xref: /openbmc/linux/fs/nilfs2/alloc.c (revision 31e67366)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * alloc.c - 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  */
25 static inline unsigned long
26 nilfs_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  */
36 static inline unsigned long
37 nilfs_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  */
47 int 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  */
82 static 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  */
99 static unsigned long
100 nilfs_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  */
115 static unsigned long
116 nilfs_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  */
129 static unsigned long
130 nilfs_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  */
147 static u32
148 nilfs_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  */
165 static unsigned long
166 nilfs_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  */
182 static 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 
196 static 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 		get_bh(prev->bh);
210 		*bhp = prev->bh;
211 		spin_unlock(lock);
212 		return 0;
213 	}
214 	spin_unlock(lock);
215 
216 	ret = nilfs_mdt_get_block(inode, blkoff, create, init_block, bhp);
217 	if (!ret) {
218 		spin_lock(lock);
219 		/*
220 		 * The following code must be safe for change of the
221 		 * cache contents during the get block call.
222 		 */
223 		brelse(prev->bh);
224 		get_bh(*bhp);
225 		prev->bh = *bhp;
226 		prev->blkoff = blkoff;
227 		spin_unlock(lock);
228 	}
229 	return ret;
230 }
231 
232 /**
233  * nilfs_palloc_delete_block - delete a block on the persistent allocator file
234  * @inode: inode of metadata file using this allocator
235  * @blkoff: block offset
236  * @prev: nilfs_bh_assoc struct of the last used buffer
237  * @lock: spin lock protecting @prev
238  */
239 static int nilfs_palloc_delete_block(struct inode *inode, unsigned long blkoff,
240 				     struct nilfs_bh_assoc *prev,
241 				     spinlock_t *lock)
242 {
243 	spin_lock(lock);
244 	if (prev->bh && blkoff == prev->blkoff) {
245 		brelse(prev->bh);
246 		prev->bh = NULL;
247 	}
248 	spin_unlock(lock);
249 	return nilfs_mdt_delete_block(inode, blkoff);
250 }
251 
252 /**
253  * nilfs_palloc_get_desc_block - get buffer head of a group descriptor block
254  * @inode: inode of metadata file using this allocator
255  * @group: group number
256  * @create: create flag
257  * @bhp: pointer to store the resultant buffer head
258  */
259 static int nilfs_palloc_get_desc_block(struct inode *inode,
260 				       unsigned long group,
261 				       int create, struct buffer_head **bhp)
262 {
263 	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
264 
265 	return nilfs_palloc_get_block(inode,
266 				      nilfs_palloc_desc_blkoff(inode, group),
267 				      create, nilfs_palloc_desc_block_init,
268 				      bhp, &cache->prev_desc, &cache->lock);
269 }
270 
271 /**
272  * nilfs_palloc_get_bitmap_block - get buffer head of a bitmap block
273  * @inode: inode of metadata file using this allocator
274  * @group: group number
275  * @create: create flag
276  * @bhp: pointer to store the resultant buffer head
277  */
278 static int nilfs_palloc_get_bitmap_block(struct inode *inode,
279 					 unsigned long group,
280 					 int create, struct buffer_head **bhp)
281 {
282 	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
283 
284 	return nilfs_palloc_get_block(inode,
285 				      nilfs_palloc_bitmap_blkoff(inode, group),
286 				      create, NULL, bhp,
287 				      &cache->prev_bitmap, &cache->lock);
288 }
289 
290 /**
291  * nilfs_palloc_delete_bitmap_block - delete a bitmap block
292  * @inode: inode of metadata file using this allocator
293  * @group: group number
294  */
295 static int nilfs_palloc_delete_bitmap_block(struct inode *inode,
296 					    unsigned long group)
297 {
298 	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
299 
300 	return nilfs_palloc_delete_block(inode,
301 					 nilfs_palloc_bitmap_blkoff(inode,
302 								    group),
303 					 &cache->prev_bitmap, &cache->lock);
304 }
305 
306 /**
307  * nilfs_palloc_get_entry_block - get buffer head of an entry block
308  * @inode: inode of metadata file using this allocator
309  * @nr: serial number of the entry (e.g. inode number)
310  * @create: create flag
311  * @bhp: pointer to store the resultant buffer head
312  */
313 int nilfs_palloc_get_entry_block(struct inode *inode, __u64 nr,
314 				 int create, struct buffer_head **bhp)
315 {
316 	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
317 
318 	return nilfs_palloc_get_block(inode,
319 				      nilfs_palloc_entry_blkoff(inode, nr),
320 				      create, NULL, bhp,
321 				      &cache->prev_entry, &cache->lock);
322 }
323 
324 /**
325  * nilfs_palloc_delete_entry_block - delete an entry block
326  * @inode: inode of metadata file using this allocator
327  * @nr: serial number of the entry
328  */
329 static int nilfs_palloc_delete_entry_block(struct inode *inode, __u64 nr)
330 {
331 	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
332 
333 	return nilfs_palloc_delete_block(inode,
334 					 nilfs_palloc_entry_blkoff(inode, nr),
335 					 &cache->prev_entry, &cache->lock);
336 }
337 
338 /**
339  * nilfs_palloc_block_get_group_desc - get kernel address of a group descriptor
340  * @inode: inode of metadata file using this allocator
341  * @group: group number
342  * @bh: buffer head of the buffer storing the group descriptor block
343  * @kaddr: kernel address mapped for the page including the buffer
344  */
345 static struct nilfs_palloc_group_desc *
346 nilfs_palloc_block_get_group_desc(const struct inode *inode,
347 				  unsigned long group,
348 				  const struct buffer_head *bh, void *kaddr)
349 {
350 	return (struct nilfs_palloc_group_desc *)(kaddr + bh_offset(bh)) +
351 		group % nilfs_palloc_groups_per_desc_block(inode);
352 }
353 
354 /**
355  * nilfs_palloc_block_get_entry - get kernel address of an entry
356  * @inode: inode of metadata file using this allocator
357  * @nr: serial number of the entry (e.g. inode number)
358  * @bh: buffer head of the buffer storing the entry block
359  * @kaddr: kernel address mapped for the page including the buffer
360  */
361 void *nilfs_palloc_block_get_entry(const struct inode *inode, __u64 nr,
362 				   const struct buffer_head *bh, void *kaddr)
363 {
364 	unsigned long entry_offset, group_offset;
365 
366 	nilfs_palloc_group(inode, nr, &group_offset);
367 	entry_offset = group_offset % NILFS_MDT(inode)->mi_entries_per_block;
368 
369 	return kaddr + bh_offset(bh) +
370 		entry_offset * NILFS_MDT(inode)->mi_entry_size;
371 }
372 
373 /**
374  * nilfs_palloc_find_available_slot - find available slot in a group
375  * @bitmap: bitmap of the group
376  * @target: offset number of an entry in the group (start point)
377  * @bsize: size in bits
378  * @lock: spin lock protecting @bitmap
379  */
380 static int nilfs_palloc_find_available_slot(unsigned char *bitmap,
381 					    unsigned long target,
382 					    unsigned int bsize,
383 					    spinlock_t *lock)
384 {
385 	int pos, end = bsize;
386 
387 	if (likely(target < bsize)) {
388 		pos = target;
389 		do {
390 			pos = nilfs_find_next_zero_bit(bitmap, end, pos);
391 			if (pos >= end)
392 				break;
393 			if (!nilfs_set_bit_atomic(lock, pos, bitmap))
394 				return pos;
395 		} while (++pos < end);
396 
397 		end = target;
398 	}
399 
400 	/* wrap around */
401 	for (pos = 0; pos < end; pos++) {
402 		pos = nilfs_find_next_zero_bit(bitmap, end, pos);
403 		if (pos >= end)
404 			break;
405 		if (!nilfs_set_bit_atomic(lock, pos, bitmap))
406 			return pos;
407 	}
408 
409 	return -ENOSPC;
410 }
411 
412 /**
413  * nilfs_palloc_rest_groups_in_desc_block - get the remaining number of groups
414  *					    in a group descriptor block
415  * @inode: inode of metadata file using this allocator
416  * @curr: current group number
417  * @max: maximum number of groups
418  */
419 static unsigned long
420 nilfs_palloc_rest_groups_in_desc_block(const struct inode *inode,
421 				       unsigned long curr, unsigned long max)
422 {
423 	return min_t(unsigned long,
424 		     nilfs_palloc_groups_per_desc_block(inode) -
425 		     curr % nilfs_palloc_groups_per_desc_block(inode),
426 		     max - curr + 1);
427 }
428 
429 /**
430  * nilfs_palloc_count_desc_blocks - count descriptor blocks number
431  * @inode: inode of metadata file using this allocator
432  * @desc_blocks: descriptor blocks number [out]
433  */
434 static int nilfs_palloc_count_desc_blocks(struct inode *inode,
435 					    unsigned long *desc_blocks)
436 {
437 	__u64 blknum;
438 	int ret;
439 
440 	ret = nilfs_bmap_last_key(NILFS_I(inode)->i_bmap, &blknum);
441 	if (likely(!ret))
442 		*desc_blocks = DIV_ROUND_UP(
443 			(unsigned long)blknum,
444 			NILFS_MDT(inode)->mi_blocks_per_desc_block);
445 	return ret;
446 }
447 
448 /**
449  * nilfs_palloc_mdt_file_can_grow - check potential opportunity for
450  *					MDT file growing
451  * @inode: inode of metadata file using this allocator
452  * @desc_blocks: known current descriptor blocks count
453  */
454 static inline bool nilfs_palloc_mdt_file_can_grow(struct inode *inode,
455 						    unsigned long desc_blocks)
456 {
457 	return (nilfs_palloc_groups_per_desc_block(inode) * desc_blocks) <
458 			nilfs_palloc_groups_count(inode);
459 }
460 
461 /**
462  * nilfs_palloc_count_max_entries - count max number of entries that can be
463  *					described by descriptor blocks count
464  * @inode: inode of metadata file using this allocator
465  * @nused: current number of used entries
466  * @nmaxp: max number of entries [out]
467  */
468 int nilfs_palloc_count_max_entries(struct inode *inode, u64 nused, u64 *nmaxp)
469 {
470 	unsigned long desc_blocks = 0;
471 	u64 entries_per_desc_block, nmax;
472 	int err;
473 
474 	err = nilfs_palloc_count_desc_blocks(inode, &desc_blocks);
475 	if (unlikely(err))
476 		return err;
477 
478 	entries_per_desc_block = (u64)nilfs_palloc_entries_per_group(inode) *
479 				nilfs_palloc_groups_per_desc_block(inode);
480 	nmax = entries_per_desc_block * desc_blocks;
481 
482 	if (nused == nmax &&
483 			nilfs_palloc_mdt_file_can_grow(inode, desc_blocks))
484 		nmax += entries_per_desc_block;
485 
486 	if (nused > nmax)
487 		return -ERANGE;
488 
489 	*nmaxp = nmax;
490 	return 0;
491 }
492 
493 /**
494  * nilfs_palloc_prepare_alloc_entry - prepare to allocate a persistent object
495  * @inode: inode of metadata file using this allocator
496  * @req: nilfs_palloc_req structure exchanged for the allocation
497  */
498 int nilfs_palloc_prepare_alloc_entry(struct inode *inode,
499 				     struct nilfs_palloc_req *req)
500 {
501 	struct buffer_head *desc_bh, *bitmap_bh;
502 	struct nilfs_palloc_group_desc *desc;
503 	unsigned char *bitmap;
504 	void *desc_kaddr, *bitmap_kaddr;
505 	unsigned long group, maxgroup, ngroups;
506 	unsigned long group_offset, maxgroup_offset;
507 	unsigned long n, entries_per_group;
508 	unsigned long i, j;
509 	spinlock_t *lock;
510 	int pos, ret;
511 
512 	ngroups = nilfs_palloc_groups_count(inode);
513 	maxgroup = ngroups - 1;
514 	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
515 	entries_per_group = nilfs_palloc_entries_per_group(inode);
516 
517 	for (i = 0; i < ngroups; i += n) {
518 		if (group >= ngroups) {
519 			/* wrap around */
520 			group = 0;
521 			maxgroup = nilfs_palloc_group(inode, req->pr_entry_nr,
522 						      &maxgroup_offset) - 1;
523 		}
524 		ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
525 		if (ret < 0)
526 			return ret;
527 		desc_kaddr = kmap(desc_bh->b_page);
528 		desc = nilfs_palloc_block_get_group_desc(
529 			inode, group, desc_bh, desc_kaddr);
530 		n = nilfs_palloc_rest_groups_in_desc_block(inode, group,
531 							   maxgroup);
532 		for (j = 0; j < n; j++, desc++, group++) {
533 			lock = nilfs_mdt_bgl_lock(inode, group);
534 			if (nilfs_palloc_group_desc_nfrees(desc, lock) > 0) {
535 				ret = nilfs_palloc_get_bitmap_block(
536 					inode, group, 1, &bitmap_bh);
537 				if (ret < 0)
538 					goto out_desc;
539 				bitmap_kaddr = kmap(bitmap_bh->b_page);
540 				bitmap = bitmap_kaddr + bh_offset(bitmap_bh);
541 				pos = nilfs_palloc_find_available_slot(
542 					bitmap, group_offset,
543 					entries_per_group, lock);
544 				if (pos >= 0) {
545 					/* found a free entry */
546 					nilfs_palloc_group_desc_add_entries(
547 						desc, lock, -1);
548 					req->pr_entry_nr =
549 						entries_per_group * group + pos;
550 					kunmap(desc_bh->b_page);
551 					kunmap(bitmap_bh->b_page);
552 
553 					req->pr_desc_bh = desc_bh;
554 					req->pr_bitmap_bh = bitmap_bh;
555 					return 0;
556 				}
557 				kunmap(bitmap_bh->b_page);
558 				brelse(bitmap_bh);
559 			}
560 
561 			group_offset = 0;
562 		}
563 
564 		kunmap(desc_bh->b_page);
565 		brelse(desc_bh);
566 	}
567 
568 	/* no entries left */
569 	return -ENOSPC;
570 
571  out_desc:
572 	kunmap(desc_bh->b_page);
573 	brelse(desc_bh);
574 	return ret;
575 }
576 
577 /**
578  * nilfs_palloc_commit_alloc_entry - finish allocation of a persistent object
579  * @inode: inode of metadata file using this allocator
580  * @req: nilfs_palloc_req structure exchanged for the allocation
581  */
582 void nilfs_palloc_commit_alloc_entry(struct inode *inode,
583 				     struct nilfs_palloc_req *req)
584 {
585 	mark_buffer_dirty(req->pr_bitmap_bh);
586 	mark_buffer_dirty(req->pr_desc_bh);
587 	nilfs_mdt_mark_dirty(inode);
588 
589 	brelse(req->pr_bitmap_bh);
590 	brelse(req->pr_desc_bh);
591 }
592 
593 /**
594  * nilfs_palloc_commit_free_entry - finish deallocating a persistent object
595  * @inode: inode of metadata file using this allocator
596  * @req: nilfs_palloc_req structure exchanged for the removal
597  */
598 void nilfs_palloc_commit_free_entry(struct inode *inode,
599 				    struct nilfs_palloc_req *req)
600 {
601 	struct nilfs_palloc_group_desc *desc;
602 	unsigned long group, group_offset;
603 	unsigned char *bitmap;
604 	void *desc_kaddr, *bitmap_kaddr;
605 	spinlock_t *lock;
606 
607 	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
608 	desc_kaddr = kmap(req->pr_desc_bh->b_page);
609 	desc = nilfs_palloc_block_get_group_desc(inode, group,
610 						 req->pr_desc_bh, desc_kaddr);
611 	bitmap_kaddr = kmap(req->pr_bitmap_bh->b_page);
612 	bitmap = bitmap_kaddr + bh_offset(req->pr_bitmap_bh);
613 	lock = nilfs_mdt_bgl_lock(inode, group);
614 
615 	if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
616 		nilfs_warn(inode->i_sb,
617 			   "%s (ino=%lu): entry number %llu already freed",
618 			   __func__, inode->i_ino,
619 			   (unsigned long long)req->pr_entry_nr);
620 	else
621 		nilfs_palloc_group_desc_add_entries(desc, lock, 1);
622 
623 	kunmap(req->pr_bitmap_bh->b_page);
624 	kunmap(req->pr_desc_bh->b_page);
625 
626 	mark_buffer_dirty(req->pr_desc_bh);
627 	mark_buffer_dirty(req->pr_bitmap_bh);
628 	nilfs_mdt_mark_dirty(inode);
629 
630 	brelse(req->pr_bitmap_bh);
631 	brelse(req->pr_desc_bh);
632 }
633 
634 /**
635  * nilfs_palloc_abort_alloc_entry - cancel allocation of a persistent object
636  * @inode: inode of metadata file using this allocator
637  * @req: nilfs_palloc_req structure exchanged for the allocation
638  */
639 void nilfs_palloc_abort_alloc_entry(struct inode *inode,
640 				    struct nilfs_palloc_req *req)
641 {
642 	struct nilfs_palloc_group_desc *desc;
643 	void *desc_kaddr, *bitmap_kaddr;
644 	unsigned char *bitmap;
645 	unsigned long group, group_offset;
646 	spinlock_t *lock;
647 
648 	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
649 	desc_kaddr = kmap(req->pr_desc_bh->b_page);
650 	desc = nilfs_palloc_block_get_group_desc(inode, group,
651 						 req->pr_desc_bh, desc_kaddr);
652 	bitmap_kaddr = kmap(req->pr_bitmap_bh->b_page);
653 	bitmap = bitmap_kaddr + bh_offset(req->pr_bitmap_bh);
654 	lock = nilfs_mdt_bgl_lock(inode, group);
655 
656 	if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
657 		nilfs_warn(inode->i_sb,
658 			   "%s (ino=%lu): entry number %llu already freed",
659 			   __func__, inode->i_ino,
660 			   (unsigned long long)req->pr_entry_nr);
661 	else
662 		nilfs_palloc_group_desc_add_entries(desc, lock, 1);
663 
664 	kunmap(req->pr_bitmap_bh->b_page);
665 	kunmap(req->pr_desc_bh->b_page);
666 
667 	brelse(req->pr_bitmap_bh);
668 	brelse(req->pr_desc_bh);
669 
670 	req->pr_entry_nr = 0;
671 	req->pr_bitmap_bh = NULL;
672 	req->pr_desc_bh = NULL;
673 }
674 
675 /**
676  * nilfs_palloc_prepare_free_entry - prepare to deallocate a persistent object
677  * @inode: inode of metadata file using this allocator
678  * @req: nilfs_palloc_req structure exchanged for the removal
679  */
680 int nilfs_palloc_prepare_free_entry(struct inode *inode,
681 				    struct nilfs_palloc_req *req)
682 {
683 	struct buffer_head *desc_bh, *bitmap_bh;
684 	unsigned long group, group_offset;
685 	int ret;
686 
687 	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
688 	ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
689 	if (ret < 0)
690 		return ret;
691 	ret = nilfs_palloc_get_bitmap_block(inode, group, 1, &bitmap_bh);
692 	if (ret < 0) {
693 		brelse(desc_bh);
694 		return ret;
695 	}
696 
697 	req->pr_desc_bh = desc_bh;
698 	req->pr_bitmap_bh = bitmap_bh;
699 	return 0;
700 }
701 
702 /**
703  * nilfs_palloc_abort_free_entry - cancel deallocating a persistent object
704  * @inode: inode of metadata file using this allocator
705  * @req: nilfs_palloc_req structure exchanged for the removal
706  */
707 void nilfs_palloc_abort_free_entry(struct inode *inode,
708 				   struct nilfs_palloc_req *req)
709 {
710 	brelse(req->pr_bitmap_bh);
711 	brelse(req->pr_desc_bh);
712 
713 	req->pr_entry_nr = 0;
714 	req->pr_bitmap_bh = NULL;
715 	req->pr_desc_bh = NULL;
716 }
717 
718 /**
719  * nilfs_palloc_freev - deallocate a set of persistent objects
720  * @inode: inode of metadata file using this allocator
721  * @entry_nrs: array of entry numbers to be deallocated
722  * @nitems: number of entries stored in @entry_nrs
723  */
724 int nilfs_palloc_freev(struct inode *inode, __u64 *entry_nrs, size_t nitems)
725 {
726 	struct buffer_head *desc_bh, *bitmap_bh;
727 	struct nilfs_palloc_group_desc *desc;
728 	unsigned char *bitmap;
729 	void *desc_kaddr, *bitmap_kaddr;
730 	unsigned long group, group_offset;
731 	__u64 group_min_nr, last_nrs[8];
732 	const unsigned long epg = nilfs_palloc_entries_per_group(inode);
733 	const unsigned int epb = NILFS_MDT(inode)->mi_entries_per_block;
734 	unsigned int entry_start, end, pos;
735 	spinlock_t *lock;
736 	int i, j, k, ret;
737 	u32 nfree;
738 
739 	for (i = 0; i < nitems; i = j) {
740 		int change_group = false;
741 		int nempties = 0, n = 0;
742 
743 		group = nilfs_palloc_group(inode, entry_nrs[i], &group_offset);
744 		ret = nilfs_palloc_get_desc_block(inode, group, 0, &desc_bh);
745 		if (ret < 0)
746 			return ret;
747 		ret = nilfs_palloc_get_bitmap_block(inode, group, 0,
748 						    &bitmap_bh);
749 		if (ret < 0) {
750 			brelse(desc_bh);
751 			return ret;
752 		}
753 
754 		/* Get the first entry number of the group */
755 		group_min_nr = (__u64)group * epg;
756 
757 		bitmap_kaddr = kmap(bitmap_bh->b_page);
758 		bitmap = bitmap_kaddr + bh_offset(bitmap_bh);
759 		lock = nilfs_mdt_bgl_lock(inode, group);
760 
761 		j = i;
762 		entry_start = rounddown(group_offset, epb);
763 		do {
764 			if (!nilfs_clear_bit_atomic(lock, group_offset,
765 						    bitmap)) {
766 				nilfs_warn(inode->i_sb,
767 					   "%s (ino=%lu): entry number %llu already freed",
768 					   __func__, inode->i_ino,
769 					   (unsigned long long)entry_nrs[j]);
770 			} else {
771 				n++;
772 			}
773 
774 			j++;
775 			if (j >= nitems || entry_nrs[j] < group_min_nr ||
776 			    entry_nrs[j] >= group_min_nr + epg) {
777 				change_group = true;
778 			} else {
779 				group_offset = entry_nrs[j] - group_min_nr;
780 				if (group_offset >= entry_start &&
781 				    group_offset < entry_start + epb) {
782 					/* This entry is in the same block */
783 					continue;
784 				}
785 			}
786 
787 			/* Test if the entry block is empty or not */
788 			end = entry_start + epb;
789 			pos = nilfs_find_next_bit(bitmap, end, entry_start);
790 			if (pos >= end) {
791 				last_nrs[nempties++] = entry_nrs[j - 1];
792 				if (nempties >= ARRAY_SIZE(last_nrs))
793 					break;
794 			}
795 
796 			if (change_group)
797 				break;
798 
799 			/* Go on to the next entry block */
800 			entry_start = rounddown(group_offset, epb);
801 		} while (true);
802 
803 		kunmap(bitmap_bh->b_page);
804 		mark_buffer_dirty(bitmap_bh);
805 		brelse(bitmap_bh);
806 
807 		for (k = 0; k < nempties; k++) {
808 			ret = nilfs_palloc_delete_entry_block(inode,
809 							      last_nrs[k]);
810 			if (ret && ret != -ENOENT)
811 				nilfs_warn(inode->i_sb,
812 					   "error %d deleting block that object (entry=%llu, ino=%lu) belongs to",
813 					   ret, (unsigned long long)last_nrs[k],
814 					   inode->i_ino);
815 		}
816 
817 		desc_kaddr = kmap_atomic(desc_bh->b_page);
818 		desc = nilfs_palloc_block_get_group_desc(
819 			inode, group, desc_bh, desc_kaddr);
820 		nfree = nilfs_palloc_group_desc_add_entries(desc, lock, n);
821 		kunmap_atomic(desc_kaddr);
822 		mark_buffer_dirty(desc_bh);
823 		nilfs_mdt_mark_dirty(inode);
824 		brelse(desc_bh);
825 
826 		if (nfree == nilfs_palloc_entries_per_group(inode)) {
827 			ret = nilfs_palloc_delete_bitmap_block(inode, group);
828 			if (ret && ret != -ENOENT)
829 				nilfs_warn(inode->i_sb,
830 					   "error %d deleting bitmap block of group=%lu, ino=%lu",
831 					   ret, group, inode->i_ino);
832 		}
833 	}
834 	return 0;
835 }
836 
837 void nilfs_palloc_setup_cache(struct inode *inode,
838 			      struct nilfs_palloc_cache *cache)
839 {
840 	NILFS_MDT(inode)->mi_palloc_cache = cache;
841 	spin_lock_init(&cache->lock);
842 }
843 
844 void nilfs_palloc_clear_cache(struct inode *inode)
845 {
846 	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
847 
848 	spin_lock(&cache->lock);
849 	brelse(cache->prev_desc.bh);
850 	brelse(cache->prev_bitmap.bh);
851 	brelse(cache->prev_entry.bh);
852 	cache->prev_desc.bh = NULL;
853 	cache->prev_bitmap.bh = NULL;
854 	cache->prev_entry.bh = NULL;
855 	spin_unlock(&cache->lock);
856 }
857 
858 void nilfs_palloc_destroy_cache(struct inode *inode)
859 {
860 	nilfs_palloc_clear_cache(inode);
861 	NILFS_MDT(inode)->mi_palloc_cache = NULL;
862 }
863