Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include "misc.h"
4#include "ctree.h"
5#include "block-rsv.h"
6#include "space-info.h"
7#include "transaction.h"
8#include "block-group.h"
9#include "disk-io.h"
10#include "fs.h"
11#include "accessors.h"
12
13/*
14 * HOW DO BLOCK RESERVES WORK
15 *
16 * Think of block_rsv's as buckets for logically grouped metadata
17 * reservations. Each block_rsv has a ->size and a ->reserved. ->size is
18 * how large we want our block rsv to be, ->reserved is how much space is
19 * currently reserved for this block reserve.
20 *
21 * ->failfast exists for the truncate case, and is described below.
22 *
23 * NORMAL OPERATION
24 *
25 * -> Reserve
26 * Entrance: btrfs_block_rsv_add, btrfs_block_rsv_refill
27 *
28 * We call into btrfs_reserve_metadata_bytes() with our bytes, which is
29 * accounted for in space_info->bytes_may_use, and then add the bytes to
30 * ->reserved, and ->size in the case of btrfs_block_rsv_add.
31 *
32 * ->size is an over-estimation of how much we may use for a particular
33 * operation.
34 *
35 * -> Use
36 * Entrance: btrfs_use_block_rsv
37 *
38 * When we do a btrfs_alloc_tree_block() we call into btrfs_use_block_rsv()
39 * to determine the appropriate block_rsv to use, and then verify that
40 * ->reserved has enough space for our tree block allocation. Once
41 * successful we subtract fs_info->nodesize from ->reserved.
42 *
43 * -> Finish
44 * Entrance: btrfs_block_rsv_release
45 *
46 * We are finished with our operation, subtract our individual reservation
47 * from ->size, and then subtract ->size from ->reserved and free up the
48 * excess if there is any.
49 *
50 * There is some logic here to refill the delayed refs rsv or the global rsv
51 * as needed, otherwise the excess is subtracted from
52 * space_info->bytes_may_use.
53 *
54 * TYPES OF BLOCK RESERVES
55 *
56 * BLOCK_RSV_TRANS, BLOCK_RSV_DELOPS, BLOCK_RSV_CHUNK
57 * These behave normally, as described above, just within the confines of the
58 * lifetime of their particular operation (transaction for the whole trans
59 * handle lifetime, for example).
60 *
61 * BLOCK_RSV_GLOBAL
62 * It is impossible to properly account for all the space that may be required
63 * to make our extent tree updates. This block reserve acts as an overflow
64 * buffer in case our delayed refs reserve does not reserve enough space to
65 * update the extent tree.
66 *
67 * We can steal from this in some cases as well, notably on evict() or
68 * truncate() in order to help users recover from ENOSPC conditions.
69 *
70 * BLOCK_RSV_DELALLOC
71 * The individual item sizes are determined by the per-inode size
72 * calculations, which are described with the delalloc code. This is pretty
73 * straightforward, it's just the calculation of ->size encodes a lot of
74 * different items, and thus it gets used when updating inodes, inserting file
75 * extents, and inserting checksums.
76 *
77 * BLOCK_RSV_DELREFS
78 * We keep a running tally of how many delayed refs we have on the system.
79 * We assume each one of these delayed refs are going to use a full
80 * reservation. We use the transaction items and pre-reserve space for every
81 * operation, and use this reservation to refill any gap between ->size and
82 * ->reserved that may exist.
83 *
84 * From there it's straightforward, removing a delayed ref means we remove its
85 * count from ->size and free up reservations as necessary. Since this is
86 * the most dynamic block reserve in the system, we will try to refill this
87 * block reserve first with any excess returned by any other block reserve.
88 *
89 * BLOCK_RSV_EMPTY
90 * This is the fallback block reserve to make us try to reserve space if we
91 * don't have a specific bucket for this allocation. It is mostly used for
92 * updating the device tree and such, since that is a separate pool we're
93 * content to just reserve space from the space_info on demand.
94 *
95 * BLOCK_RSV_TEMP
96 * This is used by things like truncate and iput. We will temporarily
97 * allocate a block reserve, set it to some size, and then truncate bytes
98 * until we have no space left. With ->failfast set we'll simply return
99 * ENOSPC from btrfs_use_block_rsv() to signal that we need to unwind and try
100 * to make a new reservation. This is because these operations are
101 * unbounded, so we want to do as much work as we can, and then back off and
102 * re-reserve.
103 */
104
105static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
106 struct btrfs_block_rsv *block_rsv,
107 struct btrfs_block_rsv *dest, u64 num_bytes,
108 u64 *qgroup_to_release_ret)
109{
110 struct btrfs_space_info *space_info = block_rsv->space_info;
111 u64 qgroup_to_release = 0;
112 u64 ret;
113
114 spin_lock(&block_rsv->lock);
115 if (num_bytes == (u64)-1) {
116 num_bytes = block_rsv->size;
117 qgroup_to_release = block_rsv->qgroup_rsv_size;
118 }
119 block_rsv->size -= num_bytes;
120 if (block_rsv->reserved >= block_rsv->size) {
121 num_bytes = block_rsv->reserved - block_rsv->size;
122 block_rsv->reserved = block_rsv->size;
123 block_rsv->full = true;
124 } else {
125 num_bytes = 0;
126 }
127 if (qgroup_to_release_ret &&
128 block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
129 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
130 block_rsv->qgroup_rsv_size;
131 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
132 } else {
133 qgroup_to_release = 0;
134 }
135 spin_unlock(&block_rsv->lock);
136
137 ret = num_bytes;
138 if (num_bytes > 0) {
139 if (dest) {
140 spin_lock(&dest->lock);
141 if (!dest->full) {
142 u64 bytes_to_add;
143
144 bytes_to_add = dest->size - dest->reserved;
145 bytes_to_add = min(num_bytes, bytes_to_add);
146 dest->reserved += bytes_to_add;
147 if (dest->reserved >= dest->size)
148 dest->full = true;
149 num_bytes -= bytes_to_add;
150 }
151 spin_unlock(&dest->lock);
152 }
153 if (num_bytes)
154 btrfs_space_info_free_bytes_may_use(fs_info,
155 space_info,
156 num_bytes);
157 }
158 if (qgroup_to_release_ret)
159 *qgroup_to_release_ret = qgroup_to_release;
160 return ret;
161}
162
163int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
164 struct btrfs_block_rsv *dst, u64 num_bytes,
165 bool update_size)
166{
167 int ret;
168
169 ret = btrfs_block_rsv_use_bytes(src, num_bytes);
170 if (ret)
171 return ret;
172
173 btrfs_block_rsv_add_bytes(dst, num_bytes, update_size);
174 return 0;
175}
176
177void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, enum btrfs_rsv_type type)
178{
179 memset(rsv, 0, sizeof(*rsv));
180 spin_lock_init(&rsv->lock);
181 rsv->type = type;
182}
183
184void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
185 struct btrfs_block_rsv *rsv,
186 enum btrfs_rsv_type type)
187{
188 btrfs_init_block_rsv(rsv, type);
189 rsv->space_info = btrfs_find_space_info(fs_info,
190 BTRFS_BLOCK_GROUP_METADATA);
191}
192
193struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
194 enum btrfs_rsv_type type)
195{
196 struct btrfs_block_rsv *block_rsv;
197
198 block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
199 if (!block_rsv)
200 return NULL;
201
202 btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
203 return block_rsv;
204}
205
206void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
207 struct btrfs_block_rsv *rsv)
208{
209 if (!rsv)
210 return;
211 btrfs_block_rsv_release(fs_info, rsv, (u64)-1, NULL);
212 kfree(rsv);
213}
214
215int btrfs_block_rsv_add(struct btrfs_fs_info *fs_info,
216 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
217 enum btrfs_reserve_flush_enum flush)
218{
219 int ret;
220
221 if (num_bytes == 0)
222 return 0;
223
224 ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv->space_info,
225 num_bytes, flush);
226 if (!ret)
227 btrfs_block_rsv_add_bytes(block_rsv, num_bytes, true);
228
229 return ret;
230}
231
232int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_percent)
233{
234 u64 num_bytes = 0;
235 int ret = -ENOSPC;
236
237 spin_lock(&block_rsv->lock);
238 num_bytes = mult_perc(block_rsv->size, min_percent);
239 if (block_rsv->reserved >= num_bytes)
240 ret = 0;
241 spin_unlock(&block_rsv->lock);
242
243 return ret;
244}
245
246int btrfs_block_rsv_refill(struct btrfs_fs_info *fs_info,
247 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
248 enum btrfs_reserve_flush_enum flush)
249{
250 int ret = -ENOSPC;
251
252 if (!block_rsv)
253 return 0;
254
255 spin_lock(&block_rsv->lock);
256 if (block_rsv->reserved >= num_bytes)
257 ret = 0;
258 else
259 num_bytes -= block_rsv->reserved;
260 spin_unlock(&block_rsv->lock);
261
262 if (!ret)
263 return 0;
264
265 ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv->space_info,
266 num_bytes, flush);
267 if (!ret) {
268 btrfs_block_rsv_add_bytes(block_rsv, num_bytes, false);
269 return 0;
270 }
271
272 return ret;
273}
274
275u64 btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
276 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
277 u64 *qgroup_to_release)
278{
279 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
280 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
281 struct btrfs_block_rsv *target = NULL;
282
283 /*
284 * If we are a delayed block reserve then push to the global rsv,
285 * otherwise dump into the global delayed reserve if it is not full.
286 */
287 if (block_rsv->type == BTRFS_BLOCK_RSV_DELOPS)
288 target = global_rsv;
289 else if (block_rsv != global_rsv && !btrfs_block_rsv_full(delayed_rsv))
290 target = delayed_rsv;
291
292 if (target && block_rsv->space_info != target->space_info)
293 target = NULL;
294
295 return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
296 qgroup_to_release);
297}
298
299int btrfs_block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv, u64 num_bytes)
300{
301 int ret = -ENOSPC;
302
303 spin_lock(&block_rsv->lock);
304 if (block_rsv->reserved >= num_bytes) {
305 block_rsv->reserved -= num_bytes;
306 if (block_rsv->reserved < block_rsv->size)
307 block_rsv->full = false;
308 ret = 0;
309 }
310 spin_unlock(&block_rsv->lock);
311 return ret;
312}
313
314void btrfs_block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
315 u64 num_bytes, bool update_size)
316{
317 spin_lock(&block_rsv->lock);
318 block_rsv->reserved += num_bytes;
319 if (update_size)
320 block_rsv->size += num_bytes;
321 else if (block_rsv->reserved >= block_rsv->size)
322 block_rsv->full = true;
323 spin_unlock(&block_rsv->lock);
324}
325
326void btrfs_update_global_block_rsv(struct btrfs_fs_info *fs_info)
327{
328 struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
329 struct btrfs_space_info *sinfo = block_rsv->space_info;
330 struct btrfs_root *root, *tmp;
331 u64 num_bytes = btrfs_root_used(&fs_info->tree_root->root_item);
332 unsigned int min_items = 1;
333
334 /*
335 * The global block rsv is based on the size of the extent tree, the
336 * checksum tree and the root tree. If the fs is empty we want to set
337 * it to a minimal amount for safety.
338 *
339 * We also are going to need to modify the minimum of the tree root and
340 * any global roots we could touch.
341 */
342 read_lock(&fs_info->global_root_lock);
343 rbtree_postorder_for_each_entry_safe(root, tmp, &fs_info->global_root_tree,
344 rb_node) {
345 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID ||
346 root->root_key.objectid == BTRFS_CSUM_TREE_OBJECTID ||
347 root->root_key.objectid == BTRFS_FREE_SPACE_TREE_OBJECTID) {
348 num_bytes += btrfs_root_used(&root->root_item);
349 min_items++;
350 }
351 }
352 read_unlock(&fs_info->global_root_lock);
353
354 if (btrfs_fs_compat_ro(fs_info, BLOCK_GROUP_TREE)) {
355 num_bytes += btrfs_root_used(&fs_info->block_group_root->root_item);
356 min_items++;
357 }
358
359 if (btrfs_fs_incompat(fs_info, RAID_STRIPE_TREE)) {
360 num_bytes += btrfs_root_used(&fs_info->stripe_root->root_item);
361 min_items++;
362 }
363
364 /*
365 * But we also want to reserve enough space so we can do the fallback
366 * global reserve for an unlink, which is an additional
367 * BTRFS_UNLINK_METADATA_UNITS items.
368 *
369 * But we also need space for the delayed ref updates from the unlink,
370 * so add BTRFS_UNLINK_METADATA_UNITS units for delayed refs, one for
371 * each unlink metadata item.
372 */
373 min_items += BTRFS_UNLINK_METADATA_UNITS;
374
375 num_bytes = max_t(u64, num_bytes,
376 btrfs_calc_insert_metadata_size(fs_info, min_items) +
377 btrfs_calc_delayed_ref_bytes(fs_info,
378 BTRFS_UNLINK_METADATA_UNITS));
379
380 spin_lock(&sinfo->lock);
381 spin_lock(&block_rsv->lock);
382
383 block_rsv->size = min_t(u64, num_bytes, SZ_512M);
384
385 if (block_rsv->reserved < block_rsv->size) {
386 num_bytes = block_rsv->size - block_rsv->reserved;
387 btrfs_space_info_update_bytes_may_use(fs_info, sinfo,
388 num_bytes);
389 block_rsv->reserved = block_rsv->size;
390 } else if (block_rsv->reserved > block_rsv->size) {
391 num_bytes = block_rsv->reserved - block_rsv->size;
392 btrfs_space_info_update_bytes_may_use(fs_info, sinfo,
393 -num_bytes);
394 block_rsv->reserved = block_rsv->size;
395 btrfs_try_granting_tickets(fs_info, sinfo);
396 }
397
398 block_rsv->full = (block_rsv->reserved == block_rsv->size);
399
400 if (block_rsv->size >= sinfo->total_bytes)
401 sinfo->force_alloc = CHUNK_ALLOC_FORCE;
402 spin_unlock(&block_rsv->lock);
403 spin_unlock(&sinfo->lock);
404}
405
406void btrfs_init_root_block_rsv(struct btrfs_root *root)
407{
408 struct btrfs_fs_info *fs_info = root->fs_info;
409
410 switch (root->root_key.objectid) {
411 case BTRFS_CSUM_TREE_OBJECTID:
412 case BTRFS_EXTENT_TREE_OBJECTID:
413 case BTRFS_FREE_SPACE_TREE_OBJECTID:
414 case BTRFS_BLOCK_GROUP_TREE_OBJECTID:
415 case BTRFS_RAID_STRIPE_TREE_OBJECTID:
416 root->block_rsv = &fs_info->delayed_refs_rsv;
417 break;
418 case BTRFS_ROOT_TREE_OBJECTID:
419 case BTRFS_DEV_TREE_OBJECTID:
420 case BTRFS_QUOTA_TREE_OBJECTID:
421 root->block_rsv = &fs_info->global_block_rsv;
422 break;
423 case BTRFS_CHUNK_TREE_OBJECTID:
424 root->block_rsv = &fs_info->chunk_block_rsv;
425 break;
426 default:
427 root->block_rsv = NULL;
428 break;
429 }
430}
431
432void btrfs_init_global_block_rsv(struct btrfs_fs_info *fs_info)
433{
434 struct btrfs_space_info *space_info;
435
436 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
437 fs_info->chunk_block_rsv.space_info = space_info;
438
439 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
440 fs_info->global_block_rsv.space_info = space_info;
441 fs_info->trans_block_rsv.space_info = space_info;
442 fs_info->empty_block_rsv.space_info = space_info;
443 fs_info->delayed_block_rsv.space_info = space_info;
444 fs_info->delayed_refs_rsv.space_info = space_info;
445
446 btrfs_update_global_block_rsv(fs_info);
447}
448
449void btrfs_release_global_block_rsv(struct btrfs_fs_info *fs_info)
450{
451 btrfs_block_rsv_release(fs_info, &fs_info->global_block_rsv, (u64)-1,
452 NULL);
453 WARN_ON(fs_info->trans_block_rsv.size > 0);
454 WARN_ON(fs_info->trans_block_rsv.reserved > 0);
455 WARN_ON(fs_info->chunk_block_rsv.size > 0);
456 WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
457 WARN_ON(fs_info->delayed_block_rsv.size > 0);
458 WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
459 WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
460 WARN_ON(fs_info->delayed_refs_rsv.size > 0);
461}
462
463static struct btrfs_block_rsv *get_block_rsv(
464 const struct btrfs_trans_handle *trans,
465 const struct btrfs_root *root)
466{
467 struct btrfs_fs_info *fs_info = root->fs_info;
468 struct btrfs_block_rsv *block_rsv = NULL;
469
470 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
471 (root == fs_info->uuid_root) ||
472 (trans->adding_csums &&
473 root->root_key.objectid == BTRFS_CSUM_TREE_OBJECTID))
474 block_rsv = trans->block_rsv;
475
476 if (!block_rsv)
477 block_rsv = root->block_rsv;
478
479 if (!block_rsv)
480 block_rsv = &fs_info->empty_block_rsv;
481
482 return block_rsv;
483}
484
485struct btrfs_block_rsv *btrfs_use_block_rsv(struct btrfs_trans_handle *trans,
486 struct btrfs_root *root,
487 u32 blocksize)
488{
489 struct btrfs_fs_info *fs_info = root->fs_info;
490 struct btrfs_block_rsv *block_rsv;
491 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
492 int ret;
493 bool global_updated = false;
494
495 block_rsv = get_block_rsv(trans, root);
496
497 if (unlikely(btrfs_block_rsv_size(block_rsv) == 0))
498 goto try_reserve;
499again:
500 ret = btrfs_block_rsv_use_bytes(block_rsv, blocksize);
501 if (!ret)
502 return block_rsv;
503
504 if (block_rsv->failfast)
505 return ERR_PTR(ret);
506
507 if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
508 global_updated = true;
509 btrfs_update_global_block_rsv(fs_info);
510 goto again;
511 }
512
513 /*
514 * The global reserve still exists to save us from ourselves, so don't
515 * warn_on if we are short on our delayed refs reserve.
516 */
517 if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
518 btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
519 static DEFINE_RATELIMIT_STATE(_rs,
520 DEFAULT_RATELIMIT_INTERVAL * 10,
521 /*DEFAULT_RATELIMIT_BURST*/ 1);
522 if (__ratelimit(&_rs))
523 WARN(1, KERN_DEBUG
524 "BTRFS: block rsv %d returned %d\n",
525 block_rsv->type, ret);
526 }
527try_reserve:
528 ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv->space_info,
529 blocksize, BTRFS_RESERVE_NO_FLUSH);
530 if (!ret)
531 return block_rsv;
532 /*
533 * If we couldn't reserve metadata bytes try and use some from
534 * the global reserve if its space type is the same as the global
535 * reservation.
536 */
537 if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
538 block_rsv->space_info == global_rsv->space_info) {
539 ret = btrfs_block_rsv_use_bytes(global_rsv, blocksize);
540 if (!ret)
541 return global_rsv;
542 }
543
544 /*
545 * All hope is lost, but of course our reservations are overly
546 * pessimistic, so instead of possibly having an ENOSPC abort here, try
547 * one last time to force a reservation if there's enough actual space
548 * on disk to make the reservation.
549 */
550 ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv->space_info, blocksize,
551 BTRFS_RESERVE_FLUSH_EMERGENCY);
552 if (!ret)
553 return block_rsv;
554
555 return ERR_PTR(ret);
556}
557
558int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
559 struct btrfs_block_rsv *rsv)
560{
561 u64 needed_bytes;
562 int ret;
563
564 /* 1 for slack space, 1 for updating the inode */
565 needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
566 btrfs_calc_metadata_size(fs_info, 1);
567
568 spin_lock(&rsv->lock);
569 if (rsv->reserved < needed_bytes)
570 ret = -ENOSPC;
571 else
572 ret = 0;
573 spin_unlock(&rsv->lock);
574 return ret;
575}
1// SPDX-License-Identifier: GPL-2.0
2
3#include "misc.h"
4#include "ctree.h"
5#include "block-rsv.h"
6#include "space-info.h"
7#include "transaction.h"
8#include "block-group.h"
9
10/*
11 * HOW DO BLOCK RESERVES WORK
12 *
13 * Think of block_rsv's as buckets for logically grouped metadata
14 * reservations. Each block_rsv has a ->size and a ->reserved. ->size is
15 * how large we want our block rsv to be, ->reserved is how much space is
16 * currently reserved for this block reserve.
17 *
18 * ->failfast exists for the truncate case, and is described below.
19 *
20 * NORMAL OPERATION
21 *
22 * -> Reserve
23 * Entrance: btrfs_block_rsv_add, btrfs_block_rsv_refill
24 *
25 * We call into btrfs_reserve_metadata_bytes() with our bytes, which is
26 * accounted for in space_info->bytes_may_use, and then add the bytes to
27 * ->reserved, and ->size in the case of btrfs_block_rsv_add.
28 *
29 * ->size is an over-estimation of how much we may use for a particular
30 * operation.
31 *
32 * -> Use
33 * Entrance: btrfs_use_block_rsv
34 *
35 * When we do a btrfs_alloc_tree_block() we call into btrfs_use_block_rsv()
36 * to determine the appropriate block_rsv to use, and then verify that
37 * ->reserved has enough space for our tree block allocation. Once
38 * successful we subtract fs_info->nodesize from ->reserved.
39 *
40 * -> Finish
41 * Entrance: btrfs_block_rsv_release
42 *
43 * We are finished with our operation, subtract our individual reservation
44 * from ->size, and then subtract ->size from ->reserved and free up the
45 * excess if there is any.
46 *
47 * There is some logic here to refill the delayed refs rsv or the global rsv
48 * as needed, otherwise the excess is subtracted from
49 * space_info->bytes_may_use.
50 *
51 * TYPES OF BLOCK RESERVES
52 *
53 * BLOCK_RSV_TRANS, BLOCK_RSV_DELOPS, BLOCK_RSV_CHUNK
54 * These behave normally, as described above, just within the confines of the
55 * lifetime of their particular operation (transaction for the whole trans
56 * handle lifetime, for example).
57 *
58 * BLOCK_RSV_GLOBAL
59 * It is impossible to properly account for all the space that may be required
60 * to make our extent tree updates. This block reserve acts as an overflow
61 * buffer in case our delayed refs reserve does not reserve enough space to
62 * update the extent tree.
63 *
64 * We can steal from this in some cases as well, notably on evict() or
65 * truncate() in order to help users recover from ENOSPC conditions.
66 *
67 * BLOCK_RSV_DELALLOC
68 * The individual item sizes are determined by the per-inode size
69 * calculations, which are described with the delalloc code. This is pretty
70 * straightforward, it's just the calculation of ->size encodes a lot of
71 * different items, and thus it gets used when updating inodes, inserting file
72 * extents, and inserting checksums.
73 *
74 * BLOCK_RSV_DELREFS
75 * We keep a running tally of how many delayed refs we have on the system.
76 * We assume each one of these delayed refs are going to use a full
77 * reservation. We use the transaction items and pre-reserve space for every
78 * operation, and use this reservation to refill any gap between ->size and
79 * ->reserved that may exist.
80 *
81 * From there it's straightforward, removing a delayed ref means we remove its
82 * count from ->size and free up reservations as necessary. Since this is
83 * the most dynamic block reserve in the system, we will try to refill this
84 * block reserve first with any excess returned by any other block reserve.
85 *
86 * BLOCK_RSV_EMPTY
87 * This is the fallback block reserve to make us try to reserve space if we
88 * don't have a specific bucket for this allocation. It is mostly used for
89 * updating the device tree and such, since that is a separate pool we're
90 * content to just reserve space from the space_info on demand.
91 *
92 * BLOCK_RSV_TEMP
93 * This is used by things like truncate and iput. We will temporarily
94 * allocate a block reserve, set it to some size, and then truncate bytes
95 * until we have no space left. With ->failfast set we'll simply return
96 * ENOSPC from btrfs_use_block_rsv() to signal that we need to unwind and try
97 * to make a new reservation. This is because these operations are
98 * unbounded, so we want to do as much work as we can, and then back off and
99 * re-reserve.
100 */
101
102static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
103 struct btrfs_block_rsv *block_rsv,
104 struct btrfs_block_rsv *dest, u64 num_bytes,
105 u64 *qgroup_to_release_ret)
106{
107 struct btrfs_space_info *space_info = block_rsv->space_info;
108 u64 qgroup_to_release = 0;
109 u64 ret;
110
111 spin_lock(&block_rsv->lock);
112 if (num_bytes == (u64)-1) {
113 num_bytes = block_rsv->size;
114 qgroup_to_release = block_rsv->qgroup_rsv_size;
115 }
116 block_rsv->size -= num_bytes;
117 if (block_rsv->reserved >= block_rsv->size) {
118 num_bytes = block_rsv->reserved - block_rsv->size;
119 block_rsv->reserved = block_rsv->size;
120 block_rsv->full = 1;
121 } else {
122 num_bytes = 0;
123 }
124 if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
125 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
126 block_rsv->qgroup_rsv_size;
127 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
128 } else {
129 qgroup_to_release = 0;
130 }
131 spin_unlock(&block_rsv->lock);
132
133 ret = num_bytes;
134 if (num_bytes > 0) {
135 if (dest) {
136 spin_lock(&dest->lock);
137 if (!dest->full) {
138 u64 bytes_to_add;
139
140 bytes_to_add = dest->size - dest->reserved;
141 bytes_to_add = min(num_bytes, bytes_to_add);
142 dest->reserved += bytes_to_add;
143 if (dest->reserved >= dest->size)
144 dest->full = 1;
145 num_bytes -= bytes_to_add;
146 }
147 spin_unlock(&dest->lock);
148 }
149 if (num_bytes)
150 btrfs_space_info_free_bytes_may_use(fs_info,
151 space_info,
152 num_bytes);
153 }
154 if (qgroup_to_release_ret)
155 *qgroup_to_release_ret = qgroup_to_release;
156 return ret;
157}
158
159int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
160 struct btrfs_block_rsv *dst, u64 num_bytes,
161 bool update_size)
162{
163 int ret;
164
165 ret = btrfs_block_rsv_use_bytes(src, num_bytes);
166 if (ret)
167 return ret;
168
169 btrfs_block_rsv_add_bytes(dst, num_bytes, update_size);
170 return 0;
171}
172
173void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
174{
175 memset(rsv, 0, sizeof(*rsv));
176 spin_lock_init(&rsv->lock);
177 rsv->type = type;
178}
179
180void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
181 struct btrfs_block_rsv *rsv,
182 unsigned short type)
183{
184 btrfs_init_block_rsv(rsv, type);
185 rsv->space_info = btrfs_find_space_info(fs_info,
186 BTRFS_BLOCK_GROUP_METADATA);
187}
188
189struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
190 unsigned short type)
191{
192 struct btrfs_block_rsv *block_rsv;
193
194 block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
195 if (!block_rsv)
196 return NULL;
197
198 btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
199 return block_rsv;
200}
201
202void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
203 struct btrfs_block_rsv *rsv)
204{
205 if (!rsv)
206 return;
207 btrfs_block_rsv_release(fs_info, rsv, (u64)-1, NULL);
208 kfree(rsv);
209}
210
211int btrfs_block_rsv_add(struct btrfs_root *root,
212 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
213 enum btrfs_reserve_flush_enum flush)
214{
215 int ret;
216
217 if (num_bytes == 0)
218 return 0;
219
220 ret = btrfs_reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
221 if (!ret)
222 btrfs_block_rsv_add_bytes(block_rsv, num_bytes, true);
223
224 return ret;
225}
226
227int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
228{
229 u64 num_bytes = 0;
230 int ret = -ENOSPC;
231
232 if (!block_rsv)
233 return 0;
234
235 spin_lock(&block_rsv->lock);
236 num_bytes = div_factor(block_rsv->size, min_factor);
237 if (block_rsv->reserved >= num_bytes)
238 ret = 0;
239 spin_unlock(&block_rsv->lock);
240
241 return ret;
242}
243
244int btrfs_block_rsv_refill(struct btrfs_root *root,
245 struct btrfs_block_rsv *block_rsv, u64 min_reserved,
246 enum btrfs_reserve_flush_enum flush)
247{
248 u64 num_bytes = 0;
249 int ret = -ENOSPC;
250
251 if (!block_rsv)
252 return 0;
253
254 spin_lock(&block_rsv->lock);
255 num_bytes = min_reserved;
256 if (block_rsv->reserved >= num_bytes)
257 ret = 0;
258 else
259 num_bytes -= block_rsv->reserved;
260 spin_unlock(&block_rsv->lock);
261
262 if (!ret)
263 return 0;
264
265 ret = btrfs_reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
266 if (!ret) {
267 btrfs_block_rsv_add_bytes(block_rsv, num_bytes, false);
268 return 0;
269 }
270
271 return ret;
272}
273
274u64 btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
275 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
276 u64 *qgroup_to_release)
277{
278 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
279 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
280 struct btrfs_block_rsv *target = NULL;
281
282 /*
283 * If we are the delayed_rsv then push to the global rsv, otherwise dump
284 * into the delayed rsv if it is not full.
285 */
286 if (block_rsv == delayed_rsv)
287 target = global_rsv;
288 else if (block_rsv != global_rsv && !delayed_rsv->full)
289 target = delayed_rsv;
290
291 if (target && block_rsv->space_info != target->space_info)
292 target = NULL;
293
294 return block_rsv_release_bytes(fs_info, block_rsv, target, num_bytes,
295 qgroup_to_release);
296}
297
298int btrfs_block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv, u64 num_bytes)
299{
300 int ret = -ENOSPC;
301
302 spin_lock(&block_rsv->lock);
303 if (block_rsv->reserved >= num_bytes) {
304 block_rsv->reserved -= num_bytes;
305 if (block_rsv->reserved < block_rsv->size)
306 block_rsv->full = 0;
307 ret = 0;
308 }
309 spin_unlock(&block_rsv->lock);
310 return ret;
311}
312
313void btrfs_block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
314 u64 num_bytes, bool update_size)
315{
316 spin_lock(&block_rsv->lock);
317 block_rsv->reserved += num_bytes;
318 if (update_size)
319 block_rsv->size += num_bytes;
320 else if (block_rsv->reserved >= block_rsv->size)
321 block_rsv->full = 1;
322 spin_unlock(&block_rsv->lock);
323}
324
325int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
326 struct btrfs_block_rsv *dest, u64 num_bytes,
327 int min_factor)
328{
329 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
330 u64 min_bytes;
331
332 if (global_rsv->space_info != dest->space_info)
333 return -ENOSPC;
334
335 spin_lock(&global_rsv->lock);
336 min_bytes = div_factor(global_rsv->size, min_factor);
337 if (global_rsv->reserved < min_bytes + num_bytes) {
338 spin_unlock(&global_rsv->lock);
339 return -ENOSPC;
340 }
341 global_rsv->reserved -= num_bytes;
342 if (global_rsv->reserved < global_rsv->size)
343 global_rsv->full = 0;
344 spin_unlock(&global_rsv->lock);
345
346 btrfs_block_rsv_add_bytes(dest, num_bytes, true);
347 return 0;
348}
349
350void btrfs_update_global_block_rsv(struct btrfs_fs_info *fs_info)
351{
352 struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
353 struct btrfs_space_info *sinfo = block_rsv->space_info;
354 u64 num_bytes;
355 unsigned min_items;
356
357 /*
358 * The global block rsv is based on the size of the extent tree, the
359 * checksum tree and the root tree. If the fs is empty we want to set
360 * it to a minimal amount for safety.
361 */
362 num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
363 btrfs_root_used(&fs_info->csum_root->root_item) +
364 btrfs_root_used(&fs_info->tree_root->root_item);
365
366 /*
367 * We at a minimum are going to modify the csum root, the tree root, and
368 * the extent root.
369 */
370 min_items = 3;
371
372 /*
373 * But we also want to reserve enough space so we can do the fallback
374 * global reserve for an unlink, which is an additional 5 items (see the
375 * comment in __unlink_start_trans for what we're modifying.)
376 *
377 * But we also need space for the delayed ref updates from the unlink,
378 * so its 10, 5 for the actual operation, and 5 for the delayed ref
379 * updates.
380 */
381 min_items += 10;
382
383 num_bytes = max_t(u64, num_bytes,
384 btrfs_calc_insert_metadata_size(fs_info, min_items));
385
386 spin_lock(&sinfo->lock);
387 spin_lock(&block_rsv->lock);
388
389 block_rsv->size = min_t(u64, num_bytes, SZ_512M);
390
391 if (block_rsv->reserved < block_rsv->size) {
392 num_bytes = block_rsv->size - block_rsv->reserved;
393 btrfs_space_info_update_bytes_may_use(fs_info, sinfo,
394 num_bytes);
395 block_rsv->reserved = block_rsv->size;
396 } else if (block_rsv->reserved > block_rsv->size) {
397 num_bytes = block_rsv->reserved - block_rsv->size;
398 btrfs_space_info_update_bytes_may_use(fs_info, sinfo,
399 -num_bytes);
400 block_rsv->reserved = block_rsv->size;
401 btrfs_try_granting_tickets(fs_info, sinfo);
402 }
403
404 if (block_rsv->reserved == block_rsv->size)
405 block_rsv->full = 1;
406 else
407 block_rsv->full = 0;
408
409 if (block_rsv->size >= sinfo->total_bytes)
410 sinfo->force_alloc = CHUNK_ALLOC_FORCE;
411 spin_unlock(&block_rsv->lock);
412 spin_unlock(&sinfo->lock);
413}
414
415void btrfs_init_global_block_rsv(struct btrfs_fs_info *fs_info)
416{
417 struct btrfs_space_info *space_info;
418
419 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
420 fs_info->chunk_block_rsv.space_info = space_info;
421
422 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
423 fs_info->global_block_rsv.space_info = space_info;
424 fs_info->trans_block_rsv.space_info = space_info;
425 fs_info->empty_block_rsv.space_info = space_info;
426 fs_info->delayed_block_rsv.space_info = space_info;
427 fs_info->delayed_refs_rsv.space_info = space_info;
428
429 fs_info->extent_root->block_rsv = &fs_info->delayed_refs_rsv;
430 fs_info->csum_root->block_rsv = &fs_info->delayed_refs_rsv;
431 fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
432 fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
433 if (fs_info->quota_root)
434 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
435 fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
436
437 btrfs_update_global_block_rsv(fs_info);
438}
439
440void btrfs_release_global_block_rsv(struct btrfs_fs_info *fs_info)
441{
442 btrfs_block_rsv_release(fs_info, &fs_info->global_block_rsv, (u64)-1,
443 NULL);
444 WARN_ON(fs_info->trans_block_rsv.size > 0);
445 WARN_ON(fs_info->trans_block_rsv.reserved > 0);
446 WARN_ON(fs_info->chunk_block_rsv.size > 0);
447 WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
448 WARN_ON(fs_info->delayed_block_rsv.size > 0);
449 WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
450 WARN_ON(fs_info->delayed_refs_rsv.reserved > 0);
451 WARN_ON(fs_info->delayed_refs_rsv.size > 0);
452}
453
454static struct btrfs_block_rsv *get_block_rsv(
455 const struct btrfs_trans_handle *trans,
456 const struct btrfs_root *root)
457{
458 struct btrfs_fs_info *fs_info = root->fs_info;
459 struct btrfs_block_rsv *block_rsv = NULL;
460
461 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
462 (root == fs_info->csum_root && trans->adding_csums) ||
463 (root == fs_info->uuid_root))
464 block_rsv = trans->block_rsv;
465
466 if (!block_rsv)
467 block_rsv = root->block_rsv;
468
469 if (!block_rsv)
470 block_rsv = &fs_info->empty_block_rsv;
471
472 return block_rsv;
473}
474
475struct btrfs_block_rsv *btrfs_use_block_rsv(struct btrfs_trans_handle *trans,
476 struct btrfs_root *root,
477 u32 blocksize)
478{
479 struct btrfs_fs_info *fs_info = root->fs_info;
480 struct btrfs_block_rsv *block_rsv;
481 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
482 int ret;
483 bool global_updated = false;
484
485 block_rsv = get_block_rsv(trans, root);
486
487 if (unlikely(block_rsv->size == 0))
488 goto try_reserve;
489again:
490 ret = btrfs_block_rsv_use_bytes(block_rsv, blocksize);
491 if (!ret)
492 return block_rsv;
493
494 if (block_rsv->failfast)
495 return ERR_PTR(ret);
496
497 if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
498 global_updated = true;
499 btrfs_update_global_block_rsv(fs_info);
500 goto again;
501 }
502
503 /*
504 * The global reserve still exists to save us from ourselves, so don't
505 * warn_on if we are short on our delayed refs reserve.
506 */
507 if (block_rsv->type != BTRFS_BLOCK_RSV_DELREFS &&
508 btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
509 static DEFINE_RATELIMIT_STATE(_rs,
510 DEFAULT_RATELIMIT_INTERVAL * 10,
511 /*DEFAULT_RATELIMIT_BURST*/ 1);
512 if (__ratelimit(&_rs))
513 WARN(1, KERN_DEBUG
514 "BTRFS: block rsv returned %d\n", ret);
515 }
516try_reserve:
517 ret = btrfs_reserve_metadata_bytes(root, block_rsv, blocksize,
518 BTRFS_RESERVE_NO_FLUSH);
519 if (!ret)
520 return block_rsv;
521 /*
522 * If we couldn't reserve metadata bytes try and use some from
523 * the global reserve if its space type is the same as the global
524 * reservation.
525 */
526 if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
527 block_rsv->space_info == global_rsv->space_info) {
528 ret = btrfs_block_rsv_use_bytes(global_rsv, blocksize);
529 if (!ret)
530 return global_rsv;
531 }
532 return ERR_PTR(ret);
533}