Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * NILFS dat/inode allocator
4 *
5 * Copyright (C) 2006-2008 Nippon Telegraph and Telephone Corporation.
6 *
7 * Originally written by Koji Sato.
8 * Two allocators were unified by Ryusuke Konishi and Amagai Yoshiji.
9 */
10
11#include <linux/types.h>
12#include <linux/buffer_head.h>
13#include <linux/fs.h>
14#include <linux/bitops.h>
15#include <linux/slab.h>
16#include "mdt.h"
17#include "alloc.h"
18
19
20/**
21 * nilfs_palloc_groups_per_desc_block - get the number of groups that a group
22 * descriptor block can maintain
23 * @inode: inode of metadata file using this allocator
24 */
25static inline unsigned long
26nilfs_palloc_groups_per_desc_block(const struct inode *inode)
27{
28 return i_blocksize(inode) /
29 sizeof(struct nilfs_palloc_group_desc);
30}
31
32/**
33 * nilfs_palloc_groups_count - get maximum number of groups
34 * @inode: inode of metadata file using this allocator
35 */
36static inline unsigned long
37nilfs_palloc_groups_count(const struct inode *inode)
38{
39 return 1UL << (BITS_PER_LONG - (inode->i_blkbits + 3 /* log2(8) */));
40}
41
42/**
43 * nilfs_palloc_init_blockgroup - initialize private variables for allocator
44 * @inode: inode of metadata file using this allocator
45 * @entry_size: size of the persistent object
46 */
47int nilfs_palloc_init_blockgroup(struct inode *inode, unsigned int entry_size)
48{
49 struct nilfs_mdt_info *mi = NILFS_MDT(inode);
50
51 mi->mi_bgl = kmalloc(sizeof(*mi->mi_bgl), GFP_NOFS);
52 if (!mi->mi_bgl)
53 return -ENOMEM;
54
55 bgl_lock_init(mi->mi_bgl);
56
57 nilfs_mdt_set_entry_size(inode, entry_size, 0);
58
59 mi->mi_blocks_per_group =
60 DIV_ROUND_UP(nilfs_palloc_entries_per_group(inode),
61 mi->mi_entries_per_block) + 1;
62 /*
63 * Number of blocks in a group including entry blocks
64 * and a bitmap block
65 */
66 mi->mi_blocks_per_desc_block =
67 nilfs_palloc_groups_per_desc_block(inode) *
68 mi->mi_blocks_per_group + 1;
69 /*
70 * Number of blocks per descriptor including the
71 * descriptor block
72 */
73 return 0;
74}
75
76/**
77 * nilfs_palloc_group - get group number and offset from an entry number
78 * @inode: inode of metadata file using this allocator
79 * @nr: serial number of the entry (e.g. inode number)
80 * @offset: pointer to store offset number in the group
81 */
82static unsigned long nilfs_palloc_group(const struct inode *inode, __u64 nr,
83 unsigned long *offset)
84{
85 __u64 group = nr;
86
87 *offset = do_div(group, nilfs_palloc_entries_per_group(inode));
88 return group;
89}
90
91/**
92 * nilfs_palloc_desc_blkoff - get block offset of a group descriptor block
93 * @inode: inode of metadata file using this allocator
94 * @group: group number
95 *
96 * nilfs_palloc_desc_blkoff() returns block offset of the descriptor
97 * block which contains a descriptor of the specified group.
98 */
99static unsigned long
100nilfs_palloc_desc_blkoff(const struct inode *inode, unsigned long group)
101{
102 unsigned long desc_block =
103 group / nilfs_palloc_groups_per_desc_block(inode);
104 return desc_block * NILFS_MDT(inode)->mi_blocks_per_desc_block;
105}
106
107/**
108 * nilfs_palloc_bitmap_blkoff - get block offset of a bitmap block
109 * @inode: inode of metadata file using this allocator
110 * @group: group number
111 *
112 * nilfs_palloc_bitmap_blkoff() returns block offset of the bitmap
113 * block used to allocate/deallocate entries in the specified group.
114 */
115static unsigned long
116nilfs_palloc_bitmap_blkoff(const struct inode *inode, unsigned long group)
117{
118 unsigned long desc_offset =
119 group % nilfs_palloc_groups_per_desc_block(inode);
120 return nilfs_palloc_desc_blkoff(inode, group) + 1 +
121 desc_offset * NILFS_MDT(inode)->mi_blocks_per_group;
122}
123
124/**
125 * nilfs_palloc_group_desc_nfrees - get the number of free entries in a group
126 * @desc: pointer to descriptor structure for the group
127 * @lock: spin lock protecting @desc
128 */
129static unsigned long
130nilfs_palloc_group_desc_nfrees(const struct nilfs_palloc_group_desc *desc,
131 spinlock_t *lock)
132{
133 unsigned long nfree;
134
135 spin_lock(lock);
136 nfree = le32_to_cpu(desc->pg_nfrees);
137 spin_unlock(lock);
138 return nfree;
139}
140
141/**
142 * nilfs_palloc_group_desc_add_entries - adjust count of free entries
143 * @desc: pointer to descriptor structure for the group
144 * @lock: spin lock protecting @desc
145 * @n: delta to be added
146 */
147static u32
148nilfs_palloc_group_desc_add_entries(struct nilfs_palloc_group_desc *desc,
149 spinlock_t *lock, u32 n)
150{
151 u32 nfree;
152
153 spin_lock(lock);
154 le32_add_cpu(&desc->pg_nfrees, n);
155 nfree = le32_to_cpu(desc->pg_nfrees);
156 spin_unlock(lock);
157 return nfree;
158}
159
160/**
161 * nilfs_palloc_entry_blkoff - get block offset of an entry block
162 * @inode: inode of metadata file using this allocator
163 * @nr: serial number of the entry (e.g. inode number)
164 */
165static unsigned long
166nilfs_palloc_entry_blkoff(const struct inode *inode, __u64 nr)
167{
168 unsigned long group, group_offset;
169
170 group = nilfs_palloc_group(inode, nr, &group_offset);
171
172 return nilfs_palloc_bitmap_blkoff(inode, group) + 1 +
173 group_offset / NILFS_MDT(inode)->mi_entries_per_block;
174}
175
176/**
177 * nilfs_palloc_desc_block_init - initialize buffer of a group descriptor block
178 * @inode: inode of metadata file
179 * @bh: buffer head of the buffer to be initialized
180 * @kaddr: kernel address mapped for the page including the buffer
181 */
182static void nilfs_palloc_desc_block_init(struct inode *inode,
183 struct buffer_head *bh, void *kaddr)
184{
185 struct nilfs_palloc_group_desc *desc = kaddr + bh_offset(bh);
186 unsigned long n = nilfs_palloc_groups_per_desc_block(inode);
187 __le32 nfrees;
188
189 nfrees = cpu_to_le32(nilfs_palloc_entries_per_group(inode));
190 while (n-- > 0) {
191 desc->pg_nfrees = nfrees;
192 desc++;
193 }
194}
195
196static int nilfs_palloc_get_block(struct inode *inode, unsigned long blkoff,
197 int create,
198 void (*init_block)(struct inode *,
199 struct buffer_head *,
200 void *),
201 struct buffer_head **bhp,
202 struct nilfs_bh_assoc *prev,
203 spinlock_t *lock)
204{
205 int ret;
206
207 spin_lock(lock);
208 if (prev->bh && blkoff == prev->blkoff &&
209 likely(buffer_uptodate(prev->bh))) {
210 get_bh(prev->bh);
211 *bhp = prev->bh;
212 spin_unlock(lock);
213 return 0;
214 }
215 spin_unlock(lock);
216
217 ret = nilfs_mdt_get_block(inode, blkoff, create, init_block, bhp);
218 if (!ret) {
219 spin_lock(lock);
220 /*
221 * The following code must be safe for change of the
222 * cache contents during the get block call.
223 */
224 brelse(prev->bh);
225 get_bh(*bhp);
226 prev->bh = *bhp;
227 prev->blkoff = blkoff;
228 spin_unlock(lock);
229 }
230 return ret;
231}
232
233/**
234 * nilfs_palloc_delete_block - delete a block on the persistent allocator file
235 * @inode: inode of metadata file using this allocator
236 * @blkoff: block offset
237 * @prev: nilfs_bh_assoc struct of the last used buffer
238 * @lock: spin lock protecting @prev
239 */
240static int nilfs_palloc_delete_block(struct inode *inode, unsigned long blkoff,
241 struct nilfs_bh_assoc *prev,
242 spinlock_t *lock)
243{
244 spin_lock(lock);
245 if (prev->bh && blkoff == prev->blkoff) {
246 brelse(prev->bh);
247 prev->bh = NULL;
248 }
249 spin_unlock(lock);
250 return nilfs_mdt_delete_block(inode, blkoff);
251}
252
253/**
254 * nilfs_palloc_get_desc_block - get buffer head of a group descriptor block
255 * @inode: inode of metadata file using this allocator
256 * @group: group number
257 * @create: create flag
258 * @bhp: pointer to store the resultant buffer head
259 */
260static int nilfs_palloc_get_desc_block(struct inode *inode,
261 unsigned long group,
262 int create, struct buffer_head **bhp)
263{
264 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
265
266 return nilfs_palloc_get_block(inode,
267 nilfs_palloc_desc_blkoff(inode, group),
268 create, nilfs_palloc_desc_block_init,
269 bhp, &cache->prev_desc, &cache->lock);
270}
271
272/**
273 * nilfs_palloc_get_bitmap_block - get buffer head of a bitmap block
274 * @inode: inode of metadata file using this allocator
275 * @group: group number
276 * @create: create flag
277 * @bhp: pointer to store the resultant buffer head
278 */
279static int nilfs_palloc_get_bitmap_block(struct inode *inode,
280 unsigned long group,
281 int create, struct buffer_head **bhp)
282{
283 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
284
285 return nilfs_palloc_get_block(inode,
286 nilfs_palloc_bitmap_blkoff(inode, group),
287 create, NULL, bhp,
288 &cache->prev_bitmap, &cache->lock);
289}
290
291/**
292 * nilfs_palloc_delete_bitmap_block - delete a bitmap block
293 * @inode: inode of metadata file using this allocator
294 * @group: group number
295 */
296static int nilfs_palloc_delete_bitmap_block(struct inode *inode,
297 unsigned long group)
298{
299 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
300
301 return nilfs_palloc_delete_block(inode,
302 nilfs_palloc_bitmap_blkoff(inode,
303 group),
304 &cache->prev_bitmap, &cache->lock);
305}
306
307/**
308 * nilfs_palloc_get_entry_block - get buffer head of an entry block
309 * @inode: inode of metadata file using this allocator
310 * @nr: serial number of the entry (e.g. inode number)
311 * @create: create flag
312 * @bhp: pointer to store the resultant buffer head
313 */
314int nilfs_palloc_get_entry_block(struct inode *inode, __u64 nr,
315 int create, struct buffer_head **bhp)
316{
317 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
318
319 return nilfs_palloc_get_block(inode,
320 nilfs_palloc_entry_blkoff(inode, nr),
321 create, NULL, bhp,
322 &cache->prev_entry, &cache->lock);
323}
324
325/**
326 * nilfs_palloc_delete_entry_block - delete an entry block
327 * @inode: inode of metadata file using this allocator
328 * @nr: serial number of the entry
329 */
330static int nilfs_palloc_delete_entry_block(struct inode *inode, __u64 nr)
331{
332 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
333
334 return nilfs_palloc_delete_block(inode,
335 nilfs_palloc_entry_blkoff(inode, nr),
336 &cache->prev_entry, &cache->lock);
337}
338
339/**
340 * nilfs_palloc_block_get_group_desc - get kernel address of a group descriptor
341 * @inode: inode of metadata file using this allocator
342 * @group: group number
343 * @bh: buffer head of the buffer storing the group descriptor block
344 * @kaddr: kernel address mapped for the page including the buffer
345 */
346static struct nilfs_palloc_group_desc *
347nilfs_palloc_block_get_group_desc(const struct inode *inode,
348 unsigned long group,
349 const struct buffer_head *bh, void *kaddr)
350{
351 return (struct nilfs_palloc_group_desc *)(kaddr + bh_offset(bh)) +
352 group % nilfs_palloc_groups_per_desc_block(inode);
353}
354
355/**
356 * nilfs_palloc_block_get_entry - get kernel address of an entry
357 * @inode: inode of metadata file using this allocator
358 * @nr: serial number of the entry (e.g. inode number)
359 * @bh: buffer head of the buffer storing the entry block
360 * @kaddr: kernel address mapped for the page including the buffer
361 */
362void *nilfs_palloc_block_get_entry(const struct inode *inode, __u64 nr,
363 const struct buffer_head *bh, void *kaddr)
364{
365 unsigned long entry_offset, group_offset;
366
367 nilfs_palloc_group(inode, nr, &group_offset);
368 entry_offset = group_offset % NILFS_MDT(inode)->mi_entries_per_block;
369
370 return kaddr + bh_offset(bh) +
371 entry_offset * NILFS_MDT(inode)->mi_entry_size;
372}
373
374/**
375 * nilfs_palloc_find_available_slot - find available slot in a group
376 * @bitmap: bitmap of the group
377 * @target: offset number of an entry in the group (start point)
378 * @bsize: size in bits
379 * @lock: spin lock protecting @bitmap
380 */
381static int nilfs_palloc_find_available_slot(unsigned char *bitmap,
382 unsigned long target,
383 unsigned int bsize,
384 spinlock_t *lock)
385{
386 int pos, end = bsize;
387
388 if (likely(target < bsize)) {
389 pos = target;
390 do {
391 pos = nilfs_find_next_zero_bit(bitmap, end, pos);
392 if (pos >= end)
393 break;
394 if (!nilfs_set_bit_atomic(lock, pos, bitmap))
395 return pos;
396 } while (++pos < end);
397
398 end = target;
399 }
400
401 /* wrap around */
402 for (pos = 0; pos < end; pos++) {
403 pos = nilfs_find_next_zero_bit(bitmap, end, pos);
404 if (pos >= end)
405 break;
406 if (!nilfs_set_bit_atomic(lock, pos, bitmap))
407 return pos;
408 }
409
410 return -ENOSPC;
411}
412
413/**
414 * nilfs_palloc_rest_groups_in_desc_block - get the remaining number of groups
415 * in a group descriptor block
416 * @inode: inode of metadata file using this allocator
417 * @curr: current group number
418 * @max: maximum number of groups
419 */
420static unsigned long
421nilfs_palloc_rest_groups_in_desc_block(const struct inode *inode,
422 unsigned long curr, unsigned long max)
423{
424 return min_t(unsigned long,
425 nilfs_palloc_groups_per_desc_block(inode) -
426 curr % nilfs_palloc_groups_per_desc_block(inode),
427 max - curr + 1);
428}
429
430/**
431 * nilfs_palloc_count_desc_blocks - count descriptor blocks number
432 * @inode: inode of metadata file using this allocator
433 * @desc_blocks: descriptor blocks number [out]
434 */
435static int nilfs_palloc_count_desc_blocks(struct inode *inode,
436 unsigned long *desc_blocks)
437{
438 __u64 blknum;
439 int ret;
440
441 ret = nilfs_bmap_last_key(NILFS_I(inode)->i_bmap, &blknum);
442 if (likely(!ret))
443 *desc_blocks = DIV_ROUND_UP(
444 (unsigned long)blknum,
445 NILFS_MDT(inode)->mi_blocks_per_desc_block);
446 return ret;
447}
448
449/**
450 * nilfs_palloc_mdt_file_can_grow - check potential opportunity for
451 * MDT file growing
452 * @inode: inode of metadata file using this allocator
453 * @desc_blocks: known current descriptor blocks count
454 */
455static inline bool nilfs_palloc_mdt_file_can_grow(struct inode *inode,
456 unsigned long desc_blocks)
457{
458 return (nilfs_palloc_groups_per_desc_block(inode) * desc_blocks) <
459 nilfs_palloc_groups_count(inode);
460}
461
462/**
463 * nilfs_palloc_count_max_entries - count max number of entries that can be
464 * described by descriptor blocks count
465 * @inode: inode of metadata file using this allocator
466 * @nused: current number of used entries
467 * @nmaxp: max number of entries [out]
468 */
469int nilfs_palloc_count_max_entries(struct inode *inode, u64 nused, u64 *nmaxp)
470{
471 unsigned long desc_blocks = 0;
472 u64 entries_per_desc_block, nmax;
473 int err;
474
475 err = nilfs_palloc_count_desc_blocks(inode, &desc_blocks);
476 if (unlikely(err))
477 return err;
478
479 entries_per_desc_block = (u64)nilfs_palloc_entries_per_group(inode) *
480 nilfs_palloc_groups_per_desc_block(inode);
481 nmax = entries_per_desc_block * desc_blocks;
482
483 if (nused == nmax &&
484 nilfs_palloc_mdt_file_can_grow(inode, desc_blocks))
485 nmax += entries_per_desc_block;
486
487 if (nused > nmax)
488 return -ERANGE;
489
490 *nmaxp = nmax;
491 return 0;
492}
493
494/**
495 * nilfs_palloc_prepare_alloc_entry - prepare to allocate a persistent object
496 * @inode: inode of metadata file using this allocator
497 * @req: nilfs_palloc_req structure exchanged for the allocation
498 */
499int nilfs_palloc_prepare_alloc_entry(struct inode *inode,
500 struct nilfs_palloc_req *req)
501{
502 struct buffer_head *desc_bh, *bitmap_bh;
503 struct nilfs_palloc_group_desc *desc;
504 unsigned char *bitmap;
505 void *desc_kaddr, *bitmap_kaddr;
506 unsigned long group, maxgroup, ngroups;
507 unsigned long group_offset, maxgroup_offset;
508 unsigned long n, entries_per_group;
509 unsigned long i, j;
510 spinlock_t *lock;
511 int pos, ret;
512
513 ngroups = nilfs_palloc_groups_count(inode);
514 maxgroup = ngroups - 1;
515 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
516 entries_per_group = nilfs_palloc_entries_per_group(inode);
517
518 for (i = 0; i < ngroups; i += n) {
519 if (group >= ngroups) {
520 /* wrap around */
521 group = 0;
522 maxgroup = nilfs_palloc_group(inode, req->pr_entry_nr,
523 &maxgroup_offset) - 1;
524 }
525 ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
526 if (ret < 0)
527 return ret;
528 desc_kaddr = kmap(desc_bh->b_page);
529 desc = nilfs_palloc_block_get_group_desc(
530 inode, group, desc_bh, desc_kaddr);
531 n = nilfs_palloc_rest_groups_in_desc_block(inode, group,
532 maxgroup);
533 for (j = 0; j < n; j++, desc++, group++) {
534 lock = nilfs_mdt_bgl_lock(inode, group);
535 if (nilfs_palloc_group_desc_nfrees(desc, lock) > 0) {
536 ret = nilfs_palloc_get_bitmap_block(
537 inode, group, 1, &bitmap_bh);
538 if (ret < 0)
539 goto out_desc;
540 bitmap_kaddr = kmap(bitmap_bh->b_page);
541 bitmap = bitmap_kaddr + bh_offset(bitmap_bh);
542 pos = nilfs_palloc_find_available_slot(
543 bitmap, group_offset,
544 entries_per_group, lock);
545 if (pos >= 0) {
546 /* found a free entry */
547 nilfs_palloc_group_desc_add_entries(
548 desc, lock, -1);
549 req->pr_entry_nr =
550 entries_per_group * group + pos;
551 kunmap(desc_bh->b_page);
552 kunmap(bitmap_bh->b_page);
553
554 req->pr_desc_bh = desc_bh;
555 req->pr_bitmap_bh = bitmap_bh;
556 return 0;
557 }
558 kunmap(bitmap_bh->b_page);
559 brelse(bitmap_bh);
560 }
561
562 group_offset = 0;
563 }
564
565 kunmap(desc_bh->b_page);
566 brelse(desc_bh);
567 }
568
569 /* no entries left */
570 return -ENOSPC;
571
572 out_desc:
573 kunmap(desc_bh->b_page);
574 brelse(desc_bh);
575 return ret;
576}
577
578/**
579 * nilfs_palloc_commit_alloc_entry - finish allocation of a persistent object
580 * @inode: inode of metadata file using this allocator
581 * @req: nilfs_palloc_req structure exchanged for the allocation
582 */
583void nilfs_palloc_commit_alloc_entry(struct inode *inode,
584 struct nilfs_palloc_req *req)
585{
586 mark_buffer_dirty(req->pr_bitmap_bh);
587 mark_buffer_dirty(req->pr_desc_bh);
588 nilfs_mdt_mark_dirty(inode);
589
590 brelse(req->pr_bitmap_bh);
591 brelse(req->pr_desc_bh);
592}
593
594/**
595 * nilfs_palloc_commit_free_entry - finish deallocating a persistent object
596 * @inode: inode of metadata file using this allocator
597 * @req: nilfs_palloc_req structure exchanged for the removal
598 */
599void nilfs_palloc_commit_free_entry(struct inode *inode,
600 struct nilfs_palloc_req *req)
601{
602 struct nilfs_palloc_group_desc *desc;
603 unsigned long group, group_offset;
604 unsigned char *bitmap;
605 void *desc_kaddr, *bitmap_kaddr;
606 spinlock_t *lock;
607
608 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
609 desc_kaddr = kmap(req->pr_desc_bh->b_page);
610 desc = nilfs_palloc_block_get_group_desc(inode, group,
611 req->pr_desc_bh, desc_kaddr);
612 bitmap_kaddr = kmap(req->pr_bitmap_bh->b_page);
613 bitmap = bitmap_kaddr + bh_offset(req->pr_bitmap_bh);
614 lock = nilfs_mdt_bgl_lock(inode, group);
615
616 if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
617 nilfs_warn(inode->i_sb,
618 "%s (ino=%lu): entry number %llu already freed",
619 __func__, inode->i_ino,
620 (unsigned long long)req->pr_entry_nr);
621 else
622 nilfs_palloc_group_desc_add_entries(desc, lock, 1);
623
624 kunmap(req->pr_bitmap_bh->b_page);
625 kunmap(req->pr_desc_bh->b_page);
626
627 mark_buffer_dirty(req->pr_desc_bh);
628 mark_buffer_dirty(req->pr_bitmap_bh);
629 nilfs_mdt_mark_dirty(inode);
630
631 brelse(req->pr_bitmap_bh);
632 brelse(req->pr_desc_bh);
633}
634
635/**
636 * nilfs_palloc_abort_alloc_entry - cancel allocation of a persistent object
637 * @inode: inode of metadata file using this allocator
638 * @req: nilfs_palloc_req structure exchanged for the allocation
639 */
640void nilfs_palloc_abort_alloc_entry(struct inode *inode,
641 struct nilfs_palloc_req *req)
642{
643 struct nilfs_palloc_group_desc *desc;
644 void *desc_kaddr, *bitmap_kaddr;
645 unsigned char *bitmap;
646 unsigned long group, group_offset;
647 spinlock_t *lock;
648
649 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
650 desc_kaddr = kmap(req->pr_desc_bh->b_page);
651 desc = nilfs_palloc_block_get_group_desc(inode, group,
652 req->pr_desc_bh, desc_kaddr);
653 bitmap_kaddr = kmap(req->pr_bitmap_bh->b_page);
654 bitmap = bitmap_kaddr + bh_offset(req->pr_bitmap_bh);
655 lock = nilfs_mdt_bgl_lock(inode, group);
656
657 if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
658 nilfs_warn(inode->i_sb,
659 "%s (ino=%lu): entry number %llu already freed",
660 __func__, inode->i_ino,
661 (unsigned long long)req->pr_entry_nr);
662 else
663 nilfs_palloc_group_desc_add_entries(desc, lock, 1);
664
665 kunmap(req->pr_bitmap_bh->b_page);
666 kunmap(req->pr_desc_bh->b_page);
667
668 brelse(req->pr_bitmap_bh);
669 brelse(req->pr_desc_bh);
670
671 req->pr_entry_nr = 0;
672 req->pr_bitmap_bh = NULL;
673 req->pr_desc_bh = NULL;
674}
675
676/**
677 * nilfs_palloc_prepare_free_entry - prepare to deallocate a persistent object
678 * @inode: inode of metadata file using this allocator
679 * @req: nilfs_palloc_req structure exchanged for the removal
680 */
681int nilfs_palloc_prepare_free_entry(struct inode *inode,
682 struct nilfs_palloc_req *req)
683{
684 struct buffer_head *desc_bh, *bitmap_bh;
685 unsigned long group, group_offset;
686 int ret;
687
688 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
689 ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
690 if (ret < 0)
691 return ret;
692 ret = nilfs_palloc_get_bitmap_block(inode, group, 1, &bitmap_bh);
693 if (ret < 0) {
694 brelse(desc_bh);
695 return ret;
696 }
697
698 req->pr_desc_bh = desc_bh;
699 req->pr_bitmap_bh = bitmap_bh;
700 return 0;
701}
702
703/**
704 * nilfs_palloc_abort_free_entry - cancel deallocating a persistent object
705 * @inode: inode of metadata file using this allocator
706 * @req: nilfs_palloc_req structure exchanged for the removal
707 */
708void nilfs_palloc_abort_free_entry(struct inode *inode,
709 struct nilfs_palloc_req *req)
710{
711 brelse(req->pr_bitmap_bh);
712 brelse(req->pr_desc_bh);
713
714 req->pr_entry_nr = 0;
715 req->pr_bitmap_bh = NULL;
716 req->pr_desc_bh = NULL;
717}
718
719/**
720 * nilfs_palloc_freev - deallocate a set of persistent objects
721 * @inode: inode of metadata file using this allocator
722 * @entry_nrs: array of entry numbers to be deallocated
723 * @nitems: number of entries stored in @entry_nrs
724 */
725int nilfs_palloc_freev(struct inode *inode, __u64 *entry_nrs, size_t nitems)
726{
727 struct buffer_head *desc_bh, *bitmap_bh;
728 struct nilfs_palloc_group_desc *desc;
729 unsigned char *bitmap;
730 void *desc_kaddr, *bitmap_kaddr;
731 unsigned long group, group_offset;
732 __u64 group_min_nr, last_nrs[8];
733 const unsigned long epg = nilfs_palloc_entries_per_group(inode);
734 const unsigned int epb = NILFS_MDT(inode)->mi_entries_per_block;
735 unsigned int entry_start, end, pos;
736 spinlock_t *lock;
737 int i, j, k, ret;
738 u32 nfree;
739
740 for (i = 0; i < nitems; i = j) {
741 int change_group = false;
742 int nempties = 0, n = 0;
743
744 group = nilfs_palloc_group(inode, entry_nrs[i], &group_offset);
745 ret = nilfs_palloc_get_desc_block(inode, group, 0, &desc_bh);
746 if (ret < 0)
747 return ret;
748 ret = nilfs_palloc_get_bitmap_block(inode, group, 0,
749 &bitmap_bh);
750 if (ret < 0) {
751 brelse(desc_bh);
752 return ret;
753 }
754
755 /* Get the first entry number of the group */
756 group_min_nr = (__u64)group * epg;
757
758 bitmap_kaddr = kmap(bitmap_bh->b_page);
759 bitmap = bitmap_kaddr + bh_offset(bitmap_bh);
760 lock = nilfs_mdt_bgl_lock(inode, group);
761
762 j = i;
763 entry_start = rounddown(group_offset, epb);
764 do {
765 if (!nilfs_clear_bit_atomic(lock, group_offset,
766 bitmap)) {
767 nilfs_warn(inode->i_sb,
768 "%s (ino=%lu): entry number %llu already freed",
769 __func__, inode->i_ino,
770 (unsigned long long)entry_nrs[j]);
771 } else {
772 n++;
773 }
774
775 j++;
776 if (j >= nitems || entry_nrs[j] < group_min_nr ||
777 entry_nrs[j] >= group_min_nr + epg) {
778 change_group = true;
779 } else {
780 group_offset = entry_nrs[j] - group_min_nr;
781 if (group_offset >= entry_start &&
782 group_offset < entry_start + epb) {
783 /* This entry is in the same block */
784 continue;
785 }
786 }
787
788 /* Test if the entry block is empty or not */
789 end = entry_start + epb;
790 pos = nilfs_find_next_bit(bitmap, end, entry_start);
791 if (pos >= end) {
792 last_nrs[nempties++] = entry_nrs[j - 1];
793 if (nempties >= ARRAY_SIZE(last_nrs))
794 break;
795 }
796
797 if (change_group)
798 break;
799
800 /* Go on to the next entry block */
801 entry_start = rounddown(group_offset, epb);
802 } while (true);
803
804 kunmap(bitmap_bh->b_page);
805 mark_buffer_dirty(bitmap_bh);
806 brelse(bitmap_bh);
807
808 for (k = 0; k < nempties; k++) {
809 ret = nilfs_palloc_delete_entry_block(inode,
810 last_nrs[k]);
811 if (ret && ret != -ENOENT)
812 nilfs_warn(inode->i_sb,
813 "error %d deleting block that object (entry=%llu, ino=%lu) belongs to",
814 ret, (unsigned long long)last_nrs[k],
815 inode->i_ino);
816 }
817
818 desc_kaddr = kmap_atomic(desc_bh->b_page);
819 desc = nilfs_palloc_block_get_group_desc(
820 inode, group, desc_bh, desc_kaddr);
821 nfree = nilfs_palloc_group_desc_add_entries(desc, lock, n);
822 kunmap_atomic(desc_kaddr);
823 mark_buffer_dirty(desc_bh);
824 nilfs_mdt_mark_dirty(inode);
825 brelse(desc_bh);
826
827 if (nfree == nilfs_palloc_entries_per_group(inode)) {
828 ret = nilfs_palloc_delete_bitmap_block(inode, group);
829 if (ret && ret != -ENOENT)
830 nilfs_warn(inode->i_sb,
831 "error %d deleting bitmap block of group=%lu, ino=%lu",
832 ret, group, inode->i_ino);
833 }
834 }
835 return 0;
836}
837
838void nilfs_palloc_setup_cache(struct inode *inode,
839 struct nilfs_palloc_cache *cache)
840{
841 NILFS_MDT(inode)->mi_palloc_cache = cache;
842 spin_lock_init(&cache->lock);
843}
844
845void nilfs_palloc_clear_cache(struct inode *inode)
846{
847 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
848
849 spin_lock(&cache->lock);
850 brelse(cache->prev_desc.bh);
851 brelse(cache->prev_bitmap.bh);
852 brelse(cache->prev_entry.bh);
853 cache->prev_desc.bh = NULL;
854 cache->prev_bitmap.bh = NULL;
855 cache->prev_entry.bh = NULL;
856 spin_unlock(&cache->lock);
857}
858
859void nilfs_palloc_destroy_cache(struct inode *inode)
860{
861 nilfs_palloc_clear_cache(inode);
862 NILFS_MDT(inode)->mi_palloc_cache = NULL;
863}
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * NILFS dat/inode allocator
4 *
5 * Copyright (C) 2006-2008 Nippon Telegraph and Telephone Corporation.
6 *
7 * Originally written by Koji Sato.
8 * Two allocators were unified by Ryusuke Konishi and Amagai Yoshiji.
9 */
10
11#include <linux/types.h>
12#include <linux/buffer_head.h>
13#include <linux/fs.h>
14#include <linux/bitops.h>
15#include <linux/slab.h>
16#include "mdt.h"
17#include "alloc.h"
18
19
20/**
21 * nilfs_palloc_groups_per_desc_block - get the number of groups that a group
22 * descriptor block can maintain
23 * @inode: inode of metadata file using this allocator
24 */
25static inline unsigned long
26nilfs_palloc_groups_per_desc_block(const struct inode *inode)
27{
28 return i_blocksize(inode) /
29 sizeof(struct nilfs_palloc_group_desc);
30}
31
32/**
33 * nilfs_palloc_groups_count - get maximum number of groups
34 * @inode: inode of metadata file using this allocator
35 */
36static inline unsigned long
37nilfs_palloc_groups_count(const struct inode *inode)
38{
39 return 1UL << (BITS_PER_LONG - (inode->i_blkbits + 3 /* log2(8) */));
40}
41
42/**
43 * nilfs_palloc_init_blockgroup - initialize private variables for allocator
44 * @inode: inode of metadata file using this allocator
45 * @entry_size: size of the persistent object
46 */
47int nilfs_palloc_init_blockgroup(struct inode *inode, unsigned int entry_size)
48{
49 struct nilfs_mdt_info *mi = NILFS_MDT(inode);
50
51 mi->mi_bgl = kmalloc(sizeof(*mi->mi_bgl), GFP_NOFS);
52 if (!mi->mi_bgl)
53 return -ENOMEM;
54
55 bgl_lock_init(mi->mi_bgl);
56
57 nilfs_mdt_set_entry_size(inode, entry_size, 0);
58
59 mi->mi_blocks_per_group =
60 DIV_ROUND_UP(nilfs_palloc_entries_per_group(inode),
61 mi->mi_entries_per_block) + 1;
62 /*
63 * Number of blocks in a group including entry blocks
64 * and a bitmap block
65 */
66 mi->mi_blocks_per_desc_block =
67 nilfs_palloc_groups_per_desc_block(inode) *
68 mi->mi_blocks_per_group + 1;
69 /*
70 * Number of blocks per descriptor including the
71 * descriptor block
72 */
73 return 0;
74}
75
76/**
77 * nilfs_palloc_group - get group number and offset from an entry number
78 * @inode: inode of metadata file using this allocator
79 * @nr: serial number of the entry (e.g. inode number)
80 * @offset: pointer to store offset number in the group
81 */
82static unsigned long nilfs_palloc_group(const struct inode *inode, __u64 nr,
83 unsigned long *offset)
84{
85 __u64 group = nr;
86
87 *offset = do_div(group, nilfs_palloc_entries_per_group(inode));
88 return group;
89}
90
91/**
92 * nilfs_palloc_desc_blkoff - get block offset of a group descriptor block
93 * @inode: inode of metadata file using this allocator
94 * @group: group number
95 *
96 * nilfs_palloc_desc_blkoff() returns block offset of the descriptor
97 * block which contains a descriptor of the specified group.
98 */
99static unsigned long
100nilfs_palloc_desc_blkoff(const struct inode *inode, unsigned long group)
101{
102 unsigned long desc_block =
103 group / nilfs_palloc_groups_per_desc_block(inode);
104 return desc_block * NILFS_MDT(inode)->mi_blocks_per_desc_block;
105}
106
107/**
108 * nilfs_palloc_bitmap_blkoff - get block offset of a bitmap block
109 * @inode: inode of metadata file using this allocator
110 * @group: group number
111 *
112 * nilfs_palloc_bitmap_blkoff() returns block offset of the bitmap
113 * block used to allocate/deallocate entries in the specified group.
114 */
115static unsigned long
116nilfs_palloc_bitmap_blkoff(const struct inode *inode, unsigned long group)
117{
118 unsigned long desc_offset =
119 group % nilfs_palloc_groups_per_desc_block(inode);
120 return nilfs_palloc_desc_blkoff(inode, group) + 1 +
121 desc_offset * NILFS_MDT(inode)->mi_blocks_per_group;
122}
123
124/**
125 * nilfs_palloc_group_desc_nfrees - get the number of free entries in a group
126 * @desc: pointer to descriptor structure for the group
127 * @lock: spin lock protecting @desc
128 */
129static unsigned long
130nilfs_palloc_group_desc_nfrees(const struct nilfs_palloc_group_desc *desc,
131 spinlock_t *lock)
132{
133 unsigned long nfree;
134
135 spin_lock(lock);
136 nfree = le32_to_cpu(desc->pg_nfrees);
137 spin_unlock(lock);
138 return nfree;
139}
140
141/**
142 * nilfs_palloc_group_desc_add_entries - adjust count of free entries
143 * @desc: pointer to descriptor structure for the group
144 * @lock: spin lock protecting @desc
145 * @n: delta to be added
146 */
147static u32
148nilfs_palloc_group_desc_add_entries(struct nilfs_palloc_group_desc *desc,
149 spinlock_t *lock, u32 n)
150{
151 u32 nfree;
152
153 spin_lock(lock);
154 le32_add_cpu(&desc->pg_nfrees, n);
155 nfree = le32_to_cpu(desc->pg_nfrees);
156 spin_unlock(lock);
157 return nfree;
158}
159
160/**
161 * nilfs_palloc_entry_blkoff - get block offset of an entry block
162 * @inode: inode of metadata file using this allocator
163 * @nr: serial number of the entry (e.g. inode number)
164 */
165static unsigned long
166nilfs_palloc_entry_blkoff(const struct inode *inode, __u64 nr)
167{
168 unsigned long group, group_offset;
169
170 group = nilfs_palloc_group(inode, nr, &group_offset);
171
172 return nilfs_palloc_bitmap_blkoff(inode, group) + 1 +
173 group_offset / NILFS_MDT(inode)->mi_entries_per_block;
174}
175
176/**
177 * nilfs_palloc_desc_block_init - initialize buffer of a group descriptor block
178 * @inode: inode of metadata file
179 * @bh: buffer head of the buffer to be initialized
180 * @from: kernel address mapped for a chunk of the block
181 *
182 * This function does not yet support the case where block size > PAGE_SIZE.
183 */
184static void nilfs_palloc_desc_block_init(struct inode *inode,
185 struct buffer_head *bh, void *from)
186{
187 struct nilfs_palloc_group_desc *desc = from;
188 unsigned long n = nilfs_palloc_groups_per_desc_block(inode);
189 __le32 nfrees;
190
191 nfrees = cpu_to_le32(nilfs_palloc_entries_per_group(inode));
192 while (n-- > 0) {
193 desc->pg_nfrees = nfrees;
194 desc++;
195 }
196}
197
198static int nilfs_palloc_get_block(struct inode *inode, unsigned long blkoff,
199 int create,
200 void (*init_block)(struct inode *,
201 struct buffer_head *,
202 void *),
203 struct buffer_head **bhp,
204 struct nilfs_bh_assoc *prev,
205 spinlock_t *lock)
206{
207 int ret;
208
209 spin_lock(lock);
210 if (prev->bh && blkoff == prev->blkoff &&
211 likely(buffer_uptodate(prev->bh))) {
212 get_bh(prev->bh);
213 *bhp = prev->bh;
214 spin_unlock(lock);
215 return 0;
216 }
217 spin_unlock(lock);
218
219 ret = nilfs_mdt_get_block(inode, blkoff, create, init_block, bhp);
220 if (!ret) {
221 spin_lock(lock);
222 /*
223 * The following code must be safe for change of the
224 * cache contents during the get block call.
225 */
226 brelse(prev->bh);
227 get_bh(*bhp);
228 prev->bh = *bhp;
229 prev->blkoff = blkoff;
230 spin_unlock(lock);
231 }
232 return ret;
233}
234
235/**
236 * nilfs_palloc_delete_block - delete a block on the persistent allocator file
237 * @inode: inode of metadata file using this allocator
238 * @blkoff: block offset
239 * @prev: nilfs_bh_assoc struct of the last used buffer
240 * @lock: spin lock protecting @prev
241 */
242static int nilfs_palloc_delete_block(struct inode *inode, unsigned long blkoff,
243 struct nilfs_bh_assoc *prev,
244 spinlock_t *lock)
245{
246 spin_lock(lock);
247 if (prev->bh && blkoff == prev->blkoff) {
248 brelse(prev->bh);
249 prev->bh = NULL;
250 }
251 spin_unlock(lock);
252 return nilfs_mdt_delete_block(inode, blkoff);
253}
254
255/**
256 * nilfs_palloc_get_desc_block - get buffer head of a group descriptor block
257 * @inode: inode of metadata file using this allocator
258 * @group: group number
259 * @create: create flag
260 * @bhp: pointer to store the resultant buffer head
261 */
262static int nilfs_palloc_get_desc_block(struct inode *inode,
263 unsigned long group,
264 int create, struct buffer_head **bhp)
265{
266 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
267
268 return nilfs_palloc_get_block(inode,
269 nilfs_palloc_desc_blkoff(inode, group),
270 create, nilfs_palloc_desc_block_init,
271 bhp, &cache->prev_desc, &cache->lock);
272}
273
274/**
275 * nilfs_palloc_get_bitmap_block - get buffer head of a bitmap block
276 * @inode: inode of metadata file using this allocator
277 * @group: group number
278 * @create: create flag
279 * @bhp: pointer to store the resultant buffer head
280 */
281static int nilfs_palloc_get_bitmap_block(struct inode *inode,
282 unsigned long group,
283 int create, struct buffer_head **bhp)
284{
285 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
286
287 return nilfs_palloc_get_block(inode,
288 nilfs_palloc_bitmap_blkoff(inode, group),
289 create, NULL, bhp,
290 &cache->prev_bitmap, &cache->lock);
291}
292
293/**
294 * nilfs_palloc_delete_bitmap_block - delete a bitmap block
295 * @inode: inode of metadata file using this allocator
296 * @group: group number
297 */
298static int nilfs_palloc_delete_bitmap_block(struct inode *inode,
299 unsigned long group)
300{
301 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
302
303 return nilfs_palloc_delete_block(inode,
304 nilfs_palloc_bitmap_blkoff(inode,
305 group),
306 &cache->prev_bitmap, &cache->lock);
307}
308
309/**
310 * nilfs_palloc_get_entry_block - get buffer head of an entry block
311 * @inode: inode of metadata file using this allocator
312 * @nr: serial number of the entry (e.g. inode number)
313 * @create: create flag
314 * @bhp: pointer to store the resultant buffer head
315 */
316int nilfs_palloc_get_entry_block(struct inode *inode, __u64 nr,
317 int create, struct buffer_head **bhp)
318{
319 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
320
321 return nilfs_palloc_get_block(inode,
322 nilfs_palloc_entry_blkoff(inode, nr),
323 create, NULL, bhp,
324 &cache->prev_entry, &cache->lock);
325}
326
327/**
328 * nilfs_palloc_delete_entry_block - delete an entry block
329 * @inode: inode of metadata file using this allocator
330 * @nr: serial number of the entry
331 */
332static int nilfs_palloc_delete_entry_block(struct inode *inode, __u64 nr)
333{
334 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
335
336 return nilfs_palloc_delete_block(inode,
337 nilfs_palloc_entry_blkoff(inode, nr),
338 &cache->prev_entry, &cache->lock);
339}
340
341/**
342 * nilfs_palloc_group_desc_offset - calculate the byte offset of a group
343 * descriptor in the folio containing it
344 * @inode: inode of metadata file using this allocator
345 * @group: group number
346 * @bh: buffer head of the group descriptor block
347 *
348 * Return: Byte offset in the folio of the group descriptor for @group.
349 */
350static size_t nilfs_palloc_group_desc_offset(const struct inode *inode,
351 unsigned long group,
352 const struct buffer_head *bh)
353{
354 return offset_in_folio(bh->b_folio, bh->b_data) +
355 sizeof(struct nilfs_palloc_group_desc) *
356 (group % nilfs_palloc_groups_per_desc_block(inode));
357}
358
359/**
360 * nilfs_palloc_bitmap_offset - calculate the byte offset of a bitmap block
361 * in the folio containing it
362 * @bh: buffer head of the bitmap block
363 *
364 * Return: Byte offset in the folio of the bitmap block for @bh.
365 */
366static size_t nilfs_palloc_bitmap_offset(const struct buffer_head *bh)
367{
368 return offset_in_folio(bh->b_folio, bh->b_data);
369}
370
371/**
372 * nilfs_palloc_entry_offset - calculate the byte offset of an entry in the
373 * folio containing it
374 * @inode: inode of metadata file using this allocator
375 * @nr: serial number of the entry (e.g. inode number)
376 * @bh: buffer head of the entry block
377 *
378 * Return: Byte offset in the folio of the entry @nr.
379 */
380size_t nilfs_palloc_entry_offset(const struct inode *inode, __u64 nr,
381 const struct buffer_head *bh)
382{
383 unsigned long entry_index_in_group, entry_index_in_block;
384
385 nilfs_palloc_group(inode, nr, &entry_index_in_group);
386 entry_index_in_block = entry_index_in_group %
387 NILFS_MDT(inode)->mi_entries_per_block;
388
389 return offset_in_folio(bh->b_folio, bh->b_data) +
390 entry_index_in_block * NILFS_MDT(inode)->mi_entry_size;
391}
392
393/**
394 * nilfs_palloc_find_available_slot - find available slot in a group
395 * @bitmap: bitmap of the group
396 * @target: offset number of an entry in the group (start point)
397 * @bsize: size in bits
398 * @lock: spin lock protecting @bitmap
399 * @wrap: whether to wrap around
400 */
401static int nilfs_palloc_find_available_slot(unsigned char *bitmap,
402 unsigned long target,
403 unsigned int bsize,
404 spinlock_t *lock, bool wrap)
405{
406 int pos, end = bsize;
407
408 if (likely(target < bsize)) {
409 pos = target;
410 do {
411 pos = nilfs_find_next_zero_bit(bitmap, end, pos);
412 if (pos >= end)
413 break;
414 if (!nilfs_set_bit_atomic(lock, pos, bitmap))
415 return pos;
416 } while (++pos < end);
417
418 end = target;
419 }
420 if (!wrap)
421 return -ENOSPC;
422
423 /* wrap around */
424 for (pos = 0; pos < end; pos++) {
425 pos = nilfs_find_next_zero_bit(bitmap, end, pos);
426 if (pos >= end)
427 break;
428 if (!nilfs_set_bit_atomic(lock, pos, bitmap))
429 return pos;
430 }
431
432 return -ENOSPC;
433}
434
435/**
436 * nilfs_palloc_rest_groups_in_desc_block - get the remaining number of groups
437 * in a group descriptor block
438 * @inode: inode of metadata file using this allocator
439 * @curr: current group number
440 * @max: maximum number of groups
441 */
442static unsigned long
443nilfs_palloc_rest_groups_in_desc_block(const struct inode *inode,
444 unsigned long curr, unsigned long max)
445{
446 return min_t(unsigned long,
447 nilfs_palloc_groups_per_desc_block(inode) -
448 curr % nilfs_palloc_groups_per_desc_block(inode),
449 max - curr + 1);
450}
451
452/**
453 * nilfs_palloc_count_desc_blocks - count descriptor blocks number
454 * @inode: inode of metadata file using this allocator
455 * @desc_blocks: descriptor blocks number [out]
456 */
457static int nilfs_palloc_count_desc_blocks(struct inode *inode,
458 unsigned long *desc_blocks)
459{
460 __u64 blknum;
461 int ret;
462
463 ret = nilfs_bmap_last_key(NILFS_I(inode)->i_bmap, &blknum);
464 if (likely(!ret))
465 *desc_blocks = DIV_ROUND_UP(
466 (unsigned long)blknum,
467 NILFS_MDT(inode)->mi_blocks_per_desc_block);
468 return ret;
469}
470
471/**
472 * nilfs_palloc_mdt_file_can_grow - check potential opportunity for
473 * MDT file growing
474 * @inode: inode of metadata file using this allocator
475 * @desc_blocks: known current descriptor blocks count
476 */
477static inline bool nilfs_palloc_mdt_file_can_grow(struct inode *inode,
478 unsigned long desc_blocks)
479{
480 return (nilfs_palloc_groups_per_desc_block(inode) * desc_blocks) <
481 nilfs_palloc_groups_count(inode);
482}
483
484/**
485 * nilfs_palloc_count_max_entries - count max number of entries that can be
486 * described by descriptor blocks count
487 * @inode: inode of metadata file using this allocator
488 * @nused: current number of used entries
489 * @nmaxp: max number of entries [out]
490 */
491int nilfs_palloc_count_max_entries(struct inode *inode, u64 nused, u64 *nmaxp)
492{
493 unsigned long desc_blocks = 0;
494 u64 entries_per_desc_block, nmax;
495 int err;
496
497 err = nilfs_palloc_count_desc_blocks(inode, &desc_blocks);
498 if (unlikely(err))
499 return err;
500
501 entries_per_desc_block = (u64)nilfs_palloc_entries_per_group(inode) *
502 nilfs_palloc_groups_per_desc_block(inode);
503 nmax = entries_per_desc_block * desc_blocks;
504
505 if (nused == nmax &&
506 nilfs_palloc_mdt_file_can_grow(inode, desc_blocks))
507 nmax += entries_per_desc_block;
508
509 if (nused > nmax)
510 return -ERANGE;
511
512 *nmaxp = nmax;
513 return 0;
514}
515
516/**
517 * nilfs_palloc_prepare_alloc_entry - prepare to allocate a persistent object
518 * @inode: inode of metadata file using this allocator
519 * @req: nilfs_palloc_req structure exchanged for the allocation
520 * @wrap: whether to wrap around
521 */
522int nilfs_palloc_prepare_alloc_entry(struct inode *inode,
523 struct nilfs_palloc_req *req, bool wrap)
524{
525 struct buffer_head *desc_bh, *bitmap_bh;
526 struct nilfs_palloc_group_desc *desc;
527 unsigned char *bitmap;
528 size_t doff, boff;
529 unsigned long group, maxgroup, ngroups;
530 unsigned long group_offset, maxgroup_offset;
531 unsigned long n, entries_per_group;
532 unsigned long i, j;
533 spinlock_t *lock;
534 int pos, ret;
535
536 ngroups = nilfs_palloc_groups_count(inode);
537 maxgroup = ngroups - 1;
538 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
539 entries_per_group = nilfs_palloc_entries_per_group(inode);
540
541 for (i = 0; i < ngroups; i += n) {
542 if (group >= ngroups && wrap) {
543 /* wrap around */
544 group = 0;
545 maxgroup = nilfs_palloc_group(inode, req->pr_entry_nr,
546 &maxgroup_offset) - 1;
547 }
548 ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
549 if (ret < 0)
550 return ret;
551
552 doff = nilfs_palloc_group_desc_offset(inode, group, desc_bh);
553 desc = kmap_local_folio(desc_bh->b_folio, doff);
554 n = nilfs_palloc_rest_groups_in_desc_block(inode, group,
555 maxgroup);
556 for (j = 0; j < n; j++, group++, group_offset = 0) {
557 lock = nilfs_mdt_bgl_lock(inode, group);
558 if (nilfs_palloc_group_desc_nfrees(&desc[j], lock) == 0)
559 continue;
560
561 kunmap_local(desc);
562 ret = nilfs_palloc_get_bitmap_block(inode, group, 1,
563 &bitmap_bh);
564 if (unlikely(ret < 0)) {
565 brelse(desc_bh);
566 return ret;
567 }
568
569 /*
570 * Re-kmap the folio containing the first (and
571 * subsequent) group descriptors.
572 */
573 desc = kmap_local_folio(desc_bh->b_folio, doff);
574
575 boff = nilfs_palloc_bitmap_offset(bitmap_bh);
576 bitmap = kmap_local_folio(bitmap_bh->b_folio, boff);
577 pos = nilfs_palloc_find_available_slot(
578 bitmap, group_offset, entries_per_group, lock,
579 wrap);
580 /*
581 * Since the search for a free slot in the second and
582 * subsequent bitmap blocks always starts from the
583 * beginning, the wrap flag only has an effect on the
584 * first search.
585 */
586 kunmap_local(bitmap);
587 if (pos >= 0)
588 goto found;
589
590 brelse(bitmap_bh);
591 }
592
593 kunmap_local(desc);
594 brelse(desc_bh);
595 }
596
597 /* no entries left */
598 return -ENOSPC;
599
600found:
601 /* found a free entry */
602 nilfs_palloc_group_desc_add_entries(&desc[j], lock, -1);
603 req->pr_entry_nr = entries_per_group * group + pos;
604 kunmap_local(desc);
605
606 req->pr_desc_bh = desc_bh;
607 req->pr_bitmap_bh = bitmap_bh;
608 return 0;
609}
610
611/**
612 * nilfs_palloc_commit_alloc_entry - finish allocation of a persistent object
613 * @inode: inode of metadata file using this allocator
614 * @req: nilfs_palloc_req structure exchanged for the allocation
615 */
616void nilfs_palloc_commit_alloc_entry(struct inode *inode,
617 struct nilfs_palloc_req *req)
618{
619 mark_buffer_dirty(req->pr_bitmap_bh);
620 mark_buffer_dirty(req->pr_desc_bh);
621 nilfs_mdt_mark_dirty(inode);
622
623 brelse(req->pr_bitmap_bh);
624 brelse(req->pr_desc_bh);
625}
626
627/**
628 * nilfs_palloc_commit_free_entry - finish deallocating a persistent object
629 * @inode: inode of metadata file using this allocator
630 * @req: nilfs_palloc_req structure exchanged for the removal
631 */
632void nilfs_palloc_commit_free_entry(struct inode *inode,
633 struct nilfs_palloc_req *req)
634{
635 unsigned long group, group_offset;
636 size_t doff, boff;
637 struct nilfs_palloc_group_desc *desc;
638 unsigned char *bitmap;
639 spinlock_t *lock;
640
641 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
642 doff = nilfs_palloc_group_desc_offset(inode, group, req->pr_desc_bh);
643 desc = kmap_local_folio(req->pr_desc_bh->b_folio, doff);
644
645 boff = nilfs_palloc_bitmap_offset(req->pr_bitmap_bh);
646 bitmap = kmap_local_folio(req->pr_bitmap_bh->b_folio, boff);
647 lock = nilfs_mdt_bgl_lock(inode, group);
648
649 if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
650 nilfs_warn(inode->i_sb,
651 "%s (ino=%lu): entry number %llu already freed",
652 __func__, inode->i_ino,
653 (unsigned long long)req->pr_entry_nr);
654 else
655 nilfs_palloc_group_desc_add_entries(desc, lock, 1);
656
657 kunmap_local(bitmap);
658 kunmap_local(desc);
659
660 mark_buffer_dirty(req->pr_desc_bh);
661 mark_buffer_dirty(req->pr_bitmap_bh);
662 nilfs_mdt_mark_dirty(inode);
663
664 brelse(req->pr_bitmap_bh);
665 brelse(req->pr_desc_bh);
666}
667
668/**
669 * nilfs_palloc_abort_alloc_entry - cancel allocation of a persistent object
670 * @inode: inode of metadata file using this allocator
671 * @req: nilfs_palloc_req structure exchanged for the allocation
672 */
673void nilfs_palloc_abort_alloc_entry(struct inode *inode,
674 struct nilfs_palloc_req *req)
675{
676 struct nilfs_palloc_group_desc *desc;
677 size_t doff, boff;
678 unsigned char *bitmap;
679 unsigned long group, group_offset;
680 spinlock_t *lock;
681
682 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
683 doff = nilfs_palloc_group_desc_offset(inode, group, req->pr_desc_bh);
684 desc = kmap_local_folio(req->pr_desc_bh->b_folio, doff);
685
686 boff = nilfs_palloc_bitmap_offset(req->pr_bitmap_bh);
687 bitmap = kmap_local_folio(req->pr_bitmap_bh->b_folio, boff);
688 lock = nilfs_mdt_bgl_lock(inode, group);
689
690 if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
691 nilfs_warn(inode->i_sb,
692 "%s (ino=%lu): entry number %llu already freed",
693 __func__, inode->i_ino,
694 (unsigned long long)req->pr_entry_nr);
695 else
696 nilfs_palloc_group_desc_add_entries(desc, lock, 1);
697
698 kunmap_local(bitmap);
699 kunmap_local(desc);
700
701 brelse(req->pr_bitmap_bh);
702 brelse(req->pr_desc_bh);
703
704 req->pr_entry_nr = 0;
705 req->pr_bitmap_bh = NULL;
706 req->pr_desc_bh = NULL;
707}
708
709/**
710 * nilfs_palloc_prepare_free_entry - prepare to deallocate a persistent object
711 * @inode: inode of metadata file using this allocator
712 * @req: nilfs_palloc_req structure exchanged for the removal
713 */
714int nilfs_palloc_prepare_free_entry(struct inode *inode,
715 struct nilfs_palloc_req *req)
716{
717 struct buffer_head *desc_bh, *bitmap_bh;
718 unsigned long group, group_offset;
719 int ret;
720
721 group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
722 ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
723 if (ret < 0)
724 return ret;
725 ret = nilfs_palloc_get_bitmap_block(inode, group, 1, &bitmap_bh);
726 if (ret < 0) {
727 brelse(desc_bh);
728 return ret;
729 }
730
731 req->pr_desc_bh = desc_bh;
732 req->pr_bitmap_bh = bitmap_bh;
733 return 0;
734}
735
736/**
737 * nilfs_palloc_abort_free_entry - cancel deallocating a persistent object
738 * @inode: inode of metadata file using this allocator
739 * @req: nilfs_palloc_req structure exchanged for the removal
740 */
741void nilfs_palloc_abort_free_entry(struct inode *inode,
742 struct nilfs_palloc_req *req)
743{
744 brelse(req->pr_bitmap_bh);
745 brelse(req->pr_desc_bh);
746
747 req->pr_entry_nr = 0;
748 req->pr_bitmap_bh = NULL;
749 req->pr_desc_bh = NULL;
750}
751
752/**
753 * nilfs_palloc_freev - deallocate a set of persistent objects
754 * @inode: inode of metadata file using this allocator
755 * @entry_nrs: array of entry numbers to be deallocated
756 * @nitems: number of entries stored in @entry_nrs
757 */
758int nilfs_palloc_freev(struct inode *inode, __u64 *entry_nrs, size_t nitems)
759{
760 struct buffer_head *desc_bh, *bitmap_bh;
761 struct nilfs_palloc_group_desc *desc;
762 unsigned char *bitmap;
763 size_t doff, boff;
764 unsigned long group, group_offset;
765 __u64 group_min_nr, last_nrs[8];
766 const unsigned long epg = nilfs_palloc_entries_per_group(inode);
767 const unsigned int epb = NILFS_MDT(inode)->mi_entries_per_block;
768 unsigned int entry_start, end, pos;
769 spinlock_t *lock;
770 int i, j, k, ret;
771 u32 nfree;
772
773 for (i = 0; i < nitems; i = j) {
774 int change_group = false;
775 int nempties = 0, n = 0;
776
777 group = nilfs_palloc_group(inode, entry_nrs[i], &group_offset);
778 ret = nilfs_palloc_get_desc_block(inode, group, 0, &desc_bh);
779 if (ret < 0)
780 return ret;
781 ret = nilfs_palloc_get_bitmap_block(inode, group, 0,
782 &bitmap_bh);
783 if (ret < 0) {
784 brelse(desc_bh);
785 return ret;
786 }
787
788 /* Get the first entry number of the group */
789 group_min_nr = (__u64)group * epg;
790
791 boff = nilfs_palloc_bitmap_offset(bitmap_bh);
792 bitmap = kmap_local_folio(bitmap_bh->b_folio, boff);
793 lock = nilfs_mdt_bgl_lock(inode, group);
794
795 j = i;
796 entry_start = rounddown(group_offset, epb);
797 do {
798 if (!nilfs_clear_bit_atomic(lock, group_offset,
799 bitmap)) {
800 nilfs_warn(inode->i_sb,
801 "%s (ino=%lu): entry number %llu already freed",
802 __func__, inode->i_ino,
803 (unsigned long long)entry_nrs[j]);
804 } else {
805 n++;
806 }
807
808 j++;
809 if (j >= nitems || entry_nrs[j] < group_min_nr ||
810 entry_nrs[j] >= group_min_nr + epg) {
811 change_group = true;
812 } else {
813 group_offset = entry_nrs[j] - group_min_nr;
814 if (group_offset >= entry_start &&
815 group_offset < entry_start + epb) {
816 /* This entry is in the same block */
817 continue;
818 }
819 }
820
821 /* Test if the entry block is empty or not */
822 end = entry_start + epb;
823 pos = nilfs_find_next_bit(bitmap, end, entry_start);
824 if (pos >= end) {
825 last_nrs[nempties++] = entry_nrs[j - 1];
826 if (nempties >= ARRAY_SIZE(last_nrs))
827 break;
828 }
829
830 if (change_group)
831 break;
832
833 /* Go on to the next entry block */
834 entry_start = rounddown(group_offset, epb);
835 } while (true);
836
837 kunmap_local(bitmap);
838 mark_buffer_dirty(bitmap_bh);
839 brelse(bitmap_bh);
840
841 for (k = 0; k < nempties; k++) {
842 ret = nilfs_palloc_delete_entry_block(inode,
843 last_nrs[k]);
844 if (ret && ret != -ENOENT)
845 nilfs_warn(inode->i_sb,
846 "error %d deleting block that object (entry=%llu, ino=%lu) belongs to",
847 ret, (unsigned long long)last_nrs[k],
848 inode->i_ino);
849 }
850
851 doff = nilfs_palloc_group_desc_offset(inode, group, desc_bh);
852 desc = kmap_local_folio(desc_bh->b_folio, doff);
853 nfree = nilfs_palloc_group_desc_add_entries(desc, lock, n);
854 kunmap_local(desc);
855 mark_buffer_dirty(desc_bh);
856 nilfs_mdt_mark_dirty(inode);
857 brelse(desc_bh);
858
859 if (nfree == nilfs_palloc_entries_per_group(inode)) {
860 ret = nilfs_palloc_delete_bitmap_block(inode, group);
861 if (ret && ret != -ENOENT)
862 nilfs_warn(inode->i_sb,
863 "error %d deleting bitmap block of group=%lu, ino=%lu",
864 ret, group, inode->i_ino);
865 }
866 }
867 return 0;
868}
869
870void nilfs_palloc_setup_cache(struct inode *inode,
871 struct nilfs_palloc_cache *cache)
872{
873 NILFS_MDT(inode)->mi_palloc_cache = cache;
874 spin_lock_init(&cache->lock);
875}
876
877void nilfs_palloc_clear_cache(struct inode *inode)
878{
879 struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
880
881 spin_lock(&cache->lock);
882 brelse(cache->prev_desc.bh);
883 brelse(cache->prev_bitmap.bh);
884 brelse(cache->prev_entry.bh);
885 cache->prev_desc.bh = NULL;
886 cache->prev_bitmap.bh = NULL;
887 cache->prev_entry.bh = NULL;
888 spin_unlock(&cache->lock);
889}
890
891void nilfs_palloc_destroy_cache(struct inode *inode)
892{
893 nilfs_palloc_clear_cache(inode);
894 NILFS_MDT(inode)->mi_palloc_cache = NULL;
895}