2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3 * Written by Alex Tomas <alex@clusterfs.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public Licens
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-
21 * mballoc.c contains the multiblocks allocation routines
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <trace/events/ext4.h>
31 * - test ext4_ext_search_left() and ext4_ext_search_right()
32 * - search for metadata in few groups
35 * - normalization should take into account whether file is still open
36 * - discard preallocations if no free space left (policy?)
37 * - don't normalize tails
39 * - reservation for superuser
42 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
43 * - track min/max extents in each group for better group selection
44 * - mb_mark_used() may allocate chunk right after splitting buddy
45 * - tree of groups sorted by number of free blocks
50 * The allocation request involve request for multiple number of blocks
51 * near to the goal(block) value specified.
53 * During initialization phase of the allocator we decide to use the
54 * group preallocation or inode preallocation depending on the size of
55 * the file. The size of the file could be the resulting file size we
56 * would have after allocation, or the current file size, which ever
57 * is larger. If the size is less than sbi->s_mb_stream_request we
58 * select to use the group preallocation. The default value of
59 * s_mb_stream_request is 16 blocks. This can also be tuned via
60 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
61 * terms of number of blocks.
63 * The main motivation for having small file use group preallocation is to
64 * ensure that we have small files closer together on the disk.
66 * First stage the allocator looks at the inode prealloc list,
67 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
68 * spaces for this particular inode. The inode prealloc space is
71 * pa_lstart -> the logical start block for this prealloc space
72 * pa_pstart -> the physical start block for this prealloc space
73 * pa_len -> length for this prealloc space
74 * pa_free -> free space available in this prealloc space
76 * The inode preallocation space is used looking at the _logical_ start
77 * block. If only the logical file block falls within the range of prealloc
78 * space we will consume the particular prealloc space. This make sure that
79 * that the we have contiguous physical blocks representing the file blocks
81 * The important thing to be noted in case of inode prealloc space is that
82 * we don't modify the values associated to inode prealloc space except
85 * If we are not able to find blocks in the inode prealloc space and if we
86 * have the group allocation flag set then we look at the locality group
87 * prealloc space. These are per CPU prealloc list repreasented as
89 * ext4_sb_info.s_locality_groups[smp_processor_id()]
91 * The reason for having a per cpu locality group is to reduce the contention
92 * between CPUs. It is possible to get scheduled at this point.
94 * The locality group prealloc space is used looking at whether we have
95 * enough free space (pa_free) withing the prealloc space.
97 * If we can't allocate blocks via inode prealloc or/and locality group
98 * prealloc then we look at the buddy cache. The buddy cache is represented
99 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
100 * mapped to the buddy and bitmap information regarding different
101 * groups. The buddy information is attached to buddy cache inode so that
102 * we can access them through the page cache. The information regarding
103 * each group is loaded via ext4_mb_load_buddy. The information involve
104 * block bitmap and buddy information. The information are stored in the
108 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
111 * one block each for bitmap and buddy information. So for each group we
112 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
113 * blocksize) blocks. So it can have information regarding groups_per_page
114 * which is blocks_per_page/2
116 * The buddy cache inode is not stored on disk. The inode is thrown
117 * away when the filesystem is unmounted.
119 * We look for count number of blocks in the buddy cache. If we were able
120 * to locate that many free blocks we return with additional information
121 * regarding rest of the contiguous physical block available
123 * Before allocating blocks via buddy cache we normalize the request
124 * blocks. This ensure we ask for more blocks that we needed. The extra
125 * blocks that we get after allocation is added to the respective prealloc
126 * list. In case of inode preallocation we follow a list of heuristics
127 * based on file size. This can be found in ext4_mb_normalize_request. If
128 * we are doing a group prealloc we try to normalize the request to
129 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
130 * 512 blocks. This can be tuned via
131 * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
132 * terms of number of blocks. If we have mounted the file system with -O
133 * stripe=<value> option the group prealloc request is normalized to the
134 * stripe value (sbi->s_stripe)
136 * The regular allocator(using the buddy cache) supports few tunables.
138 * /sys/fs/ext4/<partition>/mb_min_to_scan
139 * /sys/fs/ext4/<partition>/mb_max_to_scan
140 * /sys/fs/ext4/<partition>/mb_order2_req
142 * The regular allocator uses buddy scan only if the request len is power of
143 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
144 * value of s_mb_order2_reqs can be tuned via
145 * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to
146 * stripe size (sbi->s_stripe), we try to search for contiguous block in
147 * stripe size. This should result in better allocation on RAID setups. If
148 * not, we search in the specific group using bitmap for best extents. The
149 * tunable min_to_scan and max_to_scan control the behaviour here.
150 * min_to_scan indicate how long the mballoc __must__ look for a best
151 * extent and max_to_scan indicates how long the mballoc __can__ look for a
152 * best extent in the found extents. Searching for the blocks starts with
153 * the group specified as the goal value in allocation context via
154 * ac_g_ex. Each group is first checked based on the criteria whether it
155 * can used for allocation. ext4_mb_good_group explains how the groups are
158 * Both the prealloc space are getting populated as above. So for the first
159 * request we will hit the buddy cache which will result in this prealloc
160 * space getting filled. The prealloc space is then later used for the
161 * subsequent request.
165 * mballoc operates on the following data:
167 * - in-core buddy (actually includes buddy and bitmap)
168 * - preallocation descriptors (PAs)
170 * there are two types of preallocations:
172 * assiged to specific inode and can be used for this inode only.
173 * it describes part of inode's space preallocated to specific
174 * physical blocks. any block from that preallocated can be used
175 * independent. the descriptor just tracks number of blocks left
176 * unused. so, before taking some block from descriptor, one must
177 * make sure corresponded logical block isn't allocated yet. this
178 * also means that freeing any block within descriptor's range
179 * must discard all preallocated blocks.
181 * assigned to specific locality group which does not translate to
182 * permanent set of inodes: inode can join and leave group. space
183 * from this type of preallocation can be used for any inode. thus
184 * it's consumed from the beginning to the end.
186 * relation between them can be expressed as:
187 * in-core buddy = on-disk bitmap + preallocation descriptors
189 * this mean blocks mballoc considers used are:
190 * - allocated blocks (persistent)
191 * - preallocated blocks (non-persistent)
193 * consistency in mballoc world means that at any time a block is either
194 * free or used in ALL structures. notice: "any time" should not be read
195 * literally -- time is discrete and delimited by locks.
197 * to keep it simple, we don't use block numbers, instead we count number of
198 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
200 * all operations can be expressed as:
201 * - init buddy: buddy = on-disk + PAs
202 * - new PA: buddy += N; PA = N
203 * - use inode PA: on-disk += N; PA -= N
204 * - discard inode PA buddy -= on-disk - PA; PA = 0
205 * - use locality group PA on-disk += N; PA -= N
206 * - discard locality group PA buddy -= PA; PA = 0
207 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
208 * is used in real operation because we can't know actual used
209 * bits from PA, only from on-disk bitmap
211 * if we follow this strict logic, then all operations above should be atomic.
212 * given some of them can block, we'd have to use something like semaphores
213 * killing performance on high-end SMP hardware. let's try to relax it using
214 * the following knowledge:
215 * 1) if buddy is referenced, it's already initialized
216 * 2) while block is used in buddy and the buddy is referenced,
217 * nobody can re-allocate that block
218 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
219 * bit set and PA claims same block, it's OK. IOW, one can set bit in
220 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
223 * so, now we're building a concurrency table:
226 * blocks for PA are allocated in the buddy, buddy must be referenced
227 * until PA is linked to allocation group to avoid concurrent buddy init
229 * we need to make sure that either on-disk bitmap or PA has uptodate data
230 * given (3) we care that PA-=N operation doesn't interfere with init
232 * the simplest way would be to have buddy initialized by the discard
233 * - use locality group PA
234 * again PA-=N must be serialized with init
235 * - discard locality group PA
236 * the simplest way would be to have buddy initialized by the discard
239 * i_data_sem serializes them
241 * discard process must wait until PA isn't used by another process
242 * - use locality group PA
243 * some mutex should serialize them
244 * - discard locality group PA
245 * discard process must wait until PA isn't used by another process
248 * i_data_sem or another mutex should serializes them
250 * discard process must wait until PA isn't used by another process
251 * - use locality group PA
252 * nothing wrong here -- they're different PAs covering different blocks
253 * - discard locality group PA
254 * discard process must wait until PA isn't used by another process
256 * now we're ready to make few consequences:
257 * - PA is referenced and while it is no discard is possible
258 * - PA is referenced until block isn't marked in on-disk bitmap
259 * - PA changes only after on-disk bitmap
260 * - discard must not compete with init. either init is done before
261 * any discard or they're serialized somehow
262 * - buddy init as sum of on-disk bitmap and PAs is done atomically
264 * a special case when we've used PA to emptiness. no need to modify buddy
265 * in this case, but we should care about concurrent init
270 * Logic in few words:
275 * mark bits in on-disk bitmap
278 * - use preallocation:
279 * find proper PA (per-inode or group)
281 * mark bits in on-disk bitmap
287 * mark bits in on-disk bitmap
290 * - discard preallocations in group:
292 * move them onto local list
293 * load on-disk bitmap
295 * remove PA from object (inode or locality group)
296 * mark free blocks in-core
298 * - discard inode's preallocations:
305 * - bitlock on a group (group)
306 * - object (inode/locality) (object)
317 * - release consumed pa:
322 * - generate in-core bitmap:
326 * - discard all for given object (inode, locality group):
331 * - discard all for given group:
338 static struct kmem_cache *ext4_pspace_cachep;
339 static struct kmem_cache *ext4_ac_cachep;
340 static struct kmem_cache *ext4_free_ext_cachep;
341 static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
343 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
345 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn);
347 static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
349 #if BITS_PER_LONG == 64
350 *bit += ((unsigned long) addr & 7UL) << 3;
351 addr = (void *) ((unsigned long) addr & ~7UL);
352 #elif BITS_PER_LONG == 32
353 *bit += ((unsigned long) addr & 3UL) << 3;
354 addr = (void *) ((unsigned long) addr & ~3UL);
356 #error "how many bits you are?!"
361 static inline int mb_test_bit(int bit, void *addr)
364 * ext4_test_bit on architecture like powerpc
365 * needs unsigned long aligned address
367 addr = mb_correct_addr_and_bit(&bit, addr);
368 return ext4_test_bit(bit, addr);
371 static inline void mb_set_bit(int bit, void *addr)
373 addr = mb_correct_addr_and_bit(&bit, addr);
374 ext4_set_bit(bit, addr);
377 static inline void mb_clear_bit(int bit, void *addr)
379 addr = mb_correct_addr_and_bit(&bit, addr);
380 ext4_clear_bit(bit, addr);
383 static inline int mb_find_next_zero_bit(void *addr, int max, int start)
385 int fix = 0, ret, tmpmax;
386 addr = mb_correct_addr_and_bit(&fix, addr);
390 ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
396 static inline int mb_find_next_bit(void *addr, int max, int start)
398 int fix = 0, ret, tmpmax;
399 addr = mb_correct_addr_and_bit(&fix, addr);
403 ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
409 static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
413 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
416 if (order > e4b->bd_blkbits + 1) {
421 /* at order 0 we see each particular block */
422 *max = 1 << (e4b->bd_blkbits + 3);
424 return EXT4_MB_BITMAP(e4b);
426 bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
427 *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
433 static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
434 int first, int count)
437 struct super_block *sb = e4b->bd_sb;
439 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
441 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
442 for (i = 0; i < count; i++) {
443 if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
444 ext4_fsblk_t blocknr;
446 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
447 blocknr += first + i;
448 ext4_grp_locked_error(sb, e4b->bd_group,
449 __func__, "double-free of inode"
450 " %lu's block %llu(bit %u in group %u)",
451 inode ? inode->i_ino : 0, blocknr,
452 first + i, e4b->bd_group);
454 mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
458 static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
462 if (unlikely(e4b->bd_info->bb_bitmap == NULL))
464 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
465 for (i = 0; i < count; i++) {
466 BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
467 mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
471 static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
473 if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
474 unsigned char *b1, *b2;
476 b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
477 b2 = (unsigned char *) bitmap;
478 for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
479 if (b1[i] != b2[i]) {
480 printk(KERN_ERR "corruption in group %u "
481 "at byte %u(%u): %x in copy != %x "
482 "on disk/prealloc\n",
483 e4b->bd_group, i, i * 8, b1[i], b2[i]);
491 static inline void mb_free_blocks_double(struct inode *inode,
492 struct ext4_buddy *e4b, int first, int count)
496 static inline void mb_mark_used_double(struct ext4_buddy *e4b,
497 int first, int count)
501 static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
507 #ifdef AGGRESSIVE_CHECK
509 #define MB_CHECK_ASSERT(assert) \
513 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
514 function, file, line, # assert); \
519 static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
520 const char *function, int line)
522 struct super_block *sb = e4b->bd_sb;
523 int order = e4b->bd_blkbits + 1;
530 struct ext4_group_info *grp;
533 struct list_head *cur;
538 static int mb_check_counter;
539 if (mb_check_counter++ % 100 != 0)
544 buddy = mb_find_buddy(e4b, order, &max);
545 MB_CHECK_ASSERT(buddy);
546 buddy2 = mb_find_buddy(e4b, order - 1, &max2);
547 MB_CHECK_ASSERT(buddy2);
548 MB_CHECK_ASSERT(buddy != buddy2);
549 MB_CHECK_ASSERT(max * 2 == max2);
552 for (i = 0; i < max; i++) {
554 if (mb_test_bit(i, buddy)) {
555 /* only single bit in buddy2 may be 1 */
556 if (!mb_test_bit(i << 1, buddy2)) {
558 mb_test_bit((i<<1)+1, buddy2));
559 } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
561 mb_test_bit(i << 1, buddy2));
566 /* both bits in buddy2 must be 0 */
567 MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
568 MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
570 for (j = 0; j < (1 << order); j++) {
571 k = (i * (1 << order)) + j;
573 !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
577 MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
582 buddy = mb_find_buddy(e4b, 0, &max);
583 for (i = 0; i < max; i++) {
584 if (!mb_test_bit(i, buddy)) {
585 MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
593 /* check used bits only */
594 for (j = 0; j < e4b->bd_blkbits + 1; j++) {
595 buddy2 = mb_find_buddy(e4b, j, &max2);
597 MB_CHECK_ASSERT(k < max2);
598 MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
601 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
602 MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
604 grp = ext4_get_group_info(sb, e4b->bd_group);
605 buddy = mb_find_buddy(e4b, 0, &max);
606 list_for_each(cur, &grp->bb_prealloc_list) {
607 ext4_group_t groupnr;
608 struct ext4_prealloc_space *pa;
609 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
610 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
611 MB_CHECK_ASSERT(groupnr == e4b->bd_group);
612 for (i = 0; i < pa->pa_len; i++)
613 MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
617 #undef MB_CHECK_ASSERT
618 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
619 __FILE__, __func__, __LINE__)
621 #define mb_check_buddy(e4b)
624 /* FIXME!! need more doc */
625 static void ext4_mb_mark_free_simple(struct super_block *sb,
626 void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
627 struct ext4_group_info *grp)
629 struct ext4_sb_info *sbi = EXT4_SB(sb);
633 unsigned short border;
635 BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb));
637 border = 2 << sb->s_blocksize_bits;
640 /* find how many blocks can be covered since this position */
641 max = ffs(first | border) - 1;
643 /* find how many blocks of power 2 we need to mark */
650 /* mark multiblock chunks only */
651 grp->bb_counters[min]++;
653 mb_clear_bit(first >> min,
654 buddy + sbi->s_mb_offsets[min]);
661 static noinline_for_stack
662 void ext4_mb_generate_buddy(struct super_block *sb,
663 void *buddy, void *bitmap, ext4_group_t group)
665 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
666 ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb);
671 unsigned fragments = 0;
672 unsigned long long period = get_cycles();
674 /* initialize buddy from bitmap which is aggregation
675 * of on-disk bitmap and preallocations */
676 i = mb_find_next_zero_bit(bitmap, max, 0);
677 grp->bb_first_free = i;
681 i = mb_find_next_bit(bitmap, max, i);
685 ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
687 grp->bb_counters[0]++;
689 i = mb_find_next_zero_bit(bitmap, max, i);
691 grp->bb_fragments = fragments;
693 if (free != grp->bb_free) {
694 ext4_grp_locked_error(sb, group, __func__,
695 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
696 group, free, grp->bb_free);
698 * If we intent to continue, we consider group descritor
699 * corrupt and update bb_free using bitmap value
704 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
706 period = get_cycles() - period;
707 spin_lock(&EXT4_SB(sb)->s_bal_lock);
708 EXT4_SB(sb)->s_mb_buddies_generated++;
709 EXT4_SB(sb)->s_mb_generation_time += period;
710 spin_unlock(&EXT4_SB(sb)->s_bal_lock);
713 /* The buddy information is attached the buddy cache inode
714 * for convenience. The information regarding each group
715 * is loaded via ext4_mb_load_buddy. The information involve
716 * block bitmap and buddy information. The information are
717 * stored in the inode as
720 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
723 * one block each for bitmap and buddy information.
724 * So for each group we take up 2 blocks. A page can
725 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
726 * So it can have information regarding groups_per_page which
727 * is blocks_per_page/2
730 static int ext4_mb_init_cache(struct page *page, char *incore)
732 ext4_group_t ngroups;
738 ext4_group_t first_group;
740 struct super_block *sb;
741 struct buffer_head *bhs;
742 struct buffer_head **bh;
747 mb_debug(1, "init page %lu\n", page->index);
749 inode = page->mapping->host;
751 ngroups = ext4_get_groups_count(sb);
752 blocksize = 1 << inode->i_blkbits;
753 blocks_per_page = PAGE_CACHE_SIZE / blocksize;
755 groups_per_page = blocks_per_page >> 1;
756 if (groups_per_page == 0)
759 /* allocate buffer_heads to read bitmaps */
760 if (groups_per_page > 1) {
762 i = sizeof(struct buffer_head *) * groups_per_page;
763 bh = kzalloc(i, GFP_NOFS);
769 first_group = page->index * blocks_per_page / 2;
771 /* read all groups the page covers into the cache */
772 for (i = 0; i < groups_per_page; i++) {
773 struct ext4_group_desc *desc;
775 if (first_group + i >= ngroups)
779 desc = ext4_get_group_desc(sb, first_group + i, NULL);
784 bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
788 if (bitmap_uptodate(bh[i]))
792 if (bitmap_uptodate(bh[i])) {
793 unlock_buffer(bh[i]);
796 ext4_lock_group(sb, first_group + i);
797 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
798 ext4_init_block_bitmap(sb, bh[i],
799 first_group + i, desc);
800 set_bitmap_uptodate(bh[i]);
801 set_buffer_uptodate(bh[i]);
802 ext4_unlock_group(sb, first_group + i);
803 unlock_buffer(bh[i]);
806 ext4_unlock_group(sb, first_group + i);
807 if (buffer_uptodate(bh[i])) {
809 * if not uninit if bh is uptodate,
810 * bitmap is also uptodate
812 set_bitmap_uptodate(bh[i]);
813 unlock_buffer(bh[i]);
818 * submit the buffer_head for read. We can
819 * safely mark the bitmap as uptodate now.
820 * We do it here so the bitmap uptodate bit
821 * get set with buffer lock held.
823 set_bitmap_uptodate(bh[i]);
824 bh[i]->b_end_io = end_buffer_read_sync;
825 submit_bh(READ, bh[i]);
826 mb_debug(1, "read bitmap for group %u\n", first_group + i);
829 /* wait for I/O completion */
830 for (i = 0; i < groups_per_page && bh[i]; i++)
831 wait_on_buffer(bh[i]);
834 for (i = 0; i < groups_per_page && bh[i]; i++)
835 if (!buffer_uptodate(bh[i]))
839 first_block = page->index * blocks_per_page;
841 memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
842 for (i = 0; i < blocks_per_page; i++) {
844 struct ext4_group_info *grinfo;
846 group = (first_block + i) >> 1;
847 if (group >= ngroups)
851 * data carry information regarding this
852 * particular group in the format specified
856 data = page_address(page) + (i * blocksize);
857 bitmap = bh[group - first_group]->b_data;
860 * We place the buddy block and bitmap block
863 if ((first_block + i) & 1) {
864 /* this is block of buddy */
865 BUG_ON(incore == NULL);
866 mb_debug(1, "put buddy for group %u in page %lu/%x\n",
867 group, page->index, i * blocksize);
868 grinfo = ext4_get_group_info(sb, group);
869 grinfo->bb_fragments = 0;
870 memset(grinfo->bb_counters, 0,
871 sizeof(*grinfo->bb_counters) *
872 (sb->s_blocksize_bits+2));
874 * incore got set to the group block bitmap below
876 ext4_lock_group(sb, group);
877 ext4_mb_generate_buddy(sb, data, incore, group);
878 ext4_unlock_group(sb, group);
881 /* this is block of bitmap */
882 BUG_ON(incore != NULL);
883 mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
884 group, page->index, i * blocksize);
886 /* see comments in ext4_mb_put_pa() */
887 ext4_lock_group(sb, group);
888 memcpy(data, bitmap, blocksize);
890 /* mark all preallocated blks used in in-core bitmap */
891 ext4_mb_generate_from_pa(sb, data, group);
892 ext4_mb_generate_from_freelist(sb, data, group);
893 ext4_unlock_group(sb, group);
895 /* set incore so that the buddy information can be
896 * generated using this
901 SetPageUptodate(page);
905 for (i = 0; i < groups_per_page && bh[i]; i++)
913 static noinline_for_stack
914 int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
920 int block, pnum, poff;
921 int num_grp_locked = 0;
922 struct ext4_group_info *this_grp;
923 struct ext4_sb_info *sbi = EXT4_SB(sb);
924 struct inode *inode = sbi->s_buddy_cache;
925 struct page *page = NULL, *bitmap_page = NULL;
927 mb_debug(1, "init group %u\n", group);
928 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
929 this_grp = ext4_get_group_info(sb, group);
931 * This ensures that we don't reinit the buddy cache
932 * page which map to the group from which we are already
933 * allocating. If we are looking at the buddy cache we would
934 * have taken a reference using ext4_mb_load_buddy and that
935 * would have taken the alloc_sem lock.
937 num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group);
938 if (!EXT4_MB_GRP_NEED_INIT(this_grp)) {
940 * somebody initialized the group
941 * return without doing anything
947 * the buddy cache inode stores the block bitmap
948 * and buddy information in consecutive blocks.
949 * So for each group we need two blocks.
952 pnum = block / blocks_per_page;
953 poff = block % blocks_per_page;
954 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
956 BUG_ON(page->mapping != inode->i_mapping);
957 ret = ext4_mb_init_cache(page, NULL);
964 if (page == NULL || !PageUptodate(page)) {
968 mark_page_accessed(page);
970 bitmap = page_address(page) + (poff * sb->s_blocksize);
972 /* init buddy cache */
974 pnum = block / blocks_per_page;
975 poff = block % blocks_per_page;
976 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
977 if (page == bitmap_page) {
979 * If both the bitmap and buddy are in
980 * the same page we don't need to force
985 BUG_ON(page->mapping != inode->i_mapping);
986 ret = ext4_mb_init_cache(page, bitmap);
993 if (page == NULL || !PageUptodate(page)) {
997 mark_page_accessed(page);
999 ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked);
1001 page_cache_release(bitmap_page);
1003 page_cache_release(page);
1007 static noinline_for_stack int
1008 ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1009 struct ext4_buddy *e4b)
1011 int blocks_per_page;
1017 struct ext4_group_info *grp;
1018 struct ext4_sb_info *sbi = EXT4_SB(sb);
1019 struct inode *inode = sbi->s_buddy_cache;
1021 mb_debug(1, "load group %u\n", group);
1023 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1024 grp = ext4_get_group_info(sb, group);
1026 e4b->bd_blkbits = sb->s_blocksize_bits;
1027 e4b->bd_info = ext4_get_group_info(sb, group);
1029 e4b->bd_group = group;
1030 e4b->bd_buddy_page = NULL;
1031 e4b->bd_bitmap_page = NULL;
1032 e4b->alloc_semp = &grp->alloc_sem;
1034 /* Take the read lock on the group alloc
1035 * sem. This would make sure a parallel
1036 * ext4_mb_init_group happening on other
1037 * groups mapped by the page is blocked
1038 * till we are done with allocation
1041 down_read(e4b->alloc_semp);
1043 if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1044 /* we need to check for group need init flag
1045 * with alloc_semp held so that we can be sure
1046 * that new blocks didn't get added to the group
1047 * when we are loading the buddy cache
1049 up_read(e4b->alloc_semp);
1051 * we need full data about the group
1052 * to make a good selection
1054 ret = ext4_mb_init_group(sb, group);
1057 goto repeat_load_buddy;
1061 * the buddy cache inode stores the block bitmap
1062 * and buddy information in consecutive blocks.
1063 * So for each group we need two blocks.
1066 pnum = block / blocks_per_page;
1067 poff = block % blocks_per_page;
1069 /* we could use find_or_create_page(), but it locks page
1070 * what we'd like to avoid in fast path ... */
1071 page = find_get_page(inode->i_mapping, pnum);
1072 if (page == NULL || !PageUptodate(page)) {
1075 * drop the page reference and try
1076 * to get the page with lock. If we
1077 * are not uptodate that implies
1078 * somebody just created the page but
1079 * is yet to initialize the same. So
1080 * wait for it to initialize.
1082 page_cache_release(page);
1083 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1085 BUG_ON(page->mapping != inode->i_mapping);
1086 if (!PageUptodate(page)) {
1087 ret = ext4_mb_init_cache(page, NULL);
1092 mb_cmp_bitmaps(e4b, page_address(page) +
1093 (poff * sb->s_blocksize));
1098 if (page == NULL || !PageUptodate(page)) {
1102 e4b->bd_bitmap_page = page;
1103 e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1104 mark_page_accessed(page);
1107 pnum = block / blocks_per_page;
1108 poff = block % blocks_per_page;
1110 page = find_get_page(inode->i_mapping, pnum);
1111 if (page == NULL || !PageUptodate(page)) {
1113 page_cache_release(page);
1114 page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1116 BUG_ON(page->mapping != inode->i_mapping);
1117 if (!PageUptodate(page)) {
1118 ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1127 if (page == NULL || !PageUptodate(page)) {
1131 e4b->bd_buddy_page = page;
1132 e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1133 mark_page_accessed(page);
1135 BUG_ON(e4b->bd_bitmap_page == NULL);
1136 BUG_ON(e4b->bd_buddy_page == NULL);
1141 if (e4b->bd_bitmap_page)
1142 page_cache_release(e4b->bd_bitmap_page);
1143 if (e4b->bd_buddy_page)
1144 page_cache_release(e4b->bd_buddy_page);
1145 e4b->bd_buddy = NULL;
1146 e4b->bd_bitmap = NULL;
1148 /* Done with the buddy cache */
1149 up_read(e4b->alloc_semp);
1153 static void ext4_mb_release_desc(struct ext4_buddy *e4b)
1155 if (e4b->bd_bitmap_page)
1156 page_cache_release(e4b->bd_bitmap_page);
1157 if (e4b->bd_buddy_page)
1158 page_cache_release(e4b->bd_buddy_page);
1159 /* Done with the buddy cache */
1160 if (e4b->alloc_semp)
1161 up_read(e4b->alloc_semp);
1165 static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1170 BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
1171 BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1173 bb = EXT4_MB_BUDDY(e4b);
1174 while (order <= e4b->bd_blkbits + 1) {
1176 if (!mb_test_bit(block, bb)) {
1177 /* this block is part of buddy of order 'order' */
1180 bb += 1 << (e4b->bd_blkbits - order);
1186 static void mb_clear_bits(void *bm, int cur, int len)
1192 if ((cur & 31) == 0 && (len - cur) >= 32) {
1193 /* fast path: clear whole word at once */
1194 addr = bm + (cur >> 3);
1199 mb_clear_bit(cur, bm);
1204 static void mb_set_bits(void *bm, int cur, int len)
1210 if ((cur & 31) == 0 && (len - cur) >= 32) {
1211 /* fast path: set whole word at once */
1212 addr = bm + (cur >> 3);
1217 mb_set_bit(cur, bm);
1222 static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1223 int first, int count)
1230 struct super_block *sb = e4b->bd_sb;
1232 BUG_ON(first + count > (sb->s_blocksize << 3));
1233 assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1234 mb_check_buddy(e4b);
1235 mb_free_blocks_double(inode, e4b, first, count);
1237 e4b->bd_info->bb_free += count;
1238 if (first < e4b->bd_info->bb_first_free)
1239 e4b->bd_info->bb_first_free = first;
1241 /* let's maintain fragments counter */
1243 block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
1244 if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1245 max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
1247 e4b->bd_info->bb_fragments--;
1248 else if (!block && !max)
1249 e4b->bd_info->bb_fragments++;
1251 /* let's maintain buddy itself */
1252 while (count-- > 0) {
1256 if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
1257 ext4_fsblk_t blocknr;
1259 blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1261 ext4_grp_locked_error(sb, e4b->bd_group,
1262 __func__, "double-free of inode"
1263 " %lu's block %llu(bit %u in group %u)",
1264 inode ? inode->i_ino : 0, blocknr, block,
1267 mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
1268 e4b->bd_info->bb_counters[order]++;
1270 /* start of the buddy */
1271 buddy = mb_find_buddy(e4b, order, &max);
1275 if (mb_test_bit(block, buddy) ||
1276 mb_test_bit(block + 1, buddy))
1279 /* both the buddies are free, try to coalesce them */
1280 buddy2 = mb_find_buddy(e4b, order + 1, &max);
1286 /* for special purposes, we don't set
1287 * free bits in bitmap */
1288 mb_set_bit(block, buddy);
1289 mb_set_bit(block + 1, buddy);
1291 e4b->bd_info->bb_counters[order]--;
1292 e4b->bd_info->bb_counters[order]--;
1296 e4b->bd_info->bb_counters[order]++;
1298 mb_clear_bit(block, buddy2);
1302 mb_check_buddy(e4b);
1305 static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1306 int needed, struct ext4_free_extent *ex)
1313 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1316 buddy = mb_find_buddy(e4b, order, &max);
1317 BUG_ON(buddy == NULL);
1318 BUG_ON(block >= max);
1319 if (mb_test_bit(block, buddy)) {
1326 /* FIXME dorp order completely ? */
1327 if (likely(order == 0)) {
1328 /* find actual order */
1329 order = mb_find_order_for_block(e4b, block);
1330 block = block >> order;
1333 ex->fe_len = 1 << order;
1334 ex->fe_start = block << order;
1335 ex->fe_group = e4b->bd_group;
1337 /* calc difference from given start */
1338 next = next - ex->fe_start;
1340 ex->fe_start += next;
1342 while (needed > ex->fe_len &&
1343 (buddy = mb_find_buddy(e4b, order, &max))) {
1345 if (block + 1 >= max)
1348 next = (block + 1) * (1 << order);
1349 if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
1352 ord = mb_find_order_for_block(e4b, next);
1355 block = next >> order;
1356 ex->fe_len += 1 << order;
1359 BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1363 static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1369 int start = ex->fe_start;
1370 int len = ex->fe_len;
1375 BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1376 BUG_ON(e4b->bd_group != ex->fe_group);
1377 assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1378 mb_check_buddy(e4b);
1379 mb_mark_used_double(e4b, start, len);
1381 e4b->bd_info->bb_free -= len;
1382 if (e4b->bd_info->bb_first_free == start)
1383 e4b->bd_info->bb_first_free += len;
1385 /* let's maintain fragments counter */
1387 mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
1388 if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1389 max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
1391 e4b->bd_info->bb_fragments++;
1392 else if (!mlen && !max)
1393 e4b->bd_info->bb_fragments--;
1395 /* let's maintain buddy itself */
1397 ord = mb_find_order_for_block(e4b, start);
1399 if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1400 /* the whole chunk may be allocated at once! */
1402 buddy = mb_find_buddy(e4b, ord, &max);
1403 BUG_ON((start >> ord) >= max);
1404 mb_set_bit(start >> ord, buddy);
1405 e4b->bd_info->bb_counters[ord]--;
1412 /* store for history */
1414 ret = len | (ord << 16);
1416 /* we have to split large buddy */
1418 buddy = mb_find_buddy(e4b, ord, &max);
1419 mb_set_bit(start >> ord, buddy);
1420 e4b->bd_info->bb_counters[ord]--;
1423 cur = (start >> ord) & ~1U;
1424 buddy = mb_find_buddy(e4b, ord, &max);
1425 mb_clear_bit(cur, buddy);
1426 mb_clear_bit(cur + 1, buddy);
1427 e4b->bd_info->bb_counters[ord]++;
1428 e4b->bd_info->bb_counters[ord]++;
1431 mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
1432 mb_check_buddy(e4b);
1438 * Must be called under group lock!
1440 static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1441 struct ext4_buddy *e4b)
1443 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1446 BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1447 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1449 ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1450 ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1451 ret = mb_mark_used(e4b, &ac->ac_b_ex);
1453 /* preallocation can change ac_b_ex, thus we store actually
1454 * allocated blocks for history */
1455 ac->ac_f_ex = ac->ac_b_ex;
1457 ac->ac_status = AC_STATUS_FOUND;
1458 ac->ac_tail = ret & 0xffff;
1459 ac->ac_buddy = ret >> 16;
1462 * take the page reference. We want the page to be pinned
1463 * so that we don't get a ext4_mb_init_cache_call for this
1464 * group until we update the bitmap. That would mean we
1465 * double allocate blocks. The reference is dropped
1466 * in ext4_mb_release_context
1468 ac->ac_bitmap_page = e4b->bd_bitmap_page;
1469 get_page(ac->ac_bitmap_page);
1470 ac->ac_buddy_page = e4b->bd_buddy_page;
1471 get_page(ac->ac_buddy_page);
1472 /* on allocation we use ac to track the held semaphore */
1473 ac->alloc_semp = e4b->alloc_semp;
1474 e4b->alloc_semp = NULL;
1475 /* store last allocated for subsequent stream allocation */
1476 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1477 spin_lock(&sbi->s_md_lock);
1478 sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1479 sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1480 spin_unlock(&sbi->s_md_lock);
1485 * regular allocator, for general purposes allocation
1488 static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1489 struct ext4_buddy *e4b,
1492 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1493 struct ext4_free_extent *bex = &ac->ac_b_ex;
1494 struct ext4_free_extent *gex = &ac->ac_g_ex;
1495 struct ext4_free_extent ex;
1498 if (ac->ac_status == AC_STATUS_FOUND)
1501 * We don't want to scan for a whole year
1503 if (ac->ac_found > sbi->s_mb_max_to_scan &&
1504 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1505 ac->ac_status = AC_STATUS_BREAK;
1510 * Haven't found good chunk so far, let's continue
1512 if (bex->fe_len < gex->fe_len)
1515 if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1516 && bex->fe_group == e4b->bd_group) {
1517 /* recheck chunk's availability - we don't know
1518 * when it was found (within this lock-unlock
1520 max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1521 if (max >= gex->fe_len) {
1522 ext4_mb_use_best_found(ac, e4b);
1529 * The routine checks whether found extent is good enough. If it is,
1530 * then the extent gets marked used and flag is set to the context
1531 * to stop scanning. Otherwise, the extent is compared with the
1532 * previous found extent and if new one is better, then it's stored
1533 * in the context. Later, the best found extent will be used, if
1534 * mballoc can't find good enough extent.
1536 * FIXME: real allocation policy is to be designed yet!
1538 static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1539 struct ext4_free_extent *ex,
1540 struct ext4_buddy *e4b)
1542 struct ext4_free_extent *bex = &ac->ac_b_ex;
1543 struct ext4_free_extent *gex = &ac->ac_g_ex;
1545 BUG_ON(ex->fe_len <= 0);
1546 BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1547 BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
1548 BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1553 * The special case - take what you catch first
1555 if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1557 ext4_mb_use_best_found(ac, e4b);
1562 * Let's check whether the chuck is good enough
1564 if (ex->fe_len == gex->fe_len) {
1566 ext4_mb_use_best_found(ac, e4b);
1571 * If this is first found extent, just store it in the context
1573 if (bex->fe_len == 0) {
1579 * If new found extent is better, store it in the context
1581 if (bex->fe_len < gex->fe_len) {
1582 /* if the request isn't satisfied, any found extent
1583 * larger than previous best one is better */
1584 if (ex->fe_len > bex->fe_len)
1586 } else if (ex->fe_len > gex->fe_len) {
1587 /* if the request is satisfied, then we try to find
1588 * an extent that still satisfy the request, but is
1589 * smaller than previous one */
1590 if (ex->fe_len < bex->fe_len)
1594 ext4_mb_check_limits(ac, e4b, 0);
1597 static noinline_for_stack
1598 int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1599 struct ext4_buddy *e4b)
1601 struct ext4_free_extent ex = ac->ac_b_ex;
1602 ext4_group_t group = ex.fe_group;
1606 BUG_ON(ex.fe_len <= 0);
1607 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1611 ext4_lock_group(ac->ac_sb, group);
1612 max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1616 ext4_mb_use_best_found(ac, e4b);
1619 ext4_unlock_group(ac->ac_sb, group);
1620 ext4_mb_release_desc(e4b);
1625 static noinline_for_stack
1626 int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1627 struct ext4_buddy *e4b)
1629 ext4_group_t group = ac->ac_g_ex.fe_group;
1632 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1633 struct ext4_free_extent ex;
1635 if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1638 err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1642 ext4_lock_group(ac->ac_sb, group);
1643 max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1644 ac->ac_g_ex.fe_len, &ex);
1646 if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1649 start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1651 /* use do_div to get remainder (would be 64-bit modulo) */
1652 if (do_div(start, sbi->s_stripe) == 0) {
1655 ext4_mb_use_best_found(ac, e4b);
1657 } else if (max >= ac->ac_g_ex.fe_len) {
1658 BUG_ON(ex.fe_len <= 0);
1659 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1660 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1663 ext4_mb_use_best_found(ac, e4b);
1664 } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1665 /* Sometimes, caller may want to merge even small
1666 * number of blocks to an existing extent */
1667 BUG_ON(ex.fe_len <= 0);
1668 BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1669 BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1672 ext4_mb_use_best_found(ac, e4b);
1674 ext4_unlock_group(ac->ac_sb, group);
1675 ext4_mb_release_desc(e4b);
1681 * The routine scans buddy structures (not bitmap!) from given order
1682 * to max order and tries to find big enough chunk to satisfy the req
1684 static noinline_for_stack
1685 void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1686 struct ext4_buddy *e4b)
1688 struct super_block *sb = ac->ac_sb;
1689 struct ext4_group_info *grp = e4b->bd_info;
1695 BUG_ON(ac->ac_2order <= 0);
1696 for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1697 if (grp->bb_counters[i] == 0)
1700 buddy = mb_find_buddy(e4b, i, &max);
1701 BUG_ON(buddy == NULL);
1703 k = mb_find_next_zero_bit(buddy, max, 0);
1708 ac->ac_b_ex.fe_len = 1 << i;
1709 ac->ac_b_ex.fe_start = k << i;
1710 ac->ac_b_ex.fe_group = e4b->bd_group;
1712 ext4_mb_use_best_found(ac, e4b);
1714 BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1716 if (EXT4_SB(sb)->s_mb_stats)
1717 atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1724 * The routine scans the group and measures all found extents.
1725 * In order to optimize scanning, caller must pass number of
1726 * free blocks in the group, so the routine can know upper limit.
1728 static noinline_for_stack
1729 void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1730 struct ext4_buddy *e4b)
1732 struct super_block *sb = ac->ac_sb;
1733 void *bitmap = EXT4_MB_BITMAP(e4b);
1734 struct ext4_free_extent ex;
1738 free = e4b->bd_info->bb_free;
1741 i = e4b->bd_info->bb_first_free;
1743 while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1744 i = mb_find_next_zero_bit(bitmap,
1745 EXT4_BLOCKS_PER_GROUP(sb), i);
1746 if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
1748 * IF we have corrupt bitmap, we won't find any
1749 * free blocks even though group info says we
1750 * we have free blocks
1752 ext4_grp_locked_error(sb, e4b->bd_group,
1753 __func__, "%d free blocks as per "
1754 "group info. But bitmap says 0",
1759 mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1760 BUG_ON(ex.fe_len <= 0);
1761 if (free < ex.fe_len) {
1762 ext4_grp_locked_error(sb, e4b->bd_group,
1763 __func__, "%d free blocks as per "
1764 "group info. But got %d blocks",
1767 * The number of free blocks differs. This mostly
1768 * indicate that the bitmap is corrupt. So exit
1769 * without claiming the space.
1774 ext4_mb_measure_extent(ac, &ex, e4b);
1780 ext4_mb_check_limits(ac, e4b, 1);
1784 * This is a special case for storages like raid5
1785 * we try to find stripe-aligned chunks for stripe-size requests
1786 * XXX should do so at least for multiples of stripe size as well
1788 static noinline_for_stack
1789 void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1790 struct ext4_buddy *e4b)
1792 struct super_block *sb = ac->ac_sb;
1793 struct ext4_sb_info *sbi = EXT4_SB(sb);
1794 void *bitmap = EXT4_MB_BITMAP(e4b);
1795 struct ext4_free_extent ex;
1796 ext4_fsblk_t first_group_block;
1801 BUG_ON(sbi->s_stripe == 0);
1803 /* find first stripe-aligned block in group */
1804 first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1806 a = first_group_block + sbi->s_stripe - 1;
1807 do_div(a, sbi->s_stripe);
1808 i = (a * sbi->s_stripe) - first_group_block;
1810 while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
1811 if (!mb_test_bit(i, bitmap)) {
1812 max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1813 if (max >= sbi->s_stripe) {
1816 ext4_mb_use_best_found(ac, e4b);
1824 static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1825 ext4_group_t group, int cr)
1827 unsigned free, fragments;
1829 int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1830 struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1832 BUG_ON(cr < 0 || cr >= 4);
1833 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
1835 free = grp->bb_free;
1836 fragments = grp->bb_fragments;
1844 BUG_ON(ac->ac_2order == 0);
1846 /* Avoid using the first bg of a flexgroup for data files */
1847 if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1848 (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1849 ((group % flex_size) == 0))
1852 bits = ac->ac_sb->s_blocksize_bits + 1;
1853 for (i = ac->ac_2order; i <= bits; i++)
1854 if (grp->bb_counters[i] > 0)
1858 if ((free / fragments) >= ac->ac_g_ex.fe_len)
1862 if (free >= ac->ac_g_ex.fe_len)
1875 * lock the group_info alloc_sem of all the groups
1876 * belonging to the same buddy cache page. This
1877 * make sure other parallel operation on the buddy
1878 * cache doesn't happen whild holding the buddy cache
1881 int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group)
1885 int blocks_per_page;
1886 int groups_per_page;
1887 ext4_group_t ngroups = ext4_get_groups_count(sb);
1888 ext4_group_t first_group;
1889 struct ext4_group_info *grp;
1891 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1893 * the buddy cache inode stores the block bitmap
1894 * and buddy information in consecutive blocks.
1895 * So for each group we need two blocks.
1898 pnum = block / blocks_per_page;
1899 first_group = pnum * blocks_per_page / 2;
1901 groups_per_page = blocks_per_page >> 1;
1902 if (groups_per_page == 0)
1903 groups_per_page = 1;
1904 /* read all groups the page covers into the cache */
1905 for (i = 0; i < groups_per_page; i++) {
1907 if ((first_group + i) >= ngroups)
1909 grp = ext4_get_group_info(sb, first_group + i);
1910 /* take all groups write allocation
1911 * semaphore. This make sure there is
1912 * no block allocation going on in any
1915 down_write_nested(&grp->alloc_sem, i);
1920 void ext4_mb_put_buddy_cache_lock(struct super_block *sb,
1921 ext4_group_t group, int locked_group)
1925 int blocks_per_page;
1926 ext4_group_t first_group;
1927 struct ext4_group_info *grp;
1929 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1931 * the buddy cache inode stores the block bitmap
1932 * and buddy information in consecutive blocks.
1933 * So for each group we need two blocks.
1936 pnum = block / blocks_per_page;
1937 first_group = pnum * blocks_per_page / 2;
1938 /* release locks on all the groups */
1939 for (i = 0; i < locked_group; i++) {
1941 grp = ext4_get_group_info(sb, first_group + i);
1942 /* take all groups write allocation
1943 * semaphore. This make sure there is
1944 * no block allocation going on in any
1947 up_write(&grp->alloc_sem);
1952 static noinline_for_stack int
1953 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1955 ext4_group_t ngroups, group, i;
1959 struct ext4_sb_info *sbi;
1960 struct super_block *sb;
1961 struct ext4_buddy e4b;
1965 ngroups = ext4_get_groups_count(sb);
1966 /* non-extent files are limited to low blocks/groups */
1967 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL))
1968 ngroups = sbi->s_blockfile_groups;
1970 BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1972 /* first, try the goal */
1973 err = ext4_mb_find_by_goal(ac, &e4b);
1974 if (err || ac->ac_status == AC_STATUS_FOUND)
1977 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1981 * ac->ac2_order is set only if the fe_len is a power of 2
1982 * if ac2_order is set we also set criteria to 0 so that we
1983 * try exact allocation using buddy.
1985 i = fls(ac->ac_g_ex.fe_len);
1988 * We search using buddy data only if the order of the request
1989 * is greater than equal to the sbi_s_mb_order2_reqs
1990 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1992 if (i >= sbi->s_mb_order2_reqs) {
1994 * This should tell if fe_len is exactly power of 2
1996 if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1997 ac->ac_2order = i - 1;
2000 bsbits = ac->ac_sb->s_blocksize_bits;
2002 /* if stream allocation is enabled, use global goal */
2003 if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2004 /* TBD: may be hot point */
2005 spin_lock(&sbi->s_md_lock);
2006 ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2007 ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2008 spin_unlock(&sbi->s_md_lock);
2011 /* Let's just scan groups to find more-less suitable blocks */
2012 cr = ac->ac_2order ? 0 : 1;
2014 * cr == 0 try to get exact allocation,
2015 * cr == 3 try to get anything
2018 for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2019 ac->ac_criteria = cr;
2021 * searching for the right group start
2022 * from the goal value specified
2024 group = ac->ac_g_ex.fe_group;
2026 for (i = 0; i < ngroups; group++, i++) {
2027 struct ext4_group_info *grp;
2029 if (group == ngroups)
2032 /* quick check to skip empty groups */
2033 grp = ext4_get_group_info(sb, group);
2034 if (grp->bb_free == 0)
2037 err = ext4_mb_load_buddy(sb, group, &e4b);
2041 ext4_lock_group(sb, group);
2042 if (!ext4_mb_good_group(ac, group, cr)) {
2043 /* someone did allocation from this group */
2044 ext4_unlock_group(sb, group);
2045 ext4_mb_release_desc(&e4b);
2049 ac->ac_groups_scanned++;
2051 ext4_mb_simple_scan_group(ac, &e4b);
2053 ac->ac_g_ex.fe_len == sbi->s_stripe)
2054 ext4_mb_scan_aligned(ac, &e4b);
2056 ext4_mb_complex_scan_group(ac, &e4b);
2058 ext4_unlock_group(sb, group);
2059 ext4_mb_release_desc(&e4b);
2061 if (ac->ac_status != AC_STATUS_CONTINUE)
2066 if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2067 !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2069 * We've been searching too long. Let's try to allocate
2070 * the best chunk we've found so far
2073 ext4_mb_try_best_found(ac, &e4b);
2074 if (ac->ac_status != AC_STATUS_FOUND) {
2076 * Someone more lucky has already allocated it.
2077 * The only thing we can do is just take first
2079 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2081 ac->ac_b_ex.fe_group = 0;
2082 ac->ac_b_ex.fe_start = 0;
2083 ac->ac_b_ex.fe_len = 0;
2084 ac->ac_status = AC_STATUS_CONTINUE;
2085 ac->ac_flags |= EXT4_MB_HINT_FIRST;
2087 atomic_inc(&sbi->s_mb_lost_chunks);
2095 static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2097 struct super_block *sb = seq->private;
2100 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2103 return (void *) ((unsigned long) group);
2106 static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2108 struct super_block *sb = seq->private;
2112 if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2115 return (void *) ((unsigned long) group);
2118 static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2120 struct super_block *sb = seq->private;
2121 ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2124 struct ext4_buddy e4b;
2126 struct ext4_group_info info;
2127 ext4_grpblk_t counters[16];
2132 seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2133 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2134 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2135 "group", "free", "frags", "first",
2136 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2137 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2139 i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2140 sizeof(struct ext4_group_info);
2141 err = ext4_mb_load_buddy(sb, group, &e4b);
2143 seq_printf(seq, "#%-5u: I/O error\n", group);
2146 ext4_lock_group(sb, group);
2147 memcpy(&sg, ext4_get_group_info(sb, group), i);
2148 ext4_unlock_group(sb, group);
2149 ext4_mb_release_desc(&e4b);
2151 seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2152 sg.info.bb_fragments, sg.info.bb_first_free);
2153 for (i = 0; i <= 13; i++)
2154 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2155 sg.info.bb_counters[i] : 0);
2156 seq_printf(seq, " ]\n");
2161 static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2165 static const struct seq_operations ext4_mb_seq_groups_ops = {
2166 .start = ext4_mb_seq_groups_start,
2167 .next = ext4_mb_seq_groups_next,
2168 .stop = ext4_mb_seq_groups_stop,
2169 .show = ext4_mb_seq_groups_show,
2172 static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2174 struct super_block *sb = PDE(inode)->data;
2177 rc = seq_open(file, &ext4_mb_seq_groups_ops);
2179 struct seq_file *m = (struct seq_file *)file->private_data;
2186 static const struct file_operations ext4_mb_seq_groups_fops = {
2187 .owner = THIS_MODULE,
2188 .open = ext4_mb_seq_groups_open,
2190 .llseek = seq_lseek,
2191 .release = seq_release,
2195 /* Create and initialize ext4_group_info data for the given group. */
2196 int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2197 struct ext4_group_desc *desc)
2201 struct ext4_sb_info *sbi = EXT4_SB(sb);
2202 struct ext4_group_info **meta_group_info;
2205 * First check if this group is the first of a reserved block.
2206 * If it's true, we have to allocate a new table of pointers
2207 * to ext4_group_info structures
2209 if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2210 metalen = sizeof(*meta_group_info) <<
2211 EXT4_DESC_PER_BLOCK_BITS(sb);
2212 meta_group_info = kmalloc(metalen, GFP_KERNEL);
2213 if (meta_group_info == NULL) {
2214 printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
2216 goto exit_meta_group_info;
2218 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2223 * calculate needed size. if change bb_counters size,
2224 * don't forget about ext4_mb_generate_buddy()
2226 len = offsetof(typeof(**meta_group_info),
2227 bb_counters[sb->s_blocksize_bits + 2]);
2230 sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2231 i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2233 meta_group_info[i] = kzalloc(len, GFP_KERNEL);
2234 if (meta_group_info[i] == NULL) {
2235 printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
2236 goto exit_group_info;
2238 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2239 &(meta_group_info[i]->bb_state));
2242 * initialize bb_free to be able to skip
2243 * empty groups without initialization
2245 if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2246 meta_group_info[i]->bb_free =
2247 ext4_free_blocks_after_init(sb, group, desc);
2249 meta_group_info[i]->bb_free =
2250 ext4_free_blks_count(sb, desc);
2253 INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2254 init_rwsem(&meta_group_info[i]->alloc_sem);
2255 meta_group_info[i]->bb_free_root = RB_ROOT;
2259 struct buffer_head *bh;
2260 meta_group_info[i]->bb_bitmap =
2261 kmalloc(sb->s_blocksize, GFP_KERNEL);
2262 BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2263 bh = ext4_read_block_bitmap(sb, group);
2265 memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2274 /* If a meta_group_info table has been allocated, release it now */
2275 if (group % EXT4_DESC_PER_BLOCK(sb) == 0)
2276 kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2277 exit_meta_group_info:
2279 } /* ext4_mb_add_groupinfo */
2281 static int ext4_mb_init_backend(struct super_block *sb)
2283 ext4_group_t ngroups = ext4_get_groups_count(sb);
2285 struct ext4_sb_info *sbi = EXT4_SB(sb);
2286 struct ext4_super_block *es = sbi->s_es;
2287 int num_meta_group_infos;
2288 int num_meta_group_infos_max;
2290 struct ext4_group_desc *desc;
2292 /* This is the number of blocks used by GDT */
2293 num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2294 1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2297 * This is the total number of blocks used by GDT including
2298 * the number of reserved blocks for GDT.
2299 * The s_group_info array is allocated with this value
2300 * to allow a clean online resize without a complex
2301 * manipulation of pointer.
2302 * The drawback is the unused memory when no resize
2303 * occurs but it's very low in terms of pages
2304 * (see comments below)
2305 * Need to handle this properly when META_BG resizing is allowed
2307 num_meta_group_infos_max = num_meta_group_infos +
2308 le16_to_cpu(es->s_reserved_gdt_blocks);
2311 * array_size is the size of s_group_info array. We round it
2312 * to the next power of two because this approximation is done
2313 * internally by kmalloc so we can have some more memory
2314 * for free here (e.g. may be used for META_BG resize).
2317 while (array_size < sizeof(*sbi->s_group_info) *
2318 num_meta_group_infos_max)
2319 array_size = array_size << 1;
2320 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2321 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2322 * So a two level scheme suffices for now. */
2323 sbi->s_group_info = kmalloc(array_size, GFP_KERNEL);
2324 if (sbi->s_group_info == NULL) {
2325 printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
2328 sbi->s_buddy_cache = new_inode(sb);
2329 if (sbi->s_buddy_cache == NULL) {
2330 printk(KERN_ERR "EXT4-fs: can't get new inode\n");
2333 EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2334 for (i = 0; i < ngroups; i++) {
2335 desc = ext4_get_group_desc(sb, i, NULL);
2338 "EXT4-fs: can't read descriptor %u\n", i);
2341 if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2349 kfree(ext4_get_group_info(sb, i));
2350 i = num_meta_group_infos;
2352 kfree(sbi->s_group_info[i]);
2353 iput(sbi->s_buddy_cache);
2355 kfree(sbi->s_group_info);
2359 int ext4_mb_init(struct super_block *sb, int needs_recovery)
2361 struct ext4_sb_info *sbi = EXT4_SB(sb);
2367 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2369 sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2370 if (sbi->s_mb_offsets == NULL) {
2374 i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2375 sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2376 if (sbi->s_mb_maxs == NULL) {
2377 kfree(sbi->s_mb_offsets);
2381 /* order 0 is regular bitmap */
2382 sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2383 sbi->s_mb_offsets[0] = 0;
2387 max = sb->s_blocksize << 2;
2389 sbi->s_mb_offsets[i] = offset;
2390 sbi->s_mb_maxs[i] = max;
2391 offset += 1 << (sb->s_blocksize_bits - i);
2394 } while (i <= sb->s_blocksize_bits + 1);
2396 /* init file for buddy data */
2397 ret = ext4_mb_init_backend(sb);
2399 kfree(sbi->s_mb_offsets);
2400 kfree(sbi->s_mb_maxs);
2404 spin_lock_init(&sbi->s_md_lock);
2405 spin_lock_init(&sbi->s_bal_lock);
2407 sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2408 sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2409 sbi->s_mb_stats = MB_DEFAULT_STATS;
2410 sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2411 sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2412 sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
2414 sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2415 if (sbi->s_locality_groups == NULL) {
2416 kfree(sbi->s_mb_offsets);
2417 kfree(sbi->s_mb_maxs);
2420 for_each_possible_cpu(i) {
2421 struct ext4_locality_group *lg;
2422 lg = per_cpu_ptr(sbi->s_locality_groups, i);
2423 mutex_init(&lg->lg_mutex);
2424 for (j = 0; j < PREALLOC_TB_SIZE; j++)
2425 INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2426 spin_lock_init(&lg->lg_prealloc_lock);
2430 proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2431 &ext4_mb_seq_groups_fops, sb);
2434 sbi->s_journal->j_commit_callback = release_blocks_on_commit;
2438 /* need to called with the ext4 group lock held */
2439 static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2441 struct ext4_prealloc_space *pa;
2442 struct list_head *cur, *tmp;
2445 list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2446 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2447 list_del(&pa->pa_group_list);
2449 kmem_cache_free(ext4_pspace_cachep, pa);
2452 mb_debug(1, "mballoc: %u PAs left\n", count);
2456 int ext4_mb_release(struct super_block *sb)
2458 ext4_group_t ngroups = ext4_get_groups_count(sb);
2460 int num_meta_group_infos;
2461 struct ext4_group_info *grinfo;
2462 struct ext4_sb_info *sbi = EXT4_SB(sb);
2464 if (sbi->s_group_info) {
2465 for (i = 0; i < ngroups; i++) {
2466 grinfo = ext4_get_group_info(sb, i);
2468 kfree(grinfo->bb_bitmap);
2470 ext4_lock_group(sb, i);
2471 ext4_mb_cleanup_pa(grinfo);
2472 ext4_unlock_group(sb, i);
2475 num_meta_group_infos = (ngroups +
2476 EXT4_DESC_PER_BLOCK(sb) - 1) >>
2477 EXT4_DESC_PER_BLOCK_BITS(sb);
2478 for (i = 0; i < num_meta_group_infos; i++)
2479 kfree(sbi->s_group_info[i]);
2480 kfree(sbi->s_group_info);
2482 kfree(sbi->s_mb_offsets);
2483 kfree(sbi->s_mb_maxs);
2484 if (sbi->s_buddy_cache)
2485 iput(sbi->s_buddy_cache);
2486 if (sbi->s_mb_stats) {
2488 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2489 atomic_read(&sbi->s_bal_allocated),
2490 atomic_read(&sbi->s_bal_reqs),
2491 atomic_read(&sbi->s_bal_success));
2493 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2494 "%u 2^N hits, %u breaks, %u lost\n",
2495 atomic_read(&sbi->s_bal_ex_scanned),
2496 atomic_read(&sbi->s_bal_goals),
2497 atomic_read(&sbi->s_bal_2orders),
2498 atomic_read(&sbi->s_bal_breaks),
2499 atomic_read(&sbi->s_mb_lost_chunks));
2501 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2502 sbi->s_mb_buddies_generated++,
2503 sbi->s_mb_generation_time);
2505 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2506 atomic_read(&sbi->s_mb_preallocated),
2507 atomic_read(&sbi->s_mb_discarded));
2510 free_percpu(sbi->s_locality_groups);
2512 remove_proc_entry("mb_groups", sbi->s_proc);
2518 * This function is called by the jbd2 layer once the commit has finished,
2519 * so we know we can free the blocks that were released with that commit.
2521 static void release_blocks_on_commit(journal_t *journal, transaction_t *txn)
2523 struct super_block *sb = journal->j_private;
2524 struct ext4_buddy e4b;
2525 struct ext4_group_info *db;
2526 int err, count = 0, count2 = 0;
2527 struct ext4_free_data *entry;
2528 struct list_head *l, *ltmp;
2530 list_for_each_safe(l, ltmp, &txn->t_private_list) {
2531 entry = list_entry(l, struct ext4_free_data, list);
2533 mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2534 entry->count, entry->group, entry);
2536 if (test_opt(sb, DISCARD)) {
2537 ext4_fsblk_t discard_block;
2539 discard_block = entry->start_blk +
2540 ext4_group_first_block_no(sb, entry->group);
2541 trace_ext4_discard_blocks(sb,
2542 (unsigned long long)discard_block,
2544 sb_issue_discard(sb, discard_block, entry->count);
2547 err = ext4_mb_load_buddy(sb, entry->group, &e4b);
2548 /* we expect to find existing buddy because it's pinned */
2552 /* there are blocks to put in buddy to make them really free */
2553 count += entry->count;
2555 ext4_lock_group(sb, entry->group);
2556 /* Take it out of per group rb tree */
2557 rb_erase(&entry->node, &(db->bb_free_root));
2558 mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count);
2560 if (!db->bb_free_root.rb_node) {
2561 /* No more items in the per group rb tree
2562 * balance refcounts from ext4_mb_free_metadata()
2564 page_cache_release(e4b.bd_buddy_page);
2565 page_cache_release(e4b.bd_bitmap_page);
2567 ext4_unlock_group(sb, entry->group);
2568 kmem_cache_free(ext4_free_ext_cachep, entry);
2569 ext4_mb_release_desc(&e4b);
2572 mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2575 #ifdef CONFIG_EXT4_DEBUG
2576 u8 mb_enable_debug __read_mostly;
2578 static struct dentry *debugfs_dir;
2579 static struct dentry *debugfs_debug;
2581 static void __init ext4_create_debugfs_entry(void)
2583 debugfs_dir = debugfs_create_dir("ext4", NULL);
2585 debugfs_debug = debugfs_create_u8("mballoc-debug",
2591 static void ext4_remove_debugfs_entry(void)
2593 debugfs_remove(debugfs_debug);
2594 debugfs_remove(debugfs_dir);
2599 static void __init ext4_create_debugfs_entry(void)
2603 static void ext4_remove_debugfs_entry(void)
2609 int __init init_ext4_mballoc(void)
2611 ext4_pspace_cachep =
2612 kmem_cache_create("ext4_prealloc_space",
2613 sizeof(struct ext4_prealloc_space),
2614 0, SLAB_RECLAIM_ACCOUNT, NULL);
2615 if (ext4_pspace_cachep == NULL)
2619 kmem_cache_create("ext4_alloc_context",
2620 sizeof(struct ext4_allocation_context),
2621 0, SLAB_RECLAIM_ACCOUNT, NULL);
2622 if (ext4_ac_cachep == NULL) {
2623 kmem_cache_destroy(ext4_pspace_cachep);
2627 ext4_free_ext_cachep =
2628 kmem_cache_create("ext4_free_block_extents",
2629 sizeof(struct ext4_free_data),
2630 0, SLAB_RECLAIM_ACCOUNT, NULL);
2631 if (ext4_free_ext_cachep == NULL) {
2632 kmem_cache_destroy(ext4_pspace_cachep);
2633 kmem_cache_destroy(ext4_ac_cachep);
2636 ext4_create_debugfs_entry();
2640 void exit_ext4_mballoc(void)
2643 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2644 * before destroying the slab cache.
2647 kmem_cache_destroy(ext4_pspace_cachep);
2648 kmem_cache_destroy(ext4_ac_cachep);
2649 kmem_cache_destroy(ext4_free_ext_cachep);
2650 ext4_remove_debugfs_entry();
2655 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2656 * Returns 0 if success or error code
2658 static noinline_for_stack int
2659 ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2660 handle_t *handle, unsigned int reserv_blks)
2662 struct buffer_head *bitmap_bh = NULL;
2663 struct ext4_super_block *es;
2664 struct ext4_group_desc *gdp;
2665 struct buffer_head *gdp_bh;
2666 struct ext4_sb_info *sbi;
2667 struct super_block *sb;
2671 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2672 BUG_ON(ac->ac_b_ex.fe_len <= 0);
2680 bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2684 err = ext4_journal_get_write_access(handle, bitmap_bh);
2689 gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2693 ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2694 ext4_free_blks_count(sb, gdp));
2696 err = ext4_journal_get_write_access(handle, gdp_bh);
2700 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2702 len = ac->ac_b_ex.fe_len;
2703 if (!ext4_data_block_valid(sbi, block, len)) {
2704 ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2705 "fs metadata\n", block, block+len);
2706 /* File system mounted not to panic on error
2707 * Fix the bitmap and repeat the block allocation
2708 * We leak some of the blocks here.
2710 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2711 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2712 ac->ac_b_ex.fe_len);
2713 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2714 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2720 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2721 #ifdef AGGRESSIVE_CHECK
2724 for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2725 BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2726 bitmap_bh->b_data));
2730 mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len);
2731 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2732 gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2733 ext4_free_blks_set(sb, gdp,
2734 ext4_free_blocks_after_init(sb,
2735 ac->ac_b_ex.fe_group, gdp));
2737 len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len;
2738 ext4_free_blks_set(sb, gdp, len);
2739 gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
2741 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2742 percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
2744 * Now reduce the dirty block count also. Should not go negative
2746 if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2747 /* release all the reserved blocks if non delalloc */
2748 percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks);
2750 if (sbi->s_log_groups_per_flex) {
2751 ext4_group_t flex_group = ext4_flex_group(sbi,
2752 ac->ac_b_ex.fe_group);
2753 atomic_sub(ac->ac_b_ex.fe_len,
2754 &sbi->s_flex_groups[flex_group].free_blocks);
2757 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2760 err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2769 * here we normalize request for locality group
2770 * Group request are normalized to s_strip size if we set the same via mount
2771 * option. If not we set it to s_mb_group_prealloc which can be configured via
2772 * /sys/fs/ext4/<partition>/mb_group_prealloc
2774 * XXX: should we try to preallocate more than the group has now?
2776 static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2778 struct super_block *sb = ac->ac_sb;
2779 struct ext4_locality_group *lg = ac->ac_lg;
2782 if (EXT4_SB(sb)->s_stripe)
2783 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
2785 ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2786 mb_debug(1, "#%u: goal %u blocks for locality group\n",
2787 current->pid, ac->ac_g_ex.fe_len);
2791 * Normalization means making request better in terms of
2792 * size and alignment
2794 static noinline_for_stack void
2795 ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2796 struct ext4_allocation_request *ar)
2800 loff_t size, orig_size, start_off;
2801 ext4_lblk_t start, orig_start;
2802 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2803 struct ext4_prealloc_space *pa;
2805 /* do normalize only data requests, metadata requests
2806 do not need preallocation */
2807 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2810 /* sometime caller may want exact blocks */
2811 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2814 /* caller may indicate that preallocation isn't
2815 * required (it's a tail, for example) */
2816 if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2819 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2820 ext4_mb_normalize_group_request(ac);
2824 bsbits = ac->ac_sb->s_blocksize_bits;
2826 /* first, let's learn actual file size
2827 * given current request is allocated */
2828 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
2829 size = size << bsbits;
2830 if (size < i_size_read(ac->ac_inode))
2831 size = i_size_read(ac->ac_inode);
2833 /* max size of free chunks */
2836 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
2837 (req <= (size) || max <= (chunk_size))
2839 /* first, try to predict filesize */
2840 /* XXX: should this table be tunable? */
2842 if (size <= 16 * 1024) {
2844 } else if (size <= 32 * 1024) {
2846 } else if (size <= 64 * 1024) {
2848 } else if (size <= 128 * 1024) {
2850 } else if (size <= 256 * 1024) {
2852 } else if (size <= 512 * 1024) {
2854 } else if (size <= 1024 * 1024) {
2856 } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2857 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2858 (21 - bsbits)) << 21;
2859 size = 2 * 1024 * 1024;
2860 } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2861 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2862 (22 - bsbits)) << 22;
2863 size = 4 * 1024 * 1024;
2864 } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2865 (8<<20)>>bsbits, max, 8 * 1024)) {
2866 start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2867 (23 - bsbits)) << 23;
2868 size = 8 * 1024 * 1024;
2870 start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2871 size = ac->ac_o_ex.fe_len << bsbits;
2873 orig_size = size = size >> bsbits;
2874 orig_start = start = start_off >> bsbits;
2876 /* don't cover already allocated blocks in selected range */
2877 if (ar->pleft && start <= ar->lleft) {
2878 size -= ar->lleft + 1 - start;
2879 start = ar->lleft + 1;
2881 if (ar->pright && start + size - 1 >= ar->lright)
2882 size -= start + size - ar->lright;
2886 /* check we don't cross already preallocated blocks */
2888 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2893 spin_lock(&pa->pa_lock);
2894 if (pa->pa_deleted) {
2895 spin_unlock(&pa->pa_lock);
2899 pa_end = pa->pa_lstart + pa->pa_len;
2901 /* PA must not overlap original request */
2902 BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2903 ac->ac_o_ex.fe_logical < pa->pa_lstart));
2905 /* skip PAs this normalized request doesn't overlap with */
2906 if (pa->pa_lstart >= end || pa_end <= start) {
2907 spin_unlock(&pa->pa_lock);
2910 BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2912 /* adjust start or end to be adjacent to this pa */
2913 if (pa_end <= ac->ac_o_ex.fe_logical) {
2914 BUG_ON(pa_end < start);
2916 } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2917 BUG_ON(pa->pa_lstart > end);
2918 end = pa->pa_lstart;
2920 spin_unlock(&pa->pa_lock);
2925 /* XXX: extra loop to check we really don't overlap preallocations */
2927 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2929 spin_lock(&pa->pa_lock);
2930 if (pa->pa_deleted == 0) {
2931 pa_end = pa->pa_lstart + pa->pa_len;
2932 BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
2934 spin_unlock(&pa->pa_lock);
2938 if (start + size <= ac->ac_o_ex.fe_logical &&
2939 start > ac->ac_o_ex.fe_logical) {
2940 printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
2941 (unsigned long) start, (unsigned long) size,
2942 (unsigned long) ac->ac_o_ex.fe_logical);
2944 BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
2945 start > ac->ac_o_ex.fe_logical);
2946 BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
2948 /* now prepare goal request */
2950 /* XXX: is it better to align blocks WRT to logical
2951 * placement or satisfy big request as is */
2952 ac->ac_g_ex.fe_logical = start;
2953 ac->ac_g_ex.fe_len = size;
2955 /* define goal start in order to merge */
2956 if (ar->pright && (ar->lright == (start + size))) {
2957 /* merge to the right */
2958 ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
2959 &ac->ac_f_ex.fe_group,
2960 &ac->ac_f_ex.fe_start);
2961 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
2963 if (ar->pleft && (ar->lleft + 1 == start)) {
2964 /* merge to the left */
2965 ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
2966 &ac->ac_f_ex.fe_group,
2967 &ac->ac_f_ex.fe_start);
2968 ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
2971 mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
2972 (unsigned) orig_size, (unsigned) start);
2975 static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
2977 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2979 if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
2980 atomic_inc(&sbi->s_bal_reqs);
2981 atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
2982 if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
2983 atomic_inc(&sbi->s_bal_success);
2984 atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
2985 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
2986 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
2987 atomic_inc(&sbi->s_bal_goals);
2988 if (ac->ac_found > sbi->s_mb_max_to_scan)
2989 atomic_inc(&sbi->s_bal_breaks);
2992 if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
2993 trace_ext4_mballoc_alloc(ac);
2995 trace_ext4_mballoc_prealloc(ac);
2999 * Called on failure; free up any blocks from the inode PA for this
3000 * context. We don't need this for MB_GROUP_PA because we only change
3001 * pa_free in ext4_mb_release_context(), but on failure, we've already
3002 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3004 static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3006 struct ext4_prealloc_space *pa = ac->ac_pa;
3009 if (pa && pa->pa_type == MB_INODE_PA) {
3010 len = ac->ac_b_ex.fe_len;
3017 * use blocks preallocated to inode
3019 static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3020 struct ext4_prealloc_space *pa)
3026 /* found preallocated blocks, use them */
3027 start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3028 end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
3030 ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3031 &ac->ac_b_ex.fe_start);
3032 ac->ac_b_ex.fe_len = len;
3033 ac->ac_status = AC_STATUS_FOUND;
3036 BUG_ON(start < pa->pa_pstart);
3037 BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
3038 BUG_ON(pa->pa_free < len);
3041 mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3045 * use blocks preallocated to locality group
3047 static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3048 struct ext4_prealloc_space *pa)
3050 unsigned int len = ac->ac_o_ex.fe_len;
3052 ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3053 &ac->ac_b_ex.fe_group,
3054 &ac->ac_b_ex.fe_start);
3055 ac->ac_b_ex.fe_len = len;
3056 ac->ac_status = AC_STATUS_FOUND;
3059 /* we don't correct pa_pstart or pa_plen here to avoid
3060 * possible race when the group is being loaded concurrently
3061 * instead we correct pa later, after blocks are marked
3062 * in on-disk bitmap -- see ext4_mb_release_context()
3063 * Other CPUs are prevented from allocating from this pa by lg_mutex
3065 mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3069 * Return the prealloc space that have minimal distance
3070 * from the goal block. @cpa is the prealloc
3071 * space that is having currently known minimal distance
3072 * from the goal block.
3074 static struct ext4_prealloc_space *
3075 ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3076 struct ext4_prealloc_space *pa,
3077 struct ext4_prealloc_space *cpa)
3079 ext4_fsblk_t cur_distance, new_distance;
3082 atomic_inc(&pa->pa_count);
3085 cur_distance = abs(goal_block - cpa->pa_pstart);
3086 new_distance = abs(goal_block - pa->pa_pstart);
3088 if (cur_distance < new_distance)
3091 /* drop the previous reference */
3092 atomic_dec(&cpa->pa_count);
3093 atomic_inc(&pa->pa_count);
3098 * search goal blocks in preallocated space
3100 static noinline_for_stack int
3101 ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3104 struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3105 struct ext4_locality_group *lg;
3106 struct ext4_prealloc_space *pa, *cpa = NULL;
3107 ext4_fsblk_t goal_block;
3109 /* only data can be preallocated */
3110 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3113 /* first, try per-file preallocation */
3115 list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3117 /* all fields in this condition don't change,
3118 * so we can skip locking for them */
3119 if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3120 ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
3123 /* non-extent files can't have physical blocks past 2^32 */
3124 if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL) &&
3125 pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS)
3128 /* found preallocated blocks, use them */
3129 spin_lock(&pa->pa_lock);
3130 if (pa->pa_deleted == 0 && pa->pa_free) {
3131 atomic_inc(&pa->pa_count);
3132 ext4_mb_use_inode_pa(ac, pa);
3133 spin_unlock(&pa->pa_lock);
3134 ac->ac_criteria = 10;
3138 spin_unlock(&pa->pa_lock);
3142 /* can we use group allocation? */
3143 if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3146 /* inode may have no locality group for some reason */
3150 order = fls(ac->ac_o_ex.fe_len) - 1;
3151 if (order > PREALLOC_TB_SIZE - 1)
3152 /* The max size of hash table is PREALLOC_TB_SIZE */
3153 order = PREALLOC_TB_SIZE - 1;
3155 goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3157 * search for the prealloc space that is having
3158 * minimal distance from the goal block.
3160 for (i = order; i < PREALLOC_TB_SIZE; i++) {
3162 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3164 spin_lock(&pa->pa_lock);
3165 if (pa->pa_deleted == 0 &&
3166 pa->pa_free >= ac->ac_o_ex.fe_len) {
3168 cpa = ext4_mb_check_group_pa(goal_block,
3171 spin_unlock(&pa->pa_lock);
3176 ext4_mb_use_group_pa(ac, cpa);
3177 ac->ac_criteria = 20;
3184 * the function goes through all block freed in the group
3185 * but not yet committed and marks them used in in-core bitmap.
3186 * buddy must be generated from this bitmap
3187 * Need to be called with the ext4 group lock held
3189 static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3193 struct ext4_group_info *grp;
3194 struct ext4_free_data *entry;
3196 grp = ext4_get_group_info(sb, group);
3197 n = rb_first(&(grp->bb_free_root));
3200 entry = rb_entry(n, struct ext4_free_data, node);
3201 mb_set_bits(bitmap, entry->start_blk, entry->count);
3208 * the function goes through all preallocation in this group and marks them
3209 * used in in-core bitmap. buddy must be generated from this bitmap
3210 * Need to be called with ext4 group lock held
3212 static noinline_for_stack
3213 void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3216 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3217 struct ext4_prealloc_space *pa;
3218 struct list_head *cur;
3219 ext4_group_t groupnr;
3220 ext4_grpblk_t start;
3221 int preallocated = 0;
3225 /* all form of preallocation discards first load group,
3226 * so the only competing code is preallocation use.
3227 * we don't need any locking here
3228 * notice we do NOT ignore preallocations with pa_deleted
3229 * otherwise we could leave used blocks available for
3230 * allocation in buddy when concurrent ext4_mb_put_pa()
3231 * is dropping preallocation
3233 list_for_each(cur, &grp->bb_prealloc_list) {
3234 pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3235 spin_lock(&pa->pa_lock);
3236 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3239 spin_unlock(&pa->pa_lock);
3240 if (unlikely(len == 0))
3242 BUG_ON(groupnr != group);
3243 mb_set_bits(bitmap, start, len);
3244 preallocated += len;
3247 mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3250 static void ext4_mb_pa_callback(struct rcu_head *head)
3252 struct ext4_prealloc_space *pa;
3253 pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3254 kmem_cache_free(ext4_pspace_cachep, pa);
3258 * drops a reference to preallocated space descriptor
3259 * if this was the last reference and the space is consumed
3261 static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3262 struct super_block *sb, struct ext4_prealloc_space *pa)
3265 ext4_fsblk_t grp_blk;
3267 if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3270 /* in this short window concurrent discard can set pa_deleted */
3271 spin_lock(&pa->pa_lock);
3272 if (pa->pa_deleted == 1) {
3273 spin_unlock(&pa->pa_lock);
3278 spin_unlock(&pa->pa_lock);
3280 grp_blk = pa->pa_pstart;
3282 * If doing group-based preallocation, pa_pstart may be in the
3283 * next group when pa is used up
3285 if (pa->pa_type == MB_GROUP_PA)
3288 ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3293 * P1 (buddy init) P2 (regular allocation)
3294 * find block B in PA
3295 * copy on-disk bitmap to buddy
3296 * mark B in on-disk bitmap
3297 * drop PA from group
3298 * mark all PAs in buddy
3300 * thus, P1 initializes buddy with B available. to prevent this
3301 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3304 ext4_lock_group(sb, grp);
3305 list_del(&pa->pa_group_list);
3306 ext4_unlock_group(sb, grp);
3308 spin_lock(pa->pa_obj_lock);
3309 list_del_rcu(&pa->pa_inode_list);
3310 spin_unlock(pa->pa_obj_lock);
3312 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3316 * creates new preallocated space for given inode
3318 static noinline_for_stack int
3319 ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3321 struct super_block *sb = ac->ac_sb;
3322 struct ext4_prealloc_space *pa;
3323 struct ext4_group_info *grp;
3324 struct ext4_inode_info *ei;
3326 /* preallocate only when found space is larger then requested */
3327 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3328 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3329 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3331 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3335 if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3341 /* we can't allocate as much as normalizer wants.
3342 * so, found space must get proper lstart
3343 * to cover original request */
3344 BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3345 BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3347 /* we're limited by original request in that
3348 * logical block must be covered any way
3349 * winl is window we can move our chunk within */
3350 winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3352 /* also, we should cover whole original request */
3353 wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
3355 /* the smallest one defines real window */
3356 win = min(winl, wins);
3358 offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
3359 if (offs && offs < win)
3362 ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
3363 BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3364 BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3367 /* preallocation can change ac_b_ex, thus we store actually
3368 * allocated blocks for history */
3369 ac->ac_f_ex = ac->ac_b_ex;
3371 pa->pa_lstart = ac->ac_b_ex.fe_logical;
3372 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3373 pa->pa_len = ac->ac_b_ex.fe_len;
3374 pa->pa_free = pa->pa_len;
3375 atomic_set(&pa->pa_count, 1);
3376 spin_lock_init(&pa->pa_lock);
3377 INIT_LIST_HEAD(&pa->pa_inode_list);
3378 INIT_LIST_HEAD(&pa->pa_group_list);
3380 pa->pa_type = MB_INODE_PA;
3382 mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3383 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3384 trace_ext4_mb_new_inode_pa(ac, pa);
3386 ext4_mb_use_inode_pa(ac, pa);
3387 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3389 ei = EXT4_I(ac->ac_inode);
3390 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3392 pa->pa_obj_lock = &ei->i_prealloc_lock;
3393 pa->pa_inode = ac->ac_inode;
3395 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3396 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3397 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3399 spin_lock(pa->pa_obj_lock);
3400 list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3401 spin_unlock(pa->pa_obj_lock);
3407 * creates new preallocated space for locality group inodes belongs to
3409 static noinline_for_stack int
3410 ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3412 struct super_block *sb = ac->ac_sb;
3413 struct ext4_locality_group *lg;
3414 struct ext4_prealloc_space *pa;
3415 struct ext4_group_info *grp;
3417 /* preallocate only when found space is larger then requested */
3418 BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3419 BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3420 BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3422 BUG_ON(ext4_pspace_cachep == NULL);
3423 pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3427 /* preallocation can change ac_b_ex, thus we store actually
3428 * allocated blocks for history */
3429 ac->ac_f_ex = ac->ac_b_ex;
3431 pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3432 pa->pa_lstart = pa->pa_pstart;
3433 pa->pa_len = ac->ac_b_ex.fe_len;
3434 pa->pa_free = pa->pa_len;
3435 atomic_set(&pa->pa_count, 1);
3436 spin_lock_init(&pa->pa_lock);
3437 INIT_LIST_HEAD(&pa->pa_inode_list);
3438 INIT_LIST_HEAD(&pa->pa_group_list);
3440 pa->pa_type = MB_GROUP_PA;
3442 mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3443 pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3444 trace_ext4_mb_new_group_pa(ac, pa);
3446 ext4_mb_use_group_pa(ac, pa);
3447 atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3449 grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3453 pa->pa_obj_lock = &lg->lg_prealloc_lock;
3454 pa->pa_inode = NULL;
3456 ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3457 list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3458 ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3461 * We will later add the new pa to the right bucket
3462 * after updating the pa_free in ext4_mb_release_context
3467 static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3471 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3472 err = ext4_mb_new_group_pa(ac);
3474 err = ext4_mb_new_inode_pa(ac);
3479 * finds all unused blocks in on-disk bitmap, frees them in
3480 * in-core bitmap and buddy.
3481 * @pa must be unlinked from inode and group lists, so that
3482 * nobody else can find/use it.
3483 * the caller MUST hold group/inode locks.
3484 * TODO: optimize the case when there are no in-core structures yet
3486 static noinline_for_stack int
3487 ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3488 struct ext4_prealloc_space *pa,
3489 struct ext4_allocation_context *ac)
3491 struct super_block *sb = e4b->bd_sb;
3492 struct ext4_sb_info *sbi = EXT4_SB(sb);
3497 unsigned long long grp_blk_start;
3502 BUG_ON(pa->pa_deleted == 0);
3503 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3504 grp_blk_start = pa->pa_pstart - bit;
3505 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3506 end = bit + pa->pa_len;
3510 ac->ac_inode = pa->pa_inode;
3514 bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3517 next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3518 start = ext4_group_first_block_no(sb, group) + bit;
3519 mb_debug(1, " free preallocated %u/%u in group %u\n",
3520 (unsigned) start, (unsigned) next - bit,
3525 ac->ac_b_ex.fe_group = group;
3526 ac->ac_b_ex.fe_start = bit;
3527 ac->ac_b_ex.fe_len = next - bit;
3528 ac->ac_b_ex.fe_logical = 0;
3529 trace_ext4_mballoc_discard(ac);
3532 trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit,
3534 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3537 if (free != pa->pa_free) {
3538 printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
3539 pa, (unsigned long) pa->pa_lstart,
3540 (unsigned long) pa->pa_pstart,
3541 (unsigned long) pa->pa_len);
3542 ext4_grp_locked_error(sb, group,
3543 __func__, "free %u, pa_free %u",
3546 * pa is already deleted so we use the value obtained
3547 * from the bitmap and continue.
3550 atomic_add(free, &sbi->s_mb_discarded);
3555 static noinline_for_stack int
3556 ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3557 struct ext4_prealloc_space *pa,
3558 struct ext4_allocation_context *ac)
3560 struct super_block *sb = e4b->bd_sb;
3564 trace_ext4_mb_release_group_pa(ac, pa);
3565 BUG_ON(pa->pa_deleted == 0);
3566 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3567 BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3568 mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3569 atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3573 ac->ac_inode = NULL;
3574 ac->ac_b_ex.fe_group = group;
3575 ac->ac_b_ex.fe_start = bit;
3576 ac->ac_b_ex.fe_len = pa->pa_len;
3577 ac->ac_b_ex.fe_logical = 0;
3578 trace_ext4_mballoc_discard(ac);
3585 * releases all preallocations in given group
3587 * first, we need to decide discard policy:
3588 * - when do we discard
3590 * - how many do we discard
3591 * 1) how many requested
3593 static noinline_for_stack int
3594 ext4_mb_discard_group_preallocations(struct super_block *sb,
3595 ext4_group_t group, int needed)
3597 struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3598 struct buffer_head *bitmap_bh = NULL;
3599 struct ext4_prealloc_space *pa, *tmp;
3600 struct ext4_allocation_context *ac;
3601 struct list_head list;
3602 struct ext4_buddy e4b;
3607 mb_debug(1, "discard preallocation for group %u\n", group);
3609 if (list_empty(&grp->bb_prealloc_list))
3612 bitmap_bh = ext4_read_block_bitmap(sb, group);
3613 if (bitmap_bh == NULL) {
3614 ext4_error(sb, "Error reading block bitmap for %u", group);
3618 err = ext4_mb_load_buddy(sb, group, &e4b);
3620 ext4_error(sb, "Error loading buddy information for %u", group);
3626 needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
3628 INIT_LIST_HEAD(&list);
3629 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3633 ext4_lock_group(sb, group);
3634 list_for_each_entry_safe(pa, tmp,
3635 &grp->bb_prealloc_list, pa_group_list) {
3636 spin_lock(&pa->pa_lock);
3637 if (atomic_read(&pa->pa_count)) {
3638 spin_unlock(&pa->pa_lock);
3642 if (pa->pa_deleted) {
3643 spin_unlock(&pa->pa_lock);
3647 /* seems this one can be freed ... */
3650 /* we can trust pa_free ... */
3651 free += pa->pa_free;
3653 spin_unlock(&pa->pa_lock);
3655 list_del(&pa->pa_group_list);
3656 list_add(&pa->u.pa_tmp_list, &list);
3659 /* if we still need more blocks and some PAs were used, try again */
3660 if (free < needed && busy) {
3662 ext4_unlock_group(sb, group);
3664 * Yield the CPU here so that we don't get soft lockup
3665 * in non preempt case.
3671 /* found anything to free? */
3672 if (list_empty(&list)) {
3677 /* now free all selected PAs */
3678 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3680 /* remove from object (inode or locality group) */
3681 spin_lock(pa->pa_obj_lock);
3682 list_del_rcu(&pa->pa_inode_list);
3683 spin_unlock(pa->pa_obj_lock);
3685 if (pa->pa_type == MB_GROUP_PA)
3686 ext4_mb_release_group_pa(&e4b, pa, ac);
3688 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3690 list_del(&pa->u.pa_tmp_list);
3691 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3695 ext4_unlock_group(sb, group);
3697 kmem_cache_free(ext4_ac_cachep, ac);
3698 ext4_mb_release_desc(&e4b);
3704 * releases all non-used preallocated blocks for given inode
3706 * It's important to discard preallocations under i_data_sem
3707 * We don't want another block to be served from the prealloc
3708 * space when we are discarding the inode prealloc space.
3710 * FIXME!! Make sure it is valid at all the call sites
3712 void ext4_discard_preallocations(struct inode *inode)
3714 struct ext4_inode_info *ei = EXT4_I(inode);
3715 struct super_block *sb = inode->i_sb;
3716 struct buffer_head *bitmap_bh = NULL;
3717 struct ext4_prealloc_space *pa, *tmp;
3718 struct ext4_allocation_context *ac;
3719 ext4_group_t group = 0;
3720 struct list_head list;
3721 struct ext4_buddy e4b;
3724 if (!S_ISREG(inode->i_mode)) {
3725 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3729 mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3730 trace_ext4_discard_preallocations(inode);
3732 INIT_LIST_HEAD(&list);
3734 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
3737 ac->ac_inode = inode;
3740 /* first, collect all pa's in the inode */
3741 spin_lock(&ei->i_prealloc_lock);
3742 while (!list_empty(&ei->i_prealloc_list)) {
3743 pa = list_entry(ei->i_prealloc_list.next,
3744 struct ext4_prealloc_space, pa_inode_list);
3745 BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3746 spin_lock(&pa->pa_lock);
3747 if (atomic_read(&pa->pa_count)) {
3748 /* this shouldn't happen often - nobody should
3749 * use preallocation while we're discarding it */
3750 spin_unlock(&pa->pa_lock);
3751 spin_unlock(&ei->i_prealloc_lock);
3752 printk(KERN_ERR "uh-oh! used pa while discarding\n");
3754 schedule_timeout_uninterruptible(HZ);
3758 if (pa->pa_deleted == 0) {
3760 spin_unlock(&pa->pa_lock);
3761 list_del_rcu(&pa->pa_inode_list);
3762 list_add(&pa->u.pa_tmp_list, &list);
3766 /* someone is deleting pa right now */
3767 spin_unlock(&pa->pa_lock);
3768 spin_unlock(&ei->i_prealloc_lock);
3770 /* we have to wait here because pa_deleted
3771 * doesn't mean pa is already unlinked from
3772 * the list. as we might be called from
3773 * ->clear_inode() the inode will get freed
3774 * and concurrent thread which is unlinking
3775 * pa from inode's list may access already
3776 * freed memory, bad-bad-bad */
3778 /* XXX: if this happens too often, we can
3779 * add a flag to force wait only in case
3780 * of ->clear_inode(), but not in case of
3781 * regular truncate */
3782 schedule_timeout_uninterruptible(HZ);
3785 spin_unlock(&ei->i_prealloc_lock);
3787 list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3788 BUG_ON(pa->pa_type != MB_INODE_PA);
3789 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3791 err = ext4_mb_load_buddy(sb, group, &e4b);
3793 ext4_error(sb, "Error loading buddy information for %u",
3798 bitmap_bh = ext4_read_block_bitmap(sb, group);
3799 if (bitmap_bh == NULL) {
3800 ext4_error(sb, "Error reading block bitmap for %u",
3802 ext4_mb_release_desc(&e4b);
3806 ext4_lock_group(sb, group);
3807 list_del(&pa->pa_group_list);
3808 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
3809 ext4_unlock_group(sb, group);
3811 ext4_mb_release_desc(&e4b);
3814 list_del(&pa->u.pa_tmp_list);
3815 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3818 kmem_cache_free(ext4_ac_cachep, ac);
3822 * finds all preallocated spaces and return blocks being freed to them
3823 * if preallocated space becomes full (no block is used from the space)
3824 * then the function frees space in buddy
3825 * XXX: at the moment, truncate (which is the only way to free blocks)
3826 * discards all preallocations
3828 static void ext4_mb_return_to_preallocation(struct inode *inode,
3829 struct ext4_buddy *e4b,
3830 sector_t block, int count)
3832 BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
3834 #ifdef CONFIG_EXT4_DEBUG
3835 static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3837 struct super_block *sb = ac->ac_sb;
3838 ext4_group_t ngroups, i;
3840 printk(KERN_ERR "EXT4-fs: Can't allocate:"
3841 " Allocation context details:\n");
3842 printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
3843 ac->ac_status, ac->ac_flags);
3844 printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
3845 "best %lu/%lu/%lu@%lu cr %d\n",
3846 (unsigned long)ac->ac_o_ex.fe_group,
3847 (unsigned long)ac->ac_o_ex.fe_start,
3848 (unsigned long)ac->ac_o_ex.fe_len,
3849 (unsigned long)ac->ac_o_ex.fe_logical,
3850 (unsigned long)ac->ac_g_ex.fe_group,
3851 (unsigned long)ac->ac_g_ex.fe_start,
3852 (unsigned long)ac->ac_g_ex.fe_len,
3853 (unsigned long)ac->ac_g_ex.fe_logical,
3854 (unsigned long)ac->ac_b_ex.fe_group,
3855 (unsigned long)ac->ac_b_ex.fe_start,
3856 (unsigned long)ac->ac_b_ex.fe_len,
3857 (unsigned long)ac->ac_b_ex.fe_logical,
3858 (int)ac->ac_criteria);
3859 printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
3861 printk(KERN_ERR "EXT4-fs: groups: \n");
3862 ngroups = ext4_get_groups_count(sb);
3863 for (i = 0; i < ngroups; i++) {
3864 struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3865 struct ext4_prealloc_space *pa;
3866 ext4_grpblk_t start;
3867 struct list_head *cur;
3868 ext4_lock_group(sb, i);
3869 list_for_each(cur, &grp->bb_prealloc_list) {
3870 pa = list_entry(cur, struct ext4_prealloc_space,
3872 spin_lock(&pa->pa_lock);
3873 ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3875 spin_unlock(&pa->pa_lock);
3876 printk(KERN_ERR "PA:%u:%d:%u \n", i,
3879 ext4_unlock_group(sb, i);
3881 if (grp->bb_free == 0)
3883 printk(KERN_ERR "%u: %d/%d \n",
3884 i, grp->bb_free, grp->bb_fragments);
3886 printk(KERN_ERR "\n");
3889 static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3896 * We use locality group preallocation for small size file. The size of the
3897 * file is determined by the current size or the resulting size after
3898 * allocation which ever is larger
3900 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3902 static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3904 struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3905 int bsbits = ac->ac_sb->s_blocksize_bits;
3908 if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3911 if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3914 size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
3915 isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3918 if ((size == isize) &&
3919 !ext4_fs_is_busy(sbi) &&
3920 (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3921 ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3925 /* don't use group allocation for large files */
3926 size = max(size, isize);
3927 if (size > sbi->s_mb_stream_request) {
3928 ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3932 BUG_ON(ac->ac_lg != NULL);
3934 * locality group prealloc space are per cpu. The reason for having
3935 * per cpu locality group is to reduce the contention between block
3936 * request from multiple CPUs.
3938 ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3940 /* we're going to use group allocation */
3941 ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3943 /* serialize all allocations in the group */
3944 mutex_lock(&ac->ac_lg->lg_mutex);
3947 static noinline_for_stack int
3948 ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3949 struct ext4_allocation_request *ar)
3951 struct super_block *sb = ar->inode->i_sb;
3952 struct ext4_sb_info *sbi = EXT4_SB(sb);
3953 struct ext4_super_block *es = sbi->s_es;
3957 ext4_grpblk_t block;
3959 /* we can't allocate > group size */
3962 /* just a dirty hack to filter too big requests */
3963 if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
3964 len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
3966 /* start searching from the goal */
3968 if (goal < le32_to_cpu(es->s_first_data_block) ||
3969 goal >= ext4_blocks_count(es))
3970 goal = le32_to_cpu(es->s_first_data_block);
3971 ext4_get_group_no_and_offset(sb, goal, &group, &block);
3973 /* set up allocation goals */
3974 memset(ac, 0, sizeof(struct ext4_allocation_context));
3975 ac->ac_b_ex.fe_logical = ar->logical;
3976 ac->ac_status = AC_STATUS_CONTINUE;
3978 ac->ac_inode = ar->inode;
3979 ac->ac_o_ex.fe_logical = ar->logical;
3980 ac->ac_o_ex.fe_group = group;
3981 ac->ac_o_ex.fe_start = block;
3982 ac->ac_o_ex.fe_len = len;
3983 ac->ac_g_ex.fe_logical = ar->logical;
3984 ac->ac_g_ex.fe_group = group;
3985 ac->ac_g_ex.fe_start = block;
3986 ac->ac_g_ex.fe_len = len;
3987 ac->ac_flags = ar->flags;
3989 /* we have to define context: we'll we work with a file or
3990 * locality group. this is a policy, actually */
3991 ext4_mb_group_or_file(ac);
3993 mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
3994 "left: %u/%u, right %u/%u to %swritable\n",
3995 (unsigned) ar->len, (unsigned) ar->logical,
3996 (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
3997 (unsigned) ar->lleft, (unsigned) ar->pleft,
3998 (unsigned) ar->lright, (unsigned) ar->pright,
3999 atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4004 static noinline_for_stack void
4005 ext4_mb_discard_lg_preallocations(struct super_block *sb,
4006 struct ext4_locality_group *lg,
4007 int order, int total_entries)
4009 ext4_group_t group = 0;
4010 struct ext4_buddy e4b;
4011 struct list_head discard_list;
4012 struct ext4_prealloc_space *pa, *tmp;
4013 struct ext4_allocation_context *ac;
4015 mb_debug(1, "discard locality group preallocation\n");
4017 INIT_LIST_HEAD(&discard_list);
4018 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4022 spin_lock(&lg->lg_prealloc_lock);
4023 list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4025 spin_lock(&pa->pa_lock);
4026 if (atomic_read(&pa->pa_count)) {
4028 * This is the pa that we just used
4029 * for block allocation. So don't
4032 spin_unlock(&pa->pa_lock);
4035 if (pa->pa_deleted) {
4036 spin_unlock(&pa->pa_lock);
4039 /* only lg prealloc space */
4040 BUG_ON(pa->pa_type != MB_GROUP_PA);
4042 /* seems this one can be freed ... */
4044 spin_unlock(&pa->pa_lock);
4046 list_del_rcu(&pa->pa_inode_list);
4047 list_add(&pa->u.pa_tmp_list, &discard_list);
4050 if (total_entries <= 5) {
4052 * we want to keep only 5 entries
4053 * allowing it to grow to 8. This
4054 * mak sure we don't call discard
4055 * soon for this list.
4060 spin_unlock(&lg->lg_prealloc_lock);
4062 list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4064 ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4065 if (ext4_mb_load_buddy(sb, group, &e4b)) {
4066 ext4_error(sb, "Error loading buddy information for %u",
4070 ext4_lock_group(sb, group);
4071 list_del(&pa->pa_group_list);
4072 ext4_mb_release_group_pa(&e4b, pa, ac);
4073 ext4_unlock_group(sb, group);
4075 ext4_mb_release_desc(&e4b);
4076 list_del(&pa->u.pa_tmp_list);
4077 call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4080 kmem_cache_free(ext4_ac_cachep, ac);
4084 * We have incremented pa_count. So it cannot be freed at this
4085 * point. Also we hold lg_mutex. So no parallel allocation is
4086 * possible from this lg. That means pa_free cannot be updated.
4088 * A parallel ext4_mb_discard_group_preallocations is possible.
4089 * which can cause the lg_prealloc_list to be updated.
4092 static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4094 int order, added = 0, lg_prealloc_count = 1;
4095 struct super_block *sb = ac->ac_sb;
4096 struct ext4_locality_group *lg = ac->ac_lg;
4097 struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4099 order = fls(pa->pa_free) - 1;
4100 if (order > PREALLOC_TB_SIZE - 1)
4101 /* The max size of hash table is PREALLOC_TB_SIZE */
4102 order = PREALLOC_TB_SIZE - 1;
4103 /* Add the prealloc space to lg */
4105 list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4107 spin_lock(&tmp_pa->pa_lock);
4108 if (tmp_pa->pa_deleted) {
4109 spin_unlock(&tmp_pa->pa_lock);
4112 if (!added && pa->pa_free < tmp_pa->pa_free) {
4113 /* Add to the tail of the previous entry */
4114 list_add_tail_rcu(&pa->pa_inode_list,
4115 &tmp_pa->pa_inode_list);
4118 * we want to count the total
4119 * number of entries in the list
4122 spin_unlock(&tmp_pa->pa_lock);
4123 lg_prealloc_count++;
4126 list_add_tail_rcu(&pa->pa_inode_list,
4127 &lg->lg_prealloc_list[order]);
4130 /* Now trim the list to be not more than 8 elements */
4131 if (lg_prealloc_count > 8) {
4132 ext4_mb_discard_lg_preallocations(sb, lg,
4133 order, lg_prealloc_count);
4140 * release all resource we used in allocation
4142 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4144 struct ext4_prealloc_space *pa = ac->ac_pa;
4146 if (pa->pa_type == MB_GROUP_PA) {
4147 /* see comment in ext4_mb_use_group_pa() */
4148 spin_lock(&pa->pa_lock);
4149 pa->pa_pstart += ac->ac_b_ex.fe_len;
4150 pa->pa_lstart += ac->ac_b_ex.fe_len;
4151 pa->pa_free -= ac->ac_b_ex.fe_len;
4152 pa->pa_len -= ac->ac_b_ex.fe_len;
4153 spin_unlock(&pa->pa_lock);
4157 up_read(ac->alloc_semp);
4160 * We want to add the pa to the right bucket.
4161 * Remove it from the list and while adding
4162 * make sure the list to which we are adding
4163 * doesn't grow big. We need to release
4164 * alloc_semp before calling ext4_mb_add_n_trim()
4166 if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4167 spin_lock(pa->pa_obj_lock);
4168 list_del_rcu(&pa->pa_inode_list);
4169 spin_unlock(pa->pa_obj_lock);
4170 ext4_mb_add_n_trim(ac);
4172 ext4_mb_put_pa(ac, ac->ac_sb, pa);
4174 if (ac->ac_bitmap_page)
4175 page_cache_release(ac->ac_bitmap_page);
4176 if (ac->ac_buddy_page)
4177 page_cache_release(ac->ac_buddy_page);
4178 if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4179 mutex_unlock(&ac->ac_lg->lg_mutex);
4180 ext4_mb_collect_stats(ac);
4184 static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4186 ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4190 trace_ext4_mb_discard_preallocations(sb, needed);
4191 for (i = 0; i < ngroups && needed > 0; i++) {
4192 ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4201 * Main entry point into mballoc to allocate blocks
4202 * it tries to use preallocation first, then falls back
4203 * to usual allocation
4205 ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4206 struct ext4_allocation_request *ar, int *errp)
4209 struct ext4_allocation_context *ac = NULL;
4210 struct ext4_sb_info *sbi;
4211 struct super_block *sb;
4212 ext4_fsblk_t block = 0;
4213 unsigned int inquota = 0;
4214 unsigned int reserv_blks = 0;
4216 sb = ar->inode->i_sb;
4219 trace_ext4_request_blocks(ar);
4222 * For delayed allocation, we could skip the ENOSPC and
4223 * EDQUOT check, as blocks and quotas have been already
4224 * reserved when data being copied into pagecache.
4226 if (EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4227 ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4229 /* Without delayed allocation we need to verify
4230 * there is enough free blocks to do block allocation
4231 * and verify allocation doesn't exceed the quota limits.
4233 while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) {
4234 /* let others to free the space */
4236 ar->len = ar->len >> 1;
4242 reserv_blks = ar->len;
4243 while (ar->len && dquot_alloc_block(ar->inode, ar->len)) {
4244 ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4254 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4261 *errp = ext4_mb_initialize_context(ac, ar);
4267 ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4268 if (!ext4_mb_use_preallocated(ac)) {
4269 ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4270 ext4_mb_normalize_request(ac, ar);
4272 /* allocate space in core */
4273 ext4_mb_regular_allocator(ac);
4275 /* as we've just preallocated more space than
4276 * user requested orinally, we store allocated
4277 * space in a special descriptor */
4278 if (ac->ac_status == AC_STATUS_FOUND &&
4279 ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4280 ext4_mb_new_preallocation(ac);
4282 if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4283 *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks);
4284 if (*errp == -EAGAIN) {
4286 * drop the reference that we took
4287 * in ext4_mb_use_best_found
4289 ext4_mb_release_context(ac);
4290 ac->ac_b_ex.fe_group = 0;
4291 ac->ac_b_ex.fe_start = 0;
4292 ac->ac_b_ex.fe_len = 0;
4293 ac->ac_status = AC_STATUS_CONTINUE;
4296 ext4_discard_allocated_blocks(ac);
4297 ac->ac_b_ex.fe_len = 0;
4299 ext4_mb_show_ac(ac);
4301 block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4302 ar->len = ac->ac_b_ex.fe_len;
4305 freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4309 ac->ac_b_ex.fe_len = 0;
4311 ext4_mb_show_ac(ac);
4314 ext4_mb_release_context(ac);
4317 kmem_cache_free(ext4_ac_cachep, ac);
4319 if (inquota && ar->len < inquota)
4320 dquot_free_block(ar->inode, inquota - ar->len);
4323 if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag)
4324 /* release all the reserved blocks if non delalloc */
4325 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
4329 trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4335 * We can merge two free data extents only if the physical blocks
4336 * are contiguous, AND the extents were freed by the same transaction,
4337 * AND the blocks are associated with the same group.
4339 static int can_merge(struct ext4_free_data *entry1,
4340 struct ext4_free_data *entry2)
4342 if ((entry1->t_tid == entry2->t_tid) &&
4343 (entry1->group == entry2->group) &&
4344 ((entry1->start_blk + entry1->count) == entry2->start_blk))
4349 static noinline_for_stack int
4350 ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4351 struct ext4_free_data *new_entry)
4353 ext4_grpblk_t block;
4354 struct ext4_free_data *entry;
4355 struct ext4_group_info *db = e4b->bd_info;
4356 struct super_block *sb = e4b->bd_sb;
4357 struct ext4_sb_info *sbi = EXT4_SB(sb);
4358 struct rb_node **n = &db->bb_free_root.rb_node, *node;
4359 struct rb_node *parent = NULL, *new_node;
4361 BUG_ON(!ext4_handle_valid(handle));
4362 BUG_ON(e4b->bd_bitmap_page == NULL);
4363 BUG_ON(e4b->bd_buddy_page == NULL);
4365 new_node = &new_entry->node;
4366 block = new_entry->start_blk;
4369 /* first free block exent. We need to
4370 protect buddy cache from being freed,
4371 * otherwise we'll refresh it from
4372 * on-disk bitmap and lose not-yet-available
4374 page_cache_get(e4b->bd_buddy_page);
4375 page_cache_get(e4b->bd_bitmap_page);
4379 entry = rb_entry(parent, struct ext4_free_data, node);
4380 if (block < entry->start_blk)
4382 else if (block >= (entry->start_blk + entry->count))
4383 n = &(*n)->rb_right;
4385 ext4_grp_locked_error(sb, e4b->bd_group, __func__,
4386 "Double free of blocks %d (%d %d)",
4387 block, entry->start_blk, entry->count);
4392 rb_link_node(new_node, parent, n);
4393 rb_insert_color(new_node, &db->bb_free_root);
4395 /* Now try to see the extent can be merged to left and right */
4396 node = rb_prev(new_node);
4398 entry = rb_entry(node, struct ext4_free_data, node);
4399 if (can_merge(entry, new_entry)) {
4400 new_entry->start_blk = entry->start_blk;
4401 new_entry->count += entry->count;
4402 rb_erase(node, &(db->bb_free_root));
4403 spin_lock(&sbi->s_md_lock);
4404 list_del(&entry->list);
4405 spin_unlock(&sbi->s_md_lock);
4406 kmem_cache_free(ext4_free_ext_cachep, entry);
4410 node = rb_next(new_node);
4412 entry = rb_entry(node, struct ext4_free_data, node);
4413 if (can_merge(new_entry, entry)) {
4414 new_entry->count += entry->count;
4415 rb_erase(node, &(db->bb_free_root));
4416 spin_lock(&sbi->s_md_lock);
4417 list_del(&entry->list);
4418 spin_unlock(&sbi->s_md_lock);
4419 kmem_cache_free(ext4_free_ext_cachep, entry);
4422 /* Add the extent to transaction's private list */
4423 spin_lock(&sbi->s_md_lock);
4424 list_add(&new_entry->list, &handle->h_transaction->t_private_list);
4425 spin_unlock(&sbi->s_md_lock);
4430 * ext4_free_blocks() -- Free given blocks and update quota
4431 * @handle: handle for this transaction
4433 * @block: start physical block to free
4434 * @count: number of blocks to count
4435 * @metadata: Are these metadata blocks
4437 void ext4_free_blocks(handle_t *handle, struct inode *inode,
4438 struct buffer_head *bh, ext4_fsblk_t block,
4439 unsigned long count, int flags)
4441 struct buffer_head *bitmap_bh = NULL;
4442 struct super_block *sb = inode->i_sb;
4443 struct ext4_allocation_context *ac = NULL;
4444 struct ext4_group_desc *gdp;
4445 struct ext4_super_block *es;
4446 unsigned long freed = 0;
4447 unsigned int overflow;
4449 struct buffer_head *gd_bh;
4450 ext4_group_t block_group;
4451 struct ext4_sb_info *sbi;
4452 struct ext4_buddy e4b;
4458 BUG_ON(block != bh->b_blocknr);
4460 block = bh->b_blocknr;
4464 es = EXT4_SB(sb)->s_es;
4465 if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4466 !ext4_data_block_valid(sbi, block, count)) {
4467 ext4_error(sb, "Freeing blocks not in datazone - "
4468 "block = %llu, count = %lu", block, count);
4472 ext4_debug("freeing block %llu\n", block);
4473 trace_ext4_free_blocks(inode, block, count, flags);
4475 if (flags & EXT4_FREE_BLOCKS_FORGET) {
4476 struct buffer_head *tbh = bh;
4479 BUG_ON(bh && (count > 1));
4481 for (i = 0; i < count; i++) {
4483 tbh = sb_find_get_block(inode->i_sb,
4485 ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4486 inode, tbh, block + i);
4491 * We need to make sure we don't reuse the freed block until
4492 * after the transaction is committed, which we can do by
4493 * treating the block as metadata, below. We make an
4494 * exception if the inode is to be written in writeback mode
4495 * since writeback mode has weak data consistency guarantees.
4497 if (!ext4_should_writeback_data(inode))
4498 flags |= EXT4_FREE_BLOCKS_METADATA;
4500 ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4502 ac->ac_inode = inode;
4508 ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4511 * Check to see if we are freeing blocks across a group
4514 if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4515 overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
4518 bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4523 gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4529 if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4530 in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4531 in_range(block, ext4_inode_table(sb, gdp),
4532 EXT4_SB(sb)->s_itb_per_group) ||
4533 in_range(block + count - 1, ext4_inode_table(sb, gdp),
4534 EXT4_SB(sb)->s_itb_per_group)) {
4536 ext4_error(sb, "Freeing blocks in system zone - "
4537 "Block = %llu, count = %lu", block, count);
4538 /* err = 0. ext4_std_error should be a no op */
4542 BUFFER_TRACE(bitmap_bh, "getting write access");
4543 err = ext4_journal_get_write_access(handle, bitmap_bh);
4548 * We are about to modify some metadata. Call the journal APIs
4549 * to unshare ->b_data if a currently-committing transaction is
4552 BUFFER_TRACE(gd_bh, "get_write_access");
4553 err = ext4_journal_get_write_access(handle, gd_bh);
4556 #ifdef AGGRESSIVE_CHECK
4559 for (i = 0; i < count; i++)
4560 BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4564 ac->ac_b_ex.fe_group = block_group;
4565 ac->ac_b_ex.fe_start = bit;
4566 ac->ac_b_ex.fe_len = count;
4567 trace_ext4_mballoc_free(ac);
4570 err = ext4_mb_load_buddy(sb, block_group, &e4b);
4574 if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4575 struct ext4_free_data *new_entry;
4577 * blocks being freed are metadata. these blocks shouldn't
4578 * be used until this transaction is committed
4580 new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS);
4581 new_entry->start_blk = bit;
4582 new_entry->group = block_group;
4583 new_entry->count = count;
4584 new_entry->t_tid = handle->h_transaction->t_tid;
4586 ext4_lock_group(sb, block_group);
4587 mb_clear_bits(bitmap_bh->b_data, bit, count);
4588 ext4_mb_free_metadata(handle, &e4b, new_entry);
4590 /* need to update group_info->bb_free and bitmap
4591 * with group lock held. generate_buddy look at
4592 * them with group lock_held
4594 ext4_lock_group(sb, block_group);
4595 mb_clear_bits(bitmap_bh->b_data, bit, count);
4596 mb_free_blocks(inode, &e4b, bit, count);
4597 ext4_mb_return_to_preallocation(inode, &e4b, block, count);
4600 ret = ext4_free_blks_count(sb, gdp) + count;
4601 ext4_free_blks_set(sb, gdp, ret);
4602 gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
4603 ext4_unlock_group(sb, block_group);
4604 percpu_counter_add(&sbi->s_freeblocks_counter, count);
4606 if (sbi->s_log_groups_per_flex) {
4607 ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4608 atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks);
4611 ext4_mb_release_desc(&e4b);
4615 /* We dirtied the bitmap block */
4616 BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4617 err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4619 /* And the group descriptor block */
4620 BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4621 ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4625 if (overflow && !err) {
4634 dquot_free_block(inode, freed);
4636 ext4_std_error(sb, err);
4638 kmem_cache_free(ext4_ac_cachep, ac);