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
27 * - test ext4_ext_search_left() and ext4_ext_search_right()
28 * - search for metadata in few groups
31 * - normalization should take into account whether file is still open
32 * - discard preallocations if no free space left (policy?)
33 * - don't normalize tails
35 * - reservation for superuser
38 * - bitmap read-ahead (proposed by Oleg Drokin aka green)
39 * - track min/max extents in each group for better group selection
40 * - mb_mark_used() may allocate chunk right after splitting buddy
41 * - tree of groups sorted by number of free blocks
46 * The allocation request involve request for multiple number of blocks
47 * near to the goal(block) value specified.
49 * During initialization phase of the allocator we decide to use the group
50 * preallocation or inode preallocation depending on the size file. The
51 * size of the file could be the resulting file size we would have after
52 * allocation or the current file size which ever is larger. If the size is
53 * less that sbi->s_mb_stream_request we select the group
54 * preallocation. The default value of s_mb_stream_request is 16
55 * blocks. This can also be tuned via
56 * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
57 * of number of blocks.
59 * The main motivation for having small file use group preallocation is to
60 * ensure that we have small file closer in the disk.
62 * First stage the allocator looks at the inode prealloc list
63 * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
64 * this particular inode. The inode prealloc space is represented as:
66 * pa_lstart -> the logical start block for this prealloc space
67 * pa_pstart -> the physical start block for this prealloc space
68 * pa_len -> lenght for this prealloc space
69 * pa_free -> free space available in this prealloc space
71 * The inode preallocation space is used looking at the _logical_ start
72 * block. If only the logical file block falls within the range of prealloc
73 * space we will consume the particular prealloc space. This make sure that
74 * that the we have contiguous physical blocks representing the file blocks
76 * The important thing to be noted in case of inode prealloc space is that
77 * we don't modify the values associated to inode prealloc space except
80 * If we are not able to find blocks in the inode prealloc space and if we
81 * have the group allocation flag set then we look at the locality group
82 * prealloc space. These are per CPU prealloc list repreasented as
84 * ext4_sb_info.s_locality_groups[smp_processor_id()]
86 * The reason for having a per cpu locality group is to reduce the contention
87 * between CPUs. It is possible to get scheduled at this point.
89 * The locality group prealloc space is used looking at whether we have
90 * enough free space (pa_free) withing the prealloc space.
92 * If we can't allocate blocks via inode prealloc or/and locality group
93 * prealloc then we look at the buddy cache. The buddy cache is represented
94 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
95 * mapped to the buddy and bitmap information regarding different
96 * groups. The buddy information is attached to buddy cache inode so that
97 * we can access them through the page cache. The information regarding
98 * each group is loaded via ext4_mb_load_buddy. The information involve
99 * block bitmap and buddy information. The information are stored in the
103 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
106 * one block each for bitmap and buddy information. So for each group we
107 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
108 * blocksize) blocks. So it can have information regarding groups_per_page
109 * which is blocks_per_page/2
111 * The buddy cache inode is not stored on disk. The inode is thrown
112 * away when the filesystem is unmounted.
114 * We look for count number of blocks in the buddy cache. If we were able
115 * to locate that many free blocks we return with additional information
116 * regarding rest of the contiguous physical block available
118 * Before allocating blocks via buddy cache we normalize the request
119 * blocks. This ensure we ask for more blocks that we needed. The extra
120 * blocks that we get after allocation is added to the respective prealloc
121 * list. In case of inode preallocation we follow a list of heuristics
122 * based on file size. This can be found in ext4_mb_normalize_request. If
123 * we are doing a group prealloc we try to normalize the request to
124 * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
125 * 512 blocks. This can be tuned via
126 * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
127 * terms of number of blocks. If we have mounted the file system with -O
128 * stripe=<value> option the group prealloc request is normalized to the
129 * stripe value (sbi->s_stripe)
131 * The regular allocator(using the buddy cache) support few tunables.
133 * /proc/fs/ext4/<partition>/min_to_scan
134 * /proc/fs/ext4/<partition>/max_to_scan
135 * /proc/fs/ext4/<partition>/order2_req
137 * The regular allocator use buddy scan only if the request len is power of
138 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
139 * value of s_mb_order2_reqs can be tuned via
140 * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to
141 * stripe size (sbi->s_stripe), we try to search for contigous block in
142 * stripe size. This should result in better allocation on RAID setup. If
143 * not we search in the specific group using bitmap for best extents. The
144 * tunable min_to_scan and max_to_scan controll the behaviour here.
145 * min_to_scan indicate how long the mballoc __must__ look for a best
146 * extent and max_to_scanindicate how long the mballoc __can__ look for a
147 * best extent in the found extents. Searching for the blocks starts with
148 * the group specified as the goal value in allocation context via
149 * ac_g_ex. Each group is first checked based on the criteria whether it
150 * can used for allocation. ext4_mb_good_group explains how the groups are
153 * Both the prealloc space are getting populated as above. So for the first
154 * request we will hit the buddy cache which will result in this prealloc
155 * space getting filled. The prealloc space is then later used for the
156 * subsequent request.
160 * mballoc operates on the following data:
162 * - in-core buddy (actually includes buddy and bitmap)
163 * - preallocation descriptors (PAs)
165 * there are two types of preallocations:
167 * assiged to specific inode and can be used for this inode only.
168 * it describes part of inode's space preallocated to specific
169 * physical blocks. any block from that preallocated can be used
170 * independent. the descriptor just tracks number of blocks left
171 * unused. so, before taking some block from descriptor, one must
172 * make sure corresponded logical block isn't allocated yet. this
173 * also means that freeing any block within descriptor's range
174 * must discard all preallocated blocks.
176 * assigned to specific locality group which does not translate to
177 * permanent set of inodes: inode can join and leave group. space
178 * from this type of preallocation can be used for any inode. thus
179 * it's consumed from the beginning to the end.
181 * relation between them can be expressed as:
182 * in-core buddy = on-disk bitmap + preallocation descriptors
184 * this mean blocks mballoc considers used are:
185 * - allocated blocks (persistent)
186 * - preallocated blocks (non-persistent)
188 * consistency in mballoc world means that at any time a block is either
189 * free or used in ALL structures. notice: "any time" should not be read
190 * literally -- time is discrete and delimited by locks.
192 * to keep it simple, we don't use block numbers, instead we count number of
193 * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
195 * all operations can be expressed as:
196 * - init buddy: buddy = on-disk + PAs
197 * - new PA: buddy += N; PA = N
198 * - use inode PA: on-disk += N; PA -= N
199 * - discard inode PA buddy -= on-disk - PA; PA = 0
200 * - use locality group PA on-disk += N; PA -= N
201 * - discard locality group PA buddy -= PA; PA = 0
202 * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
203 * is used in real operation because we can't know actual used
204 * bits from PA, only from on-disk bitmap
206 * if we follow this strict logic, then all operations above should be atomic.
207 * given some of them can block, we'd have to use something like semaphores
208 * killing performance on high-end SMP hardware. let's try to relax it using
209 * the following knowledge:
210 * 1) if buddy is referenced, it's already initialized
211 * 2) while block is used in buddy and the buddy is referenced,
212 * nobody can re-allocate that block
213 * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
214 * bit set and PA claims same block, it's OK. IOW, one can set bit in
215 * on-disk bitmap if buddy has same bit set or/and PA covers corresponded
218 * so, now we're building a concurrency table:
221 * blocks for PA are allocated in the buddy, buddy must be referenced
222 * until PA is linked to allocation group to avoid concurrent buddy init
224 * we need to make sure that either on-disk bitmap or PA has uptodate data
225 * given (3) we care that PA-=N operation doesn't interfere with init
227 * the simplest way would be to have buddy initialized by the discard
228 * - use locality group PA
229 * again PA-=N must be serialized with init
230 * - discard locality group PA
231 * the simplest way would be to have buddy initialized by the discard
234 * i_data_sem serializes them
236 * discard process must wait until PA isn't used by another process
237 * - use locality group PA
238 * some mutex should serialize them
239 * - discard locality group PA
240 * discard process must wait until PA isn't used by another process
243 * i_data_sem or another mutex should serializes them
245 * discard process must wait until PA isn't used by another process
246 * - use locality group PA
247 * nothing wrong here -- they're different PAs covering different blocks
248 * - discard locality group PA
249 * discard process must wait until PA isn't used by another process
251 * now we're ready to make few consequences:
252 * - PA is referenced and while it is no discard is possible
253 * - PA is referenced until block isn't marked in on-disk bitmap
254 * - PA changes only after on-disk bitmap
255 * - discard must not compete with init. either init is done before
256 * any discard or they're serialized somehow
257 * - buddy init as sum of on-disk bitmap and PAs is done atomically
259 * a special case when we've used PA to emptiness. no need to modify buddy
260 * in this case, but we should care about concurrent init
265 * Logic in few words:
270 * mark bits in on-disk bitmap
273 * - use preallocation:
274 * find proper PA (per-inode or group)
276 * mark bits in on-disk bitmap
282 * mark bits in on-disk bitmap
285 * - discard preallocations in group:
287 * move them onto local list
288 * load on-disk bitmap
290 * remove PA from object (inode or locality group)
291 * mark free blocks in-core
293 * - discard inode's preallocations:
300 * - bitlock on a group (group)
301 * - object (inode/locality) (object)
312 * - release consumed pa:
317 * - generate in-core bitmap:
321 * - discard all for given object (inode, locality group):
326 * - discard all for given group:
333 static struct kmem_cache
*ext4_pspace_cachep
;
334 static struct kmem_cache
*ext4_ac_cachep
;
335 static struct kmem_cache
*ext4_free_ext_cachep
;
336 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
338 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
340 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
);
341 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
);
342 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
);
346 static inline void *mb_correct_addr_and_bit(int *bit
, void *addr
)
348 #if BITS_PER_LONG == 64
349 *bit
+= ((unsigned long) addr
& 7UL) << 3;
350 addr
= (void *) ((unsigned long) addr
& ~7UL);
351 #elif BITS_PER_LONG == 32
352 *bit
+= ((unsigned long) addr
& 3UL) << 3;
353 addr
= (void *) ((unsigned long) addr
& ~3UL);
355 #error "how many bits you are?!"
360 static inline int mb_test_bit(int bit
, void *addr
)
363 * ext4_test_bit on architecture like powerpc
364 * needs unsigned long aligned address
366 addr
= mb_correct_addr_and_bit(&bit
, addr
);
367 return ext4_test_bit(bit
, addr
);
370 static inline void mb_set_bit(int bit
, void *addr
)
372 addr
= mb_correct_addr_and_bit(&bit
, addr
);
373 ext4_set_bit(bit
, addr
);
376 static inline void mb_set_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
378 addr
= mb_correct_addr_and_bit(&bit
, addr
);
379 ext4_set_bit_atomic(lock
, bit
, addr
);
382 static inline void mb_clear_bit(int bit
, void *addr
)
384 addr
= mb_correct_addr_and_bit(&bit
, addr
);
385 ext4_clear_bit(bit
, addr
);
388 static inline void mb_clear_bit_atomic(spinlock_t
*lock
, int bit
, void *addr
)
390 addr
= mb_correct_addr_and_bit(&bit
, addr
);
391 ext4_clear_bit_atomic(lock
, bit
, addr
);
394 static inline int mb_find_next_zero_bit(void *addr
, int max
, int start
)
396 int fix
= 0, ret
, tmpmax
;
397 addr
= mb_correct_addr_and_bit(&fix
, addr
);
401 ret
= ext4_find_next_zero_bit(addr
, tmpmax
, start
) - fix
;
407 static inline int mb_find_next_bit(void *addr
, int max
, int start
)
409 int fix
= 0, ret
, tmpmax
;
410 addr
= mb_correct_addr_and_bit(&fix
, addr
);
414 ret
= ext4_find_next_bit(addr
, tmpmax
, start
) - fix
;
420 static void *mb_find_buddy(struct ext4_buddy
*e4b
, int order
, int *max
)
424 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
427 if (order
> e4b
->bd_blkbits
+ 1) {
432 /* at order 0 we see each particular block */
433 *max
= 1 << (e4b
->bd_blkbits
+ 3);
435 return EXT4_MB_BITMAP(e4b
);
437 bb
= EXT4_MB_BUDDY(e4b
) + EXT4_SB(e4b
->bd_sb
)->s_mb_offsets
[order
];
438 *max
= EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[order
];
444 static void mb_free_blocks_double(struct inode
*inode
, struct ext4_buddy
*e4b
,
445 int first
, int count
)
448 struct super_block
*sb
= e4b
->bd_sb
;
450 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
452 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
453 for (i
= 0; i
< count
; i
++) {
454 if (!mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
)) {
455 ext4_fsblk_t blocknr
;
456 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
457 blocknr
+= first
+ i
;
459 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
461 ext4_error(sb
, __func__
, "double-free of inode"
462 " %lu's block %llu(bit %u in group %u)",
463 inode
? inode
->i_ino
: 0, blocknr
,
464 first
+ i
, e4b
->bd_group
);
466 mb_clear_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
470 static void mb_mark_used_double(struct ext4_buddy
*e4b
, int first
, int count
)
474 if (unlikely(e4b
->bd_info
->bb_bitmap
== NULL
))
476 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
477 for (i
= 0; i
< count
; i
++) {
478 BUG_ON(mb_test_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
));
479 mb_set_bit(first
+ i
, e4b
->bd_info
->bb_bitmap
);
483 static void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
485 if (memcmp(e4b
->bd_info
->bb_bitmap
, bitmap
, e4b
->bd_sb
->s_blocksize
)) {
486 unsigned char *b1
, *b2
;
488 b1
= (unsigned char *) e4b
->bd_info
->bb_bitmap
;
489 b2
= (unsigned char *) bitmap
;
490 for (i
= 0; i
< e4b
->bd_sb
->s_blocksize
; i
++) {
491 if (b1
[i
] != b2
[i
]) {
492 printk(KERN_ERR
"corruption in group %u "
493 "at byte %u(%u): %x in copy != %x "
494 "on disk/prealloc\n",
495 e4b
->bd_group
, i
, i
* 8, b1
[i
], b2
[i
]);
503 static inline void mb_free_blocks_double(struct inode
*inode
,
504 struct ext4_buddy
*e4b
, int first
, int count
)
508 static inline void mb_mark_used_double(struct ext4_buddy
*e4b
,
509 int first
, int count
)
513 static inline void mb_cmp_bitmaps(struct ext4_buddy
*e4b
, void *bitmap
)
519 #ifdef AGGRESSIVE_CHECK
521 #define MB_CHECK_ASSERT(assert) \
525 "Assertion failure in %s() at %s:%d: \"%s\"\n", \
526 function, file, line, # assert); \
531 static int __mb_check_buddy(struct ext4_buddy
*e4b
, char *file
,
532 const char *function
, int line
)
534 struct super_block
*sb
= e4b
->bd_sb
;
535 int order
= e4b
->bd_blkbits
+ 1;
542 struct ext4_group_info
*grp
;
545 struct list_head
*cur
;
550 static int mb_check_counter
;
551 if (mb_check_counter
++ % 100 != 0)
556 buddy
= mb_find_buddy(e4b
, order
, &max
);
557 MB_CHECK_ASSERT(buddy
);
558 buddy2
= mb_find_buddy(e4b
, order
- 1, &max2
);
559 MB_CHECK_ASSERT(buddy2
);
560 MB_CHECK_ASSERT(buddy
!= buddy2
);
561 MB_CHECK_ASSERT(max
* 2 == max2
);
564 for (i
= 0; i
< max
; i
++) {
566 if (mb_test_bit(i
, buddy
)) {
567 /* only single bit in buddy2 may be 1 */
568 if (!mb_test_bit(i
<< 1, buddy2
)) {
570 mb_test_bit((i
<<1)+1, buddy2
));
571 } else if (!mb_test_bit((i
<< 1) + 1, buddy2
)) {
573 mb_test_bit(i
<< 1, buddy2
));
578 /* both bits in buddy2 must be 0 */
579 MB_CHECK_ASSERT(mb_test_bit(i
<< 1, buddy2
));
580 MB_CHECK_ASSERT(mb_test_bit((i
<< 1) + 1, buddy2
));
582 for (j
= 0; j
< (1 << order
); j
++) {
583 k
= (i
* (1 << order
)) + j
;
585 !mb_test_bit(k
, EXT4_MB_BITMAP(e4b
)));
589 MB_CHECK_ASSERT(e4b
->bd_info
->bb_counters
[order
] == count
);
594 buddy
= mb_find_buddy(e4b
, 0, &max
);
595 for (i
= 0; i
< max
; i
++) {
596 if (!mb_test_bit(i
, buddy
)) {
597 MB_CHECK_ASSERT(i
>= e4b
->bd_info
->bb_first_free
);
605 /* check used bits only */
606 for (j
= 0; j
< e4b
->bd_blkbits
+ 1; j
++) {
607 buddy2
= mb_find_buddy(e4b
, j
, &max2
);
609 MB_CHECK_ASSERT(k
< max2
);
610 MB_CHECK_ASSERT(mb_test_bit(k
, buddy2
));
613 MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b
->bd_info
));
614 MB_CHECK_ASSERT(e4b
->bd_info
->bb_fragments
== fragments
);
616 grp
= ext4_get_group_info(sb
, e4b
->bd_group
);
617 buddy
= mb_find_buddy(e4b
, 0, &max
);
618 list_for_each(cur
, &grp
->bb_prealloc_list
) {
619 ext4_group_t groupnr
;
620 struct ext4_prealloc_space
*pa
;
621 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
622 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &groupnr
, &k
);
623 MB_CHECK_ASSERT(groupnr
== e4b
->bd_group
);
624 for (i
= 0; i
< pa
->pa_len
; i
++)
625 MB_CHECK_ASSERT(mb_test_bit(k
+ i
, buddy
));
629 #undef MB_CHECK_ASSERT
630 #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \
631 __FILE__, __func__, __LINE__)
633 #define mb_check_buddy(e4b)
636 /* FIXME!! need more doc */
637 static void ext4_mb_mark_free_simple(struct super_block
*sb
,
638 void *buddy
, unsigned first
, int len
,
639 struct ext4_group_info
*grp
)
641 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
644 unsigned short chunk
;
645 unsigned short border
;
647 BUG_ON(len
> EXT4_BLOCKS_PER_GROUP(sb
));
649 border
= 2 << sb
->s_blocksize_bits
;
652 /* find how many blocks can be covered since this position */
653 max
= ffs(first
| border
) - 1;
655 /* find how many blocks of power 2 we need to mark */
662 /* mark multiblock chunks only */
663 grp
->bb_counters
[min
]++;
665 mb_clear_bit(first
>> min
,
666 buddy
+ sbi
->s_mb_offsets
[min
]);
673 static void ext4_mb_generate_buddy(struct super_block
*sb
,
674 void *buddy
, void *bitmap
, ext4_group_t group
)
676 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
677 unsigned short max
= EXT4_BLOCKS_PER_GROUP(sb
);
678 unsigned short i
= 0;
679 unsigned short first
;
682 unsigned fragments
= 0;
683 unsigned long long period
= get_cycles();
685 /* initialize buddy from bitmap which is aggregation
686 * of on-disk bitmap and preallocations */
687 i
= mb_find_next_zero_bit(bitmap
, max
, 0);
688 grp
->bb_first_free
= i
;
692 i
= mb_find_next_bit(bitmap
, max
, i
);
696 ext4_mb_mark_free_simple(sb
, buddy
, first
, len
, grp
);
698 grp
->bb_counters
[0]++;
700 i
= mb_find_next_zero_bit(bitmap
, max
, i
);
702 grp
->bb_fragments
= fragments
;
704 if (free
!= grp
->bb_free
) {
705 ext4_error(sb
, __func__
,
706 "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
707 group
, free
, grp
->bb_free
);
709 * If we intent to continue, we consider group descritor
710 * corrupt and update bb_free using bitmap value
715 clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
, &(grp
->bb_state
));
717 period
= get_cycles() - period
;
718 spin_lock(&EXT4_SB(sb
)->s_bal_lock
);
719 EXT4_SB(sb
)->s_mb_buddies_generated
++;
720 EXT4_SB(sb
)->s_mb_generation_time
+= period
;
721 spin_unlock(&EXT4_SB(sb
)->s_bal_lock
);
724 /* The buddy information is attached the buddy cache inode
725 * for convenience. The information regarding each group
726 * is loaded via ext4_mb_load_buddy. The information involve
727 * block bitmap and buddy information. The information are
728 * stored in the inode as
731 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
734 * one block each for bitmap and buddy information.
735 * So for each group we take up 2 blocks. A page can
736 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks.
737 * So it can have information regarding groups_per_page which
738 * is blocks_per_page/2
741 static int ext4_mb_init_cache(struct page
*page
, char *incore
)
748 ext4_group_t first_group
;
750 struct super_block
*sb
;
751 struct buffer_head
*bhs
;
752 struct buffer_head
**bh
;
757 mb_debug("init page %lu\n", page
->index
);
759 inode
= page
->mapping
->host
;
761 blocksize
= 1 << inode
->i_blkbits
;
762 blocks_per_page
= PAGE_CACHE_SIZE
/ blocksize
;
764 groups_per_page
= blocks_per_page
>> 1;
765 if (groups_per_page
== 0)
768 /* allocate buffer_heads to read bitmaps */
769 if (groups_per_page
> 1) {
771 i
= sizeof(struct buffer_head
*) * groups_per_page
;
772 bh
= kzalloc(i
, GFP_NOFS
);
778 first_group
= page
->index
* blocks_per_page
/ 2;
780 /* read all groups the page covers into the cache */
781 for (i
= 0; i
< groups_per_page
; i
++) {
782 struct ext4_group_desc
*desc
;
784 if (first_group
+ i
>= EXT4_SB(sb
)->s_groups_count
)
788 desc
= ext4_get_group_desc(sb
, first_group
+ i
, NULL
);
793 bh
[i
] = sb_getblk(sb
, ext4_block_bitmap(sb
, desc
));
797 if (buffer_uptodate(bh
[i
]) &&
798 !(desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)))
802 spin_lock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
803 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
804 ext4_init_block_bitmap(sb
, bh
[i
],
805 first_group
+ i
, desc
);
806 set_buffer_uptodate(bh
[i
]);
807 unlock_buffer(bh
[i
]);
808 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
811 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
813 bh
[i
]->b_end_io
= end_buffer_read_sync
;
814 submit_bh(READ
, bh
[i
]);
815 mb_debug("read bitmap for group %u\n", first_group
+ i
);
818 /* wait for I/O completion */
819 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
820 wait_on_buffer(bh
[i
]);
823 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
824 if (!buffer_uptodate(bh
[i
]))
828 first_block
= page
->index
* blocks_per_page
;
829 for (i
= 0; i
< blocks_per_page
; i
++) {
831 struct ext4_group_info
*grinfo
;
833 group
= (first_block
+ i
) >> 1;
834 if (group
>= EXT4_SB(sb
)->s_groups_count
)
838 * data carry information regarding this
839 * particular group in the format specified
843 data
= page_address(page
) + (i
* blocksize
);
844 bitmap
= bh
[group
- first_group
]->b_data
;
847 * We place the buddy block and bitmap block
850 if ((first_block
+ i
) & 1) {
851 /* this is block of buddy */
852 BUG_ON(incore
== NULL
);
853 mb_debug("put buddy for group %u in page %lu/%x\n",
854 group
, page
->index
, i
* blocksize
);
855 memset(data
, 0xff, blocksize
);
856 grinfo
= ext4_get_group_info(sb
, group
);
857 grinfo
->bb_fragments
= 0;
858 memset(grinfo
->bb_counters
, 0,
859 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
861 * incore got set to the group block bitmap below
863 ext4_lock_group(sb
, group
);
864 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
865 ext4_unlock_group(sb
, group
);
868 /* this is block of bitmap */
869 BUG_ON(incore
!= NULL
);
870 mb_debug("put bitmap for group %u in page %lu/%x\n",
871 group
, page
->index
, i
* blocksize
);
873 /* see comments in ext4_mb_put_pa() */
874 ext4_lock_group(sb
, group
);
875 memcpy(data
, bitmap
, blocksize
);
877 /* mark all preallocated blks used in in-core bitmap */
878 ext4_mb_generate_from_pa(sb
, data
, group
);
879 ext4_mb_generate_from_freelist(sb
, data
, group
);
880 ext4_unlock_group(sb
, group
);
882 /* set incore so that the buddy information can be
883 * generated using this
888 SetPageUptodate(page
);
892 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
900 static noinline_for_stack
int
901 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
902 struct ext4_buddy
*e4b
)
910 struct ext4_group_info
*grp
;
911 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
912 struct inode
*inode
= sbi
->s_buddy_cache
;
914 mb_debug("load group %u\n", group
);
916 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
917 grp
= ext4_get_group_info(sb
, group
);
919 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
920 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
922 e4b
->bd_group
= group
;
923 e4b
->bd_buddy_page
= NULL
;
924 e4b
->bd_bitmap_page
= NULL
;
925 e4b
->alloc_semp
= &grp
->alloc_sem
;
927 /* Take the read lock on the group alloc
928 * sem. This would make sure a parallel
929 * ext4_mb_init_group happening on other
930 * groups mapped by the page is blocked
931 * till we are done with allocation
933 down_read(e4b
->alloc_semp
);
936 * the buddy cache inode stores the block bitmap
937 * and buddy information in consecutive blocks.
938 * So for each group we need two blocks.
941 pnum
= block
/ blocks_per_page
;
942 poff
= block
% blocks_per_page
;
944 /* we could use find_or_create_page(), but it locks page
945 * what we'd like to avoid in fast path ... */
946 page
= find_get_page(inode
->i_mapping
, pnum
);
947 if (page
== NULL
|| !PageUptodate(page
)) {
950 * drop the page reference and try
951 * to get the page with lock. If we
952 * are not uptodate that implies
953 * somebody just created the page but
954 * is yet to initialize the same. So
955 * wait for it to initialize.
957 page_cache_release(page
);
958 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
960 BUG_ON(page
->mapping
!= inode
->i_mapping
);
961 if (!PageUptodate(page
)) {
962 ret
= ext4_mb_init_cache(page
, NULL
);
967 mb_cmp_bitmaps(e4b
, page_address(page
) +
968 (poff
* sb
->s_blocksize
));
973 if (page
== NULL
|| !PageUptodate(page
)) {
977 e4b
->bd_bitmap_page
= page
;
978 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
979 mark_page_accessed(page
);
982 pnum
= block
/ blocks_per_page
;
983 poff
= block
% blocks_per_page
;
985 page
= find_get_page(inode
->i_mapping
, pnum
);
986 if (page
== NULL
|| !PageUptodate(page
)) {
988 page_cache_release(page
);
989 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
991 BUG_ON(page
->mapping
!= inode
->i_mapping
);
992 if (!PageUptodate(page
)) {
993 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1002 if (page
== NULL
|| !PageUptodate(page
)) {
1006 e4b
->bd_buddy_page
= page
;
1007 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1008 mark_page_accessed(page
);
1010 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1011 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1016 if (e4b
->bd_bitmap_page
)
1017 page_cache_release(e4b
->bd_bitmap_page
);
1018 if (e4b
->bd_buddy_page
)
1019 page_cache_release(e4b
->bd_buddy_page
);
1020 e4b
->bd_buddy
= NULL
;
1021 e4b
->bd_bitmap
= NULL
;
1023 /* Done with the buddy cache */
1024 up_read(e4b
->alloc_semp
);
1028 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
1030 if (e4b
->bd_bitmap_page
)
1031 page_cache_release(e4b
->bd_bitmap_page
);
1032 if (e4b
->bd_buddy_page
)
1033 page_cache_release(e4b
->bd_buddy_page
);
1034 /* Done with the buddy cache */
1035 up_read(e4b
->alloc_semp
);
1039 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1044 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1045 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1047 bb
= EXT4_MB_BUDDY(e4b
);
1048 while (order
<= e4b
->bd_blkbits
+ 1) {
1050 if (!mb_test_bit(block
, bb
)) {
1051 /* this block is part of buddy of order 'order' */
1054 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1060 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1066 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1067 /* fast path: clear whole word at once */
1068 addr
= bm
+ (cur
>> 3);
1073 mb_clear_bit_atomic(lock
, cur
, bm
);
1078 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1084 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1085 /* fast path: set whole word at once */
1086 addr
= bm
+ (cur
>> 3);
1091 mb_set_bit_atomic(lock
, cur
, bm
);
1096 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1097 int first
, int count
)
1106 struct super_block
*sb
= e4b
->bd_sb
;
1108 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1109 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1110 mb_check_buddy(e4b
);
1111 mb_free_blocks_double(inode
, e4b
, first
, count
);
1113 e4b
->bd_info
->bb_free
+= count
;
1114 if (first
< e4b
->bd_info
->bb_first_free
)
1115 e4b
->bd_info
->bb_first_free
= first
;
1117 /* let's maintain fragments counter */
1119 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1120 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1121 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1123 e4b
->bd_info
->bb_fragments
--;
1124 else if (!block
&& !max
)
1125 e4b
->bd_info
->bb_fragments
++;
1127 /* let's maintain buddy itself */
1128 while (count
-- > 0) {
1132 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1133 ext4_fsblk_t blocknr
;
1134 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1137 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1138 ext4_unlock_group(sb
, e4b
->bd_group
);
1139 ext4_error(sb
, __func__
, "double-free of inode"
1140 " %lu's block %llu(bit %u in group %u)",
1141 inode
? inode
->i_ino
: 0, blocknr
, block
,
1143 ext4_lock_group(sb
, e4b
->bd_group
);
1145 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1146 e4b
->bd_info
->bb_counters
[order
]++;
1148 /* start of the buddy */
1149 buddy
= mb_find_buddy(e4b
, order
, &max
);
1153 if (mb_test_bit(block
, buddy
) ||
1154 mb_test_bit(block
+ 1, buddy
))
1157 /* both the buddies are free, try to coalesce them */
1158 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1164 /* for special purposes, we don't set
1165 * free bits in bitmap */
1166 mb_set_bit(block
, buddy
);
1167 mb_set_bit(block
+ 1, buddy
);
1169 e4b
->bd_info
->bb_counters
[order
]--;
1170 e4b
->bd_info
->bb_counters
[order
]--;
1174 e4b
->bd_info
->bb_counters
[order
]++;
1176 mb_clear_bit(block
, buddy2
);
1180 mb_check_buddy(e4b
);
1183 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1184 int needed
, struct ext4_free_extent
*ex
)
1191 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1194 buddy
= mb_find_buddy(e4b
, order
, &max
);
1195 BUG_ON(buddy
== NULL
);
1196 BUG_ON(block
>= max
);
1197 if (mb_test_bit(block
, buddy
)) {
1204 /* FIXME dorp order completely ? */
1205 if (likely(order
== 0)) {
1206 /* find actual order */
1207 order
= mb_find_order_for_block(e4b
, block
);
1208 block
= block
>> order
;
1211 ex
->fe_len
= 1 << order
;
1212 ex
->fe_start
= block
<< order
;
1213 ex
->fe_group
= e4b
->bd_group
;
1215 /* calc difference from given start */
1216 next
= next
- ex
->fe_start
;
1218 ex
->fe_start
+= next
;
1220 while (needed
> ex
->fe_len
&&
1221 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1223 if (block
+ 1 >= max
)
1226 next
= (block
+ 1) * (1 << order
);
1227 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1230 ord
= mb_find_order_for_block(e4b
, next
);
1233 block
= next
>> order
;
1234 ex
->fe_len
+= 1 << order
;
1237 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1241 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1247 int start
= ex
->fe_start
;
1248 int len
= ex
->fe_len
;
1253 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1254 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1255 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1256 mb_check_buddy(e4b
);
1257 mb_mark_used_double(e4b
, start
, len
);
1259 e4b
->bd_info
->bb_free
-= len
;
1260 if (e4b
->bd_info
->bb_first_free
== start
)
1261 e4b
->bd_info
->bb_first_free
+= len
;
1263 /* let's maintain fragments counter */
1265 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1266 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1267 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1269 e4b
->bd_info
->bb_fragments
++;
1270 else if (!mlen
&& !max
)
1271 e4b
->bd_info
->bb_fragments
--;
1273 /* let's maintain buddy itself */
1275 ord
= mb_find_order_for_block(e4b
, start
);
1277 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1278 /* the whole chunk may be allocated at once! */
1280 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1281 BUG_ON((start
>> ord
) >= max
);
1282 mb_set_bit(start
>> ord
, buddy
);
1283 e4b
->bd_info
->bb_counters
[ord
]--;
1290 /* store for history */
1292 ret
= len
| (ord
<< 16);
1294 /* we have to split large buddy */
1296 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1297 mb_set_bit(start
>> ord
, buddy
);
1298 e4b
->bd_info
->bb_counters
[ord
]--;
1301 cur
= (start
>> ord
) & ~1U;
1302 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1303 mb_clear_bit(cur
, buddy
);
1304 mb_clear_bit(cur
+ 1, buddy
);
1305 e4b
->bd_info
->bb_counters
[ord
]++;
1306 e4b
->bd_info
->bb_counters
[ord
]++;
1309 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1310 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1311 mb_check_buddy(e4b
);
1317 * Must be called under group lock!
1319 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1320 struct ext4_buddy
*e4b
)
1322 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1325 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1326 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1328 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1329 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1330 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1332 /* preallocation can change ac_b_ex, thus we store actually
1333 * allocated blocks for history */
1334 ac
->ac_f_ex
= ac
->ac_b_ex
;
1336 ac
->ac_status
= AC_STATUS_FOUND
;
1337 ac
->ac_tail
= ret
& 0xffff;
1338 ac
->ac_buddy
= ret
>> 16;
1341 * take the page reference. We want the page to be pinned
1342 * so that we don't get a ext4_mb_init_cache_call for this
1343 * group until we update the bitmap. That would mean we
1344 * double allocate blocks. The reference is dropped
1345 * in ext4_mb_release_context
1347 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1348 get_page(ac
->ac_bitmap_page
);
1349 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1350 get_page(ac
->ac_buddy_page
);
1352 /* store last allocated for subsequent stream allocation */
1353 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1354 spin_lock(&sbi
->s_md_lock
);
1355 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1356 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1357 spin_unlock(&sbi
->s_md_lock
);
1362 * regular allocator, for general purposes allocation
1365 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1366 struct ext4_buddy
*e4b
,
1369 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1370 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1371 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1372 struct ext4_free_extent ex
;
1375 if (ac
->ac_status
== AC_STATUS_FOUND
)
1378 * We don't want to scan for a whole year
1380 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1381 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1382 ac
->ac_status
= AC_STATUS_BREAK
;
1387 * Haven't found good chunk so far, let's continue
1389 if (bex
->fe_len
< gex
->fe_len
)
1392 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1393 && bex
->fe_group
== e4b
->bd_group
) {
1394 /* recheck chunk's availability - we don't know
1395 * when it was found (within this lock-unlock
1397 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1398 if (max
>= gex
->fe_len
) {
1399 ext4_mb_use_best_found(ac
, e4b
);
1406 * The routine checks whether found extent is good enough. If it is,
1407 * then the extent gets marked used and flag is set to the context
1408 * to stop scanning. Otherwise, the extent is compared with the
1409 * previous found extent and if new one is better, then it's stored
1410 * in the context. Later, the best found extent will be used, if
1411 * mballoc can't find good enough extent.
1413 * FIXME: real allocation policy is to be designed yet!
1415 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1416 struct ext4_free_extent
*ex
,
1417 struct ext4_buddy
*e4b
)
1419 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1420 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1422 BUG_ON(ex
->fe_len
<= 0);
1423 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1424 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1425 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1430 * The special case - take what you catch first
1432 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1434 ext4_mb_use_best_found(ac
, e4b
);
1439 * Let's check whether the chuck is good enough
1441 if (ex
->fe_len
== gex
->fe_len
) {
1443 ext4_mb_use_best_found(ac
, e4b
);
1448 * If this is first found extent, just store it in the context
1450 if (bex
->fe_len
== 0) {
1456 * If new found extent is better, store it in the context
1458 if (bex
->fe_len
< gex
->fe_len
) {
1459 /* if the request isn't satisfied, any found extent
1460 * larger than previous best one is better */
1461 if (ex
->fe_len
> bex
->fe_len
)
1463 } else if (ex
->fe_len
> gex
->fe_len
) {
1464 /* if the request is satisfied, then we try to find
1465 * an extent that still satisfy the request, but is
1466 * smaller than previous one */
1467 if (ex
->fe_len
< bex
->fe_len
)
1471 ext4_mb_check_limits(ac
, e4b
, 0);
1474 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1475 struct ext4_buddy
*e4b
)
1477 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1478 ext4_group_t group
= ex
.fe_group
;
1482 BUG_ON(ex
.fe_len
<= 0);
1483 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1487 ext4_lock_group(ac
->ac_sb
, group
);
1488 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1492 ext4_mb_use_best_found(ac
, e4b
);
1495 ext4_unlock_group(ac
->ac_sb
, group
);
1496 ext4_mb_release_desc(e4b
);
1501 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1502 struct ext4_buddy
*e4b
)
1504 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1507 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1508 struct ext4_super_block
*es
= sbi
->s_es
;
1509 struct ext4_free_extent ex
;
1511 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1514 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1518 ext4_lock_group(ac
->ac_sb
, group
);
1519 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1520 ac
->ac_g_ex
.fe_len
, &ex
);
1522 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1525 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1526 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1527 /* use do_div to get remainder (would be 64-bit modulo) */
1528 if (do_div(start
, sbi
->s_stripe
) == 0) {
1531 ext4_mb_use_best_found(ac
, e4b
);
1533 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1534 BUG_ON(ex
.fe_len
<= 0);
1535 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1536 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1539 ext4_mb_use_best_found(ac
, e4b
);
1540 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1541 /* Sometimes, caller may want to merge even small
1542 * number of blocks to an existing extent */
1543 BUG_ON(ex
.fe_len
<= 0);
1544 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1545 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1548 ext4_mb_use_best_found(ac
, e4b
);
1550 ext4_unlock_group(ac
->ac_sb
, group
);
1551 ext4_mb_release_desc(e4b
);
1557 * The routine scans buddy structures (not bitmap!) from given order
1558 * to max order and tries to find big enough chunk to satisfy the req
1560 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1561 struct ext4_buddy
*e4b
)
1563 struct super_block
*sb
= ac
->ac_sb
;
1564 struct ext4_group_info
*grp
= e4b
->bd_info
;
1570 BUG_ON(ac
->ac_2order
<= 0);
1571 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1572 if (grp
->bb_counters
[i
] == 0)
1575 buddy
= mb_find_buddy(e4b
, i
, &max
);
1576 BUG_ON(buddy
== NULL
);
1578 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1583 ac
->ac_b_ex
.fe_len
= 1 << i
;
1584 ac
->ac_b_ex
.fe_start
= k
<< i
;
1585 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1587 ext4_mb_use_best_found(ac
, e4b
);
1589 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1591 if (EXT4_SB(sb
)->s_mb_stats
)
1592 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1599 * The routine scans the group and measures all found extents.
1600 * In order to optimize scanning, caller must pass number of
1601 * free blocks in the group, so the routine can know upper limit.
1603 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1604 struct ext4_buddy
*e4b
)
1606 struct super_block
*sb
= ac
->ac_sb
;
1607 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1608 struct ext4_free_extent ex
;
1612 free
= e4b
->bd_info
->bb_free
;
1615 i
= e4b
->bd_info
->bb_first_free
;
1617 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1618 i
= mb_find_next_zero_bit(bitmap
,
1619 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1620 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1622 * IF we have corrupt bitmap, we won't find any
1623 * free blocks even though group info says we
1624 * we have free blocks
1626 ext4_error(sb
, __func__
, "%d free blocks as per "
1627 "group info. But bitmap says 0",
1632 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1633 BUG_ON(ex
.fe_len
<= 0);
1634 if (free
< ex
.fe_len
) {
1635 ext4_error(sb
, __func__
, "%d free blocks as per "
1636 "group info. But got %d blocks",
1639 * The number of free blocks differs. This mostly
1640 * indicate that the bitmap is corrupt. So exit
1641 * without claiming the space.
1646 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1652 ext4_mb_check_limits(ac
, e4b
, 1);
1656 * This is a special case for storages like raid5
1657 * we try to find stripe-aligned chunks for stripe-size requests
1658 * XXX should do so at least for multiples of stripe size as well
1660 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1661 struct ext4_buddy
*e4b
)
1663 struct super_block
*sb
= ac
->ac_sb
;
1664 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1665 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1666 struct ext4_free_extent ex
;
1667 ext4_fsblk_t first_group_block
;
1672 BUG_ON(sbi
->s_stripe
== 0);
1674 /* find first stripe-aligned block in group */
1675 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1676 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1677 a
= first_group_block
+ sbi
->s_stripe
- 1;
1678 do_div(a
, sbi
->s_stripe
);
1679 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1681 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1682 if (!mb_test_bit(i
, bitmap
)) {
1683 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1684 if (max
>= sbi
->s_stripe
) {
1687 ext4_mb_use_best_found(ac
, e4b
);
1695 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1696 ext4_group_t group
, int cr
)
1698 unsigned free
, fragments
;
1700 struct ext4_group_desc
*desc
;
1701 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1703 BUG_ON(cr
< 0 || cr
>= 4);
1704 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1706 free
= grp
->bb_free
;
1707 fragments
= grp
->bb_fragments
;
1715 BUG_ON(ac
->ac_2order
== 0);
1716 /* If this group is uninitialized, skip it initially */
1717 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1718 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1721 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1722 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1723 if (grp
->bb_counters
[i
] > 0)
1727 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1731 if (free
>= ac
->ac_g_ex
.fe_len
)
1744 * lock the group_info alloc_sem of all the groups
1745 * belonging to the same buddy cache page. This
1746 * make sure other parallel operation on the buddy
1747 * cache doesn't happen whild holding the buddy cache
1750 int ext4_mb_get_buddy_cache_lock(struct super_block
*sb
, ext4_group_t group
)
1754 int blocks_per_page
;
1755 int groups_per_page
;
1756 ext4_group_t first_group
;
1757 struct ext4_group_info
*grp
;
1759 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1761 * the buddy cache inode stores the block bitmap
1762 * and buddy information in consecutive blocks.
1763 * So for each group we need two blocks.
1766 pnum
= block
/ blocks_per_page
;
1767 first_group
= pnum
* blocks_per_page
/ 2;
1769 groups_per_page
= blocks_per_page
>> 1;
1770 if (groups_per_page
== 0)
1771 groups_per_page
= 1;
1772 /* read all groups the page covers into the cache */
1773 for (i
= 0; i
< groups_per_page
; i
++) {
1775 if ((first_group
+ i
) >= EXT4_SB(sb
)->s_groups_count
)
1777 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1778 /* take all groups write allocation
1779 * semaphore. This make sure there is
1780 * no block allocation going on in any
1783 down_write_nested(&grp
->alloc_sem
, i
);
1788 void ext4_mb_put_buddy_cache_lock(struct super_block
*sb
,
1789 ext4_group_t group
, int locked_group
)
1793 int blocks_per_page
;
1794 ext4_group_t first_group
;
1795 struct ext4_group_info
*grp
;
1797 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1799 * the buddy cache inode stores the block bitmap
1800 * and buddy information in consecutive blocks.
1801 * So for each group we need two blocks.
1804 pnum
= block
/ blocks_per_page
;
1805 first_group
= pnum
* blocks_per_page
/ 2;
1806 /* release locks on all the groups */
1807 for (i
= 0; i
< locked_group
; i
++) {
1809 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1810 /* take all groups write allocation
1811 * semaphore. This make sure there is
1812 * no block allocation going on in any
1815 up_write(&grp
->alloc_sem
);
1820 static int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1825 int blocks_per_page
;
1826 int block
, pnum
, poff
;
1827 int num_grp_locked
= 0;
1828 struct ext4_group_info
*this_grp
;
1829 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1830 struct inode
*inode
= sbi
->s_buddy_cache
;
1831 struct page
*page
= NULL
, *bitmap_page
= NULL
;
1833 mb_debug("init group %lu\n", group
);
1834 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1835 this_grp
= ext4_get_group_info(sb
, group
);
1837 * This ensures we don't add group
1838 * to this buddy cache via resize
1840 num_grp_locked
= ext4_mb_get_buddy_cache_lock(sb
, group
);
1841 if (!EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1843 * somebody initialized the group
1844 * return without doing anything
1850 * the buddy cache inode stores the block bitmap
1851 * and buddy information in consecutive blocks.
1852 * So for each group we need two blocks.
1855 pnum
= block
/ blocks_per_page
;
1856 poff
= block
% blocks_per_page
;
1857 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1859 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1860 ret
= ext4_mb_init_cache(page
, NULL
);
1867 if (page
== NULL
|| !PageUptodate(page
)) {
1871 mark_page_accessed(page
);
1873 bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1875 /* init buddy cache */
1877 pnum
= block
/ blocks_per_page
;
1878 poff
= block
% blocks_per_page
;
1879 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1880 if (page
== bitmap_page
) {
1882 * If both the bitmap and buddy are in
1883 * the same page we don't need to force
1888 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1889 ret
= ext4_mb_init_cache(page
, bitmap
);
1896 if (page
== NULL
|| !PageUptodate(page
)) {
1900 mark_page_accessed(page
);
1902 ext4_mb_put_buddy_cache_lock(sb
, group
, num_grp_locked
);
1904 page_cache_release(bitmap_page
);
1906 page_cache_release(page
);
1910 static noinline_for_stack
int
1911 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1918 struct ext4_sb_info
*sbi
;
1919 struct super_block
*sb
;
1920 struct ext4_buddy e4b
;
1925 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1927 /* first, try the goal */
1928 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1929 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1932 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1936 * ac->ac2_order is set only if the fe_len is a power of 2
1937 * if ac2_order is set we also set criteria to 0 so that we
1938 * try exact allocation using buddy.
1940 i
= fls(ac
->ac_g_ex
.fe_len
);
1943 * We search using buddy data only if the order of the request
1944 * is greater than equal to the sbi_s_mb_order2_reqs
1945 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1947 if (i
>= sbi
->s_mb_order2_reqs
) {
1949 * This should tell if fe_len is exactly power of 2
1951 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1952 ac
->ac_2order
= i
- 1;
1955 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
1956 /* if stream allocation is enabled, use global goal */
1957 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
1958 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
1962 if (size
< sbi
->s_mb_stream_request
&&
1963 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1964 /* TBD: may be hot point */
1965 spin_lock(&sbi
->s_md_lock
);
1966 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1967 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1968 spin_unlock(&sbi
->s_md_lock
);
1970 /* Let's just scan groups to find more-less suitable blocks */
1971 cr
= ac
->ac_2order
? 0 : 1;
1973 * cr == 0 try to get exact allocation,
1974 * cr == 3 try to get anything
1977 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
1978 ac
->ac_criteria
= cr
;
1980 * searching for the right group start
1981 * from the goal value specified
1983 group
= ac
->ac_g_ex
.fe_group
;
1985 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
1986 struct ext4_group_info
*grp
;
1987 struct ext4_group_desc
*desc
;
1989 if (group
== EXT4_SB(sb
)->s_groups_count
)
1992 /* quick check to skip empty groups */
1993 grp
= ext4_get_group_info(sb
, group
);
1994 if (grp
->bb_free
== 0)
1998 * if the group is already init we check whether it is
1999 * a good group and if not we don't load the buddy
2001 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
2003 * we need full data about the group
2004 * to make a good selection
2006 err
= ext4_mb_init_group(sb
, group
);
2012 * If the particular group doesn't satisfy our
2013 * criteria we continue with the next group
2015 if (!ext4_mb_good_group(ac
, group
, cr
))
2018 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2022 ext4_lock_group(sb
, group
);
2023 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2024 /* someone did allocation from this group */
2025 ext4_unlock_group(sb
, group
);
2026 ext4_mb_release_desc(&e4b
);
2030 ac
->ac_groups_scanned
++;
2031 desc
= ext4_get_group_desc(sb
, group
, NULL
);
2032 if (cr
== 0 || (desc
->bg_flags
&
2033 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
2034 ac
->ac_2order
!= 0))
2035 ext4_mb_simple_scan_group(ac
, &e4b
);
2037 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
2038 ext4_mb_scan_aligned(ac
, &e4b
);
2040 ext4_mb_complex_scan_group(ac
, &e4b
);
2042 ext4_unlock_group(sb
, group
);
2043 ext4_mb_release_desc(&e4b
);
2045 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2050 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2051 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2053 * We've been searching too long. Let's try to allocate
2054 * the best chunk we've found so far
2057 ext4_mb_try_best_found(ac
, &e4b
);
2058 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2060 * Someone more lucky has already allocated it.
2061 * The only thing we can do is just take first
2063 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2065 ac
->ac_b_ex
.fe_group
= 0;
2066 ac
->ac_b_ex
.fe_start
= 0;
2067 ac
->ac_b_ex
.fe_len
= 0;
2068 ac
->ac_status
= AC_STATUS_CONTINUE
;
2069 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2071 atomic_inc(&sbi
->s_mb_lost_chunks
);
2079 #ifdef EXT4_MB_HISTORY
2080 struct ext4_mb_proc_session
{
2081 struct ext4_mb_history
*history
;
2082 struct super_block
*sb
;
2087 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
2088 struct ext4_mb_history
*hs
,
2091 if (hs
== s
->history
+ s
->max
)
2093 if (!first
&& hs
== s
->history
+ s
->start
)
2095 while (hs
->orig
.fe_len
== 0) {
2097 if (hs
== s
->history
+ s
->max
)
2099 if (hs
== s
->history
+ s
->start
)
2105 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
2107 struct ext4_mb_proc_session
*s
= seq
->private;
2108 struct ext4_mb_history
*hs
;
2112 return SEQ_START_TOKEN
;
2113 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2116 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
2120 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
2123 struct ext4_mb_proc_session
*s
= seq
->private;
2124 struct ext4_mb_history
*hs
= v
;
2127 if (v
== SEQ_START_TOKEN
)
2128 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2130 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
2133 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
2135 char buf
[25], buf2
[25], buf3
[25], *fmt
;
2136 struct ext4_mb_history
*hs
= v
;
2138 if (v
== SEQ_START_TOKEN
) {
2139 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
2140 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2141 "pid", "inode", "original", "goal", "result", "found",
2142 "grps", "cr", "flags", "merge", "tail", "broken");
2146 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
2147 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2148 "%-5u %-5s %-5u %-6u\n";
2149 sprintf(buf2
, "%u/%d/%u@%u", hs
->result
.fe_group
,
2150 hs
->result
.fe_start
, hs
->result
.fe_len
,
2151 hs
->result
.fe_logical
);
2152 sprintf(buf
, "%u/%d/%u@%u", hs
->orig
.fe_group
,
2153 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2154 hs
->orig
.fe_logical
);
2155 sprintf(buf3
, "%u/%d/%u@%u", hs
->goal
.fe_group
,
2156 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
2157 hs
->goal
.fe_logical
);
2158 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
2159 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
2160 hs
->merged
? "M" : "", hs
->tail
,
2161 hs
->buddy
? 1 << hs
->buddy
: 0);
2162 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
2163 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
2164 sprintf(buf2
, "%u/%d/%u@%u", hs
->result
.fe_group
,
2165 hs
->result
.fe_start
, hs
->result
.fe_len
,
2166 hs
->result
.fe_logical
);
2167 sprintf(buf
, "%u/%d/%u@%u", hs
->orig
.fe_group
,
2168 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2169 hs
->orig
.fe_logical
);
2170 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
2171 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
2172 sprintf(buf2
, "%u/%d/%u", hs
->result
.fe_group
,
2173 hs
->result
.fe_start
, hs
->result
.fe_len
);
2174 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
2175 hs
->pid
, hs
->ino
, buf2
);
2176 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
2177 sprintf(buf2
, "%u/%d/%u", hs
->result
.fe_group
,
2178 hs
->result
.fe_start
, hs
->result
.fe_len
);
2179 seq_printf(seq
, "%-5u %-8u %-23s free\n",
2180 hs
->pid
, hs
->ino
, buf2
);
2185 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
2189 static struct seq_operations ext4_mb_seq_history_ops
= {
2190 .start
= ext4_mb_seq_history_start
,
2191 .next
= ext4_mb_seq_history_next
,
2192 .stop
= ext4_mb_seq_history_stop
,
2193 .show
= ext4_mb_seq_history_show
,
2196 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
2198 struct super_block
*sb
= PDE(inode
)->data
;
2199 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2200 struct ext4_mb_proc_session
*s
;
2204 if (unlikely(sbi
->s_mb_history
== NULL
))
2206 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
2210 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
2211 s
->history
= kmalloc(size
, GFP_KERNEL
);
2212 if (s
->history
== NULL
) {
2217 spin_lock(&sbi
->s_mb_history_lock
);
2218 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2219 s
->max
= sbi
->s_mb_history_max
;
2220 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2221 spin_unlock(&sbi
->s_mb_history_lock
);
2223 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2225 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2235 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2237 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2238 struct ext4_mb_proc_session
*s
= seq
->private;
2241 return seq_release(inode
, file
);
2244 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2245 const char __user
*buffer
,
2246 size_t count
, loff_t
*ppos
)
2248 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2249 struct ext4_mb_proc_session
*s
= seq
->private;
2250 struct super_block
*sb
= s
->sb
;
2254 if (count
>= sizeof(str
)) {
2255 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2256 "mb_history", (int)sizeof(str
));
2260 if (copy_from_user(str
, buffer
, count
))
2263 value
= simple_strtol(str
, NULL
, 0);
2266 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2271 static struct file_operations ext4_mb_seq_history_fops
= {
2272 .owner
= THIS_MODULE
,
2273 .open
= ext4_mb_seq_history_open
,
2275 .write
= ext4_mb_seq_history_write
,
2276 .llseek
= seq_lseek
,
2277 .release
= ext4_mb_seq_history_release
,
2280 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2282 struct super_block
*sb
= seq
->private;
2283 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2286 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2290 return (void *) ((unsigned long) group
);
2293 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2295 struct super_block
*sb
= seq
->private;
2296 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2300 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2303 return (void *) ((unsigned long) group
);
2306 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2308 struct super_block
*sb
= seq
->private;
2309 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2312 struct ext4_buddy e4b
;
2314 struct ext4_group_info info
;
2315 unsigned short counters
[16];
2320 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2321 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2322 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2323 "group", "free", "frags", "first",
2324 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2325 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2327 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2328 sizeof(struct ext4_group_info
);
2329 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2331 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2334 ext4_lock_group(sb
, group
);
2335 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2336 ext4_unlock_group(sb
, group
);
2337 ext4_mb_release_desc(&e4b
);
2339 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2340 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2341 for (i
= 0; i
<= 13; i
++)
2342 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2343 sg
.info
.bb_counters
[i
] : 0);
2344 seq_printf(seq
, " ]\n");
2349 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2353 static struct seq_operations ext4_mb_seq_groups_ops
= {
2354 .start
= ext4_mb_seq_groups_start
,
2355 .next
= ext4_mb_seq_groups_next
,
2356 .stop
= ext4_mb_seq_groups_stop
,
2357 .show
= ext4_mb_seq_groups_show
,
2360 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2362 struct super_block
*sb
= PDE(inode
)->data
;
2365 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2367 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2374 static struct file_operations ext4_mb_seq_groups_fops
= {
2375 .owner
= THIS_MODULE
,
2376 .open
= ext4_mb_seq_groups_open
,
2378 .llseek
= seq_lseek
,
2379 .release
= seq_release
,
2382 static void ext4_mb_history_release(struct super_block
*sb
)
2384 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2386 if (sbi
->s_proc
!= NULL
) {
2387 remove_proc_entry("mb_groups", sbi
->s_proc
);
2388 remove_proc_entry("mb_history", sbi
->s_proc
);
2390 kfree(sbi
->s_mb_history
);
2393 static void ext4_mb_history_init(struct super_block
*sb
)
2395 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2398 if (sbi
->s_proc
!= NULL
) {
2399 proc_create_data("mb_history", S_IRUGO
, sbi
->s_proc
,
2400 &ext4_mb_seq_history_fops
, sb
);
2401 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2402 &ext4_mb_seq_groups_fops
, sb
);
2405 sbi
->s_mb_history_max
= 1000;
2406 sbi
->s_mb_history_cur
= 0;
2407 spin_lock_init(&sbi
->s_mb_history_lock
);
2408 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2409 sbi
->s_mb_history
= kzalloc(i
, GFP_KERNEL
);
2410 /* if we can't allocate history, then we simple won't use it */
2413 static noinline_for_stack
void
2414 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2416 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2417 struct ext4_mb_history h
;
2419 if (unlikely(sbi
->s_mb_history
== NULL
))
2422 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2426 h
.pid
= current
->pid
;
2427 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2428 h
.orig
= ac
->ac_o_ex
;
2429 h
.result
= ac
->ac_b_ex
;
2430 h
.flags
= ac
->ac_flags
;
2431 h
.found
= ac
->ac_found
;
2432 h
.groups
= ac
->ac_groups_scanned
;
2433 h
.cr
= ac
->ac_criteria
;
2434 h
.tail
= ac
->ac_tail
;
2435 h
.buddy
= ac
->ac_buddy
;
2437 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2438 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2439 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2441 h
.goal
= ac
->ac_g_ex
;
2442 h
.result
= ac
->ac_f_ex
;
2445 spin_lock(&sbi
->s_mb_history_lock
);
2446 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2447 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2448 sbi
->s_mb_history_cur
= 0;
2449 spin_unlock(&sbi
->s_mb_history_lock
);
2453 #define ext4_mb_history_release(sb)
2454 #define ext4_mb_history_init(sb)
2458 /* Create and initialize ext4_group_info data for the given group. */
2459 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2460 struct ext4_group_desc
*desc
)
2464 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2465 struct ext4_group_info
**meta_group_info
;
2468 * First check if this group is the first of a reserved block.
2469 * If it's true, we have to allocate a new table of pointers
2470 * to ext4_group_info structures
2472 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2473 metalen
= sizeof(*meta_group_info
) <<
2474 EXT4_DESC_PER_BLOCK_BITS(sb
);
2475 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2476 if (meta_group_info
== NULL
) {
2477 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2479 goto exit_meta_group_info
;
2481 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2486 * calculate needed size. if change bb_counters size,
2487 * don't forget about ext4_mb_generate_buddy()
2489 len
= offsetof(typeof(**meta_group_info
),
2490 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2493 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2494 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2496 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2497 if (meta_group_info
[i
] == NULL
) {
2498 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2499 goto exit_group_info
;
2501 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2502 &(meta_group_info
[i
]->bb_state
));
2505 * initialize bb_free to be able to skip
2506 * empty groups without initialization
2508 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2509 meta_group_info
[i
]->bb_free
=
2510 ext4_free_blocks_after_init(sb
, group
, desc
);
2512 meta_group_info
[i
]->bb_free
=
2513 le16_to_cpu(desc
->bg_free_blocks_count
);
2516 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2517 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2518 meta_group_info
[i
]->bb_free_root
.rb_node
= NULL
;;
2522 struct buffer_head
*bh
;
2523 meta_group_info
[i
]->bb_bitmap
=
2524 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2525 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2526 bh
= ext4_read_block_bitmap(sb
, group
);
2528 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2537 /* If a meta_group_info table has been allocated, release it now */
2538 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2539 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2540 exit_meta_group_info
:
2542 } /* ext4_mb_add_groupinfo */
2545 * Update an existing group.
2546 * This function is used for online resize
2548 void ext4_mb_update_group_info(struct ext4_group_info
*grp
, ext4_grpblk_t add
)
2550 grp
->bb_free
+= add
;
2553 static int ext4_mb_init_backend(struct super_block
*sb
)
2557 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2558 struct ext4_super_block
*es
= sbi
->s_es
;
2559 int num_meta_group_infos
;
2560 int num_meta_group_infos_max
;
2562 struct ext4_group_info
**meta_group_info
;
2563 struct ext4_group_desc
*desc
;
2565 /* This is the number of blocks used by GDT */
2566 num_meta_group_infos
= (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) -
2567 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2570 * This is the total number of blocks used by GDT including
2571 * the number of reserved blocks for GDT.
2572 * The s_group_info array is allocated with this value
2573 * to allow a clean online resize without a complex
2574 * manipulation of pointer.
2575 * The drawback is the unused memory when no resize
2576 * occurs but it's very low in terms of pages
2577 * (see comments below)
2578 * Need to handle this properly when META_BG resizing is allowed
2580 num_meta_group_infos_max
= num_meta_group_infos
+
2581 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2584 * array_size is the size of s_group_info array. We round it
2585 * to the next power of two because this approximation is done
2586 * internally by kmalloc so we can have some more memory
2587 * for free here (e.g. may be used for META_BG resize).
2590 while (array_size
< sizeof(*sbi
->s_group_info
) *
2591 num_meta_group_infos_max
)
2592 array_size
= array_size
<< 1;
2593 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2594 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2595 * So a two level scheme suffices for now. */
2596 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2597 if (sbi
->s_group_info
== NULL
) {
2598 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2601 sbi
->s_buddy_cache
= new_inode(sb
);
2602 if (sbi
->s_buddy_cache
== NULL
) {
2603 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2606 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2608 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2609 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2610 if ((i
+ 1) == num_meta_group_infos
)
2611 metalen
= sizeof(*meta_group_info
) *
2612 (sbi
->s_groups_count
-
2613 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2614 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2615 if (meta_group_info
== NULL
) {
2616 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2620 sbi
->s_group_info
[i
] = meta_group_info
;
2623 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2624 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2627 "EXT4-fs: can't read descriptor %u\n", i
);
2630 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2638 kfree(ext4_get_group_info(sb
, i
));
2639 i
= num_meta_group_infos
;
2642 kfree(sbi
->s_group_info
[i
]);
2643 iput(sbi
->s_buddy_cache
);
2645 kfree(sbi
->s_group_info
);
2649 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2651 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2657 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2659 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2660 if (sbi
->s_mb_offsets
== NULL
) {
2664 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned int);
2665 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2666 if (sbi
->s_mb_maxs
== NULL
) {
2667 kfree(sbi
->s_mb_maxs
);
2671 /* order 0 is regular bitmap */
2672 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2673 sbi
->s_mb_offsets
[0] = 0;
2677 max
= sb
->s_blocksize
<< 2;
2679 sbi
->s_mb_offsets
[i
] = offset
;
2680 sbi
->s_mb_maxs
[i
] = max
;
2681 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2684 } while (i
<= sb
->s_blocksize_bits
+ 1);
2686 /* init file for buddy data */
2687 ret
= ext4_mb_init_backend(sb
);
2689 kfree(sbi
->s_mb_offsets
);
2690 kfree(sbi
->s_mb_maxs
);
2694 spin_lock_init(&sbi
->s_md_lock
);
2695 spin_lock_init(&sbi
->s_bal_lock
);
2697 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2698 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2699 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2700 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2701 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2702 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2703 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2705 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2706 if (sbi
->s_locality_groups
== NULL
) {
2707 kfree(sbi
->s_mb_offsets
);
2708 kfree(sbi
->s_mb_maxs
);
2711 for_each_possible_cpu(i
) {
2712 struct ext4_locality_group
*lg
;
2713 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2714 mutex_init(&lg
->lg_mutex
);
2715 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2716 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2717 spin_lock_init(&lg
->lg_prealloc_lock
);
2720 ext4_mb_init_per_dev_proc(sb
);
2721 ext4_mb_history_init(sb
);
2724 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2726 printk(KERN_INFO
"EXT4-fs: mballoc enabled\n");
2730 /* need to called with ext4 group lock (ext4_lock_group) */
2731 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2733 struct ext4_prealloc_space
*pa
;
2734 struct list_head
*cur
, *tmp
;
2737 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2738 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2739 list_del(&pa
->pa_group_list
);
2741 kmem_cache_free(ext4_pspace_cachep
, pa
);
2744 mb_debug("mballoc: %u PAs left\n", count
);
2748 int ext4_mb_release(struct super_block
*sb
)
2751 int num_meta_group_infos
;
2752 struct ext4_group_info
*grinfo
;
2753 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2755 if (sbi
->s_group_info
) {
2756 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2757 grinfo
= ext4_get_group_info(sb
, i
);
2759 kfree(grinfo
->bb_bitmap
);
2761 ext4_lock_group(sb
, i
);
2762 ext4_mb_cleanup_pa(grinfo
);
2763 ext4_unlock_group(sb
, i
);
2766 num_meta_group_infos
= (sbi
->s_groups_count
+
2767 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2768 EXT4_DESC_PER_BLOCK_BITS(sb
);
2769 for (i
= 0; i
< num_meta_group_infos
; i
++)
2770 kfree(sbi
->s_group_info
[i
]);
2771 kfree(sbi
->s_group_info
);
2773 kfree(sbi
->s_mb_offsets
);
2774 kfree(sbi
->s_mb_maxs
);
2775 if (sbi
->s_buddy_cache
)
2776 iput(sbi
->s_buddy_cache
);
2777 if (sbi
->s_mb_stats
) {
2779 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2780 atomic_read(&sbi
->s_bal_allocated
),
2781 atomic_read(&sbi
->s_bal_reqs
),
2782 atomic_read(&sbi
->s_bal_success
));
2784 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2785 "%u 2^N hits, %u breaks, %u lost\n",
2786 atomic_read(&sbi
->s_bal_ex_scanned
),
2787 atomic_read(&sbi
->s_bal_goals
),
2788 atomic_read(&sbi
->s_bal_2orders
),
2789 atomic_read(&sbi
->s_bal_breaks
),
2790 atomic_read(&sbi
->s_mb_lost_chunks
));
2792 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2793 sbi
->s_mb_buddies_generated
++,
2794 sbi
->s_mb_generation_time
);
2796 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2797 atomic_read(&sbi
->s_mb_preallocated
),
2798 atomic_read(&sbi
->s_mb_discarded
));
2801 free_percpu(sbi
->s_locality_groups
);
2802 ext4_mb_history_release(sb
);
2803 ext4_mb_destroy_per_dev_proc(sb
);
2809 * This function is called by the jbd2 layer once the commit has finished,
2810 * so we know we can free the blocks that were released with that commit.
2812 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2814 struct super_block
*sb
= journal
->j_private
;
2815 struct ext4_buddy e4b
;
2816 struct ext4_group_info
*db
;
2817 int err
, count
= 0, count2
= 0;
2818 struct ext4_free_data
*entry
;
2819 ext4_fsblk_t discard_block
;
2820 struct list_head
*l
, *ltmp
;
2822 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2823 entry
= list_entry(l
, struct ext4_free_data
, list
);
2825 mb_debug("gonna free %u blocks in group %u (0x%p):",
2826 entry
->count
, entry
->group
, entry
);
2828 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2829 /* we expect to find existing buddy because it's pinned */
2833 /* there are blocks to put in buddy to make them really free */
2834 count
+= entry
->count
;
2836 ext4_lock_group(sb
, entry
->group
);
2837 /* Take it out of per group rb tree */
2838 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2839 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2841 if (!db
->bb_free_root
.rb_node
) {
2842 /* No more items in the per group rb tree
2843 * balance refcounts from ext4_mb_free_metadata()
2845 page_cache_release(e4b
.bd_buddy_page
);
2846 page_cache_release(e4b
.bd_bitmap_page
);
2848 ext4_unlock_group(sb
, entry
->group
);
2849 discard_block
= (ext4_fsblk_t
) entry
->group
* EXT4_BLOCKS_PER_GROUP(sb
)
2851 + le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
2852 trace_mark(ext4_discard_blocks
, "dev %s blk %llu count %u", sb
->s_id
,
2853 (unsigned long long) discard_block
, entry
->count
);
2854 sb_issue_discard(sb
, discard_block
, entry
->count
);
2856 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2857 ext4_mb_release_desc(&e4b
);
2860 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2863 #define EXT4_MB_STATS_NAME "stats"
2864 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2865 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2866 #define EXT4_MB_ORDER2_REQ "order2_req"
2867 #define EXT4_MB_STREAM_REQ "stream_req"
2868 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2870 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2872 #ifdef CONFIG_PROC_FS
2873 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2874 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2875 struct proc_dir_entry
*proc
;
2877 if (sbi
->s_proc
== NULL
)
2880 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME
, mb_stats
);
2881 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, mb_max_to_scan
);
2882 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, mb_min_to_scan
);
2883 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, mb_order2_reqs
);
2884 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ
, mb_stream_request
);
2885 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, mb_group_prealloc
);
2889 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2890 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2891 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2892 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2893 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2894 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2901 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2903 #ifdef CONFIG_PROC_FS
2904 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2906 if (sbi
->s_proc
== NULL
)
2909 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2910 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2911 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2912 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2913 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2914 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2919 int __init
init_ext4_mballoc(void)
2921 ext4_pspace_cachep
=
2922 kmem_cache_create("ext4_prealloc_space",
2923 sizeof(struct ext4_prealloc_space
),
2924 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2925 if (ext4_pspace_cachep
== NULL
)
2929 kmem_cache_create("ext4_alloc_context",
2930 sizeof(struct ext4_allocation_context
),
2931 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2932 if (ext4_ac_cachep
== NULL
) {
2933 kmem_cache_destroy(ext4_pspace_cachep
);
2937 ext4_free_ext_cachep
=
2938 kmem_cache_create("ext4_free_block_extents",
2939 sizeof(struct ext4_free_data
),
2940 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2941 if (ext4_free_ext_cachep
== NULL
) {
2942 kmem_cache_destroy(ext4_pspace_cachep
);
2943 kmem_cache_destroy(ext4_ac_cachep
);
2949 void exit_ext4_mballoc(void)
2951 /* XXX: synchronize_rcu(); */
2952 kmem_cache_destroy(ext4_pspace_cachep
);
2953 kmem_cache_destroy(ext4_ac_cachep
);
2954 kmem_cache_destroy(ext4_free_ext_cachep
);
2959 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2960 * Returns 0 if success or error code
2962 static noinline_for_stack
int
2963 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2964 handle_t
*handle
, unsigned int reserv_blks
)
2966 struct buffer_head
*bitmap_bh
= NULL
;
2967 struct ext4_super_block
*es
;
2968 struct ext4_group_desc
*gdp
;
2969 struct buffer_head
*gdp_bh
;
2970 struct ext4_sb_info
*sbi
;
2971 struct super_block
*sb
;
2975 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
2976 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
2984 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
2988 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
2993 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
2997 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
2998 gdp
->bg_free_blocks_count
);
3000 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
3004 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
3005 + ac
->ac_b_ex
.fe_start
3006 + le32_to_cpu(es
->s_first_data_block
);
3008 len
= ac
->ac_b_ex
.fe_len
;
3009 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, len
) ||
3010 in_range(ext4_inode_bitmap(sb
, gdp
), block
, len
) ||
3011 in_range(block
, ext4_inode_table(sb
, gdp
),
3012 EXT4_SB(sb
)->s_itb_per_group
) ||
3013 in_range(block
+ len
- 1, ext4_inode_table(sb
, gdp
),
3014 EXT4_SB(sb
)->s_itb_per_group
)) {
3015 ext4_error(sb
, __func__
,
3016 "Allocating block in system zone - block = %llu",
3018 /* File system mounted not to panic on error
3019 * Fix the bitmap and repeat the block allocation
3020 * We leak some of the blocks here.
3022 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
),
3023 bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3024 ac
->ac_b_ex
.fe_len
);
3025 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3030 #ifdef AGGRESSIVE_CHECK
3033 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3034 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3035 bitmap_bh
->b_data
));
3039 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
), bitmap_bh
->b_data
,
3040 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
3042 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3043 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
3044 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3045 gdp
->bg_free_blocks_count
=
3046 cpu_to_le16(ext4_free_blocks_after_init(sb
,
3047 ac
->ac_b_ex
.fe_group
,
3050 le16_add_cpu(&gdp
->bg_free_blocks_count
, -ac
->ac_b_ex
.fe_len
);
3051 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
3052 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3053 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
3055 * Now reduce the dirty block count also. Should not go negative
3057 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
3058 /* release all the reserved blocks if non delalloc */
3059 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
3061 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
3062 ac
->ac_b_ex
.fe_len
);
3064 if (sbi
->s_log_groups_per_flex
) {
3065 ext4_group_t flex_group
= ext4_flex_group(sbi
,
3066 ac
->ac_b_ex
.fe_group
);
3067 spin_lock(sb_bgl_lock(sbi
, flex_group
));
3068 sbi
->s_flex_groups
[flex_group
].free_blocks
-= ac
->ac_b_ex
.fe_len
;
3069 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
3072 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3075 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3084 * here we normalize request for locality group
3085 * Group request are normalized to s_strip size if we set the same via mount
3086 * option. If not we set it to s_mb_group_prealloc which can be configured via
3087 * /proc/fs/ext4/<partition>/group_prealloc
3089 * XXX: should we try to preallocate more than the group has now?
3091 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3093 struct super_block
*sb
= ac
->ac_sb
;
3094 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3097 if (EXT4_SB(sb
)->s_stripe
)
3098 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
3100 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3101 mb_debug("#%u: goal %u blocks for locality group\n",
3102 current
->pid
, ac
->ac_g_ex
.fe_len
);
3106 * Normalization means making request better in terms of
3107 * size and alignment
3109 static noinline_for_stack
void
3110 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3111 struct ext4_allocation_request
*ar
)
3115 loff_t size
, orig_size
, start_off
;
3116 ext4_lblk_t start
, orig_start
;
3117 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3118 struct ext4_prealloc_space
*pa
;
3120 /* do normalize only data requests, metadata requests
3121 do not need preallocation */
3122 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3125 /* sometime caller may want exact blocks */
3126 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3129 /* caller may indicate that preallocation isn't
3130 * required (it's a tail, for example) */
3131 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3134 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3135 ext4_mb_normalize_group_request(ac
);
3139 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3141 /* first, let's learn actual file size
3142 * given current request is allocated */
3143 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3144 size
= size
<< bsbits
;
3145 if (size
< i_size_read(ac
->ac_inode
))
3146 size
= i_size_read(ac
->ac_inode
);
3148 /* max size of free chunks */
3151 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3152 (req <= (size) || max <= (chunk_size))
3154 /* first, try to predict filesize */
3155 /* XXX: should this table be tunable? */
3157 if (size
<= 16 * 1024) {
3159 } else if (size
<= 32 * 1024) {
3161 } else if (size
<= 64 * 1024) {
3163 } else if (size
<= 128 * 1024) {
3165 } else if (size
<= 256 * 1024) {
3167 } else if (size
<= 512 * 1024) {
3169 } else if (size
<= 1024 * 1024) {
3171 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3172 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3173 (21 - bsbits
)) << 21;
3174 size
= 2 * 1024 * 1024;
3175 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3176 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3177 (22 - bsbits
)) << 22;
3178 size
= 4 * 1024 * 1024;
3179 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3180 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3181 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3182 (23 - bsbits
)) << 23;
3183 size
= 8 * 1024 * 1024;
3185 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3186 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3188 orig_size
= size
= size
>> bsbits
;
3189 orig_start
= start
= start_off
>> bsbits
;
3191 /* don't cover already allocated blocks in selected range */
3192 if (ar
->pleft
&& start
<= ar
->lleft
) {
3193 size
-= ar
->lleft
+ 1 - start
;
3194 start
= ar
->lleft
+ 1;
3196 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3197 size
-= start
+ size
- ar
->lright
;
3201 /* check we don't cross already preallocated blocks */
3203 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3208 spin_lock(&pa
->pa_lock
);
3209 if (pa
->pa_deleted
) {
3210 spin_unlock(&pa
->pa_lock
);
3214 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3216 /* PA must not overlap original request */
3217 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3218 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3220 /* skip PA normalized request doesn't overlap with */
3221 if (pa
->pa_lstart
>= end
) {
3222 spin_unlock(&pa
->pa_lock
);
3225 if (pa_end
<= start
) {
3226 spin_unlock(&pa
->pa_lock
);
3229 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3231 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3232 BUG_ON(pa_end
< start
);
3236 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3237 BUG_ON(pa
->pa_lstart
> end
);
3238 end
= pa
->pa_lstart
;
3240 spin_unlock(&pa
->pa_lock
);
3245 /* XXX: extra loop to check we really don't overlap preallocations */
3247 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3249 spin_lock(&pa
->pa_lock
);
3250 if (pa
->pa_deleted
== 0) {
3251 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3252 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3254 spin_unlock(&pa
->pa_lock
);
3258 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3259 start
> ac
->ac_o_ex
.fe_logical
) {
3260 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3261 (unsigned long) start
, (unsigned long) size
,
3262 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3264 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3265 start
> ac
->ac_o_ex
.fe_logical
);
3266 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3268 /* now prepare goal request */
3270 /* XXX: is it better to align blocks WRT to logical
3271 * placement or satisfy big request as is */
3272 ac
->ac_g_ex
.fe_logical
= start
;
3273 ac
->ac_g_ex
.fe_len
= size
;
3275 /* define goal start in order to merge */
3276 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3277 /* merge to the right */
3278 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3279 &ac
->ac_f_ex
.fe_group
,
3280 &ac
->ac_f_ex
.fe_start
);
3281 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3283 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3284 /* merge to the left */
3285 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3286 &ac
->ac_f_ex
.fe_group
,
3287 &ac
->ac_f_ex
.fe_start
);
3288 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3291 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3292 (unsigned) orig_size
, (unsigned) start
);
3295 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3297 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3299 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3300 atomic_inc(&sbi
->s_bal_reqs
);
3301 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3302 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3303 atomic_inc(&sbi
->s_bal_success
);
3304 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3305 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3306 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3307 atomic_inc(&sbi
->s_bal_goals
);
3308 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3309 atomic_inc(&sbi
->s_bal_breaks
);
3312 ext4_mb_store_history(ac
);
3316 * use blocks preallocated to inode
3318 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3319 struct ext4_prealloc_space
*pa
)
3325 /* found preallocated blocks, use them */
3326 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3327 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3329 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3330 &ac
->ac_b_ex
.fe_start
);
3331 ac
->ac_b_ex
.fe_len
= len
;
3332 ac
->ac_status
= AC_STATUS_FOUND
;
3335 BUG_ON(start
< pa
->pa_pstart
);
3336 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3337 BUG_ON(pa
->pa_free
< len
);
3340 mb_debug("use %llu/%u from inode pa %p\n", start
, len
, pa
);
3344 * use blocks preallocated to locality group
3346 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3347 struct ext4_prealloc_space
*pa
)
3349 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3351 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3352 &ac
->ac_b_ex
.fe_group
,
3353 &ac
->ac_b_ex
.fe_start
);
3354 ac
->ac_b_ex
.fe_len
= len
;
3355 ac
->ac_status
= AC_STATUS_FOUND
;
3358 /* we don't correct pa_pstart or pa_plen here to avoid
3359 * possible race when the group is being loaded concurrently
3360 * instead we correct pa later, after blocks are marked
3361 * in on-disk bitmap -- see ext4_mb_release_context()
3362 * Other CPUs are prevented from allocating from this pa by lg_mutex
3364 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3368 * Return the prealloc space that have minimal distance
3369 * from the goal block. @cpa is the prealloc
3370 * space that is having currently known minimal distance
3371 * from the goal block.
3373 static struct ext4_prealloc_space
*
3374 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3375 struct ext4_prealloc_space
*pa
,
3376 struct ext4_prealloc_space
*cpa
)
3378 ext4_fsblk_t cur_distance
, new_distance
;
3381 atomic_inc(&pa
->pa_count
);
3384 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3385 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3387 if (cur_distance
< new_distance
)
3390 /* drop the previous reference */
3391 atomic_dec(&cpa
->pa_count
);
3392 atomic_inc(&pa
->pa_count
);
3397 * search goal blocks in preallocated space
3399 static noinline_for_stack
int
3400 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3403 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3404 struct ext4_locality_group
*lg
;
3405 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3406 ext4_fsblk_t goal_block
;
3408 /* only data can be preallocated */
3409 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3412 /* first, try per-file preallocation */
3414 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3416 /* all fields in this condition don't change,
3417 * so we can skip locking for them */
3418 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3419 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3422 /* found preallocated blocks, use them */
3423 spin_lock(&pa
->pa_lock
);
3424 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3425 atomic_inc(&pa
->pa_count
);
3426 ext4_mb_use_inode_pa(ac
, pa
);
3427 spin_unlock(&pa
->pa_lock
);
3428 ac
->ac_criteria
= 10;
3432 spin_unlock(&pa
->pa_lock
);
3436 /* can we use group allocation? */
3437 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3440 /* inode may have no locality group for some reason */
3444 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3445 if (order
> PREALLOC_TB_SIZE
- 1)
3446 /* The max size of hash table is PREALLOC_TB_SIZE */
3447 order
= PREALLOC_TB_SIZE
- 1;
3449 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3450 ac
->ac_g_ex
.fe_start
+
3451 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3453 * search for the prealloc space that is having
3454 * minimal distance from the goal block.
3456 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3458 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3460 spin_lock(&pa
->pa_lock
);
3461 if (pa
->pa_deleted
== 0 &&
3462 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3464 cpa
= ext4_mb_check_group_pa(goal_block
,
3467 spin_unlock(&pa
->pa_lock
);
3472 ext4_mb_use_group_pa(ac
, cpa
);
3473 ac
->ac_criteria
= 20;
3480 * the function goes through all block freed in the group
3481 * but not yet committed and marks them used in in-core bitmap.
3482 * buddy must be generated from this bitmap
3483 * Need to be called with ext4 group lock (ext4_lock_group)
3485 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3489 struct ext4_group_info
*grp
;
3490 struct ext4_free_data
*entry
;
3492 grp
= ext4_get_group_info(sb
, group
);
3493 n
= rb_first(&(grp
->bb_free_root
));
3496 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3497 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3498 bitmap
, entry
->start_blk
,
3506 * the function goes through all preallocation in this group and marks them
3507 * used in in-core bitmap. buddy must be generated from this bitmap
3508 * Need to be called with ext4 group lock (ext4_lock_group)
3510 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3513 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3514 struct ext4_prealloc_space
*pa
;
3515 struct list_head
*cur
;
3516 ext4_group_t groupnr
;
3517 ext4_grpblk_t start
;
3518 int preallocated
= 0;
3522 /* all form of preallocation discards first load group,
3523 * so the only competing code is preallocation use.
3524 * we don't need any locking here
3525 * notice we do NOT ignore preallocations with pa_deleted
3526 * otherwise we could leave used blocks available for
3527 * allocation in buddy when concurrent ext4_mb_put_pa()
3528 * is dropping preallocation
3530 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3531 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3532 spin_lock(&pa
->pa_lock
);
3533 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3536 spin_unlock(&pa
->pa_lock
);
3537 if (unlikely(len
== 0))
3539 BUG_ON(groupnr
!= group
);
3540 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3541 bitmap
, start
, len
);
3542 preallocated
+= len
;
3545 mb_debug("prellocated %u for group %u\n", preallocated
, group
);
3548 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3550 struct ext4_prealloc_space
*pa
;
3551 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3552 kmem_cache_free(ext4_pspace_cachep
, pa
);
3556 * drops a reference to preallocated space descriptor
3557 * if this was the last reference and the space is consumed
3559 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3560 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3564 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3567 /* in this short window concurrent discard can set pa_deleted */
3568 spin_lock(&pa
->pa_lock
);
3569 if (pa
->pa_deleted
== 1) {
3570 spin_unlock(&pa
->pa_lock
);
3575 spin_unlock(&pa
->pa_lock
);
3577 /* -1 is to protect from crossing allocation group */
3578 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3583 * P1 (buddy init) P2 (regular allocation)
3584 * find block B in PA
3585 * copy on-disk bitmap to buddy
3586 * mark B in on-disk bitmap
3587 * drop PA from group
3588 * mark all PAs in buddy
3590 * thus, P1 initializes buddy with B available. to prevent this
3591 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3594 ext4_lock_group(sb
, grp
);
3595 list_del(&pa
->pa_group_list
);
3596 ext4_unlock_group(sb
, grp
);
3598 spin_lock(pa
->pa_obj_lock
);
3599 list_del_rcu(&pa
->pa_inode_list
);
3600 spin_unlock(pa
->pa_obj_lock
);
3602 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3606 * creates new preallocated space for given inode
3608 static noinline_for_stack
int
3609 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3611 struct super_block
*sb
= ac
->ac_sb
;
3612 struct ext4_prealloc_space
*pa
;
3613 struct ext4_group_info
*grp
;
3614 struct ext4_inode_info
*ei
;
3616 /* preallocate only when found space is larger then requested */
3617 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3618 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3619 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3621 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3625 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3631 /* we can't allocate as much as normalizer wants.
3632 * so, found space must get proper lstart
3633 * to cover original request */
3634 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3635 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3637 /* we're limited by original request in that
3638 * logical block must be covered any way
3639 * winl is window we can move our chunk within */
3640 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3642 /* also, we should cover whole original request */
3643 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3645 /* the smallest one defines real window */
3646 win
= min(winl
, wins
);
3648 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3649 if (offs
&& offs
< win
)
3652 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3653 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3654 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3657 /* preallocation can change ac_b_ex, thus we store actually
3658 * allocated blocks for history */
3659 ac
->ac_f_ex
= ac
->ac_b_ex
;
3661 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3662 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3663 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3664 pa
->pa_free
= pa
->pa_len
;
3665 atomic_set(&pa
->pa_count
, 1);
3666 spin_lock_init(&pa
->pa_lock
);
3670 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3671 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3673 ext4_mb_use_inode_pa(ac
, pa
);
3674 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3676 ei
= EXT4_I(ac
->ac_inode
);
3677 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3679 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3680 pa
->pa_inode
= ac
->ac_inode
;
3682 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3683 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3684 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3686 spin_lock(pa
->pa_obj_lock
);
3687 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3688 spin_unlock(pa
->pa_obj_lock
);
3694 * creates new preallocated space for locality group inodes belongs to
3696 static noinline_for_stack
int
3697 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3699 struct super_block
*sb
= ac
->ac_sb
;
3700 struct ext4_locality_group
*lg
;
3701 struct ext4_prealloc_space
*pa
;
3702 struct ext4_group_info
*grp
;
3704 /* preallocate only when found space is larger then requested */
3705 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3706 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3707 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3709 BUG_ON(ext4_pspace_cachep
== NULL
);
3710 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3714 /* preallocation can change ac_b_ex, thus we store actually
3715 * allocated blocks for history */
3716 ac
->ac_f_ex
= ac
->ac_b_ex
;
3718 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3719 pa
->pa_lstart
= pa
->pa_pstart
;
3720 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3721 pa
->pa_free
= pa
->pa_len
;
3722 atomic_set(&pa
->pa_count
, 1);
3723 spin_lock_init(&pa
->pa_lock
);
3724 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3728 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3729 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3731 ext4_mb_use_group_pa(ac
, pa
);
3732 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3734 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3738 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3739 pa
->pa_inode
= NULL
;
3741 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3742 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3743 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3746 * We will later add the new pa to the right bucket
3747 * after updating the pa_free in ext4_mb_release_context
3752 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3756 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3757 err
= ext4_mb_new_group_pa(ac
);
3759 err
= ext4_mb_new_inode_pa(ac
);
3764 * finds all unused blocks in on-disk bitmap, frees them in
3765 * in-core bitmap and buddy.
3766 * @pa must be unlinked from inode and group lists, so that
3767 * nobody else can find/use it.
3768 * the caller MUST hold group/inode locks.
3769 * TODO: optimize the case when there are no in-core structures yet
3771 static noinline_for_stack
int
3772 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3773 struct ext4_prealloc_space
*pa
,
3774 struct ext4_allocation_context
*ac
)
3776 struct super_block
*sb
= e4b
->bd_sb
;
3777 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3786 BUG_ON(pa
->pa_deleted
== 0);
3787 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3788 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3789 end
= bit
+ pa
->pa_len
;
3793 ac
->ac_inode
= pa
->pa_inode
;
3794 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3798 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3801 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3802 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3803 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3804 mb_debug(" free preallocated %u/%u in group %u\n",
3805 (unsigned) start
, (unsigned) next
- bit
,
3810 ac
->ac_b_ex
.fe_group
= group
;
3811 ac
->ac_b_ex
.fe_start
= bit
;
3812 ac
->ac_b_ex
.fe_len
= next
- bit
;
3813 ac
->ac_b_ex
.fe_logical
= 0;
3814 ext4_mb_store_history(ac
);
3817 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3820 if (free
!= pa
->pa_free
) {
3821 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3822 pa
, (unsigned long) pa
->pa_lstart
,
3823 (unsigned long) pa
->pa_pstart
,
3824 (unsigned long) pa
->pa_len
);
3825 ext4_error(sb
, __func__
, "free %u, pa_free %u",
3828 * pa is already deleted so we use the value obtained
3829 * from the bitmap and continue.
3832 atomic_add(free
, &sbi
->s_mb_discarded
);
3837 static noinline_for_stack
int
3838 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3839 struct ext4_prealloc_space
*pa
,
3840 struct ext4_allocation_context
*ac
)
3842 struct super_block
*sb
= e4b
->bd_sb
;
3847 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3849 BUG_ON(pa
->pa_deleted
== 0);
3850 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3851 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3852 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3853 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3857 ac
->ac_inode
= NULL
;
3858 ac
->ac_b_ex
.fe_group
= group
;
3859 ac
->ac_b_ex
.fe_start
= bit
;
3860 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3861 ac
->ac_b_ex
.fe_logical
= 0;
3862 ext4_mb_store_history(ac
);
3869 * releases all preallocations in given group
3871 * first, we need to decide discard policy:
3872 * - when do we discard
3874 * - how many do we discard
3875 * 1) how many requested
3877 static noinline_for_stack
int
3878 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3879 ext4_group_t group
, int needed
)
3881 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3882 struct buffer_head
*bitmap_bh
= NULL
;
3883 struct ext4_prealloc_space
*pa
, *tmp
;
3884 struct ext4_allocation_context
*ac
;
3885 struct list_head list
;
3886 struct ext4_buddy e4b
;
3891 mb_debug("discard preallocation for group %u\n", group
);
3893 if (list_empty(&grp
->bb_prealloc_list
))
3896 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3897 if (bitmap_bh
== NULL
) {
3898 ext4_error(sb
, __func__
, "Error in reading block "
3899 "bitmap for %u", group
);
3903 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3905 ext4_error(sb
, __func__
, "Error in loading buddy "
3906 "information for %u", group
);
3912 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3914 INIT_LIST_HEAD(&list
);
3915 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3917 ext4_lock_group(sb
, group
);
3918 list_for_each_entry_safe(pa
, tmp
,
3919 &grp
->bb_prealloc_list
, pa_group_list
) {
3920 spin_lock(&pa
->pa_lock
);
3921 if (atomic_read(&pa
->pa_count
)) {
3922 spin_unlock(&pa
->pa_lock
);
3926 if (pa
->pa_deleted
) {
3927 spin_unlock(&pa
->pa_lock
);
3931 /* seems this one can be freed ... */
3934 /* we can trust pa_free ... */
3935 free
+= pa
->pa_free
;
3937 spin_unlock(&pa
->pa_lock
);
3939 list_del(&pa
->pa_group_list
);
3940 list_add(&pa
->u
.pa_tmp_list
, &list
);
3943 /* if we still need more blocks and some PAs were used, try again */
3944 if (free
< needed
&& busy
) {
3946 ext4_unlock_group(sb
, group
);
3948 * Yield the CPU here so that we don't get soft lockup
3949 * in non preempt case.
3955 /* found anything to free? */
3956 if (list_empty(&list
)) {
3961 /* now free all selected PAs */
3962 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3964 /* remove from object (inode or locality group) */
3965 spin_lock(pa
->pa_obj_lock
);
3966 list_del_rcu(&pa
->pa_inode_list
);
3967 spin_unlock(pa
->pa_obj_lock
);
3970 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3972 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3974 list_del(&pa
->u
.pa_tmp_list
);
3975 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3979 ext4_unlock_group(sb
, group
);
3981 kmem_cache_free(ext4_ac_cachep
, ac
);
3982 ext4_mb_release_desc(&e4b
);
3988 * releases all non-used preallocated blocks for given inode
3990 * It's important to discard preallocations under i_data_sem
3991 * We don't want another block to be served from the prealloc
3992 * space when we are discarding the inode prealloc space.
3994 * FIXME!! Make sure it is valid at all the call sites
3996 void ext4_discard_preallocations(struct inode
*inode
)
3998 struct ext4_inode_info
*ei
= EXT4_I(inode
);
3999 struct super_block
*sb
= inode
->i_sb
;
4000 struct buffer_head
*bitmap_bh
= NULL
;
4001 struct ext4_prealloc_space
*pa
, *tmp
;
4002 struct ext4_allocation_context
*ac
;
4003 ext4_group_t group
= 0;
4004 struct list_head list
;
4005 struct ext4_buddy e4b
;
4008 if (!S_ISREG(inode
->i_mode
)) {
4009 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4013 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
4015 INIT_LIST_HEAD(&list
);
4017 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4019 /* first, collect all pa's in the inode */
4020 spin_lock(&ei
->i_prealloc_lock
);
4021 while (!list_empty(&ei
->i_prealloc_list
)) {
4022 pa
= list_entry(ei
->i_prealloc_list
.next
,
4023 struct ext4_prealloc_space
, pa_inode_list
);
4024 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
4025 spin_lock(&pa
->pa_lock
);
4026 if (atomic_read(&pa
->pa_count
)) {
4027 /* this shouldn't happen often - nobody should
4028 * use preallocation while we're discarding it */
4029 spin_unlock(&pa
->pa_lock
);
4030 spin_unlock(&ei
->i_prealloc_lock
);
4031 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
4033 schedule_timeout_uninterruptible(HZ
);
4037 if (pa
->pa_deleted
== 0) {
4039 spin_unlock(&pa
->pa_lock
);
4040 list_del_rcu(&pa
->pa_inode_list
);
4041 list_add(&pa
->u
.pa_tmp_list
, &list
);
4045 /* someone is deleting pa right now */
4046 spin_unlock(&pa
->pa_lock
);
4047 spin_unlock(&ei
->i_prealloc_lock
);
4049 /* we have to wait here because pa_deleted
4050 * doesn't mean pa is already unlinked from
4051 * the list. as we might be called from
4052 * ->clear_inode() the inode will get freed
4053 * and concurrent thread which is unlinking
4054 * pa from inode's list may access already
4055 * freed memory, bad-bad-bad */
4057 /* XXX: if this happens too often, we can
4058 * add a flag to force wait only in case
4059 * of ->clear_inode(), but not in case of
4060 * regular truncate */
4061 schedule_timeout_uninterruptible(HZ
);
4064 spin_unlock(&ei
->i_prealloc_lock
);
4066 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4067 BUG_ON(pa
->pa_linear
!= 0);
4068 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4070 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4072 ext4_error(sb
, __func__
, "Error in loading buddy "
4073 "information for %u", group
);
4077 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4078 if (bitmap_bh
== NULL
) {
4079 ext4_error(sb
, __func__
, "Error in reading block "
4080 "bitmap for %u", group
);
4081 ext4_mb_release_desc(&e4b
);
4085 ext4_lock_group(sb
, group
);
4086 list_del(&pa
->pa_group_list
);
4087 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
4088 ext4_unlock_group(sb
, group
);
4090 ext4_mb_release_desc(&e4b
);
4093 list_del(&pa
->u
.pa_tmp_list
);
4094 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4097 kmem_cache_free(ext4_ac_cachep
, ac
);
4101 * finds all preallocated spaces and return blocks being freed to them
4102 * if preallocated space becomes full (no block is used from the space)
4103 * then the function frees space in buddy
4104 * XXX: at the moment, truncate (which is the only way to free blocks)
4105 * discards all preallocations
4107 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
4108 struct ext4_buddy
*e4b
,
4109 sector_t block
, int count
)
4111 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
4114 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4116 struct super_block
*sb
= ac
->ac_sb
;
4119 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
4120 " Allocation context details:\n");
4121 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
4122 ac
->ac_status
, ac
->ac_flags
);
4123 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4124 "best %lu/%lu/%lu@%lu cr %d\n",
4125 (unsigned long)ac
->ac_o_ex
.fe_group
,
4126 (unsigned long)ac
->ac_o_ex
.fe_start
,
4127 (unsigned long)ac
->ac_o_ex
.fe_len
,
4128 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4129 (unsigned long)ac
->ac_g_ex
.fe_group
,
4130 (unsigned long)ac
->ac_g_ex
.fe_start
,
4131 (unsigned long)ac
->ac_g_ex
.fe_len
,
4132 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4133 (unsigned long)ac
->ac_b_ex
.fe_group
,
4134 (unsigned long)ac
->ac_b_ex
.fe_start
,
4135 (unsigned long)ac
->ac_b_ex
.fe_len
,
4136 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4137 (int)ac
->ac_criteria
);
4138 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
4140 printk(KERN_ERR
"EXT4-fs: groups: \n");
4141 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
4142 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4143 struct ext4_prealloc_space
*pa
;
4144 ext4_grpblk_t start
;
4145 struct list_head
*cur
;
4146 ext4_lock_group(sb
, i
);
4147 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4148 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4150 spin_lock(&pa
->pa_lock
);
4151 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4153 spin_unlock(&pa
->pa_lock
);
4154 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
4157 ext4_unlock_group(sb
, i
);
4159 if (grp
->bb_free
== 0)
4161 printk(KERN_ERR
"%lu: %d/%d \n",
4162 i
, grp
->bb_free
, grp
->bb_fragments
);
4164 printk(KERN_ERR
"\n");
4167 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4174 * We use locality group preallocation for small size file. The size of the
4175 * file is determined by the current size or the resulting size after
4176 * allocation which ever is larger
4178 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4180 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4182 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4183 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4186 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4189 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
4190 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
4191 size
= max(size
, isize
);
4193 /* don't use group allocation for large files */
4194 if (size
>= sbi
->s_mb_stream_request
)
4197 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4200 BUG_ON(ac
->ac_lg
!= NULL
);
4202 * locality group prealloc space are per cpu. The reason for having
4203 * per cpu locality group is to reduce the contention between block
4204 * request from multiple CPUs.
4206 ac
->ac_lg
= per_cpu_ptr(sbi
->s_locality_groups
, raw_smp_processor_id());
4208 /* we're going to use group allocation */
4209 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4211 /* serialize all allocations in the group */
4212 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4215 static noinline_for_stack
int
4216 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4217 struct ext4_allocation_request
*ar
)
4219 struct super_block
*sb
= ar
->inode
->i_sb
;
4220 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4221 struct ext4_super_block
*es
= sbi
->s_es
;
4225 ext4_grpblk_t block
;
4227 /* we can't allocate > group size */
4230 /* just a dirty hack to filter too big requests */
4231 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4232 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4234 /* start searching from the goal */
4236 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4237 goal
>= ext4_blocks_count(es
))
4238 goal
= le32_to_cpu(es
->s_first_data_block
);
4239 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4241 /* set up allocation goals */
4242 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4243 ac
->ac_b_ex
.fe_group
= 0;
4244 ac
->ac_b_ex
.fe_start
= 0;
4245 ac
->ac_b_ex
.fe_len
= 0;
4246 ac
->ac_status
= AC_STATUS_CONTINUE
;
4247 ac
->ac_groups_scanned
= 0;
4248 ac
->ac_ex_scanned
= 0;
4251 ac
->ac_inode
= ar
->inode
;
4252 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4253 ac
->ac_o_ex
.fe_group
= group
;
4254 ac
->ac_o_ex
.fe_start
= block
;
4255 ac
->ac_o_ex
.fe_len
= len
;
4256 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4257 ac
->ac_g_ex
.fe_group
= group
;
4258 ac
->ac_g_ex
.fe_start
= block
;
4259 ac
->ac_g_ex
.fe_len
= len
;
4260 ac
->ac_f_ex
.fe_len
= 0;
4261 ac
->ac_flags
= ar
->flags
;
4263 ac
->ac_criteria
= 0;
4265 ac
->ac_bitmap_page
= NULL
;
4266 ac
->ac_buddy_page
= NULL
;
4269 /* we have to define context: we'll we work with a file or
4270 * locality group. this is a policy, actually */
4271 ext4_mb_group_or_file(ac
);
4273 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4274 "left: %u/%u, right %u/%u to %swritable\n",
4275 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4276 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4277 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4278 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4279 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4284 static noinline_for_stack
void
4285 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4286 struct ext4_locality_group
*lg
,
4287 int order
, int total_entries
)
4289 ext4_group_t group
= 0;
4290 struct ext4_buddy e4b
;
4291 struct list_head discard_list
;
4292 struct ext4_prealloc_space
*pa
, *tmp
;
4293 struct ext4_allocation_context
*ac
;
4295 mb_debug("discard locality group preallocation\n");
4297 INIT_LIST_HEAD(&discard_list
);
4298 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4300 spin_lock(&lg
->lg_prealloc_lock
);
4301 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4303 spin_lock(&pa
->pa_lock
);
4304 if (atomic_read(&pa
->pa_count
)) {
4306 * This is the pa that we just used
4307 * for block allocation. So don't
4310 spin_unlock(&pa
->pa_lock
);
4313 if (pa
->pa_deleted
) {
4314 spin_unlock(&pa
->pa_lock
);
4317 /* only lg prealloc space */
4318 BUG_ON(!pa
->pa_linear
);
4320 /* seems this one can be freed ... */
4322 spin_unlock(&pa
->pa_lock
);
4324 list_del_rcu(&pa
->pa_inode_list
);
4325 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4328 if (total_entries
<= 5) {
4330 * we want to keep only 5 entries
4331 * allowing it to grow to 8. This
4332 * mak sure we don't call discard
4333 * soon for this list.
4338 spin_unlock(&lg
->lg_prealloc_lock
);
4340 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4342 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4343 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4344 ext4_error(sb
, __func__
, "Error in loading buddy "
4345 "information for %u", group
);
4348 ext4_lock_group(sb
, group
);
4349 list_del(&pa
->pa_group_list
);
4350 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4351 ext4_unlock_group(sb
, group
);
4353 ext4_mb_release_desc(&e4b
);
4354 list_del(&pa
->u
.pa_tmp_list
);
4355 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4358 kmem_cache_free(ext4_ac_cachep
, ac
);
4362 * We have incremented pa_count. So it cannot be freed at this
4363 * point. Also we hold lg_mutex. So no parallel allocation is
4364 * possible from this lg. That means pa_free cannot be updated.
4366 * A parallel ext4_mb_discard_group_preallocations is possible.
4367 * which can cause the lg_prealloc_list to be updated.
4370 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4372 int order
, added
= 0, lg_prealloc_count
= 1;
4373 struct super_block
*sb
= ac
->ac_sb
;
4374 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4375 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4377 order
= fls(pa
->pa_free
) - 1;
4378 if (order
> PREALLOC_TB_SIZE
- 1)
4379 /* The max size of hash table is PREALLOC_TB_SIZE */
4380 order
= PREALLOC_TB_SIZE
- 1;
4381 /* Add the prealloc space to lg */
4383 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4385 spin_lock(&tmp_pa
->pa_lock
);
4386 if (tmp_pa
->pa_deleted
) {
4387 spin_unlock(&pa
->pa_lock
);
4390 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4391 /* Add to the tail of the previous entry */
4392 list_add_tail_rcu(&pa
->pa_inode_list
,
4393 &tmp_pa
->pa_inode_list
);
4396 * we want to count the total
4397 * number of entries in the list
4400 spin_unlock(&tmp_pa
->pa_lock
);
4401 lg_prealloc_count
++;
4404 list_add_tail_rcu(&pa
->pa_inode_list
,
4405 &lg
->lg_prealloc_list
[order
]);
4408 /* Now trim the list to be not more than 8 elements */
4409 if (lg_prealloc_count
> 8) {
4410 ext4_mb_discard_lg_preallocations(sb
, lg
,
4411 order
, lg_prealloc_count
);
4418 * release all resource we used in allocation
4420 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4422 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4424 if (pa
->pa_linear
) {
4425 /* see comment in ext4_mb_use_group_pa() */
4426 spin_lock(&pa
->pa_lock
);
4427 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4428 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4429 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4430 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4431 spin_unlock(&pa
->pa_lock
);
4433 * We want to add the pa to the right bucket.
4434 * Remove it from the list and while adding
4435 * make sure the list to which we are adding
4438 if (likely(pa
->pa_free
)) {
4439 spin_lock(pa
->pa_obj_lock
);
4440 list_del_rcu(&pa
->pa_inode_list
);
4441 spin_unlock(pa
->pa_obj_lock
);
4442 ext4_mb_add_n_trim(ac
);
4445 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4447 if (ac
->ac_bitmap_page
)
4448 page_cache_release(ac
->ac_bitmap_page
);
4449 if (ac
->ac_buddy_page
)
4450 page_cache_release(ac
->ac_buddy_page
);
4451 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4452 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4453 ext4_mb_collect_stats(ac
);
4457 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4463 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4464 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4473 * Main entry point into mballoc to allocate blocks
4474 * it tries to use preallocation first, then falls back
4475 * to usual allocation
4477 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4478 struct ext4_allocation_request
*ar
, int *errp
)
4481 struct ext4_allocation_context
*ac
= NULL
;
4482 struct ext4_sb_info
*sbi
;
4483 struct super_block
*sb
;
4484 ext4_fsblk_t block
= 0;
4485 unsigned int inquota
;
4486 unsigned int reserv_blks
= 0;
4488 sb
= ar
->inode
->i_sb
;
4491 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
) {
4493 * With delalloc we already reserved the blocks
4495 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4496 /* let others to free the space */
4498 ar
->len
= ar
->len
>> 1;
4504 reserv_blks
= ar
->len
;
4506 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4507 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4516 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4517 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4519 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4526 *errp
= ext4_mb_initialize_context(ac
, ar
);
4532 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4533 if (!ext4_mb_use_preallocated(ac
)) {
4534 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4535 ext4_mb_normalize_request(ac
, ar
);
4537 /* allocate space in core */
4538 ext4_mb_regular_allocator(ac
);
4540 /* as we've just preallocated more space than
4541 * user requested orinally, we store allocated
4542 * space in a special descriptor */
4543 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4544 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4545 ext4_mb_new_preallocation(ac
);
4548 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4549 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4550 if (*errp
== -EAGAIN
) {
4551 ac
->ac_b_ex
.fe_group
= 0;
4552 ac
->ac_b_ex
.fe_start
= 0;
4553 ac
->ac_b_ex
.fe_len
= 0;
4554 ac
->ac_status
= AC_STATUS_CONTINUE
;
4557 ac
->ac_b_ex
.fe_len
= 0;
4559 ext4_mb_show_ac(ac
);
4561 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4562 ar
->len
= ac
->ac_b_ex
.fe_len
;
4565 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4569 ac
->ac_b_ex
.fe_len
= 0;
4571 ext4_mb_show_ac(ac
);
4574 ext4_mb_release_context(ac
);
4577 kmem_cache_free(ext4_ac_cachep
, ac
);
4579 if (ar
->len
< inquota
)
4580 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4586 * We can merge two free data extents only if the physical blocks
4587 * are contiguous, AND the extents were freed by the same transaction,
4588 * AND the blocks are associated with the same group.
4590 static int can_merge(struct ext4_free_data
*entry1
,
4591 struct ext4_free_data
*entry2
)
4593 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4594 (entry1
->group
== entry2
->group
) &&
4595 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4600 static noinline_for_stack
int
4601 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4602 struct ext4_free_data
*new_entry
)
4604 ext4_grpblk_t block
;
4605 struct ext4_free_data
*entry
;
4606 struct ext4_group_info
*db
= e4b
->bd_info
;
4607 struct super_block
*sb
= e4b
->bd_sb
;
4608 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4609 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4610 struct rb_node
*parent
= NULL
, *new_node
;
4612 BUG_ON(!ext4_handle_valid(handle
));
4613 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4614 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4616 new_node
= &new_entry
->node
;
4617 block
= new_entry
->start_blk
;
4620 /* first free block exent. We need to
4621 protect buddy cache from being freed,
4622 * otherwise we'll refresh it from
4623 * on-disk bitmap and lose not-yet-available
4625 page_cache_get(e4b
->bd_buddy_page
);
4626 page_cache_get(e4b
->bd_bitmap_page
);
4630 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4631 if (block
< entry
->start_blk
)
4633 else if (block
>= (entry
->start_blk
+ entry
->count
))
4634 n
= &(*n
)->rb_right
;
4636 ext4_error(sb
, __func__
,
4637 "Double free of blocks %d (%d %d)",
4638 block
, entry
->start_blk
, entry
->count
);
4643 rb_link_node(new_node
, parent
, n
);
4644 rb_insert_color(new_node
, &db
->bb_free_root
);
4646 /* Now try to see the extent can be merged to left and right */
4647 node
= rb_prev(new_node
);
4649 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4650 if (can_merge(entry
, new_entry
)) {
4651 new_entry
->start_blk
= entry
->start_blk
;
4652 new_entry
->count
+= entry
->count
;
4653 rb_erase(node
, &(db
->bb_free_root
));
4654 spin_lock(&sbi
->s_md_lock
);
4655 list_del(&entry
->list
);
4656 spin_unlock(&sbi
->s_md_lock
);
4657 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4661 node
= rb_next(new_node
);
4663 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4664 if (can_merge(new_entry
, entry
)) {
4665 new_entry
->count
+= entry
->count
;
4666 rb_erase(node
, &(db
->bb_free_root
));
4667 spin_lock(&sbi
->s_md_lock
);
4668 list_del(&entry
->list
);
4669 spin_unlock(&sbi
->s_md_lock
);
4670 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4673 /* Add the extent to transaction's private list */
4674 spin_lock(&sbi
->s_md_lock
);
4675 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4676 spin_unlock(&sbi
->s_md_lock
);
4681 * Main entry point into mballoc to free blocks
4683 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4684 unsigned long block
, unsigned long count
,
4685 int metadata
, unsigned long *freed
)
4687 struct buffer_head
*bitmap_bh
= NULL
;
4688 struct super_block
*sb
= inode
->i_sb
;
4689 struct ext4_allocation_context
*ac
= NULL
;
4690 struct ext4_group_desc
*gdp
;
4691 struct ext4_super_block
*es
;
4692 unsigned int overflow
;
4694 struct buffer_head
*gd_bh
;
4695 ext4_group_t block_group
;
4696 struct ext4_sb_info
*sbi
;
4697 struct ext4_buddy e4b
;
4704 es
= EXT4_SB(sb
)->s_es
;
4705 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4706 block
+ count
< block
||
4707 block
+ count
> ext4_blocks_count(es
)) {
4708 ext4_error(sb
, __func__
,
4709 "Freeing blocks not in datazone - "
4710 "block = %lu, count = %lu", block
, count
);
4714 ext4_debug("freeing block %lu\n", block
);
4716 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4718 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4719 ac
->ac_inode
= inode
;
4725 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4728 * Check to see if we are freeing blocks across a group
4731 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4732 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4735 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4740 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4746 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4747 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4748 in_range(block
, ext4_inode_table(sb
, gdp
),
4749 EXT4_SB(sb
)->s_itb_per_group
) ||
4750 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4751 EXT4_SB(sb
)->s_itb_per_group
)) {
4753 ext4_error(sb
, __func__
,
4754 "Freeing blocks in system zone - "
4755 "Block = %lu, count = %lu", block
, count
);
4756 /* err = 0. ext4_std_error should be a no op */
4760 BUFFER_TRACE(bitmap_bh
, "getting write access");
4761 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4766 * We are about to modify some metadata. Call the journal APIs
4767 * to unshare ->b_data if a currently-committing transaction is
4770 BUFFER_TRACE(gd_bh
, "get_write_access");
4771 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4774 #ifdef AGGRESSIVE_CHECK
4777 for (i
= 0; i
< count
; i
++)
4778 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4782 ac
->ac_b_ex
.fe_group
= block_group
;
4783 ac
->ac_b_ex
.fe_start
= bit
;
4784 ac
->ac_b_ex
.fe_len
= count
;
4785 ext4_mb_store_history(ac
);
4788 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4791 if (metadata
&& ext4_handle_valid(handle
)) {
4792 struct ext4_free_data
*new_entry
;
4794 * blocks being freed are metadata. these blocks shouldn't
4795 * be used until this transaction is committed
4797 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4798 new_entry
->start_blk
= bit
;
4799 new_entry
->group
= block_group
;
4800 new_entry
->count
= count
;
4801 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4802 ext4_lock_group(sb
, block_group
);
4803 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4805 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4806 ext4_unlock_group(sb
, block_group
);
4808 ext4_lock_group(sb
, block_group
);
4809 /* need to update group_info->bb_free and bitmap
4810 * with group lock held. generate_buddy look at
4811 * them with group lock_held
4813 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4815 mb_free_blocks(inode
, &e4b
, bit
, count
);
4816 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4817 ext4_unlock_group(sb
, block_group
);
4820 spin_lock(sb_bgl_lock(sbi
, block_group
));
4821 le16_add_cpu(&gdp
->bg_free_blocks_count
, count
);
4822 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4823 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4824 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4826 if (sbi
->s_log_groups_per_flex
) {
4827 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4828 spin_lock(sb_bgl_lock(sbi
, flex_group
));
4829 sbi
->s_flex_groups
[flex_group
].free_blocks
+= count
;
4830 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
4833 ext4_mb_release_desc(&e4b
);
4837 /* We dirtied the bitmap block */
4838 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4839 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4841 /* And the group descriptor block */
4842 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4843 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4847 if (overflow
&& !err
) {
4856 ext4_std_error(sb
, err
);
4858 kmem_cache_free(ext4_ac_cachep
, ac
);