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_local_page(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++, group_offset = 0) {
534			lock = nilfs_mdt_bgl_lock(inode, group);
535			if (nilfs_palloc_group_desc_nfrees(desc, lock) == 0)
536				continue;
537
538			kunmap_local(desc_kaddr);
539			ret = nilfs_palloc_get_bitmap_block(inode, group, 1,
540							    &bitmap_bh);
541			if (unlikely(ret < 0)) {
542				brelse(desc_bh);
543				return ret;
544			}
545
546			desc_kaddr = kmap_local_page(desc_bh->b_page);
547			desc = nilfs_palloc_block_get_group_desc(
548				inode, group, desc_bh, desc_kaddr);
549
550			bitmap_kaddr = kmap_local_page(bitmap_bh->b_page);
551			bitmap = bitmap_kaddr + bh_offset(bitmap_bh);
552			pos = nilfs_palloc_find_available_slot(
553				bitmap, group_offset, entries_per_group, lock);
554			kunmap_local(bitmap_kaddr);
555			if (pos >= 0)
556				goto found;
557
558			brelse(bitmap_bh);
559		}
560
561		kunmap_local(desc_kaddr);
562		brelse(desc_bh);
563	}
564
565	/* no entries left */
566	return -ENOSPC;
567
568found:
569	/* found a free entry */
570	nilfs_palloc_group_desc_add_entries(desc, lock, -1);
571	req->pr_entry_nr = entries_per_group * group + pos;
572	kunmap_local(desc_kaddr);
573
574	req->pr_desc_bh = desc_bh;
575	req->pr_bitmap_bh = bitmap_bh;
576	return 0;
577}
578
579/**
580 * nilfs_palloc_commit_alloc_entry - finish allocation of a persistent object
581 * @inode: inode of metadata file using this allocator
582 * @req: nilfs_palloc_req structure exchanged for the allocation
583 */
584void nilfs_palloc_commit_alloc_entry(struct inode *inode,
585				     struct nilfs_palloc_req *req)
586{
587	mark_buffer_dirty(req->pr_bitmap_bh);
588	mark_buffer_dirty(req->pr_desc_bh);
589	nilfs_mdt_mark_dirty(inode);
590
591	brelse(req->pr_bitmap_bh);
592	brelse(req->pr_desc_bh);
593}
594
595/**
596 * nilfs_palloc_commit_free_entry - finish deallocating a persistent object
597 * @inode: inode of metadata file using this allocator
598 * @req: nilfs_palloc_req structure exchanged for the removal
599 */
600void nilfs_palloc_commit_free_entry(struct inode *inode,
601				    struct nilfs_palloc_req *req)
602{
603	struct nilfs_palloc_group_desc *desc;
604	unsigned long group, group_offset;
605	unsigned char *bitmap;
606	void *desc_kaddr, *bitmap_kaddr;
607	spinlock_t *lock;
608
609	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
610	desc_kaddr = kmap_local_page(req->pr_desc_bh->b_page);
611	desc = nilfs_palloc_block_get_group_desc(inode, group,
612						 req->pr_desc_bh, desc_kaddr);
613	bitmap_kaddr = kmap_local_page(req->pr_bitmap_bh->b_page);
614	bitmap = bitmap_kaddr + bh_offset(req->pr_bitmap_bh);
615	lock = nilfs_mdt_bgl_lock(inode, group);
616
617	if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
618		nilfs_warn(inode->i_sb,
619			   "%s (ino=%lu): entry number %llu already freed",
620			   __func__, inode->i_ino,
621			   (unsigned long long)req->pr_entry_nr);
622	else
623		nilfs_palloc_group_desc_add_entries(desc, lock, 1);
624
625	kunmap_local(bitmap_kaddr);
626	kunmap_local(desc_kaddr);
627
628	mark_buffer_dirty(req->pr_desc_bh);
629	mark_buffer_dirty(req->pr_bitmap_bh);
630	nilfs_mdt_mark_dirty(inode);
631
632	brelse(req->pr_bitmap_bh);
633	brelse(req->pr_desc_bh);
634}
635
636/**
637 * nilfs_palloc_abort_alloc_entry - cancel allocation of a persistent object
638 * @inode: inode of metadata file using this allocator
639 * @req: nilfs_palloc_req structure exchanged for the allocation
640 */
641void nilfs_palloc_abort_alloc_entry(struct inode *inode,
642				    struct nilfs_palloc_req *req)
643{
644	struct nilfs_palloc_group_desc *desc;
645	void *desc_kaddr, *bitmap_kaddr;
646	unsigned char *bitmap;
647	unsigned long group, group_offset;
648	spinlock_t *lock;
649
650	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
651	desc_kaddr = kmap_local_page(req->pr_desc_bh->b_page);
652	desc = nilfs_palloc_block_get_group_desc(inode, group,
653						 req->pr_desc_bh, desc_kaddr);
654	bitmap_kaddr = kmap_local_page(req->pr_bitmap_bh->b_page);
655	bitmap = bitmap_kaddr + bh_offset(req->pr_bitmap_bh);
656	lock = nilfs_mdt_bgl_lock(inode, group);
657
658	if (!nilfs_clear_bit_atomic(lock, group_offset, bitmap))
659		nilfs_warn(inode->i_sb,
660			   "%s (ino=%lu): entry number %llu already freed",
661			   __func__, inode->i_ino,
662			   (unsigned long long)req->pr_entry_nr);
663	else
664		nilfs_palloc_group_desc_add_entries(desc, lock, 1);
665
666	kunmap_local(bitmap_kaddr);
667	kunmap_local(desc_kaddr);
668
669	brelse(req->pr_bitmap_bh);
670	brelse(req->pr_desc_bh);
671
672	req->pr_entry_nr = 0;
673	req->pr_bitmap_bh = NULL;
674	req->pr_desc_bh = NULL;
675}
676
677/**
678 * nilfs_palloc_prepare_free_entry - prepare to deallocate a persistent object
679 * @inode: inode of metadata file using this allocator
680 * @req: nilfs_palloc_req structure exchanged for the removal
681 */
682int nilfs_palloc_prepare_free_entry(struct inode *inode,
683				    struct nilfs_palloc_req *req)
684{
685	struct buffer_head *desc_bh, *bitmap_bh;
686	unsigned long group, group_offset;
687	int ret;
688
689	group = nilfs_palloc_group(inode, req->pr_entry_nr, &group_offset);
690	ret = nilfs_palloc_get_desc_block(inode, group, 1, &desc_bh);
691	if (ret < 0)
692		return ret;
693	ret = nilfs_palloc_get_bitmap_block(inode, group, 1, &bitmap_bh);
694	if (ret < 0) {
695		brelse(desc_bh);
696		return ret;
697	}
698
699	req->pr_desc_bh = desc_bh;
700	req->pr_bitmap_bh = bitmap_bh;
701	return 0;
702}
703
704/**
705 * nilfs_palloc_abort_free_entry - cancel deallocating a persistent object
706 * @inode: inode of metadata file using this allocator
707 * @req: nilfs_palloc_req structure exchanged for the removal
708 */
709void nilfs_palloc_abort_free_entry(struct inode *inode,
710				   struct nilfs_palloc_req *req)
711{
712	brelse(req->pr_bitmap_bh);
713	brelse(req->pr_desc_bh);
714
715	req->pr_entry_nr = 0;
716	req->pr_bitmap_bh = NULL;
717	req->pr_desc_bh = NULL;
718}
719
720/**
721 * nilfs_palloc_freev - deallocate a set of persistent objects
722 * @inode: inode of metadata file using this allocator
723 * @entry_nrs: array of entry numbers to be deallocated
724 * @nitems: number of entries stored in @entry_nrs
725 */
726int nilfs_palloc_freev(struct inode *inode, __u64 *entry_nrs, size_t nitems)
727{
728	struct buffer_head *desc_bh, *bitmap_bh;
729	struct nilfs_palloc_group_desc *desc;
730	unsigned char *bitmap;
731	void *desc_kaddr, *bitmap_kaddr;
732	unsigned long group, group_offset;
733	__u64 group_min_nr, last_nrs[8];
734	const unsigned long epg = nilfs_palloc_entries_per_group(inode);
735	const unsigned int epb = NILFS_MDT(inode)->mi_entries_per_block;
736	unsigned int entry_start, end, pos;
737	spinlock_t *lock;
738	int i, j, k, ret;
739	u32 nfree;
740
741	for (i = 0; i < nitems; i = j) {
742		int change_group = false;
743		int nempties = 0, n = 0;
744
745		group = nilfs_palloc_group(inode, entry_nrs[i], &group_offset);
746		ret = nilfs_palloc_get_desc_block(inode, group, 0, &desc_bh);
747		if (ret < 0)
748			return ret;
749		ret = nilfs_palloc_get_bitmap_block(inode, group, 0,
750						    &bitmap_bh);
751		if (ret < 0) {
752			brelse(desc_bh);
753			return ret;
754		}
755
756		/* Get the first entry number of the group */
757		group_min_nr = (__u64)group * epg;
758
759		bitmap_kaddr = kmap_local_page(bitmap_bh->b_page);
760		bitmap = bitmap_kaddr + bh_offset(bitmap_bh);
761		lock = nilfs_mdt_bgl_lock(inode, group);
762
763		j = i;
764		entry_start = rounddown(group_offset, epb);
765		do {
766			if (!nilfs_clear_bit_atomic(lock, group_offset,
767						    bitmap)) {
768				nilfs_warn(inode->i_sb,
769					   "%s (ino=%lu): entry number %llu already freed",
770					   __func__, inode->i_ino,
771					   (unsigned long long)entry_nrs[j]);
772			} else {
773				n++;
774			}
775
776			j++;
777			if (j >= nitems || entry_nrs[j] < group_min_nr ||
778			    entry_nrs[j] >= group_min_nr + epg) {
779				change_group = true;
780			} else {
781				group_offset = entry_nrs[j] - group_min_nr;
782				if (group_offset >= entry_start &&
783				    group_offset < entry_start + epb) {
784					/* This entry is in the same block */
785					continue;
786				}
787			}
788
789			/* Test if the entry block is empty or not */
790			end = entry_start + epb;
791			pos = nilfs_find_next_bit(bitmap, end, entry_start);
792			if (pos >= end) {
793				last_nrs[nempties++] = entry_nrs[j - 1];
794				if (nempties >= ARRAY_SIZE(last_nrs))
795					break;
796			}
797
798			if (change_group)
799				break;
800
801			/* Go on to the next entry block */
802			entry_start = rounddown(group_offset, epb);
803		} while (true);
804
805		kunmap_local(bitmap_kaddr);
806		mark_buffer_dirty(bitmap_bh);
807		brelse(bitmap_bh);
808
809		for (k = 0; k < nempties; k++) {
810			ret = nilfs_palloc_delete_entry_block(inode,
811							      last_nrs[k]);
812			if (ret && ret != -ENOENT)
813				nilfs_warn(inode->i_sb,
814					   "error %d deleting block that object (entry=%llu, ino=%lu) belongs to",
815					   ret, (unsigned long long)last_nrs[k],
816					   inode->i_ino);
817		}
818
819		desc_kaddr = kmap_local_page(desc_bh->b_page);
820		desc = nilfs_palloc_block_get_group_desc(
821			inode, group, desc_bh, desc_kaddr);
822		nfree = nilfs_palloc_group_desc_add_entries(desc, lock, n);
823		kunmap_local(desc_kaddr);
824		mark_buffer_dirty(desc_bh);
825		nilfs_mdt_mark_dirty(inode);
826		brelse(desc_bh);
827
828		if (nfree == nilfs_palloc_entries_per_group(inode)) {
829			ret = nilfs_palloc_delete_bitmap_block(inode, group);
830			if (ret && ret != -ENOENT)
831				nilfs_warn(inode->i_sb,
832					   "error %d deleting bitmap block of group=%lu, ino=%lu",
833					   ret, group, inode->i_ino);
834		}
835	}
836	return 0;
837}
838
839void nilfs_palloc_setup_cache(struct inode *inode,
840			      struct nilfs_palloc_cache *cache)
841{
842	NILFS_MDT(inode)->mi_palloc_cache = cache;
843	spin_lock_init(&cache->lock);
844}
845
846void nilfs_palloc_clear_cache(struct inode *inode)
847{
848	struct nilfs_palloc_cache *cache = NILFS_MDT(inode)->mi_palloc_cache;
849
850	spin_lock(&cache->lock);
851	brelse(cache->prev_desc.bh);
852	brelse(cache->prev_bitmap.bh);
853	brelse(cache->prev_entry.bh);
854	cache->prev_desc.bh = NULL;
855	cache->prev_bitmap.bh = NULL;
856	cache->prev_entry.bh = NULL;
857	spin_unlock(&cache->lock);
858}
859
860void nilfs_palloc_destroy_cache(struct inode *inode)
861{
862	nilfs_palloc_clear_cache(inode);
863	NILFS_MDT(inode)->mi_palloc_cache = NULL;
864}
865