5966 nr_extents = count_max_extents(num_bytes);
5967 btrfs_mod_outstanding_extents(inode, -nr_extents);
5968 inode->csum_bytes -= num_bytes;
5969 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5970 spin_unlock(&inode->lock);
5971
5972 btrfs_inode_rsv_release(inode, true);
5973 if (delalloc_lock)
5974 mutex_unlock(&inode->delalloc_mutex);
5975 return ret;
5976 }
5977
5978 /**
5979 * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5980 * @inode: the inode to release the reservation for.
5981 * @num_bytes: the number of bytes we are releasing.
5982 * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
5983 *
5984 * This will release the metadata reservation for an inode. This can be called
5985 * once we complete IO for a given set of bytes to release their metadata
5986 * reservations, or on error for the same reason.
5987 */
5988 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
5989 bool qgroup_free)
5990 {
5991 struct btrfs_fs_info *fs_info = inode->root->fs_info;
5992
5993 num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
5994 spin_lock(&inode->lock);
5995 inode->csum_bytes -= num_bytes;
5996 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
5997 spin_unlock(&inode->lock);
5998
5999 if (btrfs_is_testing(fs_info))
6000 return;
6001
6002 btrfs_inode_rsv_release(inode, qgroup_free);
6003 }
6004
6005 /**
6006 * btrfs_delalloc_release_extents - release our outstanding_extents
6007 * @inode: the inode to balance the reservation for.
6008 * @num_bytes: the number of bytes we originally reserved with
6009 * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6010 *
6011 * When we reserve space we increase outstanding_extents for the extents we may
6012 * add. Once we've set the range as delalloc or created our ordered extents we
6013 * have outstanding_extents to track the real usage, so we use this to free our
6014 * temporarily tracked outstanding_extents. This _must_ be used in conjunction
6015 * with btrfs_delalloc_reserve_metadata.
6016 */
6017 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6018 bool qgroup_free)
6019 {
6020 struct btrfs_fs_info *fs_info = inode->root->fs_info;
6021 unsigned num_extents;
6022
6023 spin_lock(&inode->lock);
6024 num_extents = count_max_extents(num_bytes);
6025 btrfs_mod_outstanding_extents(inode, -num_extents);
6026 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6027 spin_unlock(&inode->lock);
6028
6029 if (btrfs_is_testing(fs_info))
6030 return;
6031
6032 btrfs_inode_rsv_release(inode, qgroup_free);
6033 }
6034
6035 /**
6036 * btrfs_delalloc_reserve_space - reserve data and metadata space for
6037 * delalloc
6038 * @inode: inode we're writing to
6039 * @start: start range we are writing to
6040 * @len: how long the range we are writing to
6041 * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6042 * current reservation.
6043 *
6044 * This will do the following things
6045 *
6046 * o reserve space in data space info for num bytes
6047 * and reserve precious corresponding qgroup space
6048 * (Done in check_data_free_space)
6049 *
6050 * o reserve space for metadata space, based on the number of outstanding
6051 * extents and how much csums will be needed
6052 * also reserve metadata space in a per root over-reserve method.
6053 * o add to the inodes->delalloc_bytes
6054 * o add it to the fs_info's delalloc inodes list.
6055 * (Above 3 all done in delalloc_reserve_metadata)
6056 *
6057 * Return 0 for success
6058 * Return <0 for error(-ENOSPC or -EQUOT)
6059 */
6060 int btrfs_delalloc_reserve_space(struct inode *inode,
6061 struct extent_changeset **reserved, u64 start, u64 len)
6062 {
6063 int ret;
6064
6065 ret = btrfs_check_data_free_space(inode, reserved, start, len);
6066 if (ret < 0)
6067 return ret;
6068 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6069 if (ret < 0)
6070 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6071 return ret;
6072 }
6073
6074 /**
6075 * btrfs_delalloc_release_space - release data and metadata space for delalloc
6076 * @inode: inode we're releasing space for
6077 * @start: start position of the space already reserved
6078 * @len: the len of the space already reserved
6079 * @release_bytes: the len of the space we consumed or didn't use
6080 *
6081 * This function will release the metadata space that was not used and will
6082 * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6083 * list if there are no delalloc bytes left.
6084 * Also it will handle the qgroup reserved space.
6085 */
6086 void btrfs_delalloc_release_space(struct inode *inode,
6087 struct extent_changeset *reserved,
6088 u64 start, u64 len, bool qgroup_free)
6089 {
6090 btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6091 btrfs_free_reserved_data_space(inode, reserved, start, len);
6092 }
6093
6094 static int update_block_group(struct btrfs_trans_handle *trans,
6095 struct btrfs_fs_info *info, u64 bytenr,
6096 u64 num_bytes, int alloc)
6097 {
6098 struct btrfs_block_group_cache *cache = NULL;
6099 u64 total = num_bytes;
6100 u64 old_val;
6101 u64 byte_in_group;
6102 int factor;
6103
6104 /* block accounting for super block */
6105 spin_lock(&info->delalloc_root_lock);
6106 old_val = btrfs_super_bytes_used(info->super_copy);
6107 if (alloc)
6108 old_val += num_bytes;
6109 else
6110 old_val -= num_bytes;
6111 btrfs_set_super_bytes_used(info->super_copy, old_val);
6112 spin_unlock(&info->delalloc_root_lock);
6113
6114 while (total) {
6115 cache = btrfs_lookup_block_group(info, bytenr);
6116 if (!cache)
6117 return -ENOENT;
6118 factor = btrfs_bg_type_to_factor(cache->flags);
6119
6120 /*
6121 * If this block group has free space cache written out, we
6122 * need to make sure to load it if we are removing space. This
6123 * is because we need the unpinning stage to actually add the
6124 * space back to the block group, otherwise we will leak space.
6125 */
6126 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6127 cache_block_group(cache, 1);
6128
6129 byte_in_group = bytenr - cache->key.objectid;
6130 WARN_ON(byte_in_group > cache->key.offset);
6131
6132 spin_lock(&cache->space_info->lock);
6133 spin_lock(&cache->lock);
6134
6135 if (btrfs_test_opt(info, SPACE_CACHE) &&
6136 cache->disk_cache_state < BTRFS_DC_CLEAR)
6137 cache->disk_cache_state = BTRFS_DC_CLEAR;
6138
6139 old_val = btrfs_block_group_used(&cache->item);
6140 num_bytes = min(total, cache->key.offset - byte_in_group);
6141 if (alloc) {
6142 old_val += num_bytes;
6143 btrfs_set_block_group_used(&cache->item, old_val);
6144 cache->reserved -= num_bytes;
6145 cache->space_info->bytes_reserved -= num_bytes;
6146 cache->space_info->bytes_used += num_bytes;
6147 cache->space_info->disk_used += num_bytes * factor;
6148 spin_unlock(&cache->lock);
6149 spin_unlock(&cache->space_info->lock);
6150 } else {
6151 old_val -= num_bytes;
6152 btrfs_set_block_group_used(&cache->item, old_val);
6153 cache->pinned += num_bytes;
6154 cache->space_info->bytes_pinned += num_bytes;
6155 cache->space_info->bytes_used -= num_bytes;
6156 cache->space_info->disk_used -= num_bytes * factor;
6157 spin_unlock(&cache->lock);
6158 spin_unlock(&cache->space_info->lock);
6159
6160 trace_btrfs_space_reservation(info, "pinned",
6161 cache->space_info->flags,
6162 num_bytes, 1);
6163 percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6164 num_bytes,
6165 BTRFS_TOTAL_BYTES_PINNED_BATCH);
6166 set_extent_dirty(info->pinned_extents,
6167 bytenr, bytenr + num_bytes - 1,
6168 GFP_NOFS | __GFP_NOFAIL);
6169 }
6170
6171 spin_lock(&trans->transaction->dirty_bgs_lock);
6172 if (list_empty(&cache->dirty_list)) {
6173 list_add_tail(&cache->dirty_list,
6174 &trans->transaction->dirty_bgs);
6175 trans->transaction->num_dirty_bgs++;
6176 btrfs_get_block_group(cache);
6177 }
6178 spin_unlock(&trans->transaction->dirty_bgs_lock);
6179
6180 /*
6181 * No longer have used bytes in this block group, queue it for
6182 * deletion. We do this after adding the block group to the
6183 * dirty list to avoid races between cleaner kthread and space
6184 * cache writeout.
6185 */
6186 if (!alloc && old_val == 0)
6187 btrfs_mark_bg_unused(cache);
6188
6189 btrfs_put_block_group(cache);
6190 total -= num_bytes;
6191 bytenr += num_bytes;
6192 }
6193 return 0;
6194 }
6195
6196 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6197 {
6198 struct btrfs_block_group_cache *cache;
6199 u64 bytenr;
6200
6201 spin_lock(&fs_info->block_group_cache_lock);
6202 bytenr = fs_info->first_logical_byte;
6203 spin_unlock(&fs_info->block_group_cache_lock);
6204
6205 if (bytenr < (u64)-1)
6206 return bytenr;
6207
6208 cache = btrfs_lookup_first_block_group(fs_info, search_start);
6209 if (!cache)
6210 return 0;
6211
6212 bytenr = cache->key.objectid;
6213 btrfs_put_block_group(cache);
6214
6215 return bytenr;
6216 }
6217
6218 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6219 struct btrfs_block_group_cache *cache,
6220 u64 bytenr, u64 num_bytes, int reserved)
6221 {
6222 spin_lock(&cache->space_info->lock);
6223 spin_lock(&cache->lock);
6224 cache->pinned += num_bytes;
6225 cache->space_info->bytes_pinned += num_bytes;
6226 if (reserved) {
6227 cache->reserved -= num_bytes;
6228 cache->space_info->bytes_reserved -= num_bytes;
6229 }
6230 spin_unlock(&cache->lock);
6231 spin_unlock(&cache->space_info->lock);
6232
6233 trace_btrfs_space_reservation(fs_info, "pinned",
6234 cache->space_info->flags, num_bytes, 1);
6235 percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6236 num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6237 set_extent_dirty(fs_info->pinned_extents, bytenr,
6238 bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6239 return 0;
6240 }
6241
6242 /*
6243 * this function must be called within transaction
6244 */
6245 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6246 u64 bytenr, u64 num_bytes, int reserved)
6247 {
6248 struct btrfs_block_group_cache *cache;
6249
6250 cache = btrfs_lookup_block_group(fs_info, bytenr);
6251 BUG_ON(!cache); /* Logic error */
6252
6253 pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6254
6255 btrfs_put_block_group(cache);
6256 return 0;
6257 }
6258
6259 /*
6260 * this function must be called within transaction
6261 */
6262 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6263 u64 bytenr, u64 num_bytes)
6264 {
6265 struct btrfs_block_group_cache *cache;
6266 int ret;
6267
6268 cache = btrfs_lookup_block_group(fs_info, bytenr);
6269 if (!cache)
6270 return -EINVAL;
6271
6272 /*
6273 * pull in the free space cache (if any) so that our pin
6274 * removes the free space from the cache. We have load_only set
6275 * to one because the slow code to read in the free extents does check
6276 * the pinned extents.
6277 */
6278 cache_block_group(cache, 1);
6279
6280 pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6281
6282 /* remove us from the free space cache (if we're there at all) */
6283 ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6284 btrfs_put_block_group(cache);
6285 return ret;
6286 }
6287
6288 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6289 u64 start, u64 num_bytes)
6290 {
6291 int ret;
6292 struct btrfs_block_group_cache *block_group;
6293 struct btrfs_caching_control *caching_ctl;
6294
6295 block_group = btrfs_lookup_block_group(fs_info, start);
6296 if (!block_group)
6297 return -EINVAL;
6298
6299 cache_block_group(block_group, 0);
6300 caching_ctl = get_caching_control(block_group);
6301
6302 if (!caching_ctl) {
6303 /* Logic error */
6304 BUG_ON(!block_group_cache_done(block_group));
6305 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6306 } else {
6307 mutex_lock(&caching_ctl->mutex);
6308
6309 if (start >= caching_ctl->progress) {
6310 ret = add_excluded_extent(fs_info, start, num_bytes);
6311 } else if (start + num_bytes <= caching_ctl->progress) {
6312 ret = btrfs_remove_free_space(block_group,
6313 start, num_bytes);
6314 } else {
6315 num_bytes = caching_ctl->progress - start;
6316 ret = btrfs_remove_free_space(block_group,
6317 start, num_bytes);
6318 if (ret)
6319 goto out_lock;
6320
6321 num_bytes = (start + num_bytes) -
6322 caching_ctl->progress;
6323 start = caching_ctl->progress;
6324 ret = add_excluded_extent(fs_info, start, num_bytes);
6325 }
6326 out_lock:
6327 mutex_unlock(&caching_ctl->mutex);
6328 put_caching_control(caching_ctl);
6329 }
6330 btrfs_put_block_group(block_group);
6331 return ret;
6332 }
6333
6334 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6335 struct extent_buffer *eb)
6336 {
6337 struct btrfs_file_extent_item *item;
6338 struct btrfs_key key;
6339 int found_type;
6340 int i;
6341 int ret = 0;
6342
6343 if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6344 return 0;
6345
6346 for (i = 0; i < btrfs_header_nritems(eb); i++) {
6347 btrfs_item_key_to_cpu(eb, &key, i);
6348 if (key.type != BTRFS_EXTENT_DATA_KEY)
6349 continue;
6350 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6351 found_type = btrfs_file_extent_type(eb, item);
6352 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6353 continue;
6354 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6355 continue;
6356 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6357 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6358 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6359 if (ret)
6360 break;
6361 }
6362
6363 return ret;
6364 }
6365
6366 static void
6367 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6368 {
6369 atomic_inc(&bg->reservations);
6370 }
6371
6372 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6373 const u64 start)
6374 {
6375 struct btrfs_block_group_cache *bg;
6376
6377 bg = btrfs_lookup_block_group(fs_info, start);
6378 ASSERT(bg);
6379 if (atomic_dec_and_test(&bg->reservations))
6380 wake_up_var(&bg->reservations);
6381 btrfs_put_block_group(bg);
6382 }
6383
6384 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6385 {
6386 struct btrfs_space_info *space_info = bg->space_info;
6387
6388 ASSERT(bg->ro);
6389
6390 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6391 return;
6392
6393 /*
6394 * Our block group is read only but before we set it to read only,
6395 * some task might have had allocated an extent from it already, but it
6396 * has not yet created a respective ordered extent (and added it to a
6397 * root's list of ordered extents).
6398 * Therefore wait for any task currently allocating extents, since the
6399 * block group's reservations counter is incremented while a read lock
6400 * on the groups' semaphore is held and decremented after releasing
6401 * the read access on that semaphore and creating the ordered extent.
6402 */
6403 down_write(&space_info->groups_sem);
6404 up_write(&space_info->groups_sem);
6405
6406 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6407 }
6408
6409 /**
6410 * btrfs_add_reserved_bytes - update the block_group and space info counters
6411 * @cache: The cache we are manipulating
6412 * @ram_bytes: The number of bytes of file content, and will be same to
6413 * @num_bytes except for the compress path.
6414 * @num_bytes: The number of bytes in question
6415 * @delalloc: The blocks are allocated for the delalloc write
6416 *
6417 * This is called by the allocator when it reserves space. If this is a
6418 * reservation and the block group has become read only we cannot make the
6419 * reservation and return -EAGAIN, otherwise this function always succeeds.
6420 */
6421 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6422 u64 ram_bytes, u64 num_bytes, int delalloc)
6423 {
6424 struct btrfs_space_info *space_info = cache->space_info;
6425 int ret = 0;
6426
6427 spin_lock(&space_info->lock);
6428 spin_lock(&cache->lock);
6429 if (cache->ro) {
6430 ret = -EAGAIN;
6431 } else {
6432 cache->reserved += num_bytes;
6433 space_info->bytes_reserved += num_bytes;
6434 space_info->bytes_may_use -= ram_bytes;
6435 if (delalloc)
6436 cache->delalloc_bytes += num_bytes;
6437 }
6438 spin_unlock(&cache->lock);
6439 spin_unlock(&space_info->lock);
6440 return ret;
6441 }
6442
6443 /**
6444 * btrfs_free_reserved_bytes - update the block_group and space info counters
6445 * @cache: The cache we are manipulating
6446 * @num_bytes: The number of bytes in question
6447 * @delalloc: The blocks are allocated for the delalloc write
6448 *
6449 * This is called by somebody who is freeing space that was never actually used
6450 * on disk. For example if you reserve some space for a new leaf in transaction
6451 * A and before transaction A commits you free that leaf, you call this with
6452 * reserve set to 0 in order to clear the reservation.
6453 */
6454
6455 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6456 u64 num_bytes, int delalloc)
6457 {
6458 struct btrfs_space_info *space_info = cache->space_info;
6459
6460 spin_lock(&space_info->lock);
6461 spin_lock(&cache->lock);
6462 if (cache->ro)
6463 space_info->bytes_readonly += num_bytes;
6464 cache->reserved -= num_bytes;
6465 space_info->bytes_reserved -= num_bytes;
6466 space_info->max_extent_size = 0;
6467
6468 if (delalloc)
6469 cache->delalloc_bytes -= num_bytes;
6470 spin_unlock(&cache->lock);
6471 spin_unlock(&space_info->lock);
6472 }
6473 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6474 {
6475 struct btrfs_caching_control *next;
6476 struct btrfs_caching_control *caching_ctl;
6477 struct btrfs_block_group_cache *cache;
6478
6479 down_write(&fs_info->commit_root_sem);
6480
6481 list_for_each_entry_safe(caching_ctl, next,
6482 &fs_info->caching_block_groups, list) {
6483 cache = caching_ctl->block_group;
6484 if (block_group_cache_done(cache)) {
6485 cache->last_byte_to_unpin = (u64)-1;
6486 list_del_init(&caching_ctl->list);
6487 put_caching_control(caching_ctl);
6488 } else {
6489 cache->last_byte_to_unpin = caching_ctl->progress;
6490 }
6491 }
6492
6493 if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6494 fs_info->pinned_extents = &fs_info->freed_extents[1];
6495 else
6496 fs_info->pinned_extents = &fs_info->freed_extents[0];
6497
6498 up_write(&fs_info->commit_root_sem);
6499
6500 update_global_block_rsv(fs_info);
6501 }
6502
6503 /*
6504 * Returns the free cluster for the given space info and sets empty_cluster to
6505 * what it should be based on the mount options.
6506 */
6507 static struct btrfs_free_cluster *
6508 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6509 struct btrfs_space_info *space_info, u64 *empty_cluster)
6510 {
6511 struct btrfs_free_cluster *ret = NULL;
6512
6513 *empty_cluster = 0;
6514 if (btrfs_mixed_space_info(space_info))
6515 return ret;
6516
6517 if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6518 ret = &fs_info->meta_alloc_cluster;
6519 if (btrfs_test_opt(fs_info, SSD))
6520 *empty_cluster = SZ_2M;
6521 else
6522 *empty_cluster = SZ_64K;
6523 } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6524 btrfs_test_opt(fs_info, SSD_SPREAD)) {
6525 *empty_cluster = SZ_2M;
6526 ret = &fs_info->data_alloc_cluster;
6527 }
6528
6529 return ret;
6530 }
6531
6532 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6533 u64 start, u64 end,
6534 const bool return_free_space)
6535 {
6536 struct btrfs_block_group_cache *cache = NULL;
6537 struct btrfs_space_info *space_info;
6538 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6539 struct btrfs_free_cluster *cluster = NULL;
6540 u64 len;
6541 u64 total_unpinned = 0;
6542 u64 empty_cluster = 0;
6543 bool readonly;
6544
6545 while (start <= end) {
6546 readonly = false;
6547 if (!cache ||
6548 start >= cache->key.objectid + cache->key.offset) {
6549 if (cache)
6550 btrfs_put_block_group(cache);
6551 total_unpinned = 0;
6552 cache = btrfs_lookup_block_group(fs_info, start);
6553 BUG_ON(!cache); /* Logic error */
6554
6555 cluster = fetch_cluster_info(fs_info,
6556 cache->space_info,
6557 &empty_cluster);
6558 empty_cluster <<= 1;
6559 }
6560
6561 len = cache->key.objectid + cache->key.offset - start;
6562 len = min(len, end + 1 - start);
6563
6564 if (start < cache->last_byte_to_unpin) {
6565 len = min(len, cache->last_byte_to_unpin - start);
6566 if (return_free_space)
6567 btrfs_add_free_space(cache, start, len);
6568 }
6569
6570 start += len;
6571 total_unpinned += len;
6572 space_info = cache->space_info;
6573
6574 /*
6575 * If this space cluster has been marked as fragmented and we've
6576 * unpinned enough in this block group to potentially allow a
6577 * cluster to be created inside of it go ahead and clear the
6578 * fragmented check.
6579 */
6580 if (cluster && cluster->fragmented &&
6581 total_unpinned > empty_cluster) {
6582 spin_lock(&cluster->lock);
6583 cluster->fragmented = 0;
6584 spin_unlock(&cluster->lock);
6585 }
6586
6587 spin_lock(&space_info->lock);
6588 spin_lock(&cache->lock);
6589 cache->pinned -= len;
6590 space_info->bytes_pinned -= len;
6591
6592 trace_btrfs_space_reservation(fs_info, "pinned",
6593 space_info->flags, len, 0);
6594 space_info->max_extent_size = 0;
6595 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6596 -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6597 if (cache->ro) {
6598 space_info->bytes_readonly += len;
6599 readonly = true;
6600 }
6601 spin_unlock(&cache->lock);
6602 if (!readonly && return_free_space &&
6603 global_rsv->space_info == space_info) {
6604 u64 to_add = len;
6605
6606 spin_lock(&global_rsv->lock);
6607 if (!global_rsv->full) {
6608 to_add = min(len, global_rsv->size -
6609 global_rsv->reserved);
6610 global_rsv->reserved += to_add;
6611 space_info->bytes_may_use += to_add;
6612 if (global_rsv->reserved >= global_rsv->size)
6613 global_rsv->full = 1;
6614 trace_btrfs_space_reservation(fs_info,
6615 "space_info",
6616 space_info->flags,
6617 to_add, 1);
6618 len -= to_add;
6619 }
6620 spin_unlock(&global_rsv->lock);
6621 /* Add to any tickets we may have */
6622 if (len)
6623 space_info_add_new_bytes(fs_info, space_info,
6624 len);
6625 }
6626 spin_unlock(&space_info->lock);
6627 }
6628
6629 if (cache)
6630 btrfs_put_block_group(cache);
6631 return 0;
6632 }
6633
6634 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6635 {
6636 struct btrfs_fs_info *fs_info = trans->fs_info;
6637 struct btrfs_block_group_cache *block_group, *tmp;
6638 struct list_head *deleted_bgs;
6639 struct extent_io_tree *unpin;
6640 u64 start;
6641 u64 end;
6642 int ret;
6643
6644 if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6645 unpin = &fs_info->freed_extents[1];
6646 else
6647 unpin = &fs_info->freed_extents[0];
6648
6649 while (!trans->aborted) {
6650 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6651 ret = find_first_extent_bit(unpin, 0, &start, &end,
6652 EXTENT_DIRTY, NULL);
6653 if (ret) {
6654 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6655 break;
6656 }
6657
6658 if (btrfs_test_opt(fs_info, DISCARD))
6659 ret = btrfs_discard_extent(fs_info, start,
6660 end + 1 - start, NULL);
6661
6662 clear_extent_dirty(unpin, start, end);
6663 unpin_extent_range(fs_info, start, end, true);
6664 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6665 cond_resched();
6666 }
6667
6668 /*
6669 * Transaction is finished. We don't need the lock anymore. We
6670 * do need to clean up the block groups in case of a transaction
6671 * abort.
6672 */
6673 deleted_bgs = &trans->transaction->deleted_bgs;
6674 list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6675 u64 trimmed = 0;
6676
6677 ret = -EROFS;
6678 if (!trans->aborted)
6679 ret = btrfs_discard_extent(fs_info,
6680 block_group->key.objectid,
6681 block_group->key.offset,
6682 &trimmed);
6683
6684 list_del_init(&block_group->bg_list);
6685 btrfs_put_block_group_trimming(block_group);
6686 btrfs_put_block_group(block_group);
6687
6688 if (ret) {
6689 const char *errstr = btrfs_decode_error(ret);
6690 btrfs_warn(fs_info,
6691 "discard failed while removing blockgroup: errno=%d %s",
6692 ret, errstr);
6693 }
6694 }
6695
6696 return 0;
6697 }
6698
6699 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6700 struct btrfs_delayed_ref_node *node, u64 parent,
6701 u64 root_objectid, u64 owner_objectid,
6702 u64 owner_offset, int refs_to_drop,
6703 struct btrfs_delayed_extent_op *extent_op)
6704 {
6705 struct btrfs_fs_info *info = trans->fs_info;
6706 struct btrfs_key key;
6707 struct btrfs_path *path;
6708 struct btrfs_root *extent_root = info->extent_root;
6709 struct extent_buffer *leaf;
6710 struct btrfs_extent_item *ei;
6711 struct btrfs_extent_inline_ref *iref;
6712 int ret;
6713 int is_data;
6714 int extent_slot = 0;
6715 int found_extent = 0;
6716 int num_to_del = 1;
6717 u32 item_size;
6718 u64 refs;
6719 u64 bytenr = node->bytenr;
6720 u64 num_bytes = node->num_bytes;
6721 int last_ref = 0;
6722 bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6723
6724 path = btrfs_alloc_path();
6725 if (!path)
6726 return -ENOMEM;
6727
6728 path->reada = READA_FORWARD;
6729 path->leave_spinning = 1;
6730
6731 is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6732 BUG_ON(!is_data && refs_to_drop != 1);
6733
6734 if (is_data)
6735 skinny_metadata = false;
6736
6737 ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6738 parent, root_objectid, owner_objectid,
6739 owner_offset);
6740 if (ret == 0) {
6741 extent_slot = path->slots[0];
6742 while (extent_slot >= 0) {
6743 btrfs_item_key_to_cpu(path->nodes[0], &key,
6744 extent_slot);
6745 if (key.objectid != bytenr)
6746 break;
6747 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6748 key.offset == num_bytes) {
6749 found_extent = 1;
6750 break;
6751 }
6752 if (key.type == BTRFS_METADATA_ITEM_KEY &&
6753 key.offset == owner_objectid) {
6754 found_extent = 1;
6755 break;
6756 }
6757 if (path->slots[0] - extent_slot > 5)
6758 break;
6759 extent_slot--;
6760 }
6761
6762 if (!found_extent) {
6763 BUG_ON(iref);
6764 ret = remove_extent_backref(trans, path, NULL,
6765 refs_to_drop,
6766 is_data, &last_ref);
6767 if (ret) {
6768 btrfs_abort_transaction(trans, ret);
6769 goto out;
6770 }
6771 btrfs_release_path(path);
6772 path->leave_spinning = 1;
6773
6774 key.objectid = bytenr;
6775 key.type = BTRFS_EXTENT_ITEM_KEY;
6776 key.offset = num_bytes;
6777
6778 if (!is_data && skinny_metadata) {
6779 key.type = BTRFS_METADATA_ITEM_KEY;
6780 key.offset = owner_objectid;
6781 }
6782
6783 ret = btrfs_search_slot(trans, extent_root,
6784 &key, path, -1, 1);
6785 if (ret > 0 && skinny_metadata && path->slots[0]) {
6786 /*
6787 * Couldn't find our skinny metadata item,
6788 * see if we have ye olde extent item.
6789 */
6790 path->slots[0]--;
6791 btrfs_item_key_to_cpu(path->nodes[0], &key,
6792 path->slots[0]);
6793 if (key.objectid == bytenr &&
6794 key.type == BTRFS_EXTENT_ITEM_KEY &&
6795 key.offset == num_bytes)
6796 ret = 0;
6797 }
6798
6799 if (ret > 0 && skinny_metadata) {
6800 skinny_metadata = false;
6801 key.objectid = bytenr;
6802 key.type = BTRFS_EXTENT_ITEM_KEY;
6803 key.offset = num_bytes;
6804 btrfs_release_path(path);
6805 ret = btrfs_search_slot(trans, extent_root,
6806 &key, path, -1, 1);
6807 }
6808
6809 if (ret) {
6810 btrfs_err(info,
6811 "umm, got %d back from search, was looking for %llu",
6812 ret, bytenr);
6813 if (ret > 0)
6814 btrfs_print_leaf(path->nodes[0]);
6815 }
6816 if (ret < 0) {
6817 btrfs_abort_transaction(trans, ret);
6818 goto out;
6819 }
6820 extent_slot = path->slots[0];
6821 }
6822 } else if (WARN_ON(ret == -ENOENT)) {
6823 btrfs_print_leaf(path->nodes[0]);
6824 btrfs_err(info,
6825 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu",
6826 bytenr, parent, root_objectid, owner_objectid,
6827 owner_offset);
6828 btrfs_abort_transaction(trans, ret);
6829 goto out;
6830 } else {
6831 btrfs_abort_transaction(trans, ret);
6832 goto out;
6833 }
6834
6835 leaf = path->nodes[0];
6836 item_size = btrfs_item_size_nr(leaf, extent_slot);
6837 if (unlikely(item_size < sizeof(*ei))) {
6838 ret = -EINVAL;
6839 btrfs_print_v0_err(info);
6840 btrfs_abort_transaction(trans, ret);
6841 goto out;
6842 }
6843 ei = btrfs_item_ptr(leaf, extent_slot,
6844 struct btrfs_extent_item);
6845 if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6846 key.type == BTRFS_EXTENT_ITEM_KEY) {
6847 struct btrfs_tree_block_info *bi;
6848 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6849 bi = (struct btrfs_tree_block_info *)(ei + 1);
6850 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6851 }
6852
6853 refs = btrfs_extent_refs(leaf, ei);
6854 if (refs < refs_to_drop) {
6855 btrfs_err(info,
6856 "trying to drop %d refs but we only have %Lu for bytenr %Lu",
6857 refs_to_drop, refs, bytenr);
6858 ret = -EINVAL;
6859 btrfs_abort_transaction(trans, ret);
6860 goto out;
6861 }
6862 refs -= refs_to_drop;
6863
6864 if (refs > 0) {
6865 if (extent_op)
6866 __run_delayed_extent_op(extent_op, leaf, ei);
6867 /*
6868 * In the case of inline back ref, reference count will
6869 * be updated by remove_extent_backref
6870 */
6871 if (iref) {
6872 BUG_ON(!found_extent);
6873 } else {
6874 btrfs_set_extent_refs(leaf, ei, refs);
6875 btrfs_mark_buffer_dirty(leaf);
6876 }
6877 if (found_extent) {
6878 ret = remove_extent_backref(trans, path, iref,
6879 refs_to_drop, is_data,
6880 &last_ref);
6881 if (ret) {
6882 btrfs_abort_transaction(trans, ret);
6883 goto out;
6884 }
6885 }
6886 } else {
6887 if (found_extent) {
6888 BUG_ON(is_data && refs_to_drop !=
6889 extent_data_ref_count(path, iref));
6890 if (iref) {
6891 BUG_ON(path->slots[0] != extent_slot);
6892 } else {
6893 BUG_ON(path->slots[0] != extent_slot + 1);
6894 path->slots[0] = extent_slot;
6895 num_to_del = 2;
6896 }
6897 }
6898
6899 last_ref = 1;
6900 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6901 num_to_del);
6902 if (ret) {
6903 btrfs_abort_transaction(trans, ret);
6904 goto out;
6905 }
6906 btrfs_release_path(path);
6907
6908 if (is_data) {
6909 ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
6910 if (ret) {
6911 btrfs_abort_transaction(trans, ret);
6912 goto out;
6913 }
6914 }
6915
6916 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
6917 if (ret) {
6918 btrfs_abort_transaction(trans, ret);
6919 goto out;
6920 }
6921
6922 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
6923 if (ret) {
6924 btrfs_abort_transaction(trans, ret);
6925 goto out;
6926 }
6927 }
6928 btrfs_release_path(path);
6929
6930 out:
6931 btrfs_free_path(path);
6932 return ret;
6933 }
6934
6935 /*
6936 * when we free an block, it is possible (and likely) that we free the last
6937 * delayed ref for that extent as well. This searches the delayed ref tree for
6938 * a given extent, and if there are no other delayed refs to be processed, it
6939 * removes it from the tree.
6940 */
6941 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6942 u64 bytenr)
6943 {
6944 struct btrfs_delayed_ref_head *head;
6945 struct btrfs_delayed_ref_root *delayed_refs;
6946 int ret = 0;
6947
6948 delayed_refs = &trans->transaction->delayed_refs;
6949 spin_lock(&delayed_refs->lock);
6950 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
6951 if (!head)
6952 goto out_delayed_unlock;
6953
6954 spin_lock(&head->lock);
6955 if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
6956 goto out;
6957
6958 if (head->extent_op) {
6959 if (!head->must_insert_reserved)
6960 goto out;
6961 btrfs_free_delayed_extent_op(head->extent_op);
6962 head->extent_op = NULL;
6963 }
6964
6965 /*
6966 * waiting for the lock here would deadlock. If someone else has it
6967 * locked they are already in the process of dropping it anyway
6968 */
6969 if (!mutex_trylock(&head->mutex))
6970 goto out;
6971
6972 /*
6973 * at this point we have a head with no other entries. Go
6974 * ahead and process it.
6975 */
6976 rb_erase_cached(&head->href_node, &delayed_refs->href_root);
6977 RB_CLEAR_NODE(&head->href_node);
6978 atomic_dec(&delayed_refs->num_entries);
6979
6980 /*
6981 * we don't take a ref on the node because we're removing it from the
6982 * tree, so we just steal the ref the tree was holding.
6983 */
6984 delayed_refs->num_heads--;
6985 if (head->processing == 0)
6986 delayed_refs->num_heads_ready--;
6987 head->processing = 0;
6988 spin_unlock(&head->lock);
6989 spin_unlock(&delayed_refs->lock);
6990
6991 BUG_ON(head->extent_op);
6992 if (head->must_insert_reserved)
6993 ret = 1;
6994
6995 mutex_unlock(&head->mutex);
6996 btrfs_put_delayed_ref_head(head);
6997 return ret;
6998 out:
6999 spin_unlock(&head->lock);
7000
7001 out_delayed_unlock:
7002 spin_unlock(&delayed_refs->lock);
7003 return 0;
7004 }
7005
7006 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7007 struct btrfs_root *root,
7008 struct extent_buffer *buf,
7009 u64 parent, int last_ref)
7010 {
7011 struct btrfs_fs_info *fs_info = root->fs_info;
7012 int pin = 1;
7013 int ret;
7014
7015 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7016 int old_ref_mod, new_ref_mod;
7017
7018 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7019 root->root_key.objectid,
7020 btrfs_header_level(buf), 0,
7021 BTRFS_DROP_DELAYED_REF);
7022 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7023 buf->len, parent,
7024 root->root_key.objectid,
7025 btrfs_header_level(buf),
7026 BTRFS_DROP_DELAYED_REF, NULL,
7027 &old_ref_mod, &new_ref_mod);
7028 BUG_ON(ret); /* -ENOMEM */
7029 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7030 }
7031
7032 if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7033 struct btrfs_block_group_cache *cache;
7034
7035 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7036 ret = check_ref_cleanup(trans, buf->start);
7037 if (!ret)
7038 goto out;
7039 }
7040
7041 pin = 0;
7042 cache = btrfs_lookup_block_group(fs_info, buf->start);
7043
7044 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7045 pin_down_extent(fs_info, cache, buf->start,
7046 buf->len, 1);
7047 btrfs_put_block_group(cache);
7048 goto out;
7049 }
7050
7051 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7052
7053 btrfs_add_free_space(cache, buf->start, buf->len);
7054 btrfs_free_reserved_bytes(cache, buf->len, 0);
7055 btrfs_put_block_group(cache);
7056 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7057 }
7058 out:
7059 if (pin)
7060 add_pinned_bytes(fs_info, buf->len, true,
7061 root->root_key.objectid);
7062
7063 if (last_ref) {
7064 /*
7065 * Deleting the buffer, clear the corrupt flag since it doesn't
7066 * matter anymore.
7067 */
7068 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7069 }
7070 }
7071
7072 /* Can return -ENOMEM */
7073 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7074 struct btrfs_root *root,
7075 u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7076 u64 owner, u64 offset)
7077 {
7078 struct btrfs_fs_info *fs_info = root->fs_info;
7079 int old_ref_mod, new_ref_mod;
7080 int ret;
7081
7082 if (btrfs_is_testing(fs_info))
7083 return 0;
7084
7085 if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7086 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7087 root_objectid, owner, offset,
7088 BTRFS_DROP_DELAYED_REF);
7089
7090 /*
7091 * tree log blocks never actually go into the extent allocation
7092 * tree, just update pinning info and exit early.
7093 */
7094 if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7095 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7096 /* unlocks the pinned mutex */
7097 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7098 old_ref_mod = new_ref_mod = 0;
7099 ret = 0;
7100 } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7101 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7102 num_bytes, parent,
7103 root_objectid, (int)owner,
7104 BTRFS_DROP_DELAYED_REF, NULL,
7105 &old_ref_mod, &new_ref_mod);
7106 } else {
7107 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7108 num_bytes, parent,
7109 root_objectid, owner, offset,
7110 0, BTRFS_DROP_DELAYED_REF,
7111 &old_ref_mod, &new_ref_mod);
7112 }
7113
7114 if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7115 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7116
7117 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7118 }
7119
7120 return ret;
7121 }
7122
7123 /*
7124 * when we wait for progress in the block group caching, its because
7125 * our allocation attempt failed at least once. So, we must sleep
7126 * and let some progress happen before we try again.
7127 *
7128 * This function will sleep at least once waiting for new free space to
7129 * show up, and then it will check the block group free space numbers
7130 * for our min num_bytes. Another option is to have it go ahead
7131 * and look in the rbtree for a free extent of a given size, but this
7132 * is a good start.
7133 *
7134 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7135 * any of the information in this block group.
7136 */
7137 static noinline void
7138 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7139 u64 num_bytes)
7140 {
7141 struct btrfs_caching_control *caching_ctl;
7142
7143 caching_ctl = get_caching_control(cache);
7144 if (!caching_ctl)
7145 return;
7146
7147 wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7148 (cache->free_space_ctl->free_space >= num_bytes));
7149
7150 put_caching_control(caching_ctl);
7151 }
7152
7153 static noinline int
7154 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7155 {
7156 struct btrfs_caching_control *caching_ctl;
7157 int ret = 0;
7158
7159 caching_ctl = get_caching_control(cache);
7160 if (!caching_ctl)
7161 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7162
7163 wait_event(caching_ctl->wait, block_group_cache_done(cache));
7164 if (cache->cached == BTRFS_CACHE_ERROR)
7165 ret = -EIO;
7166 put_caching_control(caching_ctl);
7167 return ret;
7168 }
7169
7170 enum btrfs_loop_type {
7171 LOOP_CACHING_NOWAIT = 0,
7172 LOOP_CACHING_WAIT = 1,
7173 LOOP_ALLOC_CHUNK = 2,
7174 LOOP_NO_EMPTY_SIZE = 3,
7175 };
7176
7177 static inline void
7178 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7179 int delalloc)
7180 {
7181 if (delalloc)
7182 down_read(&cache->data_rwsem);
7183 }
7184
7185 static inline void
7186 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7187 int delalloc)
7188 {
7189 btrfs_get_block_group(cache);
7190 if (delalloc)
7191 down_read(&cache->data_rwsem);
7192 }
7193
7194 static struct btrfs_block_group_cache *
7195 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7196 struct btrfs_free_cluster *cluster,
7197 int delalloc)
7198 {
7199 struct btrfs_block_group_cache *used_bg = NULL;
7200
7201 spin_lock(&cluster->refill_lock);
7202 while (1) {
7203 used_bg = cluster->block_group;
7204 if (!used_bg)
7205 return NULL;
7206
7207 if (used_bg == block_group)
7208 return used_bg;
7209
7210 btrfs_get_block_group(used_bg);
7211
7212 if (!delalloc)
7213 return used_bg;
7214
7215 if (down_read_trylock(&used_bg->data_rwsem))
7216 return used_bg;
7217
7218 spin_unlock(&cluster->refill_lock);
7219
7220 /* We should only have one-level nested. */
7221 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7222
7223 spin_lock(&cluster->refill_lock);
7224 if (used_bg == cluster->block_group)
7225 return used_bg;
7226
7227 up_read(&used_bg->data_rwsem);
7228 btrfs_put_block_group(used_bg);
7229 }
7230 }
7231
7232 static inline void
7233 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7234 int delalloc)
7235 {
7236 if (delalloc)
7237 up_read(&cache->data_rwsem);
7238 btrfs_put_block_group(cache);
7239 }
7240
7241 /*
7242 * walks the btree of allocated extents and find a hole of a given size.
7243 * The key ins is changed to record the hole:
7244 * ins->objectid == start position
7245 * ins->flags = BTRFS_EXTENT_ITEM_KEY
7246 * ins->offset == the size of the hole.
7247 * Any available blocks before search_start are skipped.
7248 *
7249 * If there is no suitable free space, we will record the max size of
7250 * the free space extent currently.
7251 */
7252 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7253 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7254 u64 hint_byte, struct btrfs_key *ins,
7255 u64 flags, int delalloc)
7256 {
7257 int ret = 0;
7258 struct btrfs_root *root = fs_info->extent_root;
7259 struct btrfs_free_cluster *last_ptr = NULL;
7260 struct btrfs_block_group_cache *block_group = NULL;
7261 u64 search_start = 0;
7262 u64 max_extent_size = 0;
7263 u64 max_free_space = 0;
7264 u64 empty_cluster = 0;
7265 struct btrfs_space_info *space_info;
7266 int loop = 0;
7267 int index = btrfs_bg_flags_to_raid_index(flags);
7268 bool failed_cluster_refill = false;
7269 bool failed_alloc = false;
7270 bool use_cluster = true;
7271 bool have_caching_bg = false;
7272 bool orig_have_caching_bg = false;
7273 bool full_search = false;
7274
7275 WARN_ON(num_bytes < fs_info->sectorsize);
7276 ins->type = BTRFS_EXTENT_ITEM_KEY;
7277 ins->objectid = 0;
7278 ins->offset = 0;
7279
7280 trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7281
7282 space_info = __find_space_info(fs_info, flags);
7283 if (!space_info) {
7284 btrfs_err(fs_info, "No space info for %llu", flags);
7285 return -ENOSPC;
7286 }
7287
7288 /*
7289 * If our free space is heavily fragmented we may not be able to make
7290 * big contiguous allocations, so instead of doing the expensive search
7291 * for free space, simply return ENOSPC with our max_extent_size so we
7292 * can go ahead and search for a more manageable chunk.
7293 *
7294 * If our max_extent_size is large enough for our allocation simply
7295 * disable clustering since we will likely not be able to find enough
7296 * space to create a cluster and induce latency trying.
7297 */
7298 if (unlikely(space_info->max_extent_size)) {
7299 spin_lock(&space_info->lock);
7300 if (space_info->max_extent_size &&
7301 num_bytes > space_info->max_extent_size) {
7302 ins->offset = space_info->max_extent_size;
7303 spin_unlock(&space_info->lock);
7304 return -ENOSPC;
7305 } else if (space_info->max_extent_size) {
7306 use_cluster = false;
7307 }
7308 spin_unlock(&space_info->lock);
7309 }
7310
7311 last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7312 if (last_ptr) {
7313 spin_lock(&last_ptr->lock);
7314 if (last_ptr->block_group)
7315 hint_byte = last_ptr->window_start;
7316 if (last_ptr->fragmented) {
7317 /*
7318 * We still set window_start so we can keep track of the
7319 * last place we found an allocation to try and save
7320 * some time.
7321 */
7322 hint_byte = last_ptr->window_start;
7323 use_cluster = false;
7324 }
7325 spin_unlock(&last_ptr->lock);
7326 }
7327
7328 search_start = max(search_start, first_logical_byte(fs_info, 0));
7329 search_start = max(search_start, hint_byte);
7330 if (search_start == hint_byte) {
7331 block_group = btrfs_lookup_block_group(fs_info, search_start);
7332 /*
7333 * we don't want to use the block group if it doesn't match our
7334 * allocation bits, or if its not cached.
7335 *
7336 * However if we are re-searching with an ideal block group
7337 * picked out then we don't care that the block group is cached.
7338 */
7339 if (block_group && block_group_bits(block_group, flags) &&
7340 block_group->cached != BTRFS_CACHE_NO) {
7341 down_read(&space_info->groups_sem);
7342 if (list_empty(&block_group->list) ||
7343 block_group->ro) {
7344 /*
7345 * someone is removing this block group,
7346 * we can't jump into the have_block_group
7347 * target because our list pointers are not
7348 * valid
7349 */
7350 btrfs_put_block_group(block_group);
7351 up_read(&space_info->groups_sem);
7352 } else {
7353 index = btrfs_bg_flags_to_raid_index(
7354 block_group->flags);
7355 btrfs_lock_block_group(block_group, delalloc);
7356 goto have_block_group;
7357 }
7358 } else if (block_group) {
7359 btrfs_put_block_group(block_group);
7360 }
7361 }
7362 search:
7363 have_caching_bg = false;
7364 if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7365 full_search = true;
7366 down_read(&space_info->groups_sem);
7367 list_for_each_entry(block_group, &space_info->block_groups[index],
7368 list) {
7369 u64 offset;
7370 int cached;
7371
7372 /* If the block group is read-only, we can skip it entirely. */
7373 if (unlikely(block_group->ro))
7374 continue;
7375
7376 btrfs_grab_block_group(block_group, delalloc);
7377 search_start = block_group->key.objectid;
7378
7379 /*
7380 * this can happen if we end up cycling through all the
7381 * raid types, but we want to make sure we only allocate
7382 * for the proper type.
7383 */
7384 if (!block_group_bits(block_group, flags)) {
7385 u64 extra = BTRFS_BLOCK_GROUP_DUP |
7386 BTRFS_BLOCK_GROUP_RAID1 |
7387 BTRFS_BLOCK_GROUP_RAID5 |
7388 BTRFS_BLOCK_GROUP_RAID6 |
7389 BTRFS_BLOCK_GROUP_RAID10;
7390
7391 /*
7392 * if they asked for extra copies and this block group
7393 * doesn't provide them, bail. This does allow us to
7394 * fill raid0 from raid1.
7395 */
7396 if ((flags & extra) && !(block_group->flags & extra))
7397 goto loop;
7398 }
7399
7400 have_block_group:
7401 cached = block_group_cache_done(block_group);
7402 if (unlikely(!cached)) {
7403 have_caching_bg = true;
7404 ret = cache_block_group(block_group, 0);
7405 BUG_ON(ret < 0);
7406 ret = 0;
7407 }
7408
7409 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7410 goto loop;
7411
7412 /*
7413 * Ok we want to try and use the cluster allocator, so
7414 * lets look there
7415 */
7416 if (last_ptr && use_cluster) {
7417 struct btrfs_block_group_cache *used_block_group;
7418 unsigned long aligned_cluster;
7419 /*
7420 * the refill lock keeps out other
7421 * people trying to start a new cluster
7422 */
7423 used_block_group = btrfs_lock_cluster(block_group,
7424 last_ptr,
7425 delalloc);
7426 if (!used_block_group)
7427 goto refill_cluster;
7428
7429 if (used_block_group != block_group &&
7430 (used_block_group->ro ||
7431 !block_group_bits(used_block_group, flags)))
7432 goto release_cluster;
7433
7434 offset = btrfs_alloc_from_cluster(used_block_group,
7435 last_ptr,
7436 num_bytes,
7437 used_block_group->key.objectid,
7438 &max_extent_size);
7439 if (offset) {
7440 /* we have a block, we're done */
7441 spin_unlock(&last_ptr->refill_lock);
7442 trace_btrfs_reserve_extent_cluster(
7443 used_block_group,
7444 search_start, num_bytes);
7445 if (used_block_group != block_group) {
7446 btrfs_release_block_group(block_group,
7447 delalloc);
7448 block_group = used_block_group;
7449 }
7450 goto checks;
7451 }
7452
7453 WARN_ON(last_ptr->block_group != used_block_group);
7454 release_cluster:
7455 /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7456 * set up a new clusters, so lets just skip it
7457 * and let the allocator find whatever block
7458 * it can find. If we reach this point, we
7459 * will have tried the cluster allocator
7460 * plenty of times and not have found
7461 * anything, so we are likely way too
7462 * fragmented for the clustering stuff to find
7463 * anything.
7464 *
7465 * However, if the cluster is taken from the
7466 * current block group, release the cluster
7467 * first, so that we stand a better chance of
7468 * succeeding in the unclustered
7469 * allocation. */
7470 if (loop >= LOOP_NO_EMPTY_SIZE &&
7471 used_block_group != block_group) {
7472 spin_unlock(&last_ptr->refill_lock);
7473 btrfs_release_block_group(used_block_group,
7474 delalloc);
7475 goto unclustered_alloc;
7476 }
7477
7478 /*
7479 * this cluster didn't work out, free it and
7480 * start over
7481 */
7482 btrfs_return_cluster_to_free_space(NULL, last_ptr);
7483
7484 if (used_block_group != block_group)
7485 btrfs_release_block_group(used_block_group,
7486 delalloc);
7487 refill_cluster:
7488 if (loop >= LOOP_NO_EMPTY_SIZE) {
7489 spin_unlock(&last_ptr->refill_lock);
7490 goto unclustered_alloc;
7491 }
7492
7493 aligned_cluster = max_t(unsigned long,
7494 empty_cluster + empty_size,
7495 block_group->full_stripe_len);
7496
7497 /* allocate a cluster in this block group */
7498 ret = btrfs_find_space_cluster(fs_info, block_group,
7499 last_ptr, search_start,
7500 num_bytes,
7501 aligned_cluster);
7502 if (ret == 0) {
7503 /*
7504 * now pull our allocation out of this
7505 * cluster
7506 */
7507 offset = btrfs_alloc_from_cluster(block_group,
7508 last_ptr,
7509 num_bytes,
7510 search_start,
7511 &max_extent_size);
7512 if (offset) {
7513 /* we found one, proceed */
7514 spin_unlock(&last_ptr->refill_lock);
7515 trace_btrfs_reserve_extent_cluster(
7516 block_group, search_start,
7517 num_bytes);
7518 goto checks;
7519 }
7520 } else if (!cached && loop > LOOP_CACHING_NOWAIT
7521 && !failed_cluster_refill) {
7522 spin_unlock(&last_ptr->refill_lock);
7523
7524 failed_cluster_refill = true;
7525 wait_block_group_cache_progress(block_group,
7526 num_bytes + empty_cluster + empty_size);
7527 goto have_block_group;
7528 }
7529
7530 /*
7531 * at this point we either didn't find a cluster
7532 * or we weren't able to allocate a block from our
7533 * cluster. Free the cluster we've been trying
7534 * to use, and go to the next block group
7535 */
7536 btrfs_return_cluster_to_free_space(NULL, last_ptr);
7537 spin_unlock(&last_ptr->refill_lock);
7538 goto loop;
7539 }
7540
7541 unclustered_alloc:
7542 /*
7543 * We are doing an unclustered alloc, set the fragmented flag so
7544 * we don't bother trying to setup a cluster again until we get
7545 * more space.
7546 */
7547 if (unlikely(last_ptr)) {
7548 spin_lock(&last_ptr->lock);
7549 last_ptr->fragmented = 1;
7550 spin_unlock(&last_ptr->lock);
7551 }
7552 if (cached) {
7553 struct btrfs_free_space_ctl *ctl =
7554 block_group->free_space_ctl;
7555
7556 spin_lock(&ctl->tree_lock);
7557 if (ctl->free_space <
7558 num_bytes + empty_cluster + empty_size) {
7559 max_free_space = max(max_free_space,
7560 ctl->free_space);
7561 spin_unlock(&ctl->tree_lock);
7562 goto loop;
7563 }
7564 spin_unlock(&ctl->tree_lock);
7565 }
7566
7567 offset = btrfs_find_space_for_alloc(block_group, search_start,
7568 num_bytes, empty_size,
7569 &max_extent_size);
7570 /*
7571 * If we didn't find a chunk, and we haven't failed on this
7572 * block group before, and this block group is in the middle of
7573 * caching and we are ok with waiting, then go ahead and wait
7574 * for progress to be made, and set failed_alloc to true.
7575 *
7576 * If failed_alloc is true then we've already waited on this
7577 * block group once and should move on to the next block group.
7578 */
7579 if (!offset && !failed_alloc && !cached &&
7580 loop > LOOP_CACHING_NOWAIT) {
7581 wait_block_group_cache_progress(block_group,
7582 num_bytes + empty_size);
7583 failed_alloc = true;
7584 goto have_block_group;
7585 } else if (!offset) {
7586 goto loop;
7587 }
7588 checks:
7589 search_start = round_up(offset, fs_info->stripesize);
7590
7591 /* move on to the next group */
7592 if (search_start + num_bytes >
7593 block_group->key.objectid + block_group->key.offset) {
7594 btrfs_add_free_space(block_group, offset, num_bytes);
7595 goto loop;
7596 }
7597
7598 if (offset < search_start)
7599 btrfs_add_free_space(block_group, offset,
7600 search_start - offset);
7601
7602 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7603 num_bytes, delalloc);
7604 if (ret == -EAGAIN) {
7605 btrfs_add_free_space(block_group, offset, num_bytes);
7606 goto loop;
7607 }
7608 btrfs_inc_block_group_reservations(block_group);
7609
7610 /* we are all good, lets return */
7611 ins->objectid = search_start;
7612 ins->offset = num_bytes;
7613
7614 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7615 btrfs_release_block_group(block_group, delalloc);
7616 break;
7617 loop:
7618 failed_cluster_refill = false;
7619 failed_alloc = false;
7620 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7621 index);
7622 btrfs_release_block_group(block_group, delalloc);
7623 cond_resched();
7624 }
7625 up_read(&space_info->groups_sem);
7626
7627 if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7628 && !orig_have_caching_bg)
7629 orig_have_caching_bg = true;
7630
7631 if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7632 goto search;
7633
7634 if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7635 goto search;
7636
7637 /*
7638 * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7639 * caching kthreads as we move along
7640 * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7641 * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7642 * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7643 * again
7644 */
7645 if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7646 index = 0;
7647 if (loop == LOOP_CACHING_NOWAIT) {
7648 /*
7649 * We want to skip the LOOP_CACHING_WAIT step if we
7650 * don't have any uncached bgs and we've already done a
7651 * full search through.
7652 */
7653 if (orig_have_caching_bg || !full_search)
7654 loop = LOOP_CACHING_WAIT;
7655 else
7656 loop = LOOP_ALLOC_CHUNK;
7657 } else {
7658 loop++;
7659 }
7660
7661 if (loop == LOOP_ALLOC_CHUNK) {
7662 struct btrfs_trans_handle *trans;
7663 int exist = 0;
7664
7665 trans = current->journal_info;
7666 if (trans)
7667 exist = 1;
7668 else
7669 trans = btrfs_join_transaction(root);
7670
7671 if (IS_ERR(trans)) {
7672 ret = PTR_ERR(trans);
7673 goto out;
7674 }
7675
7676 ret = do_chunk_alloc(trans, flags, CHUNK_ALLOC_FORCE);
7677
7678 /*
7679 * If we can't allocate a new chunk we've already looped
7680 * through at least once, move on to the NO_EMPTY_SIZE
7681 * case.
7682 */
7683 if (ret == -ENOSPC)
7684 loop = LOOP_NO_EMPTY_SIZE;
7685
7686 /*
7687 * Do not bail out on ENOSPC since we
7688 * can do more things.
7689 */
7690 if (ret < 0 && ret != -ENOSPC)
7691 btrfs_abort_transaction(trans, ret);
7692 else
7693 ret = 0;
7694 if (!exist)
7695 btrfs_end_transaction(trans);
7696 if (ret)
7697 goto out;
7698 }
7699
7700 if (loop == LOOP_NO_EMPTY_SIZE) {
7701 /*
7702 * Don't loop again if we already have no empty_size and
7703 * no empty_cluster.
7704 */
7705 if (empty_size == 0 &&
7706 empty_cluster == 0) {
7707 ret = -ENOSPC;
7708 goto out;
7709 }
7710 empty_size = 0;
7711 empty_cluster = 0;
7712 }
7713
7714 goto search;
7715 } else if (!ins->objectid) {
7716 ret = -ENOSPC;
7717 } else if (ins->objectid) {
7718 if (!use_cluster && last_ptr) {
7719 spin_lock(&last_ptr->lock);
7720 last_ptr->window_start = ins->objectid;
7721 spin_unlock(&last_ptr->lock);
7722 }
7723 ret = 0;
7724 }
7725 out:
7726 if (ret == -ENOSPC) {
7727 if (!max_extent_size)
7728 max_extent_size = max_free_space;
7729 spin_lock(&space_info->lock);
7730 space_info->max_extent_size = max_extent_size;
7731 spin_unlock(&space_info->lock);
7732 ins->offset = max_extent_size;
7733 }
7734 return ret;
7735 }
7736
7737 static void dump_space_info(struct btrfs_fs_info *fs_info,
7738 struct btrfs_space_info *info, u64 bytes,
7739 int dump_block_groups)
7740 {
7741 struct btrfs_block_group_cache *cache;
7742 int index = 0;
7743
7744 spin_lock(&info->lock);
7745 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7746 info->flags,
7747 info->total_bytes - btrfs_space_info_used(info, true),
7748 info->full ? "" : "not ");
7749 btrfs_info(fs_info,
7750 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7751 info->total_bytes, info->bytes_used, info->bytes_pinned,
7752 info->bytes_reserved, info->bytes_may_use,
7753 info->bytes_readonly);
7754 spin_unlock(&info->lock);
7755
7756 if (!dump_block_groups)
7757 return;
7758
7759 down_read(&info->groups_sem);
7760 again:
7761 list_for_each_entry(cache, &info->block_groups[index], list) {
7762 spin_lock(&cache->lock);
7763 btrfs_info(fs_info,
7764 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7765 cache->key.objectid, cache->key.offset,
7766 btrfs_block_group_used(&cache->item), cache->pinned,
7767 cache->reserved, cache->ro ? "[readonly]" : "");
7768 btrfs_dump_free_space(cache, bytes);
7769 spin_unlock(&cache->lock);
7770 }
7771 if (++index < BTRFS_NR_RAID_TYPES)
7772 goto again;
7773 up_read(&info->groups_sem);
7774 }
7775
7776 /*
7777 * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7778 * hole that is at least as big as @num_bytes.
7779 *
7780 * @root - The root that will contain this extent
7781 *
7782 * @ram_bytes - The amount of space in ram that @num_bytes take. This
7783 * is used for accounting purposes. This value differs
7784 * from @num_bytes only in the case of compressed extents.
7785 *
7786 * @num_bytes - Number of bytes to allocate on-disk.
7787 *
7788 * @min_alloc_size - Indicates the minimum amount of space that the
7789 * allocator should try to satisfy. In some cases
7790 * @num_bytes may be larger than what is required and if
7791 * the filesystem is fragmented then allocation fails.
7792 * However, the presence of @min_alloc_size gives a
7793 * chance to try and satisfy the smaller allocation.
7794 *
7795 * @empty_size - A hint that you plan on doing more COW. This is the
7796 * size in bytes the allocator should try to find free
7797 * next to the block it returns. This is just a hint and
7798 * may be ignored by the allocator.
7799 *
7800 * @hint_byte - Hint to the allocator to start searching above the byte
7801 * address passed. It might be ignored.
7802 *
7803 * @ins - This key is modified to record the found hole. It will
7804 * have the following values:
7805 * ins->objectid == start position
7806 * ins->flags = BTRFS_EXTENT_ITEM_KEY
7807 * ins->offset == the size of the hole.
7808 *
7809 * @is_data - Boolean flag indicating whether an extent is
7810 * allocated for data (true) or metadata (false)
7811 *
7812 * @delalloc - Boolean flag indicating whether this allocation is for
7813 * delalloc or not. If 'true' data_rwsem of block groups
7814 * is going to be acquired.
7815 *
7816 *
7817 * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7818 * case -ENOSPC is returned then @ins->offset will contain the size of the
7819 * largest available hole the allocator managed to find.
7820 */
7821 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7822 u64 num_bytes, u64 min_alloc_size,
7823 u64 empty_size, u64 hint_byte,
7824 struct btrfs_key *ins, int is_data, int delalloc)
7825 {
7826 struct btrfs_fs_info *fs_info = root->fs_info;
7827 bool final_tried = num_bytes == min_alloc_size;
7828 u64 flags;
7829 int ret;
7830
7831 flags = get_alloc_profile_by_root(root, is_data);
7832 again:
7833 WARN_ON(num_bytes < fs_info->sectorsize);
7834 ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7835 hint_byte, ins, flags, delalloc);
7836 if (!ret && !is_data) {
7837 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7838 } else if (ret == -ENOSPC) {
7839 if (!final_tried && ins->offset) {
7840 num_bytes = min(num_bytes >> 1, ins->offset);
7841 num_bytes = round_down(num_bytes,
7842 fs_info->sectorsize);
7843 num_bytes = max(num_bytes, min_alloc_size);
7844 ram_bytes = num_bytes;
7845 if (num_bytes == min_alloc_size)
7846 final_tried = true;
7847 goto again;
7848 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7849 struct btrfs_space_info *sinfo;
7850
7851 sinfo = __find_space_info(fs_info, flags);
7852 btrfs_err(fs_info,
7853 "allocation failed flags %llu, wanted %llu",
7854 flags, num_bytes);
7855 if (sinfo)
7856 dump_space_info(fs_info, sinfo, num_bytes, 1);
7857 }
7858 }
7859
7860 return ret;
7861 }
7862
7863 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7864 u64 start, u64 len,
7865 int pin, int delalloc)
7866 {
7867 struct btrfs_block_group_cache *cache;
7868 int ret = 0;
7869
7870 cache = btrfs_lookup_block_group(fs_info, start);
7871 if (!cache) {
7872 btrfs_err(fs_info, "Unable to find block group for %llu",
7873 start);
7874 return -ENOSPC;
7875 }
7876
7877 if (pin)
7878 pin_down_extent(fs_info, cache, start, len, 1);
7879 else {
7880 if (btrfs_test_opt(fs_info, DISCARD))
7881 ret = btrfs_discard_extent(fs_info, start, len, NULL);
7882 btrfs_add_free_space(cache, start, len);
7883 btrfs_free_reserved_bytes(cache, len, delalloc);
7884 trace_btrfs_reserved_extent_free(fs_info, start, len);
7885 }
7886
7887 btrfs_put_block_group(cache);
7888 return ret;
7889 }
7890
7891 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7892 u64 start, u64 len, int delalloc)
7893 {
7894 return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
7895 }
7896
7897 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
7898 u64 start, u64 len)
7899 {
7900 return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
7901 }
7902
7903 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7904 u64 parent, u64 root_objectid,
7905 u64 flags, u64 owner, u64 offset,
7906 struct btrfs_key *ins, int ref_mod)
7907 {
7908 struct btrfs_fs_info *fs_info = trans->fs_info;
7909 int ret;
7910 struct btrfs_extent_item *extent_item;
7911 struct btrfs_extent_inline_ref *iref;
7912 struct btrfs_path *path;
7913 struct extent_buffer *leaf;
7914 int type;
7915 u32 size;
7916
7917 if (parent > 0)
7918 type = BTRFS_SHARED_DATA_REF_KEY;
7919 else
7920 type = BTRFS_EXTENT_DATA_REF_KEY;
7921
7922 size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7923
7924 path = btrfs_alloc_path();
7925 if (!path)
7926 return -ENOMEM;
7927
7928 path->leave_spinning = 1;
7929 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7930 ins, size);
7931 if (ret) {
7932 btrfs_free_path(path);
7933 return ret;
7934 }
7935
7936 leaf = path->nodes[0];
7937 extent_item = btrfs_item_ptr(leaf, path->slots[0],
7938 struct btrfs_extent_item);
7939 btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7940 btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7941 btrfs_set_extent_flags(leaf, extent_item,
7942 flags | BTRFS_EXTENT_FLAG_DATA);
7943
7944 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7945 btrfs_set_extent_inline_ref_type(leaf, iref, type);
7946 if (parent > 0) {
7947 struct btrfs_shared_data_ref *ref;
7948 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7949 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7950 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7951 } else {
7952 struct btrfs_extent_data_ref *ref;
7953 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7954 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7955 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7956 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7957 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7958 }
7959
7960 btrfs_mark_buffer_dirty(path->nodes[0]);
7961 btrfs_free_path(path);
7962
7963 ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
7964 if (ret)
7965 return ret;
7966
7967 ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
7968 if (ret) { /* -ENOENT, logic error */
7969 btrfs_err(fs_info, "update block group failed for %llu %llu",
7970 ins->objectid, ins->offset);
7971 BUG();
7972 }
7973 trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
7974 return ret;
7975 }
7976
7977 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7978 struct btrfs_delayed_ref_node *node,
7979 struct btrfs_delayed_extent_op *extent_op)
7980 {
7981 struct btrfs_fs_info *fs_info = trans->fs_info;
7982 int ret;
7983 struct btrfs_extent_item *extent_item;
7984 struct btrfs_key extent_key;
7985 struct btrfs_tree_block_info *block_info;
7986 struct btrfs_extent_inline_ref *iref;
7987 struct btrfs_path *path;
7988 struct extent_buffer *leaf;
7989 struct btrfs_delayed_tree_ref *ref;
7990 u32 size = sizeof(*extent_item) + sizeof(*iref);
7991 u64 num_bytes;
7992 u64 flags = extent_op->flags_to_set;
7993 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
7994
7995 ref = btrfs_delayed_node_to_tree_ref(node);
7996
7997 extent_key.objectid = node->bytenr;
7998 if (skinny_metadata) {
7999 extent_key.offset = ref->level;
8000 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8001 num_bytes = fs_info->nodesize;
8002 } else {
8003 extent_key.offset = node->num_bytes;
8004 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8005 size += sizeof(*block_info);
8006 num_bytes = node->num_bytes;
8007 }
8008
8009 path = btrfs_alloc_path();
8010 if (!path)
8011 return -ENOMEM;
8012
8013 path->leave_spinning = 1;
8014 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8015 &extent_key, size);
8016 if (ret) {
8017 btrfs_free_path(path);
8018 return ret;
8019 }
8020
8021 leaf = path->nodes[0];
8022 extent_item = btrfs_item_ptr(leaf, path->slots[0],
8023 struct btrfs_extent_item);
8024 btrfs_set_extent_refs(leaf, extent_item, 1);
8025 btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8026 btrfs_set_extent_flags(leaf, extent_item,
8027 flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8028
8029 if (skinny_metadata) {
8030 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8031 } else {
8032 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8033 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8034 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8035 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8036 }
8037
8038 if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8039 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8040 btrfs_set_extent_inline_ref_type(leaf, iref,
8041 BTRFS_SHARED_BLOCK_REF_KEY);
8042 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8043 } else {
8044 btrfs_set_extent_inline_ref_type(leaf, iref,
8045 BTRFS_TREE_BLOCK_REF_KEY);
8046 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8047 }
8048
8049 btrfs_mark_buffer_dirty(leaf);
8050 btrfs_free_path(path);
8051
8052 ret = remove_from_free_space_tree(trans, extent_key.objectid,
8053 num_bytes);
8054 if (ret)
8055 return ret;
8056
8057 ret = update_block_group(trans, fs_info, extent_key.objectid,
8058 fs_info->nodesize, 1);
8059 if (ret) { /* -ENOENT, logic error */
8060 btrfs_err(fs_info, "update block group failed for %llu %llu",
8061 extent_key.objectid, extent_key.offset);
8062 BUG();
8063 }
8064
8065 trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8066 fs_info->nodesize);
8067 return ret;
8068 }
8069
8070 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8071 struct btrfs_root *root, u64 owner,
8072 u64 offset, u64 ram_bytes,
8073 struct btrfs_key *ins)
8074 {
8075 int ret;
8076
8077 BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8078
8079 btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8080 root->root_key.objectid, owner, offset,
8081 BTRFS_ADD_DELAYED_EXTENT);
8082
8083 ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8084 ins->offset, 0,
8085 root->root_key.objectid, owner,
8086 offset, ram_bytes,
8087 BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8088 return ret;
8089 }
8090
8091 /*
8092 * this is used by the tree logging recovery code. It records that
8093 * an extent has been allocated and makes sure to clear the free
8094 * space cache bits as well
8095 */
8096 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8097 u64 root_objectid, u64 owner, u64 offset,
8098 struct btrfs_key *ins)
8099 {
8100 struct btrfs_fs_info *fs_info = trans->fs_info;
8101 int ret;
8102 struct btrfs_block_group_cache *block_group;
8103 struct btrfs_space_info *space_info;
8104
8105 /*
8106 * Mixed block groups will exclude before processing the log so we only
8107 * need to do the exclude dance if this fs isn't mixed.
8108 */
8109 if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8110 ret = __exclude_logged_extent(fs_info, ins->objectid,
8111 ins->offset);
8112 if (ret)
8113 return ret;
8114 }
8115
8116 block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8117 if (!block_group)
8118 return -EINVAL;
8119
8120 space_info = block_group->space_info;
8121 spin_lock(&space_info->lock);
8122 spin_lock(&block_group->lock);
8123 space_info->bytes_reserved += ins->offset;
8124 block_group->reserved += ins->offset;
8125 spin_unlock(&block_group->lock);
8126 spin_unlock(&space_info->lock);
8127
8128 ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8129 offset, ins, 1);
8130 btrfs_put_block_group(block_group);
8131 return ret;
8132 }
8133
8134 static struct extent_buffer *
8135 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8136 u64 bytenr, int level, u64 owner)
8137 {
8138 struct btrfs_fs_info *fs_info = root->fs_info;
8139 struct extent_buffer *buf;
8140
8141 buf = btrfs_find_create_tree_block(fs_info, bytenr);
8142 if (IS_ERR(buf))
8143 return buf;
8144
8145 /*
8146 * Extra safety check in case the extent tree is corrupted and extent
8147 * allocator chooses to use a tree block which is already used and
8148 * locked.
8149 */
8150 if (buf->lock_owner == current->pid) {
8151 btrfs_err_rl(fs_info,
8152 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8153 buf->start, btrfs_header_owner(buf), current->pid);
8154 free_extent_buffer(buf);
8155 return ERR_PTR(-EUCLEAN);
8156 }
8157
8158 btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8159 btrfs_tree_lock(buf);
8160 clean_tree_block(fs_info, buf);
8161 clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8162
8163 btrfs_set_lock_blocking(buf);
8164 set_extent_buffer_uptodate(buf);
8165
8166 memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8167 btrfs_set_header_level(buf, level);
8168 btrfs_set_header_bytenr(buf, buf->start);
8169 btrfs_set_header_generation(buf, trans->transid);
8170 btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8171 btrfs_set_header_owner(buf, owner);
8172 write_extent_buffer_fsid(buf, fs_info->fsid);
8173 write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8174 if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8175 buf->log_index = root->log_transid % 2;
8176 /*
8177 * we allow two log transactions at a time, use different
8178 * EXENT bit to differentiate dirty pages.
8179 */
8180 if (buf->log_index == 0)
8181 set_extent_dirty(&root->dirty_log_pages, buf->start,
8182 buf->start + buf->len - 1, GFP_NOFS);
8183 else
8184 set_extent_new(&root->dirty_log_pages, buf->start,
8185 buf->start + buf->len - 1);
8186 } else {
8187 buf->log_index = -1;
8188 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8189 buf->start + buf->len - 1, GFP_NOFS);
8190 }
8191 trans->dirty = true;
8192 /* this returns a buffer locked for blocking */
8193 return buf;
8194 }
8195
8196 static struct btrfs_block_rsv *
8197 use_block_rsv(struct btrfs_trans_handle *trans,
8198 struct btrfs_root *root, u32 blocksize)
8199 {
8200 struct btrfs_fs_info *fs_info = root->fs_info;
8201 struct btrfs_block_rsv *block_rsv;
8202 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8203 int ret;
8204 bool global_updated = false;
8205
8206 block_rsv = get_block_rsv(trans, root);
8207
8208 if (unlikely(block_rsv->size == 0))
8209 goto try_reserve;
8210 again:
8211 ret = block_rsv_use_bytes(block_rsv, blocksize);
8212 if (!ret)
8213 return block_rsv;
8214
8215 if (block_rsv->failfast)
8216 return ERR_PTR(ret);
8217
8218 if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8219 global_updated = true;
8220 update_global_block_rsv(fs_info);
8221 goto again;
8222 }
8223
8224 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8225 static DEFINE_RATELIMIT_STATE(_rs,
8226 DEFAULT_RATELIMIT_INTERVAL * 10,
8227 /*DEFAULT_RATELIMIT_BURST*/ 1);
8228 if (__ratelimit(&_rs))
8229 WARN(1, KERN_DEBUG
8230 "BTRFS: block rsv returned %d\n", ret);
8231 }
8232 try_reserve:
8233 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8234 BTRFS_RESERVE_NO_FLUSH);
8235 if (!ret)
8236 return block_rsv;
8237 /*
8238 * If we couldn't reserve metadata bytes try and use some from
8239 * the global reserve if its space type is the same as the global
8240 * reservation.
8241 */
8242 if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8243 block_rsv->space_info == global_rsv->space_info) {
8244 ret = block_rsv_use_bytes(global_rsv, blocksize);
8245 if (!ret)
8246 return global_rsv;
8247 }
8248 return ERR_PTR(ret);
8249 }
8250
8251 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8252 struct btrfs_block_rsv *block_rsv, u32 blocksize)
8253 {
8254 block_rsv_add_bytes(block_rsv, blocksize, false);
8255 block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8256 }
8257
8258 /*
8259 * finds a free extent and does all the dirty work required for allocation
8260 * returns the tree buffer or an ERR_PTR on error.
8261 */
8262 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8263 struct btrfs_root *root,
8264 u64 parent, u64 root_objectid,
8265 const struct btrfs_disk_key *key,
8266 int level, u64 hint,
8267 u64 empty_size)
8268 {
8269 struct btrfs_fs_info *fs_info = root->fs_info;
8270 struct btrfs_key ins;
8271 struct btrfs_block_rsv *block_rsv;
8272 struct extent_buffer *buf;
8273 struct btrfs_delayed_extent_op *extent_op;
8274 u64 flags = 0;
8275 int ret;
8276 u32 blocksize = fs_info->nodesize;
8277 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8278
8279 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8280 if (btrfs_is_testing(fs_info)) {
8281 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8282 level, root_objectid);
8283 if (!IS_ERR(buf))
8284 root->alloc_bytenr += blocksize;
8285 return buf;
8286 }
8287 #endif
8288
8289 block_rsv = use_block_rsv(trans, root, blocksize);
8290 if (IS_ERR(block_rsv))
8291 return ERR_CAST(block_rsv);
8292
8293 ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8294 empty_size, hint, &ins, 0, 0);
8295 if (ret)
8296 goto out_unuse;
8297
8298 buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8299 root_objectid);
8300 if (IS_ERR(buf)) {
8301 ret = PTR_ERR(buf);
8302 goto out_free_reserved;
8303 }
8304
8305 if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8306 if (parent == 0)
8307 parent = ins.objectid;
8308 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8309 } else
8310 BUG_ON(parent > 0);
8311
8312 if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8313 extent_op = btrfs_alloc_delayed_extent_op();
8314 if (!extent_op) {
8315 ret = -ENOMEM;
8316 goto out_free_buf;
8317 }
8318 if (key)
8319 memcpy(&extent_op->key, key, sizeof(extent_op->key));
8320 else
8321 memset(&extent_op->key, 0, sizeof(extent_op->key));
8322 extent_op->flags_to_set = flags;
8323 extent_op->update_key = skinny_metadata ? false : true;
8324 extent_op->update_flags = true;
8325 extent_op->is_data = false;
8326 extent_op->level = level;
8327
8328 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8329 root_objectid, level, 0,
8330 BTRFS_ADD_DELAYED_EXTENT);
8331 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8332 ins.offset, parent,
8333 root_objectid, level,
8334 BTRFS_ADD_DELAYED_EXTENT,
8335 extent_op, NULL, NULL);
8336 if (ret)
8337 goto out_free_delayed;
8338 }
8339 return buf;
8340
8341 out_free_delayed:
8342 btrfs_free_delayed_extent_op(extent_op);
8343 out_free_buf:
8344 free_extent_buffer(buf);
8345 out_free_reserved:
8346 btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8347 out_unuse:
8348 unuse_block_rsv(fs_info, block_rsv, blocksize);
8349 return ERR_PTR(ret);
8350 }
8351
8352 struct walk_control {
8353 u64 refs[BTRFS_MAX_LEVEL];
8354 u64 flags[BTRFS_MAX_LEVEL];
8355 struct btrfs_key update_progress;
8356 int stage;
8357 int level;
8358 int shared_level;
8359 int update_ref;
8360 int keep_locks;
8361 int reada_slot;
8362 int reada_count;
8363 };
8364
8365 #define DROP_REFERENCE 1
8366 #define UPDATE_BACKREF 2
8367
8368 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8369 struct btrfs_root *root,
8370 struct walk_control *wc,
8371 struct btrfs_path *path)
8372 {
8373 struct btrfs_fs_info *fs_info = root->fs_info;
8374 u64 bytenr;
8375 u64 generation;
8376 u64 refs;
8377 u64 flags;
8378 u32 nritems;
8379 struct btrfs_key key;
8380 struct extent_buffer *eb;
8381 int ret;
8382 int slot;
8383 int nread = 0;
8384
8385 if (path->slots[wc->level] < wc->reada_slot) {
8386 wc->reada_count = wc->reada_count * 2 / 3;
8387 wc->reada_count = max(wc->reada_count, 2);
8388 } else {
8389 wc->reada_count = wc->reada_count * 3 / 2;
8390 wc->reada_count = min_t(int, wc->reada_count,
8391 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8392 }
8393
8394 eb = path->nodes[wc->level];
8395 nritems = btrfs_header_nritems(eb);
8396
8397 for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8398 if (nread >= wc->reada_count)
8399 break;
8400
8401 cond_resched();
8402 bytenr = btrfs_node_blockptr(eb, slot);
8403 generation = btrfs_node_ptr_generation(eb, slot);
8404
8405 if (slot == path->slots[wc->level])
8406 goto reada;
8407
8408 if (wc->stage == UPDATE_BACKREF &&
8409 generation <= root->root_key.offset)
8410 continue;
8411
8412 /* We don't lock the tree block, it's OK to be racy here */
8413 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8414 wc->level - 1, 1, &refs,
8415 &flags);
8416 /* We don't care about errors in readahead. */
8417 if (ret < 0)
8418 continue;
8419 BUG_ON(refs == 0);
8420
8421 if (wc->stage == DROP_REFERENCE) {
8422 if (refs == 1)
8423 goto reada;
8424
8425 if (wc->level == 1 &&
8426 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8427 continue;
8428 if (!wc->update_ref ||
8429 generation <= root->root_key.offset)
8430 continue;
8431 btrfs_node_key_to_cpu(eb, &key, slot);
8432 ret = btrfs_comp_cpu_keys(&key,
8433 &wc->update_progress);
8434 if (ret < 0)
8435 continue;
8436 } else {
8437 if (wc->level == 1 &&
8438 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8439 continue;
8440 }
8441 reada:
8442 readahead_tree_block(fs_info, bytenr);
8443 nread++;
8444 }
8445 wc->reada_slot = slot;
8446 }
8447
8448 /*
8449 * helper to process tree block while walking down the tree.
8450 *
8451 * when wc->stage == UPDATE_BACKREF, this function updates
8452 * back refs for pointers in the block.
8453 *
8454 * NOTE: return value 1 means we should stop walking down.
8455 */
8456 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8457 struct btrfs_root *root,
8458 struct btrfs_path *path,
8459 struct walk_control *wc, int lookup_info)
8460 {
8461 struct btrfs_fs_info *fs_info = root->fs_info;
8462 int level = wc->level;
8463 struct extent_buffer *eb = path->nodes[level];
8464 u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8465 int ret;
8466
8467 if (wc->stage == UPDATE_BACKREF &&
8468 btrfs_header_owner(eb) != root->root_key.objectid)
8469 return 1;
8470
8471 /*
8472 * when reference count of tree block is 1, it won't increase
8473 * again. once full backref flag is set, we never clear it.
8474 */
8475 if (lookup_info &&
8476 ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8477 (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8478 BUG_ON(!path->locks[level]);
8479 ret = btrfs_lookup_extent_info(trans, fs_info,
8480 eb->start, level, 1,
8481 &wc->refs[level],
8482 &wc->flags[level]);
8483 BUG_ON(ret == -ENOMEM);
8484 if (ret)
8485 return ret;
8486 BUG_ON(wc->refs[level] == 0);
8487 }
8488
8489 if (wc->stage == DROP_REFERENCE) {
8490 if (wc->refs[level] > 1)
8491 return 1;
8492
8493 if (path->locks[level] && !wc->keep_locks) {
8494 btrfs_tree_unlock_rw(eb, path->locks[level]);
8495 path->locks[level] = 0;
8496 }
8497 return 0;
8498 }
8499
8500 /* wc->stage == UPDATE_BACKREF */
8501 if (!(wc->flags[level] & flag)) {
8502 BUG_ON(!path->locks[level]);
8503 ret = btrfs_inc_ref(trans, root, eb, 1);
8504 BUG_ON(ret); /* -ENOMEM */
8505 ret = btrfs_dec_ref(trans, root, eb, 0);
8506 BUG_ON(ret); /* -ENOMEM */
8507 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8508 eb->len, flag,
8509 btrfs_header_level(eb), 0);
8510 BUG_ON(ret); /* -ENOMEM */
8511 wc->flags[level] |= flag;
8512 }
8513
8514 /*
8515 * the block is shared by multiple trees, so it's not good to
8516 * keep the tree lock
8517 */
8518 if (path->locks[level] && level > 0) {
8519 btrfs_tree_unlock_rw(eb, path->locks[level]);
8520 path->locks[level] = 0;
8521 }
8522 return 0;
8523 }
8524
8525 /*
8526 * helper to process tree block pointer.
8527 *
8528 * when wc->stage == DROP_REFERENCE, this function checks
8529 * reference count of the block pointed to. if the block
8530 * is shared and we need update back refs for the subtree
8531 * rooted at the block, this function changes wc->stage to
8532 * UPDATE_BACKREF. if the block is shared and there is no
8533 * need to update back, this function drops the reference
8534 * to the block.
8535 *
8536 * NOTE: return value 1 means we should stop walking down.
8537 */
8538 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8539 struct btrfs_root *root,
8540 struct btrfs_path *path,
8541 struct walk_control *wc, int *lookup_info)
8542 {
8543 struct btrfs_fs_info *fs_info = root->fs_info;
8544 u64 bytenr;
8545 u64 generation;
8546 u64 parent;
8547 u32 blocksize;
8548 struct btrfs_key key;
8549 struct btrfs_key first_key;
8550 struct extent_buffer *next;
8551 int level = wc->level;
8552 int reada = 0;
8553 int ret = 0;
8554 bool need_account = false;
8555
8556 generation = btrfs_node_ptr_generation(path->nodes[level],
8557 path->slots[level]);
8558 /*
8559 * if the lower level block was created before the snapshot
8560 * was created, we know there is no need to update back refs
8561 * for the subtree
8562 */
8563 if (wc->stage == UPDATE_BACKREF &&
8564 generation <= root->root_key.offset) {
8565 *lookup_info = 1;
8566 return 1;
8567 }
8568
8569 bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8570 btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8571 path->slots[level]);
8572 blocksize = fs_info->nodesize;
8573
8574 next = find_extent_buffer(fs_info, bytenr);
8575 if (!next) {
8576 next = btrfs_find_create_tree_block(fs_info, bytenr);
8577 if (IS_ERR(next))
8578 return PTR_ERR(next);
8579
8580 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8581 level - 1);
8582 reada = 1;
8583 }
8584 btrfs_tree_lock(next);
8585 btrfs_set_lock_blocking(next);
8586
8587 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8588 &wc->refs[level - 1],
8589 &wc->flags[level - 1]);
8590 if (ret < 0)
8591 goto out_unlock;
8592
8593 if (unlikely(wc->refs[level - 1] == 0)) {
8594 btrfs_err(fs_info, "Missing references.");
8595 ret = -EIO;
8596 goto out_unlock;
8597 }
8598 *lookup_info = 0;
8599
8600 if (wc->stage == DROP_REFERENCE) {
8601 if (wc->refs[level - 1] > 1) {
8602 need_account = true;
8603 if (level == 1 &&
8604 (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8605 goto skip;
8606
8607 if (!wc->update_ref ||
8608 generation <= root->root_key.offset)
8609 goto skip;
8610
8611 btrfs_node_key_to_cpu(path->nodes[level], &key,
8612 path->slots[level]);
8613 ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8614 if (ret < 0)
8615 goto skip;
8616
8617 wc->stage = UPDATE_BACKREF;
8618 wc->shared_level = level - 1;
8619 }
8620 } else {
8621 if (level == 1 &&
8622 (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8623 goto skip;
8624 }
8625
8626 if (!btrfs_buffer_uptodate(next, generation, 0)) {
8627 btrfs_tree_unlock(next);
8628 free_extent_buffer(next);
8629 next = NULL;
8630 *lookup_info = 1;
8631 }
8632
8633 if (!next) {
8634 if (reada && level == 1)
8635 reada_walk_down(trans, root, wc, path);
8636 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8637 &first_key);
8638 if (IS_ERR(next)) {
8639 return PTR_ERR(next);
8640 } else if (!extent_buffer_uptodate(next)) {
8641 free_extent_buffer(next);
8642 return -EIO;
8643 }
8644 btrfs_tree_lock(next);
8645 btrfs_set_lock_blocking(next);
8646 }
8647
8648 level--;
8649 ASSERT(level == btrfs_header_level(next));
8650 if (level != btrfs_header_level(next)) {
8651 btrfs_err(root->fs_info, "mismatched level");
8652 ret = -EIO;
8653 goto out_unlock;
8654 }
8655 path->nodes[level] = next;
8656 path->slots[level] = 0;
8657 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8658 wc->level = level;
8659 if (wc->level == 1)
8660 wc->reada_slot = 0;
8661 return 0;
8662 skip:
8663 wc->refs[level - 1] = 0;
8664 wc->flags[level - 1] = 0;
8665 if (wc->stage == DROP_REFERENCE) {
8666 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8667 parent = path->nodes[level]->start;
8668 } else {
8669 ASSERT(root->root_key.objectid ==
8670 btrfs_header_owner(path->nodes[level]));
8671 if (root->root_key.objectid !=
8672 btrfs_header_owner(path->nodes[level])) {
8673 btrfs_err(root->fs_info,
8674 "mismatched block owner");
8675 ret = -EIO;
8676 goto out_unlock;
8677 }
8678 parent = 0;
8679 }
8680
8681 /*
8682 * Reloc tree doesn't contribute to qgroup numbers, and we have
8683 * already accounted them at merge time (replace_path),
8684 * thus we could skip expensive subtree trace here.
8685 */
8686 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
8687 need_account) {
8688 ret = btrfs_qgroup_trace_subtree(trans, next,
8689 generation, level - 1);
8690 if (ret) {
8691 btrfs_err_rl(fs_info,
8692 "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8693 ret);
8694 }
8695 }
8696 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8697 parent, root->root_key.objectid,
8698 level - 1, 0);
8699 if (ret)
8700 goto out_unlock;
8701 }
8702
8703 *lookup_info = 1;
8704 ret = 1;
8705
8706 out_unlock:
8707 btrfs_tree_unlock(next);
8708 free_extent_buffer(next);
8709
8710 return ret;
8711 }
8712
8713 /*
8714 * helper to process tree block while walking up the tree.
8715 *
8716 * when wc->stage == DROP_REFERENCE, this function drops
8717 * reference count on the block.
8718 *
8719 * when wc->stage == UPDATE_BACKREF, this function changes
8720 * wc->stage back to DROP_REFERENCE if we changed wc->stage
8721 * to UPDATE_BACKREF previously while processing the block.
8722 *
8723 * NOTE: return value 1 means we should stop walking up.
8724 */
8725 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8726 struct btrfs_root *root,
8727 struct btrfs_path *path,
8728 struct walk_control *wc)
8729 {
8730 struct btrfs_fs_info *fs_info = root->fs_info;
8731 int ret;
8732 int level = wc->level;
8733 struct extent_buffer *eb = path->nodes[level];
8734 u64 parent = 0;
8735
8736 if (wc->stage == UPDATE_BACKREF) {
8737 BUG_ON(wc->shared_level < level);
8738 if (level < wc->shared_level)
8739 goto out;
8740
8741 ret = find_next_key(path, level + 1, &wc->update_progress);
8742 if (ret > 0)
8743 wc->update_ref = 0;
8744
8745 wc->stage = DROP_REFERENCE;
8746 wc->shared_level = -1;
8747 path->slots[level] = 0;
8748
8749 /*
8750 * check reference count again if the block isn't locked.
8751 * we should start walking down the tree again if reference
8752 * count is one.
8753 */
8754 if (!path->locks[level]) {
8755 BUG_ON(level == 0);
8756 btrfs_tree_lock(eb);
8757 btrfs_set_lock_blocking(eb);
8758 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8759
8760 ret = btrfs_lookup_extent_info(trans, fs_info,
8761 eb->start, level, 1,
8762 &wc->refs[level],
8763 &wc->flags[level]);
8764 if (ret < 0) {
8765 btrfs_tree_unlock_rw(eb, path->locks[level]);
8766 path->locks[level] = 0;
8767 return ret;
8768 }
8769 BUG_ON(wc->refs[level] == 0);
8770 if (wc->refs[level] == 1) {
8771 btrfs_tree_unlock_rw(eb, path->locks[level]);
8772 path->locks[level] = 0;
8773 return 1;
8774 }
8775 }
8776 }
8777
8778 /* wc->stage == DROP_REFERENCE */
8779 BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8780
8781 if (wc->refs[level] == 1) {
8782 if (level == 0) {
8783 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8784 ret = btrfs_dec_ref(trans, root, eb, 1);
8785 else
8786 ret = btrfs_dec_ref(trans, root, eb, 0);
8787 BUG_ON(ret); /* -ENOMEM */
8788 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
8789 if (ret) {
8790 btrfs_err_rl(fs_info,
8791 "error %d accounting leaf items. Quota is out of sync, rescan required.",
8792 ret);
8793 }
8794 }
8795 /* make block locked assertion in clean_tree_block happy */
8796 if (!path->locks[level] &&
8797 btrfs_header_generation(eb) == trans->transid) {
8798 btrfs_tree_lock(eb);
8799 btrfs_set_lock_blocking(eb);
8800 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8801 }
8802 clean_tree_block(fs_info, eb);
8803 }
8804
8805 if (eb == root->node) {
8806 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8807 parent = eb->start;
8808 else if (root->root_key.objectid != btrfs_header_owner(eb))
8809 goto owner_mismatch;
8810 } else {
8811 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8812 parent = path->nodes[level + 1]->start;
8813 else if (root->root_key.objectid !=
8814 btrfs_header_owner(path->nodes[level + 1]))
8815 goto owner_mismatch;
8816 }
8817
8818 btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8819 out:
8820 wc->refs[level] = 0;
8821 wc->flags[level] = 0;
8822 return 0;
8823
8824 owner_mismatch:
8825 btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
8826 btrfs_header_owner(eb), root->root_key.objectid);
8827 return -EUCLEAN;
8828 }
8829
8830 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8831 struct btrfs_root *root,
8832 struct btrfs_path *path,
8833 struct walk_control *wc)
8834 {
8835 int level = wc->level;
8836 int lookup_info = 1;
8837 int ret;
8838
8839 while (level >= 0) {
8840 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8841 if (ret > 0)
8842 break;
8843
8844 if (level == 0)
8845 break;
8846
8847 if (path->slots[level] >=
8848 btrfs_header_nritems(path->nodes[level]))
8849 break;
8850
8851 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8852 if (ret > 0) {
8853 path->slots[level]++;
8854 continue;
8855 } else if (ret < 0)
8856 return ret;
8857 level = wc->level;
8858 }
8859 return 0;
8860 }
8861
8862 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8863 struct btrfs_root *root,
8864 struct btrfs_path *path,
8865 struct walk_control *wc, int max_level)
8866 {
8867 int level = wc->level;
8868 int ret;
8869
8870 path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8871 while (level < max_level && path->nodes[level]) {
8872 wc->level = level;
8873 if (path->slots[level] + 1 <
8874 btrfs_header_nritems(path->nodes[level])) {
8875 path->slots[level]++;
8876 return 0;
8877 } else {
8878 ret = walk_up_proc(trans, root, path, wc);
8879 if (ret > 0)
8880 return 0;
8881 if (ret < 0)
8882 return ret;
8883
8884 if (path->locks[level]) {
8885 btrfs_tree_unlock_rw(path->nodes[level],
8886 path->locks[level]);
8887 path->locks[level] = 0;
8888 }
8889 free_extent_buffer(path->nodes[level]);
8890 path->nodes[level] = NULL;
8891 level++;
8892 }
8893 }
8894 return 1;
8895 }
8896
8897 /*
8898 * drop a subvolume tree.
8899 *
8900 * this function traverses the tree freeing any blocks that only
8901 * referenced by the tree.
8902 *
8903 * when a shared tree block is found. this function decreases its
8904 * reference count by one. if update_ref is true, this function
8905 * also make sure backrefs for the shared block and all lower level
8906 * blocks are properly updated.
8907 *
8908 * If called with for_reloc == 0, may exit early with -EAGAIN
8909 */
8910 int btrfs_drop_snapshot(struct btrfs_root *root,
8911 struct btrfs_block_rsv *block_rsv, int update_ref,
8912 int for_reloc)
8913 {
8914 struct btrfs_fs_info *fs_info = root->fs_info;
8915 struct btrfs_path *path;
8916 struct btrfs_trans_handle *trans;
8917 struct btrfs_root *tree_root = fs_info->tree_root;
8918 struct btrfs_root_item *root_item = &root->root_item;
8919 struct walk_control *wc;
8920 struct btrfs_key key;
8921 int err = 0;
8922 int ret;
8923 int level;
8924 bool root_dropped = false;
8925
8926 btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
8927
8928 path = btrfs_alloc_path();
8929 if (!path) {
8930 err = -ENOMEM;
8931 goto out;
8932 }
8933
8934 wc = kzalloc(sizeof(*wc), GFP_NOFS);
8935 if (!wc) {
8936 btrfs_free_path(path);
8937 err = -ENOMEM;
8938 goto out;
8939 }
8940
8941 trans = btrfs_start_transaction(tree_root, 0);
8942 if (IS_ERR(trans)) {
8943 err = PTR_ERR(trans);
8944 goto out_free;
8945 }
8946
8947 if (block_rsv)
8948 trans->block_rsv = block_rsv;
8949
8950 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8951 level = btrfs_header_level(root->node);
8952 path->nodes[level] = btrfs_lock_root_node(root);
8953 btrfs_set_lock_blocking(path->nodes[level]);
8954 path->slots[level] = 0;
8955 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8956 memset(&wc->update_progress, 0,
8957 sizeof(wc->update_progress));
8958 } else {
8959 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8960 memcpy(&wc->update_progress, &key,
8961 sizeof(wc->update_progress));
8962
8963 level = root_item->drop_level;
8964 BUG_ON(level == 0);
8965 path->lowest_level = level;
8966 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8967 path->lowest_level = 0;
8968 if (ret < 0) {
8969 err = ret;
8970 goto out_end_trans;
8971 }
8972 WARN_ON(ret > 0);
8973
8974 /*
8975 * unlock our path, this is safe because only this
8976 * function is allowed to delete this snapshot
8977 */
8978 btrfs_unlock_up_safe(path, 0);
8979
8980 level = btrfs_header_level(root->node);
8981 while (1) {
8982 btrfs_tree_lock(path->nodes[level]);
8983 btrfs_set_lock_blocking(path->nodes[level]);
8984 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8985
8986 ret = btrfs_lookup_extent_info(trans, fs_info,
8987 path->nodes[level]->start,
8988 level, 1, &wc->refs[level],
8989 &wc->flags[level]);
8990 if (ret < 0) {
8991 err = ret;
8992 goto out_end_trans;
8993 }
8994 BUG_ON(wc->refs[level] == 0);
8995
8996 if (level == root_item->drop_level)
8997 break;
8998
8999 btrfs_tree_unlock(path->nodes[level]);
9000 path->locks[level] = 0;
9001 WARN_ON(wc->refs[level] != 1);
9002 level--;
9003 }
9004 }
9005
9006 wc->level = level;
9007 wc->shared_level = -1;
9008 wc->stage = DROP_REFERENCE;
9009 wc->update_ref = update_ref;
9010 wc->keep_locks = 0;
9011 wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9012
9013 while (1) {
9014
9015 ret = walk_down_tree(trans, root, path, wc);
9016 if (ret < 0) {
9017 err = ret;
9018 break;
9019 }
9020
9021 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9022 if (ret < 0) {
9023 err = ret;
9024 break;
9025 }
9026
9027 if (ret > 0) {
9028 BUG_ON(wc->stage != DROP_REFERENCE);
9029 break;
9030 }
9031
9032 if (wc->stage == DROP_REFERENCE) {
9033 level = wc->level;
9034 btrfs_node_key(path->nodes[level],
9035 &root_item->drop_progress,
9036 path->slots[level]);
9037 root_item->drop_level = level;
9038 }
9039
9040 BUG_ON(wc->level == 0);
9041 if (btrfs_should_end_transaction(trans) ||
9042 (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9043 ret = btrfs_update_root(trans, tree_root,
9044 &root->root_key,
9045 root_item);
9046 if (ret) {
9047 btrfs_abort_transaction(trans, ret);
9048 err = ret;
9049 goto out_end_trans;
9050 }
9051
9052 btrfs_end_transaction_throttle(trans);
9053 if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9054 btrfs_debug(fs_info,
9055 "drop snapshot early exit");
9056 err = -EAGAIN;
9057 goto out_free;
9058 }
9059
9060 trans = btrfs_start_transaction(tree_root, 0);
9061 if (IS_ERR(trans)) {
9062 err = PTR_ERR(trans);
9063 goto out_free;
9064 }
9065 if (block_rsv)
9066 trans->block_rsv = block_rsv;
9067 }
9068 }
9069 btrfs_release_path(path);
9070 if (err)
9071 goto out_end_trans;
9072
9073 ret = btrfs_del_root(trans, &root->root_key);
9074 if (ret) {
9075 btrfs_abort_transaction(trans, ret);
9076 err = ret;
9077 goto out_end_trans;
9078 }
9079
9080 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9081 ret = btrfs_find_root(tree_root, &root->root_key, path,
9082 NULL, NULL);
9083 if (ret < 0) {
9084 btrfs_abort_transaction(trans, ret);
9085 err = ret;
9086 goto out_end_trans;
9087 } else if (ret > 0) {
9088 /* if we fail to delete the orphan item this time
9089 * around, it'll get picked up the next time.
9090 *
9091 * The most common failure here is just -ENOENT.
9092 */
9093 btrfs_del_orphan_item(trans, tree_root,
9094 root->root_key.objectid);
9095 }
9096 }
9097
9098 if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9099 btrfs_add_dropped_root(trans, root);
9100 } else {
9101 free_extent_buffer(root->node);
9102 free_extent_buffer(root->commit_root);
9103 btrfs_put_fs_root(root);
9104 }
9105 root_dropped = true;
9106 out_end_trans:
9107 btrfs_end_transaction_throttle(trans);
9108 out_free:
9109 kfree(wc);
9110 btrfs_free_path(path);
9111 out:
9112 /*
9113 * So if we need to stop dropping the snapshot for whatever reason we
9114 * need to make sure to add it back to the dead root list so that we
9115 * keep trying to do the work later. This also cleans up roots if we
9116 * don't have it in the radix (like when we recover after a power fail
9117 * or unmount) so we don't leak memory.
9118 */
9119 if (!for_reloc && !root_dropped)
9120 btrfs_add_dead_root(root);
9121 if (err && err != -EAGAIN)
9122 btrfs_handle_fs_error(fs_info, err, NULL);
9123 return err;
9124 }
9125
9126 /*
9127 * drop subtree rooted at tree block 'node'.
9128 *
9129 * NOTE: this function will unlock and release tree block 'node'
9130 * only used by relocation code
9131 */
9132 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9133 struct btrfs_root *root,
9134 struct extent_buffer *node,
9135 struct extent_buffer *parent)
9136 {
9137 struct btrfs_fs_info *fs_info = root->fs_info;
9138 struct btrfs_path *path;
9139 struct walk_control *wc;
9140 int level;
9141 int parent_level;
9142 int ret = 0;
9143 int wret;
9144
9145 BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9146
9147 path = btrfs_alloc_path();
9148 if (!path)
9149 return -ENOMEM;
9150
9151 wc = kzalloc(sizeof(*wc), GFP_NOFS);
9152 if (!wc) {
9153 btrfs_free_path(path);
9154 return -ENOMEM;
9155 }
9156
9157 btrfs_assert_tree_locked(parent);
9158 parent_level = btrfs_header_level(parent);
9159 extent_buffer_get(parent);
9160 path->nodes[parent_level] = parent;
9161 path->slots[parent_level] = btrfs_header_nritems(parent);
9162
9163 btrfs_assert_tree_locked(node);
9164 level = btrfs_header_level(node);
9165 path->nodes[level] = node;
9166 path->slots[level] = 0;
9167 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9168
9169 wc->refs[parent_level] = 1;
9170 wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9171 wc->level = level;
9172 wc->shared_level = -1;
9173 wc->stage = DROP_REFERENCE;
9174 wc->update_ref = 0;
9175 wc->keep_locks = 1;
9176 wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9177
9178 while (1) {
9179 wret = walk_down_tree(trans, root, path, wc);
9180 if (wret < 0) {
9181 ret = wret;
9182 break;
9183 }
9184
9185 wret = walk_up_tree(trans, root, path, wc, parent_level);
9186 if (wret < 0)
9187 ret = wret;
9188 if (wret != 0)
9189 break;
9190 }
9191
9192 kfree(wc);
9193 btrfs_free_path(path);
9194 return ret;
9195 }
9196
9197 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9198 {
9199 u64 num_devices;
9200 u64 stripped;
9201
9202 /*
9203 * if restripe for this chunk_type is on pick target profile and
9204 * return, otherwise do the usual balance
9205 */
9206 stripped = get_restripe_target(fs_info, flags);
9207 if (stripped)
9208 return extended_to_chunk(stripped);
9209
9210 num_devices = fs_info->fs_devices->rw_devices;
9211
9212 stripped = BTRFS_BLOCK_GROUP_RAID0 |
9213 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9214 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9215
9216 if (num_devices == 1) {
9217 stripped |= BTRFS_BLOCK_GROUP_DUP;
9218 stripped = flags & ~stripped;
9219
9220 /* turn raid0 into single device chunks */
9221 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9222 return stripped;
9223
9224 /* turn mirroring into duplication */
9225 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9226 BTRFS_BLOCK_GROUP_RAID10))
9227 return stripped | BTRFS_BLOCK_GROUP_DUP;
9228 } else {
9229 /* they already had raid on here, just return */
9230 if (flags & stripped)
9231 return flags;
9232
9233 stripped |= BTRFS_BLOCK_GROUP_DUP;
9234 stripped = flags & ~stripped;
9235
9236 /* switch duplicated blocks with raid1 */
9237 if (flags & BTRFS_BLOCK_GROUP_DUP)
9238 return stripped | BTRFS_BLOCK_GROUP_RAID1;
9239
9240 /* this is drive concat, leave it alone */
9241 }
9242
9243 return flags;
9244 }
9245
9246 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9247 {
9248 struct btrfs_space_info *sinfo = cache->space_info;
9249 u64 num_bytes;
9250 u64 min_allocable_bytes;
9251 int ret = -ENOSPC;
9252
9253 /*
9254 * We need some metadata space and system metadata space for
9255 * allocating chunks in some corner cases until we force to set
9256 * it to be readonly.
9257 */
9258 if ((sinfo->flags &
9259 (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9260 !force)
9261 min_allocable_bytes = SZ_1M;
9262 else
9263 min_allocable_bytes = 0;
9264
9265 spin_lock(&sinfo->lock);
9266 spin_lock(&cache->lock);
9267
9268 if (cache->ro) {
9269 cache->ro++;
9270 ret = 0;
9271 goto out;
9272 }
9273
9274 num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9275 cache->bytes_super - btrfs_block_group_used(&cache->item);
9276
9277 if (btrfs_space_info_used(sinfo, true) + num_bytes +
9278 min_allocable_bytes <= sinfo->total_bytes) {
9279 sinfo->bytes_readonly += num_bytes;
9280 cache->ro++;
9281 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9282 ret = 0;
9283 }
9284 out:
9285 spin_unlock(&cache->lock);
9286 spin_unlock(&sinfo->lock);
9287 return ret;
9288 }
9289
9290 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9291
9292 {
9293 struct btrfs_fs_info *fs_info = cache->fs_info;
9294 struct btrfs_trans_handle *trans;
9295 u64 alloc_flags;
9296 int ret;
9297
9298 again:
9299 trans = btrfs_join_transaction(fs_info->extent_root);
9300 if (IS_ERR(trans))
9301 return PTR_ERR(trans);
9302
9303 /*
9304 * we're not allowed to set block groups readonly after the dirty
9305 * block groups cache has started writing. If it already started,
9306 * back off and let this transaction commit
9307 */
9308 mutex_lock(&fs_info->ro_block_group_mutex);
9309 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9310 u64 transid = trans->transid;
9311
9312 mutex_unlock(&fs_info->ro_block_group_mutex);
9313 btrfs_end_transaction(trans);
9314
9315 ret = btrfs_wait_for_commit(fs_info, transid);
9316 if (ret)
9317 return ret;
9318 goto again;
9319 }
9320
9321 /*
9322 * if we are changing raid levels, try to allocate a corresponding
9323 * block group with the new raid level.
9324 */
9325 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9326 if (alloc_flags != cache->flags) {
9327 ret = do_chunk_alloc(trans, alloc_flags,
9328 CHUNK_ALLOC_FORCE);
9329 /*
9330 * ENOSPC is allowed here, we may have enough space
9331 * already allocated at the new raid level to
9332 * carry on
9333 */
9334 if (ret == -ENOSPC)
9335 ret = 0;
9336 if (ret < 0)
9337 goto out;
9338 }
9339
9340 ret = inc_block_group_ro(cache, 0);
9341 if (!ret)
9342 goto out;
9343 alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9344 ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9345 if (ret < 0)
9346 goto out;
9347 ret = inc_block_group_ro(cache, 0);
9348 out:
9349 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9350 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9351 mutex_lock(&fs_info->chunk_mutex);
9352 check_system_chunk(trans, alloc_flags);
9353 mutex_unlock(&fs_info->chunk_mutex);
9354 }
9355 mutex_unlock(&fs_info->ro_block_group_mutex);
9356
9357 btrfs_end_transaction(trans);
9358 return ret;
9359 }
9360
9361 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9362 {
9363 u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9364
9365 return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9366 }
9367
9368 /*
9369 * helper to account the unused space of all the readonly block group in the
9370 * space_info. takes mirrors into account.
9371 */
9372 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9373 {
9374 struct btrfs_block_group_cache *block_group;
9375 u64 free_bytes = 0;
9376 int factor;
9377
9378 /* It's df, we don't care if it's racy */
9379 if (list_empty(&sinfo->ro_bgs))
9380 return 0;
9381
9382 spin_lock(&sinfo->lock);
9383 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9384 spin_lock(&block_group->lock);
9385
9386 if (!block_group->ro) {
9387 spin_unlock(&block_group->lock);
9388 continue;
9389 }
9390
9391 factor = btrfs_bg_type_to_factor(block_group->flags);
9392 free_bytes += (block_group->key.offset -
9393 btrfs_block_group_used(&block_group->item)) *
9394 factor;
9395
9396 spin_unlock(&block_group->lock);
9397 }
9398 spin_unlock(&sinfo->lock);
9399
9400 return free_bytes;
9401 }
9402
9403 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9404 {
9405 struct btrfs_space_info *sinfo = cache->space_info;
9406 u64 num_bytes;
9407
9408 BUG_ON(!cache->ro);
9409
9410 spin_lock(&sinfo->lock);
9411 spin_lock(&cache->lock);
9412 if (!--cache->ro) {
9413 num_bytes = cache->key.offset - cache->reserved -
9414 cache->pinned - cache->bytes_super -
9415 btrfs_block_group_used(&cache->item);
9416 sinfo->bytes_readonly -= num_bytes;
9417 list_del_init(&cache->ro_list);
9418 }
9419 spin_unlock(&cache->lock);
9420 spin_unlock(&sinfo->lock);
9421 }
9422
9423 /*
9424 * checks to see if its even possible to relocate this block group.
9425 *
9426 * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9427 * ok to go ahead and try.
9428 */
9429 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9430 {
9431 struct btrfs_root *root = fs_info->extent_root;
9432 struct btrfs_block_group_cache *block_group;
9433 struct btrfs_space_info *space_info;
9434 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9435 struct btrfs_device *device;
9436 struct btrfs_trans_handle *trans;
9437 u64 min_free;
9438 u64 dev_min = 1;
9439 u64 dev_nr = 0;
9440 u64 target;
9441 int debug;
9442 int index;
9443 int full = 0;
9444 int ret = 0;
9445
9446 debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9447
9448 block_group = btrfs_lookup_block_group(fs_info, bytenr);
9449
9450 /* odd, couldn't find the block group, leave it alone */
9451 if (!block_group) {
9452 if (debug)
9453 btrfs_warn(fs_info,
9454 "can't find block group for bytenr %llu",
9455 bytenr);
9456 return -1;
9457 }
9458
9459 min_free = btrfs_block_group_used(&block_group->item);
9460
9461 /* no bytes used, we're good */
9462 if (!min_free)
9463 goto out;
9464
9465 space_info = block_group->space_info;
9466 spin_lock(&space_info->lock);
9467
9468 full = space_info->full;
9469
9470 /*
9471 * if this is the last block group we have in this space, we can't
9472 * relocate it unless we're able to allocate a new chunk below.
9473 *
9474 * Otherwise, we need to make sure we have room in the space to handle
9475 * all of the extents from this block group. If we can, we're good
9476 */
9477 if ((space_info->total_bytes != block_group->key.offset) &&
9478 (btrfs_space_info_used(space_info, false) + min_free <
9479 space_info->total_bytes)) {
9480 spin_unlock(&space_info->lock);
9481 goto out;
9482 }
9483 spin_unlock(&space_info->lock);
9484
9485 /*
9486 * ok we don't have enough space, but maybe we have free space on our
9487 * devices to allocate new chunks for relocation, so loop through our
9488 * alloc devices and guess if we have enough space. if this block
9489 * group is going to be restriped, run checks against the target
9490 * profile instead of the current one.
9491 */
9492 ret = -1;
9493
9494 /*
9495 * index:
9496 * 0: raid10
9497 * 1: raid1
9498 * 2: dup
9499 * 3: raid0
9500 * 4: single
9501 */
9502 target = get_restripe_target(fs_info, block_group->flags);
9503 if (target) {
9504 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9505 } else {
9506 /*
9507 * this is just a balance, so if we were marked as full
9508 * we know there is no space for a new chunk
9509 */
9510 if (full) {
9511 if (debug)
9512 btrfs_warn(fs_info,
9513 "no space to alloc new chunk for block group %llu",
9514 block_group->key.objectid);
9515 goto out;
9516 }
9517
9518 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9519 }
9520
9521 if (index == BTRFS_RAID_RAID10) {
9522 dev_min = 4;
9523 /* Divide by 2 */
9524 min_free >>= 1;
9525 } else if (index == BTRFS_RAID_RAID1) {
9526 dev_min = 2;
9527 } else if (index == BTRFS_RAID_DUP) {
9528 /* Multiply by 2 */
9529 min_free <<= 1;
9530 } else if (index == BTRFS_RAID_RAID0) {
9531 dev_min = fs_devices->rw_devices;
9532 min_free = div64_u64(min_free, dev_min);
9533 }
9534
9535 /* We need to do this so that we can look at pending chunks */
9536 trans = btrfs_join_transaction(root);
9537 if (IS_ERR(trans)) {
9538 ret = PTR_ERR(trans);
9539 goto out;
9540 }
9541
9542 mutex_lock(&fs_info->chunk_mutex);
9543 list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9544 u64 dev_offset;
9545
9546 /*
9547 * check to make sure we can actually find a chunk with enough
9548 * space to fit our block group in.
9549 */
9550 if (device->total_bytes > device->bytes_used + min_free &&
9551 !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9552 ret = find_free_dev_extent(trans, device, min_free,
9553 &dev_offset, NULL);
9554 if (!ret)
9555 dev_nr++;
9556
9557 if (dev_nr >= dev_min)
9558 break;
9559
9560 ret = -1;
9561 }
9562 }
9563 if (debug && ret == -1)
9564 btrfs_warn(fs_info,
9565 "no space to allocate a new chunk for block group %llu",
9566 block_group->key.objectid);
9567 mutex_unlock(&fs_info->chunk_mutex);
9568 btrfs_end_transaction(trans);
9569 out:
9570 btrfs_put_block_group(block_group);
9571 return ret;
9572 }
9573
9574 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9575 struct btrfs_path *path,
9576 struct btrfs_key *key)
9577 {
9578 struct btrfs_root *root = fs_info->extent_root;
9579 int ret = 0;
9580 struct btrfs_key found_key;
9581 struct extent_buffer *leaf;
9582 struct btrfs_block_group_item bg;
9583 u64 flags;
9584 int slot;
9585
9586 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9587 if (ret < 0)
9588 goto out;
9589
9590 while (1) {
9591 slot = path->slots[0];
9592 leaf = path->nodes[0];
9593 if (slot >= btrfs_header_nritems(leaf)) {
9594 ret = btrfs_next_leaf(root, path);
9595 if (ret == 0)
9596 continue;
9597 if (ret < 0)
9598 goto out;
9599 break;
9600 }
9601 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9602
9603 if (found_key.objectid >= key->objectid &&
9604 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9605 struct extent_map_tree *em_tree;
9606 struct extent_map *em;
9607
9608 em_tree = &root->fs_info->mapping_tree.map_tree;
9609 read_lock(&em_tree->lock);
9610 em = lookup_extent_mapping(em_tree, found_key.objectid,
9611 found_key.offset);
9612 read_unlock(&em_tree->lock);
9613 if (!em) {
9614 btrfs_err(fs_info,
9615 "logical %llu len %llu found bg but no related chunk",
9616 found_key.objectid, found_key.offset);
9617 ret = -ENOENT;
9618 } else if (em->start != found_key.objectid ||
9619 em->len != found_key.offset) {
9620 btrfs_err(fs_info,
9621 "block group %llu len %llu mismatch with chunk %llu len %llu",
9622 found_key.objectid, found_key.offset,
9623 em->start, em->len);
9624 ret = -EUCLEAN;
9625 } else {
9626 read_extent_buffer(leaf, &bg,
9627 btrfs_item_ptr_offset(leaf, slot),
9628 sizeof(bg));
9629 flags = btrfs_block_group_flags(&bg) &
9630 BTRFS_BLOCK_GROUP_TYPE_MASK;
9631
9632 if (flags != (em->map_lookup->type &
9633 BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9634 btrfs_err(fs_info,
9635 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9636 found_key.objectid,
9637 found_key.offset, flags,
9638 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9639 em->map_lookup->type));
9640 ret = -EUCLEAN;
9641 } else {
9642 ret = 0;
9643 }
9644 }
9645 free_extent_map(em);
9646 goto out;
9647 }
9648 path->slots[0]++;
9649 }
9650 out:
9651 return ret;
9652 }
9653
9654 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9655 {
9656 struct btrfs_block_group_cache *block_group;
9657 u64 last = 0;
9658
9659 while (1) {
9660 struct inode *inode;
9661
9662 block_group = btrfs_lookup_first_block_group(info, last);
9663 while (block_group) {
9664 wait_block_group_cache_done(block_group);
9665 spin_lock(&block_group->lock);
9666 if (block_group->iref)
9667 break;
9668 spin_unlock(&block_group->lock);
9669 block_group = next_block_group(info, block_group);
9670 }
9671 if (!block_group) {
9672 if (last == 0)
9673 break;
9674 last = 0;
9675 continue;
9676 }
9677
9678 inode = block_group->inode;
9679 block_group->iref = 0;
9680 block_group->inode = NULL;
9681 spin_unlock(&block_group->lock);
9682 ASSERT(block_group->io_ctl.inode == NULL);
9683 iput(inode);
9684 last = block_group->key.objectid + block_group->key.offset;
9685 btrfs_put_block_group(block_group);
9686 }
9687 }
9688
9689 /*
9690 * Must be called only after stopping all workers, since we could have block
9691 * group caching kthreads running, and therefore they could race with us if we
9692 * freed the block groups before stopping them.
9693 */
9694 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9695 {
9696 struct btrfs_block_group_cache *block_group;
9697 struct btrfs_space_info *space_info;
9698 struct btrfs_caching_control *caching_ctl;
9699 struct rb_node *n;
9700
9701 down_write(&info->commit_root_sem);
9702 while (!list_empty(&info->caching_block_groups)) {
9703 caching_ctl = list_entry(info->caching_block_groups.next,
9704 struct btrfs_caching_control, list);
9705 list_del(&caching_ctl->list);
9706 put_caching_control(caching_ctl);
9707 }
9708 up_write(&info->commit_root_sem);
9709
9710 spin_lock(&info->unused_bgs_lock);
9711 while (!list_empty(&info->unused_bgs)) {
9712 block_group = list_first_entry(&info->unused_bgs,
9713 struct btrfs_block_group_cache,
9714 bg_list);
9715 list_del_init(&block_group->bg_list);
9716 btrfs_put_block_group(block_group);
9717 }
9718 spin_unlock(&info->unused_bgs_lock);
9719
9720 spin_lock(&info->block_group_cache_lock);
9721 while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9722 block_group = rb_entry(n, struct btrfs_block_group_cache,
9723 cache_node);
9724 rb_erase(&block_group->cache_node,
9725 &info->block_group_cache_tree);
9726 RB_CLEAR_NODE(&block_group->cache_node);
9727 spin_unlock(&info->block_group_cache_lock);
9728
9729 down_write(&block_group->space_info->groups_sem);
9730 list_del(&block_group->list);
9731 up_write(&block_group->space_info->groups_sem);
9732
9733 /*
9734 * We haven't cached this block group, which means we could
9735 * possibly have excluded extents on this block group.
9736 */
9737 if (block_group->cached == BTRFS_CACHE_NO ||
9738 block_group->cached == BTRFS_CACHE_ERROR)
9739 free_excluded_extents(block_group);
9740
9741 btrfs_remove_free_space_cache(block_group);
9742 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9743 ASSERT(list_empty(&block_group->dirty_list));
9744 ASSERT(list_empty(&block_group->io_list));
9745 ASSERT(list_empty(&block_group->bg_list));
9746 ASSERT(atomic_read(&block_group->count) == 1);
9747 btrfs_put_block_group(block_group);
9748
9749 spin_lock(&info->block_group_cache_lock);
9750 }
9751 spin_unlock(&info->block_group_cache_lock);
9752
9753 /* now that all the block groups are freed, go through and
9754 * free all the space_info structs. This is only called during
9755 * the final stages of unmount, and so we know nobody is
9756 * using them. We call synchronize_rcu() once before we start,
9757 * just to be on the safe side.
9758 */
9759 synchronize_rcu();
9760
9761 release_global_block_rsv(info);
9762
9763 while (!list_empty(&info->space_info)) {
9764 int i;
9765
9766 space_info = list_entry(info->space_info.next,
9767 struct btrfs_space_info,
9768 list);
9769
9770 /*
9771 * Do not hide this behind enospc_debug, this is actually
9772 * important and indicates a real bug if this happens.
9773 */
9774 if (WARN_ON(space_info->bytes_pinned > 0 ||
9775 space_info->bytes_reserved > 0 ||
9776 space_info->bytes_may_use > 0))
9777 dump_space_info(info, space_info, 0, 0);
9778 list_del(&space_info->list);
9779 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9780 struct kobject *kobj;
9781 kobj = space_info->block_group_kobjs[i];
9782 space_info->block_group_kobjs[i] = NULL;
9783 if (kobj) {
9784 kobject_del(kobj);
9785 kobject_put(kobj);
9786 }
9787 }
9788 kobject_del(&space_info->kobj);
9789 kobject_put(&space_info->kobj);
9790 }
9791 return 0;
9792 }
9793
9794 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9795 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9796 {
9797 struct btrfs_space_info *space_info;
9798 struct raid_kobject *rkobj;
9799 LIST_HEAD(list);
9800 int index;
9801 int ret = 0;
9802
9803 spin_lock(&fs_info->pending_raid_kobjs_lock);
9804 list_splice_init(&fs_info->pending_raid_kobjs, &list);
9805 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9806
9807 list_for_each_entry(rkobj, &list, list) {
9808 space_info = __find_space_info(fs_info, rkobj->flags);
9809 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9810
9811 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9812 "%s", get_raid_name(index));
9813 if (ret) {
9814 kobject_put(&rkobj->kobj);
9815 break;
9816 }
9817 }
9818 if (ret)
9819 btrfs_warn(fs_info,
9820 "failed to add kobject for block cache, ignoring");
9821 }
9822
9823 static void link_block_group(struct btrfs_block_group_cache *cache)
9824 {
9825 struct btrfs_space_info *space_info = cache->space_info;
9826 struct btrfs_fs_info *fs_info = cache->fs_info;
9827 int index = btrfs_bg_flags_to_raid_index(cache->flags);
9828 bool first = false;
9829
9830 down_write(&space_info->groups_sem);
9831 if (list_empty(&space_info->block_groups[index]))
9832 first = true;
9833 list_add_tail(&cache->list, &space_info->block_groups[index]);
9834 up_write(&space_info->groups_sem);
9835
9836 if (first) {
9837 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9838 if (!rkobj) {
9839 btrfs_warn(cache->fs_info,
9840 "couldn't alloc memory for raid level kobject");
9841 return;
9842 }
9843 rkobj->flags = cache->flags;
9844 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9845
9846 spin_lock(&fs_info->pending_raid_kobjs_lock);
9847 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9848 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9849 space_info->block_group_kobjs[index] = &rkobj->kobj;
9850 }
9851 }
9852
9853 static struct btrfs_block_group_cache *
9854 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9855 u64 start, u64 size)
9856 {
9857 struct btrfs_block_group_cache *cache;
9858
9859 cache = kzalloc(sizeof(*cache), GFP_NOFS);
9860 if (!cache)
9861 return NULL;
9862
9863 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9864 GFP_NOFS);
9865 if (!cache->free_space_ctl) {
9866 kfree(cache);
9867 return NULL;
9868 }
9869
9870 cache->key.objectid = start;
9871 cache->key.offset = size;
9872 cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9873
9874 cache->fs_info = fs_info;
9875 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9876 set_free_space_tree_thresholds(cache);
9877
9878 atomic_set(&cache->count, 1);
9879 spin_lock_init(&cache->lock);
9880 init_rwsem(&cache->data_rwsem);
9881 INIT_LIST_HEAD(&cache->list);
9882 INIT_LIST_HEAD(&cache->cluster_list);
9883 INIT_LIST_HEAD(&cache->bg_list);
9884 INIT_LIST_HEAD(&cache->ro_list);
9885 INIT_LIST_HEAD(&cache->dirty_list);
9886 INIT_LIST_HEAD(&cache->io_list);
9887 btrfs_init_free_space_ctl(cache);
9888 atomic_set(&cache->trimming, 0);
9889 mutex_init(&cache->free_space_lock);
9890 btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9891
9892 return cache;
9893 }
9894
9895
9896 /*
9897 * Iterate all chunks and verify that each of them has the corresponding block
9898 * group
9899 */
9900 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
9901 {
9902 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
9903 struct extent_map *em;
9904 struct btrfs_block_group_cache *bg;
9905 u64 start = 0;
9906 int ret = 0;
9907
9908 while (1) {
9909 read_lock(&map_tree->map_tree.lock);
9910 /*
9911 * lookup_extent_mapping will return the first extent map
9912 * intersecting the range, so setting @len to 1 is enough to
9913 * get the first chunk.
9914 */
9915 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
9916 read_unlock(&map_tree->map_tree.lock);
9917 if (!em)
9918 break;
9919
9920 bg = btrfs_lookup_block_group(fs_info, em->start);
9921 if (!bg) {
9922 btrfs_err(fs_info,
9923 "chunk start=%llu len=%llu doesn't have corresponding block group",
9924 em->start, em->len);
9925 ret = -EUCLEAN;
9926 free_extent_map(em);
9927 break;
9928 }
9929 if (bg->key.objectid != em->start ||
9930 bg->key.offset != em->len ||
9931 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
9932 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9933 btrfs_err(fs_info,
9934 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
9935 em->start, em->len,
9936 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
9937 bg->key.objectid, bg->key.offset,
9938 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
9939 ret = -EUCLEAN;
9940 free_extent_map(em);
9941 btrfs_put_block_group(bg);
9942 break;
9943 }
9944 start = em->start + em->len;
9945 free_extent_map(em);
9946 btrfs_put_block_group(bg);
9947 }
9948 return ret;
9949 }
9950
9951 int btrfs_read_block_groups(struct btrfs_fs_info *info)
9952 {
9953 struct btrfs_path *path;
9954 int ret;
9955 struct btrfs_block_group_cache *cache;
9956 struct btrfs_space_info *space_info;
9957 struct btrfs_key key;
9958 struct btrfs_key found_key;
9959 struct extent_buffer *leaf;
9960 int need_clear = 0;
9961 u64 cache_gen;
9962 u64 feature;
9963 int mixed;
9964
9965 feature = btrfs_super_incompat_flags(info->super_copy);
9966 mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
9967
9968 key.objectid = 0;
9969 key.offset = 0;
9970 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9971 path = btrfs_alloc_path();
9972 if (!path)
9973 return -ENOMEM;
9974 path->reada = READA_FORWARD;
9975
9976 cache_gen = btrfs_super_cache_generation(info->super_copy);
9977 if (btrfs_test_opt(info, SPACE_CACHE) &&
9978 btrfs_super_generation(info->super_copy) != cache_gen)
9979 need_clear = 1;
9980 if (btrfs_test_opt(info, CLEAR_CACHE))
9981 need_clear = 1;
9982
9983 while (1) {
9984 ret = find_first_block_group(info, path, &key);
9985 if (ret > 0)
9986 break;
9987 if (ret != 0)
9988 goto error;
9989
9990 leaf = path->nodes[0];
9991 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9992
9993 cache = btrfs_create_block_group_cache(info, found_key.objectid,
9994 found_key.offset);
9995 if (!cache) {
9996 ret = -ENOMEM;
9997 goto error;
9998 }
9999
10000 if (need_clear) {
10001 /*
10002 * When we mount with old space cache, we need to
10003 * set BTRFS_DC_CLEAR and set dirty flag.
10004 *
10005 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10006 * truncate the old free space cache inode and
10007 * setup a new one.
10008 * b) Setting 'dirty flag' makes sure that we flush
10009 * the new space cache info onto disk.
10010 */
10011 if (btrfs_test_opt(info, SPACE_CACHE))
10012 cache->disk_cache_state = BTRFS_DC_CLEAR;
10013 }
10014
10015 read_extent_buffer(leaf, &cache->item,
10016 btrfs_item_ptr_offset(leaf, path->slots[0]),
10017 sizeof(cache->item));
10018 cache->flags = btrfs_block_group_flags(&cache->item);
10019 if (!mixed &&
10020 ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10021 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10022 btrfs_err(info,
10023 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10024 cache->key.objectid);
10025 ret = -EINVAL;
10026 goto error;
10027 }
10028
10029 key.objectid = found_key.objectid + found_key.offset;
10030 btrfs_release_path(path);
10031
10032 /*
10033 * We need to exclude the super stripes now so that the space
10034 * info has super bytes accounted for, otherwise we'll think
10035 * we have more space than we actually do.
10036 */
10037 ret = exclude_super_stripes(cache);
10038 if (ret) {
10039 /*
10040 * We may have excluded something, so call this just in
10041 * case.
10042 */
10043 free_excluded_extents(cache);
10044 btrfs_put_block_group(cache);
10045 goto error;
10046 }
10047
10048 /*
10049 * check for two cases, either we are full, and therefore
10050 * don't need to bother with the caching work since we won't
10051 * find any space, or we are empty, and we can just add all
10052 * the space in and be done with it. This saves us _alot_ of
10053 * time, particularly in the full case.
10054 */
10055 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10056 cache->last_byte_to_unpin = (u64)-1;
10057 cache->cached = BTRFS_CACHE_FINISHED;
10058 free_excluded_extents(cache);
10059 } else if (btrfs_block_group_used(&cache->item) == 0) {
10060 cache->last_byte_to_unpin = (u64)-1;
10061 cache->cached = BTRFS_CACHE_FINISHED;
10062 add_new_free_space(cache, found_key.objectid,
10063 found_key.objectid +
10064 found_key.offset);
10065 free_excluded_extents(cache);
10066 }
10067
10068 ret = btrfs_add_block_group_cache(info, cache);
10069 if (ret) {
10070 btrfs_remove_free_space_cache(cache);
10071 btrfs_put_block_group(cache);
10072 goto error;
10073 }
10074
10075 trace_btrfs_add_block_group(info, cache, 0);
10076 update_space_info(info, cache->flags, found_key.offset,
10077 btrfs_block_group_used(&cache->item),
10078 cache->bytes_super, &space_info);
10079
10080 cache->space_info = space_info;
10081
10082 link_block_group(cache);
10083
10084 set_avail_alloc_bits(info, cache->flags);
10085 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10086 inc_block_group_ro(cache, 1);
10087 } else if (btrfs_block_group_used(&cache->item) == 0) {
10088 ASSERT(list_empty(&cache->bg_list));
10089 btrfs_mark_bg_unused(cache);
10090 }
10091 }
10092
10093 list_for_each_entry_rcu(space_info, &info->space_info, list) {
10094 if (!(get_alloc_profile(info, space_info->flags) &
10095 (BTRFS_BLOCK_GROUP_RAID10 |
10096 BTRFS_BLOCK_GROUP_RAID1 |
10097 BTRFS_BLOCK_GROUP_RAID5 |
10098 BTRFS_BLOCK_GROUP_RAID6 |
10099 BTRFS_BLOCK_GROUP_DUP)))
10100 continue;
10101 /*
10102 * avoid allocating from un-mirrored block group if there are
10103 * mirrored block groups.
10104 */
10105 list_for_each_entry(cache,
10106 &space_info->block_groups[BTRFS_RAID_RAID0],
10107 list)
10108 inc_block_group_ro(cache, 1);
10109 list_for_each_entry(cache,
10110 &space_info->block_groups[BTRFS_RAID_SINGLE],
10111 list)
10112 inc_block_group_ro(cache, 1);
10113 }
10114
10115 btrfs_add_raid_kobjects(info);
10116 init_global_block_rsv(info);
10117 ret = check_chunk_block_group_mappings(info);
10118 error:
10119 btrfs_free_path(path);
10120 return ret;
10121 }
10122
10123 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10124 {
10125 struct btrfs_fs_info *fs_info = trans->fs_info;
10126 struct btrfs_block_group_cache *block_group;
10127 struct btrfs_root *extent_root = fs_info->extent_root;
10128 struct btrfs_block_group_item item;
10129 struct btrfs_key key;
10130 int ret = 0;
10131
10132 if (!trans->can_flush_pending_bgs)
10133 return;
10134
10135 while (!list_empty(&trans->new_bgs)) {
10136 block_group = list_first_entry(&trans->new_bgs,
10137 struct btrfs_block_group_cache,
10138 bg_list);
10139 if (ret)
10140 goto next;
10141
10142 spin_lock(&block_group->lock);
10143 memcpy(&item, &block_group->item, sizeof(item));
10144 memcpy(&key, &block_group->key, sizeof(key));
10145 spin_unlock(&block_group->lock);
10146
10147 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10148 sizeof(item));
10149 if (ret)
10150 btrfs_abort_transaction(trans, ret);
10151 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10152 if (ret)
10153 btrfs_abort_transaction(trans, ret);
10154 add_block_group_free_space(trans, block_group);
10155 /* already aborted the transaction if it failed. */
10156 next:
10157 list_del_init(&block_group->bg_list);
10158 }
10159 btrfs_trans_release_chunk_metadata(trans);
10160 }
10161
10162 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10163 u64 type, u64 chunk_offset, u64 size)
10164 {
10165 struct btrfs_fs_info *fs_info = trans->fs_info;
10166 struct btrfs_block_group_cache *cache;
10167 int ret;
10168
10169 btrfs_set_log_full_commit(fs_info, trans);
10170
10171 cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10172 if (!cache)
10173 return -ENOMEM;
10174
10175 btrfs_set_block_group_used(&cache->item, bytes_used);
10176 btrfs_set_block_group_chunk_objectid(&cache->item,
10177 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10178 btrfs_set_block_group_flags(&cache->item, type);
10179
10180 cache->flags = type;
10181 cache->last_byte_to_unpin = (u64)-1;
10182 cache->cached = BTRFS_CACHE_FINISHED;
10183 cache->needs_free_space = 1;
10184 ret = exclude_super_stripes(cache);
10185 if (ret) {
10186 /*
10187 * We may have excluded something, so call this just in
10188 * case.
10189 */
10190 free_excluded_extents(cache);
10191 btrfs_put_block_group(cache);
10192 return ret;
10193 }
10194
10195 add_new_free_space(cache, chunk_offset, chunk_offset + size);
10196
10197 free_excluded_extents(cache);
10198
10199 #ifdef CONFIG_BTRFS_DEBUG
10200 if (btrfs_should_fragment_free_space(cache)) {
10201 u64 new_bytes_used = size - bytes_used;
10202
10203 bytes_used += new_bytes_used >> 1;
10204 fragment_free_space(cache);
10205 }
10206 #endif
10207 /*
10208 * Ensure the corresponding space_info object is created and
10209 * assigned to our block group. We want our bg to be added to the rbtree
10210 * with its ->space_info set.
10211 */
10212 cache->space_info = __find_space_info(fs_info, cache->flags);
10213 ASSERT(cache->space_info);
10214
10215 ret = btrfs_add_block_group_cache(fs_info, cache);
10216 if (ret) {
10217 btrfs_remove_free_space_cache(cache);
10218 btrfs_put_block_group(cache);
10219 return ret;
10220 }
10221
10222 /*
10223 * Now that our block group has its ->space_info set and is inserted in
10224 * the rbtree, update the space info's counters.
10225 */
10226 trace_btrfs_add_block_group(fs_info, cache, 1);
10227 update_space_info(fs_info, cache->flags, size, bytes_used,
10228 cache->bytes_super, &cache->space_info);
10229 update_global_block_rsv(fs_info);
10230
10231 link_block_group(cache);
10232
10233 list_add_tail(&cache->bg_list, &trans->new_bgs);
10234
10235 set_avail_alloc_bits(fs_info, type);
10236 return 0;
10237 }
10238
10239 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10240 {
10241 u64 extra_flags = chunk_to_extended(flags) &
10242 BTRFS_EXTENDED_PROFILE_MASK;
10243
10244 write_seqlock(&fs_info->profiles_lock);
10245 if (flags & BTRFS_BLOCK_GROUP_DATA)
10246 fs_info->avail_data_alloc_bits &= ~extra_flags;
10247 if (flags & BTRFS_BLOCK_GROUP_METADATA)
10248 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10249 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10250 fs_info->avail_system_alloc_bits &= ~extra_flags;
10251 write_sequnlock(&fs_info->profiles_lock);
10252 }
10253
10254 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10255 u64 group_start, struct extent_map *em)
10256 {
10257 struct btrfs_fs_info *fs_info = trans->fs_info;
10258 struct btrfs_root *root = fs_info->extent_root;
10259 struct btrfs_path *path;
10260 struct btrfs_block_group_cache *block_group;
10261 struct btrfs_free_cluster *cluster;
10262 struct btrfs_root *tree_root = fs_info->tree_root;
10263 struct btrfs_key key;
10264 struct inode *inode;
10265 struct kobject *kobj = NULL;
10266 int ret;
10267 int index;
10268 int factor;
10269 struct btrfs_caching_control *caching_ctl = NULL;
10270 bool remove_em;
10271
10272 block_group = btrfs_lookup_block_group(fs_info, group_start);
10273 BUG_ON(!block_group);
10274 BUG_ON(!block_group->ro);
10275
10276 trace_btrfs_remove_block_group(block_group);
10277 /*
10278 * Free the reserved super bytes from this block group before
10279 * remove it.
10280 */
10281 free_excluded_extents(block_group);
10282 btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10283 block_group->key.offset);
10284
10285 memcpy(&key, &block_group->key, sizeof(key));
10286 index = btrfs_bg_flags_to_raid_index(block_group->flags);
10287 factor = btrfs_bg_type_to_factor(block_group->flags);
10288
10289 /* make sure this block group isn't part of an allocation cluster */
10290 cluster = &fs_info->data_alloc_cluster;
10291 spin_lock(&cluster->refill_lock);
10292 btrfs_return_cluster_to_free_space(block_group, cluster);
10293 spin_unlock(&cluster->refill_lock);
10294
10295 /*
10296 * make sure this block group isn't part of a metadata
10297 * allocation cluster
10298 */
10299 cluster = &fs_info->meta_alloc_cluster;
10300 spin_lock(&cluster->refill_lock);
10301 btrfs_return_cluster_to_free_space(block_group, cluster);
10302 spin_unlock(&cluster->refill_lock);
10303
10304 path = btrfs_alloc_path();
10305 if (!path) {
10306 ret = -ENOMEM;
10307 goto out;
10308 }
10309
10310 /*
10311 * get the inode first so any iput calls done for the io_list
10312 * aren't the final iput (no unlinks allowed now)
10313 */
10314 inode = lookup_free_space_inode(fs_info, block_group, path);
10315
10316 mutex_lock(&trans->transaction->cache_write_mutex);
10317 /*
10318 * make sure our free spache cache IO is done before remove the
10319 * free space inode
10320 */
10321 spin_lock(&trans->transaction->dirty_bgs_lock);
10322 if (!list_empty(&block_group->io_list)) {
10323 list_del_init(&block_group->io_list);
10324
10325 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10326
10327 spin_unlock(&trans->transaction->dirty_bgs_lock);
10328 btrfs_wait_cache_io(trans, block_group, path);
10329 btrfs_put_block_group(block_group);
10330 spin_lock(&trans->transaction->dirty_bgs_lock);
10331 }
10332
10333 if (!list_empty(&block_group->dirty_list)) {
10334 list_del_init(&block_group->dirty_list);
10335 btrfs_put_block_group(block_group);
10336 }
10337 spin_unlock(&trans->transaction->dirty_bgs_lock);
10338 mutex_unlock(&trans->transaction->cache_write_mutex);
10339
10340 if (!IS_ERR(inode)) {
10341 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10342 if (ret) {
10343 btrfs_add_delayed_iput(inode);
10344 goto out;
10345 }
10346 clear_nlink(inode);
10347 /* One for the block groups ref */
10348 spin_lock(&block_group->lock);
10349 if (block_group->iref) {
10350 block_group->iref = 0;
10351 block_group->inode = NULL;
10352 spin_unlock(&block_group->lock);
10353 iput(inode);
10354 } else {
10355 spin_unlock(&block_group->lock);
10356 }
10357 /* One for our lookup ref */
10358 btrfs_add_delayed_iput(inode);
10359 }
10360
10361 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10362 key.offset = block_group->key.objectid;
10363 key.type = 0;
10364
10365 ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10366 if (ret < 0)
10367 goto out;
10368 if (ret > 0)
10369 btrfs_release_path(path);
10370 if (ret == 0) {
10371 ret = btrfs_del_item(trans, tree_root, path);
10372 if (ret)
10373 goto out;
10374 btrfs_release_path(path);
10375 }
10376
10377 spin_lock(&fs_info->block_group_cache_lock);
10378 rb_erase(&block_group->cache_node,
10379 &fs_info->block_group_cache_tree);
10380 RB_CLEAR_NODE(&block_group->cache_node);
10381
10382 if (fs_info->first_logical_byte == block_group->key.objectid)
10383 fs_info->first_logical_byte = (u64)-1;
10384 spin_unlock(&fs_info->block_group_cache_lock);
10385
10386 down_write(&block_group->space_info->groups_sem);
10387 /*
10388 * we must use list_del_init so people can check to see if they
10389 * are still on the list after taking the semaphore
10390 */
10391 list_del_init(&block_group->list);
10392 if (list_empty(&block_group->space_info->block_groups[index])) {
10393 kobj = block_group->space_info->block_group_kobjs[index];
10394 block_group->space_info->block_group_kobjs[index] = NULL;
10395 clear_avail_alloc_bits(fs_info, block_group->flags);
10396 }
10397 up_write(&block_group->space_info->groups_sem);
10398 if (kobj) {
10399 kobject_del(kobj);
10400 kobject_put(kobj);
10401 }
10402
10403 if (block_group->has_caching_ctl)
10404 caching_ctl = get_caching_control(block_group);
10405 if (block_group->cached == BTRFS_CACHE_STARTED)
10406 wait_block_group_cache_done(block_group);
10407 if (block_group->has_caching_ctl) {
10408 down_write(&fs_info->commit_root_sem);
10409 if (!caching_ctl) {
10410 struct btrfs_caching_control *ctl;
10411
10412 list_for_each_entry(ctl,
10413 &fs_info->caching_block_groups, list)
10414 if (ctl->block_group == block_group) {
10415 caching_ctl = ctl;
10416 refcount_inc(&caching_ctl->count);
10417 break;
10418 }
10419 }
10420 if (caching_ctl)
10421 list_del_init(&caching_ctl->list);
10422 up_write(&fs_info->commit_root_sem);
10423 if (caching_ctl) {
10424 /* Once for the caching bgs list and once for us. */
10425 put_caching_control(caching_ctl);
10426 put_caching_control(caching_ctl);
10427 }
10428 }
10429
10430 spin_lock(&trans->transaction->dirty_bgs_lock);
10431 if (!list_empty(&block_group->dirty_list)) {
10432 WARN_ON(1);
10433 }
10434 if (!list_empty(&block_group->io_list)) {
10435 WARN_ON(1);
10436 }
10437 spin_unlock(&trans->transaction->dirty_bgs_lock);
10438 btrfs_remove_free_space_cache(block_group);
10439
10440 spin_lock(&block_group->space_info->lock);
10441 list_del_init(&block_group->ro_list);
10442
10443 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10444 WARN_ON(block_group->space_info->total_bytes
10445 < block_group->key.offset);
10446 WARN_ON(block_group->space_info->bytes_readonly
10447 < block_group->key.offset);
10448 WARN_ON(block_group->space_info->disk_total
10449 < block_group->key.offset * factor);
10450 }
10451 block_group->space_info->total_bytes -= block_group->key.offset;
10452 block_group->space_info->bytes_readonly -= block_group->key.offset;
10453 block_group->space_info->disk_total -= block_group->key.offset * factor;
10454
10455 spin_unlock(&block_group->space_info->lock);
10456
10457 memcpy(&key, &block_group->key, sizeof(key));
10458
10459 mutex_lock(&fs_info->chunk_mutex);
10460 if (!list_empty(&em->list)) {
10461 /* We're in the transaction->pending_chunks list. */
10462 free_extent_map(em);
10463 }
10464 spin_lock(&block_group->lock);
10465 block_group->removed = 1;
10466 /*
10467 * At this point trimming can't start on this block group, because we
10468 * removed the block group from the tree fs_info->block_group_cache_tree
10469 * so no one can't find it anymore and even if someone already got this
10470 * block group before we removed it from the rbtree, they have already
10471 * incremented block_group->trimming - if they didn't, they won't find
10472 * any free space entries because we already removed them all when we
10473 * called btrfs_remove_free_space_cache().
10474 *
10475 * And we must not remove the extent map from the fs_info->mapping_tree
10476 * to prevent the same logical address range and physical device space
10477 * ranges from being reused for a new block group. This is because our
10478 * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10479 * completely transactionless, so while it is trimming a range the
10480 * currently running transaction might finish and a new one start,
10481 * allowing for new block groups to be created that can reuse the same
10482 * physical device locations unless we take this special care.
10483 *
10484 * There may also be an implicit trim operation if the file system
10485 * is mounted with -odiscard. The same protections must remain
10486 * in place until the extents have been discarded completely when
10487 * the transaction commit has completed.
10488 */
10489 remove_em = (atomic_read(&block_group->trimming) == 0);
10490 /*
10491 * Make sure a trimmer task always sees the em in the pinned_chunks list
10492 * if it sees block_group->removed == 1 (needs to lock block_group->lock
10493 * before checking block_group->removed).
10494 */
10495 if (!remove_em) {
10496 /*
10497 * Our em might be in trans->transaction->pending_chunks which
10498 * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10499 * and so is the fs_info->pinned_chunks list.
10500 *
10501 * So at this point we must be holding the chunk_mutex to avoid
10502 * any races with chunk allocation (more specifically at
10503 * volumes.c:contains_pending_extent()), to ensure it always
10504 * sees the em, either in the pending_chunks list or in the
10505 * pinned_chunks list.
10506 */
10507 list_move_tail(&em->list, &fs_info->pinned_chunks);
10508 }
10509 spin_unlock(&block_group->lock);
10510
10511 if (remove_em) {
10512 struct extent_map_tree *em_tree;
10513
10514 em_tree = &fs_info->mapping_tree.map_tree;
10515 write_lock(&em_tree->lock);
10516 /*
10517 * The em might be in the pending_chunks list, so make sure the
10518 * chunk mutex is locked, since remove_extent_mapping() will
10519 * delete us from that list.
10520 */
10521 remove_extent_mapping(em_tree, em);
10522 write_unlock(&em_tree->lock);
10523 /* once for the tree */
10524 free_extent_map(em);
10525 }
10526
10527 mutex_unlock(&fs_info->chunk_mutex);
10528
10529 ret = remove_block_group_free_space(trans, block_group);
10530 if (ret)
10531 goto out;
10532
10533 btrfs_put_block_group(block_group);
10534 btrfs_put_block_group(block_group);
10535
10536 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10537 if (ret > 0)
10538 ret = -EIO;
10539 if (ret < 0)
10540 goto out;
10541
10542 ret = btrfs_del_item(trans, root, path);
10543 out:
10544 btrfs_free_path(path);
10545 return ret;
10546 }
10547
10548 struct btrfs_trans_handle *
10549 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10550 const u64 chunk_offset)
10551 {
10552 struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10553 struct extent_map *em;
10554 struct map_lookup *map;
10555 unsigned int num_items;
10556
10557 read_lock(&em_tree->lock);
10558 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10559 read_unlock(&em_tree->lock);
10560 ASSERT(em && em->start == chunk_offset);
10561
10562 /*
10563 * We need to reserve 3 + N units from the metadata space info in order
10564 * to remove a block group (done at btrfs_remove_chunk() and at
10565 * btrfs_remove_block_group()), which are used for:
10566 *
10567 * 1 unit for adding the free space inode's orphan (located in the tree
10568 * of tree roots).
10569 * 1 unit for deleting the block group item (located in the extent
10570 * tree).
10571 * 1 unit for deleting the free space item (located in tree of tree
10572 * roots).
10573 * N units for deleting N device extent items corresponding to each
10574 * stripe (located in the device tree).
10575 *
10576 * In order to remove a block group we also need to reserve units in the
10577 * system space info in order to update the chunk tree (update one or
10578 * more device items and remove one chunk item), but this is done at
10579 * btrfs_remove_chunk() through a call to check_system_chunk().
10580 */
10581 map = em->map_lookup;
10582 num_items = 3 + map->num_stripes;
10583 free_extent_map(em);
10584
10585 return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10586 num_items, 1);
10587 }
10588
10589 /*
10590 * Process the unused_bgs list and remove any that don't have any allocated
10591 * space inside of them.
10592 */
10593 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10594 {
10595 struct btrfs_block_group_cache *block_group;
10596 struct btrfs_space_info *space_info;
10597 struct btrfs_trans_handle *trans;
10598 int ret = 0;
10599
10600 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10601 return;
10602
10603 spin_lock(&fs_info->unused_bgs_lock);
10604 while (!list_empty(&fs_info->unused_bgs)) {
10605 u64 start, end;
10606 int trimming;
10607
10608 block_group = list_first_entry(&fs_info->unused_bgs,
10609 struct btrfs_block_group_cache,
10610 bg_list);
10611 list_del_init(&block_group->bg_list);
10612
10613 space_info = block_group->space_info;
10614
10615 if (ret || btrfs_mixed_space_info(space_info)) {
10616 btrfs_put_block_group(block_group);
10617 continue;
10618 }
10619 spin_unlock(&fs_info->unused_bgs_lock);
10620
10621 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10622
10623 /* Don't want to race with allocators so take the groups_sem */
10624 down_write(&space_info->groups_sem);
10625 spin_lock(&block_group->lock);
10626 if (block_group->reserved || block_group->pinned ||
10627 btrfs_block_group_used(&block_group->item) ||
10628 block_group->ro ||
10629 list_is_singular(&block_group->list)) {
10630 /*
10631 * We want to bail if we made new allocations or have
10632 * outstanding allocations in this block group. We do
10633 * the ro check in case balance is currently acting on
10634 * this block group.
10635 */
10636 trace_btrfs_skip_unused_block_group(block_group);
10637 spin_unlock(&block_group->lock);
10638 up_write(&space_info->groups_sem);
10639 goto next;
10640 }
10641 spin_unlock(&block_group->lock);
10642
10643 /* We don't want to force the issue, only flip if it's ok. */
10644 ret = inc_block_group_ro(block_group, 0);
10645 up_write(&space_info->groups_sem);
10646 if (ret < 0) {
10647 ret = 0;
10648 goto next;
10649 }
10650
10651 /*
10652 * Want to do this before we do anything else so we can recover
10653 * properly if we fail to join the transaction.
10654 */
10655 trans = btrfs_start_trans_remove_block_group(fs_info,
10656 block_group->key.objectid);
10657 if (IS_ERR(trans)) {
10658 btrfs_dec_block_group_ro(block_group);
10659 ret = PTR_ERR(trans);
10660 goto next;
10661 }
10662
10663 /*
10664 * We could have pending pinned extents for this block group,
10665 * just delete them, we don't care about them anymore.
10666 */
10667 start = block_group->key.objectid;
10668 end = start + block_group->key.offset - 1;
10669 /*
10670 * Hold the unused_bg_unpin_mutex lock to avoid racing with
10671 * btrfs_finish_extent_commit(). If we are at transaction N,
10672 * another task might be running finish_extent_commit() for the
10673 * previous transaction N - 1, and have seen a range belonging
10674 * to the block group in freed_extents[] before we were able to
10675 * clear the whole block group range from freed_extents[]. This
10676 * means that task can lookup for the block group after we
10677 * unpinned it from freed_extents[] and removed it, leading to
10678 * a BUG_ON() at btrfs_unpin_extent_range().
10679 */
10680 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10681 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10682 EXTENT_DIRTY);
10683 if (ret) {
10684 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10685 btrfs_dec_block_group_ro(block_group);
10686 goto end_trans;
10687 }
10688 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10689 EXTENT_DIRTY);
10690 if (ret) {
10691 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10692 btrfs_dec_block_group_ro(block_group);
10693 goto end_trans;
10694 }
10695 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10696
10697 /* Reset pinned so btrfs_put_block_group doesn't complain */
10698 spin_lock(&space_info->lock);
10699 spin_lock(&block_group->lock);
10700
10701 space_info->bytes_pinned -= block_group->pinned;
10702 space_info->bytes_readonly += block_group->pinned;
10703 percpu_counter_add_batch(&space_info->total_bytes_pinned,
10704 -block_group->pinned,
10705 BTRFS_TOTAL_BYTES_PINNED_BATCH);
10706 block_group->pinned = 0;
10707
10708 spin_unlock(&block_group->lock);
10709 spin_unlock(&space_info->lock);
10710
10711 /* DISCARD can flip during remount */
10712 trimming = btrfs_test_opt(fs_info, DISCARD);
10713
10714 /* Implicit trim during transaction commit. */
10715 if (trimming)
10716 btrfs_get_block_group_trimming(block_group);
10717
10718 /*
10719 * Btrfs_remove_chunk will abort the transaction if things go
10720 * horribly wrong.
10721 */
10722 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
10723
10724 if (ret) {
10725 if (trimming)
10726 btrfs_put_block_group_trimming(block_group);
10727 goto end_trans;
10728 }
10729
10730 /*
10731 * If we're not mounted with -odiscard, we can just forget
10732 * about this block group. Otherwise we'll need to wait
10733 * until transaction commit to do the actual discard.
10734 */
10735 if (trimming) {
10736 spin_lock(&fs_info->unused_bgs_lock);
10737 /*
10738 * A concurrent scrub might have added us to the list
10739 * fs_info->unused_bgs, so use a list_move operation
10740 * to add the block group to the deleted_bgs list.
10741 */
10742 list_move(&block_group->bg_list,
10743 &trans->transaction->deleted_bgs);
10744 spin_unlock(&fs_info->unused_bgs_lock);
10745 btrfs_get_block_group(block_group);
10746 }
10747 end_trans:
10748 btrfs_end_transaction(trans);
10749 next:
10750 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10751 btrfs_put_block_group(block_group);
10752 spin_lock(&fs_info->unused_bgs_lock);
10753 }
10754 spin_unlock(&fs_info->unused_bgs_lock);
10755 }
10756
10757 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10758 {
10759 struct btrfs_super_block *disk_super;
10760 u64 features;
10761 u64 flags;
10762 int mixed = 0;
10763 int ret;
10764
10765 disk_super = fs_info->super_copy;
10766 if (!btrfs_super_root(disk_super))
10767 return -EINVAL;
10768
10769 features = btrfs_super_incompat_flags(disk_super);
10770 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10771 mixed = 1;
10772
10773 flags = BTRFS_BLOCK_GROUP_SYSTEM;
10774 ret = create_space_info(fs_info, flags);
10775 if (ret)
10776 goto out;
10777
10778 if (mixed) {
10779 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10780 ret = create_space_info(fs_info, flags);
10781 } else {
10782 flags = BTRFS_BLOCK_GROUP_METADATA;
10783 ret = create_space_info(fs_info, flags);
10784 if (ret)
10785 goto out;
10786
10787 flags = BTRFS_BLOCK_GROUP_DATA;
10788 ret = create_space_info(fs_info, flags);
10789 }
10790 out:
10791 return ret;
10792 }
10793
10794 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10795 u64 start, u64 end)
10796 {
10797 return unpin_extent_range(fs_info, start, end, false);
10798 }
10799
10800 /*
10801 * It used to be that old block groups would be left around forever.
10802 * Iterating over them would be enough to trim unused space. Since we
10803 * now automatically remove them, we also need to iterate over unallocated
10804 * space.
10805 *
10806 * We don't want a transaction for this since the discard may take a
10807 * substantial amount of time. We don't require that a transaction be
10808 * running, but we do need to take a running transaction into account
10809 * to ensure that we're not discarding chunks that were released or
10810 * allocated in the current transaction.
10811 *
10812 * Holding the chunks lock will prevent other threads from allocating
10813 * or releasing chunks, but it won't prevent a running transaction
10814 * from committing and releasing the memory that the pending chunks
10815 * list head uses. For that, we need to take a reference to the
10816 * transaction and hold the commit root sem. We only need to hold
10817 * it while performing the free space search since we have already
10818 * held back allocations.
10819 */
10820 static int btrfs_trim_free_extents(struct btrfs_device *device,
10821 u64 minlen, u64 *trimmed)
10822 {
10823 u64 start = 0, len = 0;
10824 int ret;
10825
10826 *trimmed = 0;
10827
10828 /* Discard not supported = nothing to do. */
10829 if (!blk_queue_discard(bdev_get_queue(device->bdev)))
10830 return 0;
10831
10832 /* Not writeable = nothing to do. */
10833 if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10834 return 0;
10835
10836 /* No free space = nothing to do. */
10837 if (device->total_bytes <= device->bytes_used)
10838 return 0;
10839
10840 ret = 0;
10841
10842 while (1) {
10843 struct btrfs_fs_info *fs_info = device->fs_info;
10844 struct btrfs_transaction *trans;
10845 u64 bytes;
10846
10847 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10848 if (ret)
10849 break;
10850
10851 ret = down_read_killable(&fs_info->commit_root_sem);
10852 if (ret) {
10853 mutex_unlock(&fs_info->chunk_mutex);
10854 break;
10855 }
10856
10857 spin_lock(&fs_info->trans_lock);
10858 trans = fs_info->running_transaction;
10859 if (trans)
10860 refcount_inc(&trans->use_count);
10861 spin_unlock(&fs_info->trans_lock);
10862
10863 if (!trans)
10864 up_read(&fs_info->commit_root_sem);
10865
10866 ret = find_free_dev_extent_start(trans, device, minlen, start,
10867 &start, &len);
10868 if (trans) {
10869 up_read(&fs_info->commit_root_sem);
10870 btrfs_put_transaction(trans);
10871 }
10872
10873 if (ret) {
10874 mutex_unlock(&fs_info->chunk_mutex);
10875 if (ret == -ENOSPC)
10876 ret = 0;
10877 break;
10878 }
10879
10880 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10881 mutex_unlock(&fs_info->chunk_mutex);
10882
10883 if (ret)
10884 break;
10885
10886 start += len;
10887 *trimmed += bytes;
10888
10889 if (fatal_signal_pending(current)) {
10890 ret = -ERESTARTSYS;
10891 break;
10892 }
10893
10894 cond_resched();
10895 }
10896
10897 return ret;
10898 }
10899
10900 /*
10901 * Trim the whole filesystem by:
10902 * 1) trimming the free space in each block group
10903 * 2) trimming the unallocated space on each device
10904 *
10905 * This will also continue trimming even if a block group or device encounters
10906 * an error. The return value will be the last error, or 0 if nothing bad
10907 * happens.
10908 */
10909 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10910 {
10911 struct btrfs_block_group_cache *cache = NULL;
10912 struct btrfs_device *device;
10913 struct list_head *devices;
10914 u64 group_trimmed;
10915 u64 start;
10916 u64 end;
10917 u64 trimmed = 0;
10918 u64 bg_failed = 0;
10919 u64 dev_failed = 0;
10920 int bg_ret = 0;
10921 int dev_ret = 0;
10922 int ret = 0;
10923
10924 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10925 for (; cache; cache = next_block_group(fs_info, cache)) {
10926 if (cache->key.objectid >= (range->start + range->len)) {
10927 btrfs_put_block_group(cache);
10928 break;
10929 }
10930
10931 start = max(range->start, cache->key.objectid);
10932 end = min(range->start + range->len,
10933 cache->key.objectid + cache->key.offset);
10934
10935 if (end - start >= range->minlen) {
10936 if (!block_group_cache_done(cache)) {
10937 ret = cache_block_group(cache, 0);
10938 if (ret) {
10939 bg_failed++;
10940 bg_ret = ret;
10941 continue;
10942 }
10943 ret = wait_block_group_cache_done(cache);
10944 if (ret) {
10945 bg_failed++;
10946 bg_ret = ret;
10947 continue;
10948 }
10949 }
10950 ret = btrfs_trim_block_group(cache,
10951 &group_trimmed,
10952 start,
10953 end,
10954 range->minlen);
10955
10956 trimmed += group_trimmed;
10957 if (ret) {
10958 bg_failed++;
10959 bg_ret = ret;
10960 continue;
10961 }
10962 }
10963 }
10964
10965 if (bg_failed)
10966 btrfs_warn(fs_info,
10967 "failed to trim %llu block group(s), last error %d",
10968 bg_failed, bg_ret);
10969 mutex_lock(&fs_info->fs_devices->device_list_mutex);
10970 devices = &fs_info->fs_devices->devices;
10971 list_for_each_entry(device, devices, dev_list) {
10972 ret = btrfs_trim_free_extents(device, range->minlen,
10973 &group_trimmed);
10974 if (ret) {
10975 dev_failed++;
10976 dev_ret = ret;
10977 break;
10978 }
10979
10980 trimmed += group_trimmed;
10981 }
10982 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
10983
10984 if (dev_failed)
10985 btrfs_warn(fs_info,
10986 "failed to trim %llu device(s), last error %d",
10987 dev_failed, dev_ret);
10988 range->len = trimmed;
10989 if (bg_ret)
10990 return bg_ret;
10991 return dev_ret;
10992 }
10993
10994 /*
10995 * btrfs_{start,end}_write_no_snapshotting() are similar to
10996 * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10997 * data into the page cache through nocow before the subvolume is snapshoted,
10998 * but flush the data into disk after the snapshot creation, or to prevent
10999 * operations while snapshotting is ongoing and that cause the snapshot to be
11000 * inconsistent (writes followed by expanding truncates for example).
11001 */
11002 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11003 {
11004 percpu_counter_dec(&root->subv_writers->counter);
11005 cond_wake_up(&root->subv_writers->wait);
11006 }
11007
11008 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11009 {
11010 if (atomic_read(&root->will_be_snapshotted))
11011 return 0;
11012
11013 percpu_counter_inc(&root->subv_writers->counter);
11014 /*
11015 * Make sure counter is updated before we check for snapshot creation.
11016 */
11017 smp_mb();
11018 if (atomic_read(&root->will_be_snapshotted)) {
11019 btrfs_end_write_no_snapshotting(root);
11020 return 0;
11021 }
11022 return 1;
11023 }
11024
11025 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11026 {
11027 while (true) {
11028 int ret;
11029
11030 ret = btrfs_start_write_no_snapshotting(root);
11031 if (ret)
11032 break;
11033 wait_var_event(&root->will_be_snapshotted,
11034 !atomic_read(&root->will_be_snapshotted));
11035 }
11036 }
11037
11038 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11039 {
11040 struct btrfs_fs_info *fs_info = bg->fs_info;
11041
11042 spin_lock(&fs_info->unused_bgs_lock);
11043 if (list_empty(&bg->bg_list)) {
11044 btrfs_get_block_group(bg);
11045 trace_btrfs_add_unused_block_group(bg);
11046 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11047 }
11048 spin_unlock(&fs_info->unused_bgs_lock);
11049 }
Linux with added FPC-III support