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
);
460 ext4_grp_locked_error(sb
, e4b
->bd_group
,
461 __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_grp_locked_error(sb
, group
, __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 (bitmap_uptodate(bh
[i
]))
801 if (bitmap_uptodate(bh
[i
])) {
802 unlock_buffer(bh
[i
]);
805 spin_lock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
806 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
807 ext4_init_block_bitmap(sb
, bh
[i
],
808 first_group
+ i
, desc
);
809 set_bitmap_uptodate(bh
[i
]);
810 set_buffer_uptodate(bh
[i
]);
811 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
812 unlock_buffer(bh
[i
]);
815 spin_unlock(sb_bgl_lock(EXT4_SB(sb
), first_group
+ i
));
816 if (buffer_uptodate(bh
[i
])) {
818 * if not uninit if bh is uptodate,
819 * bitmap is also uptodate
821 set_bitmap_uptodate(bh
[i
]);
822 unlock_buffer(bh
[i
]);
827 * submit the buffer_head for read. We can
828 * safely mark the bitmap as uptodate now.
829 * We do it here so the bitmap uptodate bit
830 * get set with buffer lock held.
832 set_bitmap_uptodate(bh
[i
]);
833 bh
[i
]->b_end_io
= end_buffer_read_sync
;
834 submit_bh(READ
, bh
[i
]);
835 mb_debug("read bitmap for group %u\n", first_group
+ i
);
838 /* wait for I/O completion */
839 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
840 wait_on_buffer(bh
[i
]);
843 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
844 if (!buffer_uptodate(bh
[i
]))
848 first_block
= page
->index
* blocks_per_page
;
849 for (i
= 0; i
< blocks_per_page
; i
++) {
851 struct ext4_group_info
*grinfo
;
853 group
= (first_block
+ i
) >> 1;
854 if (group
>= EXT4_SB(sb
)->s_groups_count
)
858 * data carry information regarding this
859 * particular group in the format specified
863 data
= page_address(page
) + (i
* blocksize
);
864 bitmap
= bh
[group
- first_group
]->b_data
;
867 * We place the buddy block and bitmap block
870 if ((first_block
+ i
) & 1) {
871 /* this is block of buddy */
872 BUG_ON(incore
== NULL
);
873 mb_debug("put buddy for group %u in page %lu/%x\n",
874 group
, page
->index
, i
* blocksize
);
875 memset(data
, 0xff, blocksize
);
876 grinfo
= ext4_get_group_info(sb
, group
);
877 grinfo
->bb_fragments
= 0;
878 memset(grinfo
->bb_counters
, 0,
879 sizeof(unsigned short)*(sb
->s_blocksize_bits
+2));
881 * incore got set to the group block bitmap below
883 ext4_lock_group(sb
, group
);
884 ext4_mb_generate_buddy(sb
, data
, incore
, group
);
885 ext4_unlock_group(sb
, group
);
888 /* this is block of bitmap */
889 BUG_ON(incore
!= NULL
);
890 mb_debug("put bitmap for group %u in page %lu/%x\n",
891 group
, page
->index
, i
* blocksize
);
893 /* see comments in ext4_mb_put_pa() */
894 ext4_lock_group(sb
, group
);
895 memcpy(data
, bitmap
, blocksize
);
897 /* mark all preallocated blks used in in-core bitmap */
898 ext4_mb_generate_from_pa(sb
, data
, group
);
899 ext4_mb_generate_from_freelist(sb
, data
, group
);
900 ext4_unlock_group(sb
, group
);
902 /* set incore so that the buddy information can be
903 * generated using this
908 SetPageUptodate(page
);
912 for (i
= 0; i
< groups_per_page
&& bh
[i
]; i
++)
920 static noinline_for_stack
int
921 ext4_mb_load_buddy(struct super_block
*sb
, ext4_group_t group
,
922 struct ext4_buddy
*e4b
)
930 struct ext4_group_info
*grp
;
931 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
932 struct inode
*inode
= sbi
->s_buddy_cache
;
934 mb_debug("load group %u\n", group
);
936 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
937 grp
= ext4_get_group_info(sb
, group
);
939 e4b
->bd_blkbits
= sb
->s_blocksize_bits
;
940 e4b
->bd_info
= ext4_get_group_info(sb
, group
);
942 e4b
->bd_group
= group
;
943 e4b
->bd_buddy_page
= NULL
;
944 e4b
->bd_bitmap_page
= NULL
;
945 e4b
->alloc_semp
= &grp
->alloc_sem
;
947 /* Take the read lock on the group alloc
948 * sem. This would make sure a parallel
949 * ext4_mb_init_group happening on other
950 * groups mapped by the page is blocked
951 * till we are done with allocation
953 down_read(e4b
->alloc_semp
);
956 * the buddy cache inode stores the block bitmap
957 * and buddy information in consecutive blocks.
958 * So for each group we need two blocks.
961 pnum
= block
/ blocks_per_page
;
962 poff
= block
% blocks_per_page
;
964 /* we could use find_or_create_page(), but it locks page
965 * what we'd like to avoid in fast path ... */
966 page
= find_get_page(inode
->i_mapping
, pnum
);
967 if (page
== NULL
|| !PageUptodate(page
)) {
970 * drop the page reference and try
971 * to get the page with lock. If we
972 * are not uptodate that implies
973 * somebody just created the page but
974 * is yet to initialize the same. So
975 * wait for it to initialize.
977 page_cache_release(page
);
978 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
980 BUG_ON(page
->mapping
!= inode
->i_mapping
);
981 if (!PageUptodate(page
)) {
982 ret
= ext4_mb_init_cache(page
, NULL
);
987 mb_cmp_bitmaps(e4b
, page_address(page
) +
988 (poff
* sb
->s_blocksize
));
993 if (page
== NULL
|| !PageUptodate(page
)) {
997 e4b
->bd_bitmap_page
= page
;
998 e4b
->bd_bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
999 mark_page_accessed(page
);
1002 pnum
= block
/ blocks_per_page
;
1003 poff
= block
% blocks_per_page
;
1005 page
= find_get_page(inode
->i_mapping
, pnum
);
1006 if (page
== NULL
|| !PageUptodate(page
)) {
1008 page_cache_release(page
);
1009 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1011 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1012 if (!PageUptodate(page
)) {
1013 ret
= ext4_mb_init_cache(page
, e4b
->bd_bitmap
);
1022 if (page
== NULL
|| !PageUptodate(page
)) {
1026 e4b
->bd_buddy_page
= page
;
1027 e4b
->bd_buddy
= page_address(page
) + (poff
* sb
->s_blocksize
);
1028 mark_page_accessed(page
);
1030 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
1031 BUG_ON(e4b
->bd_buddy_page
== NULL
);
1036 if (e4b
->bd_bitmap_page
)
1037 page_cache_release(e4b
->bd_bitmap_page
);
1038 if (e4b
->bd_buddy_page
)
1039 page_cache_release(e4b
->bd_buddy_page
);
1040 e4b
->bd_buddy
= NULL
;
1041 e4b
->bd_bitmap
= NULL
;
1043 /* Done with the buddy cache */
1044 up_read(e4b
->alloc_semp
);
1048 static void ext4_mb_release_desc(struct ext4_buddy
*e4b
)
1050 if (e4b
->bd_bitmap_page
)
1051 page_cache_release(e4b
->bd_bitmap_page
);
1052 if (e4b
->bd_buddy_page
)
1053 page_cache_release(e4b
->bd_buddy_page
);
1054 /* Done with the buddy cache */
1055 up_read(e4b
->alloc_semp
);
1059 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1064 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1065 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1067 bb
= EXT4_MB_BUDDY(e4b
);
1068 while (order
<= e4b
->bd_blkbits
+ 1) {
1070 if (!mb_test_bit(block
, bb
)) {
1071 /* this block is part of buddy of order 'order' */
1074 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1080 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1086 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1087 /* fast path: clear whole word at once */
1088 addr
= bm
+ (cur
>> 3);
1094 mb_clear_bit_atomic(lock
, cur
, bm
);
1096 mb_clear_bit(cur
, bm
);
1101 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1107 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1108 /* fast path: set whole word at once */
1109 addr
= bm
+ (cur
>> 3);
1115 mb_set_bit_atomic(lock
, cur
, bm
);
1117 mb_set_bit(cur
, bm
);
1122 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1123 int first
, int count
)
1130 struct super_block
*sb
= e4b
->bd_sb
;
1132 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1133 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1134 mb_check_buddy(e4b
);
1135 mb_free_blocks_double(inode
, e4b
, first
, count
);
1137 e4b
->bd_info
->bb_free
+= count
;
1138 if (first
< e4b
->bd_info
->bb_first_free
)
1139 e4b
->bd_info
->bb_first_free
= first
;
1141 /* let's maintain fragments counter */
1143 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1144 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1145 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1147 e4b
->bd_info
->bb_fragments
--;
1148 else if (!block
&& !max
)
1149 e4b
->bd_info
->bb_fragments
++;
1151 /* let's maintain buddy itself */
1152 while (count
-- > 0) {
1156 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1157 ext4_fsblk_t blocknr
;
1158 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1161 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1162 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1163 __func__
, "double-free of inode"
1164 " %lu's block %llu(bit %u in group %u)",
1165 inode
? inode
->i_ino
: 0, blocknr
, block
,
1168 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1169 e4b
->bd_info
->bb_counters
[order
]++;
1171 /* start of the buddy */
1172 buddy
= mb_find_buddy(e4b
, order
, &max
);
1176 if (mb_test_bit(block
, buddy
) ||
1177 mb_test_bit(block
+ 1, buddy
))
1180 /* both the buddies are free, try to coalesce them */
1181 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1187 /* for special purposes, we don't set
1188 * free bits in bitmap */
1189 mb_set_bit(block
, buddy
);
1190 mb_set_bit(block
+ 1, buddy
);
1192 e4b
->bd_info
->bb_counters
[order
]--;
1193 e4b
->bd_info
->bb_counters
[order
]--;
1197 e4b
->bd_info
->bb_counters
[order
]++;
1199 mb_clear_bit(block
, buddy2
);
1203 mb_check_buddy(e4b
);
1206 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1207 int needed
, struct ext4_free_extent
*ex
)
1214 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1217 buddy
= mb_find_buddy(e4b
, order
, &max
);
1218 BUG_ON(buddy
== NULL
);
1219 BUG_ON(block
>= max
);
1220 if (mb_test_bit(block
, buddy
)) {
1227 /* FIXME dorp order completely ? */
1228 if (likely(order
== 0)) {
1229 /* find actual order */
1230 order
= mb_find_order_for_block(e4b
, block
);
1231 block
= block
>> order
;
1234 ex
->fe_len
= 1 << order
;
1235 ex
->fe_start
= block
<< order
;
1236 ex
->fe_group
= e4b
->bd_group
;
1238 /* calc difference from given start */
1239 next
= next
- ex
->fe_start
;
1241 ex
->fe_start
+= next
;
1243 while (needed
> ex
->fe_len
&&
1244 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1246 if (block
+ 1 >= max
)
1249 next
= (block
+ 1) * (1 << order
);
1250 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1253 ord
= mb_find_order_for_block(e4b
, next
);
1256 block
= next
>> order
;
1257 ex
->fe_len
+= 1 << order
;
1260 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1264 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1270 int start
= ex
->fe_start
;
1271 int len
= ex
->fe_len
;
1276 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1277 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1278 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1279 mb_check_buddy(e4b
);
1280 mb_mark_used_double(e4b
, start
, len
);
1282 e4b
->bd_info
->bb_free
-= len
;
1283 if (e4b
->bd_info
->bb_first_free
== start
)
1284 e4b
->bd_info
->bb_first_free
+= len
;
1286 /* let's maintain fragments counter */
1288 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1289 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1290 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1292 e4b
->bd_info
->bb_fragments
++;
1293 else if (!mlen
&& !max
)
1294 e4b
->bd_info
->bb_fragments
--;
1296 /* let's maintain buddy itself */
1298 ord
= mb_find_order_for_block(e4b
, start
);
1300 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1301 /* the whole chunk may be allocated at once! */
1303 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1304 BUG_ON((start
>> ord
) >= max
);
1305 mb_set_bit(start
>> ord
, buddy
);
1306 e4b
->bd_info
->bb_counters
[ord
]--;
1313 /* store for history */
1315 ret
= len
| (ord
<< 16);
1317 /* we have to split large buddy */
1319 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1320 mb_set_bit(start
>> ord
, buddy
);
1321 e4b
->bd_info
->bb_counters
[ord
]--;
1324 cur
= (start
>> ord
) & ~1U;
1325 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1326 mb_clear_bit(cur
, buddy
);
1327 mb_clear_bit(cur
+ 1, buddy
);
1328 e4b
->bd_info
->bb_counters
[ord
]++;
1329 e4b
->bd_info
->bb_counters
[ord
]++;
1332 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1333 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1334 mb_check_buddy(e4b
);
1340 * Must be called under group lock!
1342 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1343 struct ext4_buddy
*e4b
)
1345 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1348 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1349 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1351 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1352 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1353 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1355 /* preallocation can change ac_b_ex, thus we store actually
1356 * allocated blocks for history */
1357 ac
->ac_f_ex
= ac
->ac_b_ex
;
1359 ac
->ac_status
= AC_STATUS_FOUND
;
1360 ac
->ac_tail
= ret
& 0xffff;
1361 ac
->ac_buddy
= ret
>> 16;
1364 * take the page reference. We want the page to be pinned
1365 * so that we don't get a ext4_mb_init_cache_call for this
1366 * group until we update the bitmap. That would mean we
1367 * double allocate blocks. The reference is dropped
1368 * in ext4_mb_release_context
1370 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1371 get_page(ac
->ac_bitmap_page
);
1372 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1373 get_page(ac
->ac_buddy_page
);
1375 /* store last allocated for subsequent stream allocation */
1376 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1377 spin_lock(&sbi
->s_md_lock
);
1378 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1379 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1380 spin_unlock(&sbi
->s_md_lock
);
1385 * regular allocator, for general purposes allocation
1388 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1389 struct ext4_buddy
*e4b
,
1392 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1393 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1394 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1395 struct ext4_free_extent ex
;
1398 if (ac
->ac_status
== AC_STATUS_FOUND
)
1401 * We don't want to scan for a whole year
1403 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1404 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1405 ac
->ac_status
= AC_STATUS_BREAK
;
1410 * Haven't found good chunk so far, let's continue
1412 if (bex
->fe_len
< gex
->fe_len
)
1415 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1416 && bex
->fe_group
== e4b
->bd_group
) {
1417 /* recheck chunk's availability - we don't know
1418 * when it was found (within this lock-unlock
1420 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1421 if (max
>= gex
->fe_len
) {
1422 ext4_mb_use_best_found(ac
, e4b
);
1429 * The routine checks whether found extent is good enough. If it is,
1430 * then the extent gets marked used and flag is set to the context
1431 * to stop scanning. Otherwise, the extent is compared with the
1432 * previous found extent and if new one is better, then it's stored
1433 * in the context. Later, the best found extent will be used, if
1434 * mballoc can't find good enough extent.
1436 * FIXME: real allocation policy is to be designed yet!
1438 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1439 struct ext4_free_extent
*ex
,
1440 struct ext4_buddy
*e4b
)
1442 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1443 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1445 BUG_ON(ex
->fe_len
<= 0);
1446 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1447 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1448 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1453 * The special case - take what you catch first
1455 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1457 ext4_mb_use_best_found(ac
, e4b
);
1462 * Let's check whether the chuck is good enough
1464 if (ex
->fe_len
== gex
->fe_len
) {
1466 ext4_mb_use_best_found(ac
, e4b
);
1471 * If this is first found extent, just store it in the context
1473 if (bex
->fe_len
== 0) {
1479 * If new found extent is better, store it in the context
1481 if (bex
->fe_len
< gex
->fe_len
) {
1482 /* if the request isn't satisfied, any found extent
1483 * larger than previous best one is better */
1484 if (ex
->fe_len
> bex
->fe_len
)
1486 } else if (ex
->fe_len
> gex
->fe_len
) {
1487 /* if the request is satisfied, then we try to find
1488 * an extent that still satisfy the request, but is
1489 * smaller than previous one */
1490 if (ex
->fe_len
< bex
->fe_len
)
1494 ext4_mb_check_limits(ac
, e4b
, 0);
1497 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1498 struct ext4_buddy
*e4b
)
1500 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1501 ext4_group_t group
= ex
.fe_group
;
1505 BUG_ON(ex
.fe_len
<= 0);
1506 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1510 ext4_lock_group(ac
->ac_sb
, group
);
1511 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1515 ext4_mb_use_best_found(ac
, e4b
);
1518 ext4_unlock_group(ac
->ac_sb
, group
);
1519 ext4_mb_release_desc(e4b
);
1524 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1525 struct ext4_buddy
*e4b
)
1527 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1530 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1531 struct ext4_super_block
*es
= sbi
->s_es
;
1532 struct ext4_free_extent ex
;
1534 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1537 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1541 ext4_lock_group(ac
->ac_sb
, group
);
1542 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1543 ac
->ac_g_ex
.fe_len
, &ex
);
1545 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1548 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1549 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1550 /* use do_div to get remainder (would be 64-bit modulo) */
1551 if (do_div(start
, sbi
->s_stripe
) == 0) {
1554 ext4_mb_use_best_found(ac
, e4b
);
1556 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1557 BUG_ON(ex
.fe_len
<= 0);
1558 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1559 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1562 ext4_mb_use_best_found(ac
, e4b
);
1563 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1564 /* Sometimes, caller may want to merge even small
1565 * number of blocks to an existing extent */
1566 BUG_ON(ex
.fe_len
<= 0);
1567 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1568 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1571 ext4_mb_use_best_found(ac
, e4b
);
1573 ext4_unlock_group(ac
->ac_sb
, group
);
1574 ext4_mb_release_desc(e4b
);
1580 * The routine scans buddy structures (not bitmap!) from given order
1581 * to max order and tries to find big enough chunk to satisfy the req
1583 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1584 struct ext4_buddy
*e4b
)
1586 struct super_block
*sb
= ac
->ac_sb
;
1587 struct ext4_group_info
*grp
= e4b
->bd_info
;
1593 BUG_ON(ac
->ac_2order
<= 0);
1594 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1595 if (grp
->bb_counters
[i
] == 0)
1598 buddy
= mb_find_buddy(e4b
, i
, &max
);
1599 BUG_ON(buddy
== NULL
);
1601 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1606 ac
->ac_b_ex
.fe_len
= 1 << i
;
1607 ac
->ac_b_ex
.fe_start
= k
<< i
;
1608 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1610 ext4_mb_use_best_found(ac
, e4b
);
1612 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1614 if (EXT4_SB(sb
)->s_mb_stats
)
1615 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1622 * The routine scans the group and measures all found extents.
1623 * In order to optimize scanning, caller must pass number of
1624 * free blocks in the group, so the routine can know upper limit.
1626 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1627 struct ext4_buddy
*e4b
)
1629 struct super_block
*sb
= ac
->ac_sb
;
1630 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1631 struct ext4_free_extent ex
;
1635 free
= e4b
->bd_info
->bb_free
;
1638 i
= e4b
->bd_info
->bb_first_free
;
1640 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1641 i
= mb_find_next_zero_bit(bitmap
,
1642 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1643 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1645 * IF we have corrupt bitmap, we won't find any
1646 * free blocks even though group info says we
1647 * we have free blocks
1649 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1650 __func__
, "%d free blocks as per "
1651 "group info. But bitmap says 0",
1656 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1657 BUG_ON(ex
.fe_len
<= 0);
1658 if (free
< ex
.fe_len
) {
1659 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1660 __func__
, "%d free blocks as per "
1661 "group info. But got %d blocks",
1664 * The number of free blocks differs. This mostly
1665 * indicate that the bitmap is corrupt. So exit
1666 * without claiming the space.
1671 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1677 ext4_mb_check_limits(ac
, e4b
, 1);
1681 * This is a special case for storages like raid5
1682 * we try to find stripe-aligned chunks for stripe-size requests
1683 * XXX should do so at least for multiples of stripe size as well
1685 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1686 struct ext4_buddy
*e4b
)
1688 struct super_block
*sb
= ac
->ac_sb
;
1689 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1690 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1691 struct ext4_free_extent ex
;
1692 ext4_fsblk_t first_group_block
;
1697 BUG_ON(sbi
->s_stripe
== 0);
1699 /* find first stripe-aligned block in group */
1700 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1701 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1702 a
= first_group_block
+ sbi
->s_stripe
- 1;
1703 do_div(a
, sbi
->s_stripe
);
1704 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1706 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1707 if (!mb_test_bit(i
, bitmap
)) {
1708 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1709 if (max
>= sbi
->s_stripe
) {
1712 ext4_mb_use_best_found(ac
, e4b
);
1720 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1721 ext4_group_t group
, int cr
)
1723 unsigned free
, fragments
;
1725 struct ext4_group_desc
*desc
;
1726 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1728 BUG_ON(cr
< 0 || cr
>= 4);
1729 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1731 free
= grp
->bb_free
;
1732 fragments
= grp
->bb_fragments
;
1740 BUG_ON(ac
->ac_2order
== 0);
1741 /* If this group is uninitialized, skip it initially */
1742 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1743 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1746 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1747 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1748 if (grp
->bb_counters
[i
] > 0)
1752 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1756 if (free
>= ac
->ac_g_ex
.fe_len
)
1769 * lock the group_info alloc_sem of all the groups
1770 * belonging to the same buddy cache page. This
1771 * make sure other parallel operation on the buddy
1772 * cache doesn't happen whild holding the buddy cache
1775 int ext4_mb_get_buddy_cache_lock(struct super_block
*sb
, ext4_group_t group
)
1779 int blocks_per_page
;
1780 int groups_per_page
;
1781 ext4_group_t first_group
;
1782 struct ext4_group_info
*grp
;
1784 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1786 * the buddy cache inode stores the block bitmap
1787 * and buddy information in consecutive blocks.
1788 * So for each group we need two blocks.
1791 pnum
= block
/ blocks_per_page
;
1792 first_group
= pnum
* blocks_per_page
/ 2;
1794 groups_per_page
= blocks_per_page
>> 1;
1795 if (groups_per_page
== 0)
1796 groups_per_page
= 1;
1797 /* read all groups the page covers into the cache */
1798 for (i
= 0; i
< groups_per_page
; i
++) {
1800 if ((first_group
+ i
) >= EXT4_SB(sb
)->s_groups_count
)
1802 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1803 /* take all groups write allocation
1804 * semaphore. This make sure there is
1805 * no block allocation going on in any
1808 down_write_nested(&grp
->alloc_sem
, i
);
1813 void ext4_mb_put_buddy_cache_lock(struct super_block
*sb
,
1814 ext4_group_t group
, int locked_group
)
1818 int blocks_per_page
;
1819 ext4_group_t first_group
;
1820 struct ext4_group_info
*grp
;
1822 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1824 * the buddy cache inode stores the block bitmap
1825 * and buddy information in consecutive blocks.
1826 * So for each group we need two blocks.
1829 pnum
= block
/ blocks_per_page
;
1830 first_group
= pnum
* blocks_per_page
/ 2;
1831 /* release locks on all the groups */
1832 for (i
= 0; i
< locked_group
; i
++) {
1834 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1835 /* take all groups write allocation
1836 * semaphore. This make sure there is
1837 * no block allocation going on in any
1840 up_write(&grp
->alloc_sem
);
1845 static int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1850 int blocks_per_page
;
1851 int block
, pnum
, poff
;
1852 int num_grp_locked
= 0;
1853 struct ext4_group_info
*this_grp
;
1854 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1855 struct inode
*inode
= sbi
->s_buddy_cache
;
1856 struct page
*page
= NULL
, *bitmap_page
= NULL
;
1858 mb_debug("init group %lu\n", group
);
1859 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1860 this_grp
= ext4_get_group_info(sb
, group
);
1862 * This ensures we don't add group
1863 * to this buddy cache via resize
1865 num_grp_locked
= ext4_mb_get_buddy_cache_lock(sb
, group
);
1866 if (!EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1868 * somebody initialized the group
1869 * return without doing anything
1875 * the buddy cache inode stores the block bitmap
1876 * and buddy information in consecutive blocks.
1877 * So for each group we need two blocks.
1880 pnum
= block
/ blocks_per_page
;
1881 poff
= block
% blocks_per_page
;
1882 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1884 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1885 ret
= ext4_mb_init_cache(page
, NULL
);
1892 if (page
== NULL
|| !PageUptodate(page
)) {
1896 mark_page_accessed(page
);
1898 bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1900 /* init buddy cache */
1902 pnum
= block
/ blocks_per_page
;
1903 poff
= block
% blocks_per_page
;
1904 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1905 if (page
== bitmap_page
) {
1907 * If both the bitmap and buddy are in
1908 * the same page we don't need to force
1913 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1914 ret
= ext4_mb_init_cache(page
, bitmap
);
1921 if (page
== NULL
|| !PageUptodate(page
)) {
1925 mark_page_accessed(page
);
1927 ext4_mb_put_buddy_cache_lock(sb
, group
, num_grp_locked
);
1929 page_cache_release(bitmap_page
);
1931 page_cache_release(page
);
1935 static noinline_for_stack
int
1936 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1943 struct ext4_sb_info
*sbi
;
1944 struct super_block
*sb
;
1945 struct ext4_buddy e4b
;
1950 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1952 /* first, try the goal */
1953 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1954 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1957 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1961 * ac->ac2_order is set only if the fe_len is a power of 2
1962 * if ac2_order is set we also set criteria to 0 so that we
1963 * try exact allocation using buddy.
1965 i
= fls(ac
->ac_g_ex
.fe_len
);
1968 * We search using buddy data only if the order of the request
1969 * is greater than equal to the sbi_s_mb_order2_reqs
1970 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1972 if (i
>= sbi
->s_mb_order2_reqs
) {
1974 * This should tell if fe_len is exactly power of 2
1976 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1977 ac
->ac_2order
= i
- 1;
1980 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
1981 /* if stream allocation is enabled, use global goal */
1982 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
1983 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
1987 if (size
< sbi
->s_mb_stream_request
&&
1988 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1989 /* TBD: may be hot point */
1990 spin_lock(&sbi
->s_md_lock
);
1991 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1992 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1993 spin_unlock(&sbi
->s_md_lock
);
1995 /* Let's just scan groups to find more-less suitable blocks */
1996 cr
= ac
->ac_2order
? 0 : 1;
1998 * cr == 0 try to get exact allocation,
1999 * cr == 3 try to get anything
2002 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2003 ac
->ac_criteria
= cr
;
2005 * searching for the right group start
2006 * from the goal value specified
2008 group
= ac
->ac_g_ex
.fe_group
;
2010 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
2011 struct ext4_group_info
*grp
;
2012 struct ext4_group_desc
*desc
;
2014 if (group
== EXT4_SB(sb
)->s_groups_count
)
2017 /* quick check to skip empty groups */
2018 grp
= ext4_get_group_info(sb
, group
);
2019 if (grp
->bb_free
== 0)
2023 * if the group is already init we check whether it is
2024 * a good group and if not we don't load the buddy
2026 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
2028 * we need full data about the group
2029 * to make a good selection
2031 err
= ext4_mb_init_group(sb
, group
);
2037 * If the particular group doesn't satisfy our
2038 * criteria we continue with the next group
2040 if (!ext4_mb_good_group(ac
, group
, cr
))
2043 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2047 ext4_lock_group(sb
, group
);
2048 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2049 /* someone did allocation from this group */
2050 ext4_unlock_group(sb
, group
);
2051 ext4_mb_release_desc(&e4b
);
2055 ac
->ac_groups_scanned
++;
2056 desc
= ext4_get_group_desc(sb
, group
, NULL
);
2057 if (cr
== 0 || (desc
->bg_flags
&
2058 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
2059 ac
->ac_2order
!= 0))
2060 ext4_mb_simple_scan_group(ac
, &e4b
);
2062 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
2063 ext4_mb_scan_aligned(ac
, &e4b
);
2065 ext4_mb_complex_scan_group(ac
, &e4b
);
2067 ext4_unlock_group(sb
, group
);
2068 ext4_mb_release_desc(&e4b
);
2070 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2075 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2076 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2078 * We've been searching too long. Let's try to allocate
2079 * the best chunk we've found so far
2082 ext4_mb_try_best_found(ac
, &e4b
);
2083 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2085 * Someone more lucky has already allocated it.
2086 * The only thing we can do is just take first
2088 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2090 ac
->ac_b_ex
.fe_group
= 0;
2091 ac
->ac_b_ex
.fe_start
= 0;
2092 ac
->ac_b_ex
.fe_len
= 0;
2093 ac
->ac_status
= AC_STATUS_CONTINUE
;
2094 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2096 atomic_inc(&sbi
->s_mb_lost_chunks
);
2104 #ifdef EXT4_MB_HISTORY
2105 struct ext4_mb_proc_session
{
2106 struct ext4_mb_history
*history
;
2107 struct super_block
*sb
;
2112 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
2113 struct ext4_mb_history
*hs
,
2116 if (hs
== s
->history
+ s
->max
)
2118 if (!first
&& hs
== s
->history
+ s
->start
)
2120 while (hs
->orig
.fe_len
== 0) {
2122 if (hs
== s
->history
+ s
->max
)
2124 if (hs
== s
->history
+ s
->start
)
2130 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
2132 struct ext4_mb_proc_session
*s
= seq
->private;
2133 struct ext4_mb_history
*hs
;
2137 return SEQ_START_TOKEN
;
2138 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2141 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
2145 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
2148 struct ext4_mb_proc_session
*s
= seq
->private;
2149 struct ext4_mb_history
*hs
= v
;
2152 if (v
== SEQ_START_TOKEN
)
2153 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2155 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
2158 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
2160 char buf
[25], buf2
[25], buf3
[25], *fmt
;
2161 struct ext4_mb_history
*hs
= v
;
2163 if (v
== SEQ_START_TOKEN
) {
2164 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
2165 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2166 "pid", "inode", "original", "goal", "result", "found",
2167 "grps", "cr", "flags", "merge", "tail", "broken");
2171 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
2172 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2173 "%-5u %-5s %-5u %-6u\n";
2174 sprintf(buf2
, "%u/%d/%u@%u", hs
->result
.fe_group
,
2175 hs
->result
.fe_start
, hs
->result
.fe_len
,
2176 hs
->result
.fe_logical
);
2177 sprintf(buf
, "%u/%d/%u@%u", hs
->orig
.fe_group
,
2178 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2179 hs
->orig
.fe_logical
);
2180 sprintf(buf3
, "%u/%d/%u@%u", hs
->goal
.fe_group
,
2181 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
2182 hs
->goal
.fe_logical
);
2183 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
2184 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
2185 hs
->merged
? "M" : "", hs
->tail
,
2186 hs
->buddy
? 1 << hs
->buddy
: 0);
2187 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
2188 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
2189 sprintf(buf2
, "%u/%d/%u@%u", hs
->result
.fe_group
,
2190 hs
->result
.fe_start
, hs
->result
.fe_len
,
2191 hs
->result
.fe_logical
);
2192 sprintf(buf
, "%u/%d/%u@%u", hs
->orig
.fe_group
,
2193 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2194 hs
->orig
.fe_logical
);
2195 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
2196 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
2197 sprintf(buf2
, "%u/%d/%u", hs
->result
.fe_group
,
2198 hs
->result
.fe_start
, hs
->result
.fe_len
);
2199 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
2200 hs
->pid
, hs
->ino
, buf2
);
2201 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
2202 sprintf(buf2
, "%u/%d/%u", hs
->result
.fe_group
,
2203 hs
->result
.fe_start
, hs
->result
.fe_len
);
2204 seq_printf(seq
, "%-5u %-8u %-23s free\n",
2205 hs
->pid
, hs
->ino
, buf2
);
2210 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
2214 static struct seq_operations ext4_mb_seq_history_ops
= {
2215 .start
= ext4_mb_seq_history_start
,
2216 .next
= ext4_mb_seq_history_next
,
2217 .stop
= ext4_mb_seq_history_stop
,
2218 .show
= ext4_mb_seq_history_show
,
2221 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
2223 struct super_block
*sb
= PDE(inode
)->data
;
2224 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2225 struct ext4_mb_proc_session
*s
;
2229 if (unlikely(sbi
->s_mb_history
== NULL
))
2231 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
2235 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
2236 s
->history
= kmalloc(size
, GFP_KERNEL
);
2237 if (s
->history
== NULL
) {
2242 spin_lock(&sbi
->s_mb_history_lock
);
2243 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2244 s
->max
= sbi
->s_mb_history_max
;
2245 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2246 spin_unlock(&sbi
->s_mb_history_lock
);
2248 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2250 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2260 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2262 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2263 struct ext4_mb_proc_session
*s
= seq
->private;
2266 return seq_release(inode
, file
);
2269 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2270 const char __user
*buffer
,
2271 size_t count
, loff_t
*ppos
)
2273 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2274 struct ext4_mb_proc_session
*s
= seq
->private;
2275 struct super_block
*sb
= s
->sb
;
2279 if (count
>= sizeof(str
)) {
2280 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2281 "mb_history", (int)sizeof(str
));
2285 if (copy_from_user(str
, buffer
, count
))
2288 value
= simple_strtol(str
, NULL
, 0);
2291 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2296 static struct file_operations ext4_mb_seq_history_fops
= {
2297 .owner
= THIS_MODULE
,
2298 .open
= ext4_mb_seq_history_open
,
2300 .write
= ext4_mb_seq_history_write
,
2301 .llseek
= seq_lseek
,
2302 .release
= ext4_mb_seq_history_release
,
2305 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2307 struct super_block
*sb
= seq
->private;
2308 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2311 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2315 return (void *) ((unsigned long) group
);
2318 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2320 struct super_block
*sb
= seq
->private;
2321 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2325 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2328 return (void *) ((unsigned long) group
);
2331 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2333 struct super_block
*sb
= seq
->private;
2334 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2337 struct ext4_buddy e4b
;
2339 struct ext4_group_info info
;
2340 unsigned short counters
[16];
2345 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2346 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2347 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2348 "group", "free", "frags", "first",
2349 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2350 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2352 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2353 sizeof(struct ext4_group_info
);
2354 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2356 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2359 ext4_lock_group(sb
, group
);
2360 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2361 ext4_unlock_group(sb
, group
);
2362 ext4_mb_release_desc(&e4b
);
2364 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2365 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2366 for (i
= 0; i
<= 13; i
++)
2367 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2368 sg
.info
.bb_counters
[i
] : 0);
2369 seq_printf(seq
, " ]\n");
2374 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2378 static struct seq_operations ext4_mb_seq_groups_ops
= {
2379 .start
= ext4_mb_seq_groups_start
,
2380 .next
= ext4_mb_seq_groups_next
,
2381 .stop
= ext4_mb_seq_groups_stop
,
2382 .show
= ext4_mb_seq_groups_show
,
2385 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2387 struct super_block
*sb
= PDE(inode
)->data
;
2390 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2392 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2399 static struct file_operations ext4_mb_seq_groups_fops
= {
2400 .owner
= THIS_MODULE
,
2401 .open
= ext4_mb_seq_groups_open
,
2403 .llseek
= seq_lseek
,
2404 .release
= seq_release
,
2407 static void ext4_mb_history_release(struct super_block
*sb
)
2409 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2411 if (sbi
->s_proc
!= NULL
) {
2412 remove_proc_entry("mb_groups", sbi
->s_proc
);
2413 remove_proc_entry("mb_history", sbi
->s_proc
);
2415 kfree(sbi
->s_mb_history
);
2418 static void ext4_mb_history_init(struct super_block
*sb
)
2420 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2423 if (sbi
->s_proc
!= NULL
) {
2424 proc_create_data("mb_history", S_IRUGO
, sbi
->s_proc
,
2425 &ext4_mb_seq_history_fops
, sb
);
2426 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2427 &ext4_mb_seq_groups_fops
, sb
);
2430 sbi
->s_mb_history_max
= 1000;
2431 sbi
->s_mb_history_cur
= 0;
2432 spin_lock_init(&sbi
->s_mb_history_lock
);
2433 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2434 sbi
->s_mb_history
= kzalloc(i
, GFP_KERNEL
);
2435 /* if we can't allocate history, then we simple won't use it */
2438 static noinline_for_stack
void
2439 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2441 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2442 struct ext4_mb_history h
;
2444 if (unlikely(sbi
->s_mb_history
== NULL
))
2447 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2451 h
.pid
= current
->pid
;
2452 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2453 h
.orig
= ac
->ac_o_ex
;
2454 h
.result
= ac
->ac_b_ex
;
2455 h
.flags
= ac
->ac_flags
;
2456 h
.found
= ac
->ac_found
;
2457 h
.groups
= ac
->ac_groups_scanned
;
2458 h
.cr
= ac
->ac_criteria
;
2459 h
.tail
= ac
->ac_tail
;
2460 h
.buddy
= ac
->ac_buddy
;
2462 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2463 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2464 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2466 h
.goal
= ac
->ac_g_ex
;
2467 h
.result
= ac
->ac_f_ex
;
2470 spin_lock(&sbi
->s_mb_history_lock
);
2471 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2472 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2473 sbi
->s_mb_history_cur
= 0;
2474 spin_unlock(&sbi
->s_mb_history_lock
);
2478 #define ext4_mb_history_release(sb)
2479 #define ext4_mb_history_init(sb)
2483 /* Create and initialize ext4_group_info data for the given group. */
2484 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2485 struct ext4_group_desc
*desc
)
2489 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2490 struct ext4_group_info
**meta_group_info
;
2493 * First check if this group is the first of a reserved block.
2494 * If it's true, we have to allocate a new table of pointers
2495 * to ext4_group_info structures
2497 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2498 metalen
= sizeof(*meta_group_info
) <<
2499 EXT4_DESC_PER_BLOCK_BITS(sb
);
2500 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2501 if (meta_group_info
== NULL
) {
2502 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2504 goto exit_meta_group_info
;
2506 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2511 * calculate needed size. if change bb_counters size,
2512 * don't forget about ext4_mb_generate_buddy()
2514 len
= offsetof(typeof(**meta_group_info
),
2515 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2518 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2519 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2521 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2522 if (meta_group_info
[i
] == NULL
) {
2523 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2524 goto exit_group_info
;
2526 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2527 &(meta_group_info
[i
]->bb_state
));
2530 * initialize bb_free to be able to skip
2531 * empty groups without initialization
2533 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2534 meta_group_info
[i
]->bb_free
=
2535 ext4_free_blocks_after_init(sb
, group
, desc
);
2537 meta_group_info
[i
]->bb_free
=
2538 ext4_free_blks_count(sb
, desc
);
2541 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2542 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2543 meta_group_info
[i
]->bb_free_root
.rb_node
= NULL
;;
2547 struct buffer_head
*bh
;
2548 meta_group_info
[i
]->bb_bitmap
=
2549 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2550 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2551 bh
= ext4_read_block_bitmap(sb
, group
);
2553 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2562 /* If a meta_group_info table has been allocated, release it now */
2563 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2564 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2565 exit_meta_group_info
:
2567 } /* ext4_mb_add_groupinfo */
2570 * Update an existing group.
2571 * This function is used for online resize
2573 void ext4_mb_update_group_info(struct ext4_group_info
*grp
, ext4_grpblk_t add
)
2575 grp
->bb_free
+= add
;
2578 static int ext4_mb_init_backend(struct super_block
*sb
)
2582 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2583 struct ext4_super_block
*es
= sbi
->s_es
;
2584 int num_meta_group_infos
;
2585 int num_meta_group_infos_max
;
2587 struct ext4_group_info
**meta_group_info
;
2588 struct ext4_group_desc
*desc
;
2590 /* This is the number of blocks used by GDT */
2591 num_meta_group_infos
= (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) -
2592 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2595 * This is the total number of blocks used by GDT including
2596 * the number of reserved blocks for GDT.
2597 * The s_group_info array is allocated with this value
2598 * to allow a clean online resize without a complex
2599 * manipulation of pointer.
2600 * The drawback is the unused memory when no resize
2601 * occurs but it's very low in terms of pages
2602 * (see comments below)
2603 * Need to handle this properly when META_BG resizing is allowed
2605 num_meta_group_infos_max
= num_meta_group_infos
+
2606 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2609 * array_size is the size of s_group_info array. We round it
2610 * to the next power of two because this approximation is done
2611 * internally by kmalloc so we can have some more memory
2612 * for free here (e.g. may be used for META_BG resize).
2615 while (array_size
< sizeof(*sbi
->s_group_info
) *
2616 num_meta_group_infos_max
)
2617 array_size
= array_size
<< 1;
2618 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2619 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2620 * So a two level scheme suffices for now. */
2621 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2622 if (sbi
->s_group_info
== NULL
) {
2623 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2626 sbi
->s_buddy_cache
= new_inode(sb
);
2627 if (sbi
->s_buddy_cache
== NULL
) {
2628 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2631 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2633 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2634 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2635 if ((i
+ 1) == num_meta_group_infos
)
2636 metalen
= sizeof(*meta_group_info
) *
2637 (sbi
->s_groups_count
-
2638 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2639 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2640 if (meta_group_info
== NULL
) {
2641 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2645 sbi
->s_group_info
[i
] = meta_group_info
;
2648 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2649 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2652 "EXT4-fs: can't read descriptor %u\n", i
);
2655 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2663 kfree(ext4_get_group_info(sb
, i
));
2664 i
= num_meta_group_infos
;
2667 kfree(sbi
->s_group_info
[i
]);
2668 iput(sbi
->s_buddy_cache
);
2670 kfree(sbi
->s_group_info
);
2674 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2676 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2682 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2684 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2685 if (sbi
->s_mb_offsets
== NULL
) {
2689 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned int);
2690 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2691 if (sbi
->s_mb_maxs
== NULL
) {
2692 kfree(sbi
->s_mb_maxs
);
2696 /* order 0 is regular bitmap */
2697 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2698 sbi
->s_mb_offsets
[0] = 0;
2702 max
= sb
->s_blocksize
<< 2;
2704 sbi
->s_mb_offsets
[i
] = offset
;
2705 sbi
->s_mb_maxs
[i
] = max
;
2706 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2709 } while (i
<= sb
->s_blocksize_bits
+ 1);
2711 /* init file for buddy data */
2712 ret
= ext4_mb_init_backend(sb
);
2714 kfree(sbi
->s_mb_offsets
);
2715 kfree(sbi
->s_mb_maxs
);
2719 spin_lock_init(&sbi
->s_md_lock
);
2720 spin_lock_init(&sbi
->s_bal_lock
);
2722 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2723 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2724 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2725 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2726 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2727 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2728 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2730 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2731 if (sbi
->s_locality_groups
== NULL
) {
2732 kfree(sbi
->s_mb_offsets
);
2733 kfree(sbi
->s_mb_maxs
);
2736 for_each_possible_cpu(i
) {
2737 struct ext4_locality_group
*lg
;
2738 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2739 mutex_init(&lg
->lg_mutex
);
2740 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2741 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2742 spin_lock_init(&lg
->lg_prealloc_lock
);
2745 ext4_mb_init_per_dev_proc(sb
);
2746 ext4_mb_history_init(sb
);
2749 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2751 printk(KERN_INFO
"EXT4-fs: mballoc enabled\n");
2755 /* need to called with ext4 group lock (ext4_lock_group) */
2756 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2758 struct ext4_prealloc_space
*pa
;
2759 struct list_head
*cur
, *tmp
;
2762 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2763 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2764 list_del(&pa
->pa_group_list
);
2766 kmem_cache_free(ext4_pspace_cachep
, pa
);
2769 mb_debug("mballoc: %u PAs left\n", count
);
2773 int ext4_mb_release(struct super_block
*sb
)
2776 int num_meta_group_infos
;
2777 struct ext4_group_info
*grinfo
;
2778 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2780 if (sbi
->s_group_info
) {
2781 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2782 grinfo
= ext4_get_group_info(sb
, i
);
2784 kfree(grinfo
->bb_bitmap
);
2786 ext4_lock_group(sb
, i
);
2787 ext4_mb_cleanup_pa(grinfo
);
2788 ext4_unlock_group(sb
, i
);
2791 num_meta_group_infos
= (sbi
->s_groups_count
+
2792 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2793 EXT4_DESC_PER_BLOCK_BITS(sb
);
2794 for (i
= 0; i
< num_meta_group_infos
; i
++)
2795 kfree(sbi
->s_group_info
[i
]);
2796 kfree(sbi
->s_group_info
);
2798 kfree(sbi
->s_mb_offsets
);
2799 kfree(sbi
->s_mb_maxs
);
2800 if (sbi
->s_buddy_cache
)
2801 iput(sbi
->s_buddy_cache
);
2802 if (sbi
->s_mb_stats
) {
2804 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2805 atomic_read(&sbi
->s_bal_allocated
),
2806 atomic_read(&sbi
->s_bal_reqs
),
2807 atomic_read(&sbi
->s_bal_success
));
2809 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2810 "%u 2^N hits, %u breaks, %u lost\n",
2811 atomic_read(&sbi
->s_bal_ex_scanned
),
2812 atomic_read(&sbi
->s_bal_goals
),
2813 atomic_read(&sbi
->s_bal_2orders
),
2814 atomic_read(&sbi
->s_bal_breaks
),
2815 atomic_read(&sbi
->s_mb_lost_chunks
));
2817 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2818 sbi
->s_mb_buddies_generated
++,
2819 sbi
->s_mb_generation_time
);
2821 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2822 atomic_read(&sbi
->s_mb_preallocated
),
2823 atomic_read(&sbi
->s_mb_discarded
));
2826 free_percpu(sbi
->s_locality_groups
);
2827 ext4_mb_history_release(sb
);
2828 ext4_mb_destroy_per_dev_proc(sb
);
2834 * This function is called by the jbd2 layer once the commit has finished,
2835 * so we know we can free the blocks that were released with that commit.
2837 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2839 struct super_block
*sb
= journal
->j_private
;
2840 struct ext4_buddy e4b
;
2841 struct ext4_group_info
*db
;
2842 int err
, count
= 0, count2
= 0;
2843 struct ext4_free_data
*entry
;
2844 ext4_fsblk_t discard_block
;
2845 struct list_head
*l
, *ltmp
;
2847 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2848 entry
= list_entry(l
, struct ext4_free_data
, list
);
2850 mb_debug("gonna free %u blocks in group %u (0x%p):",
2851 entry
->count
, entry
->group
, entry
);
2853 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2854 /* we expect to find existing buddy because it's pinned */
2858 /* there are blocks to put in buddy to make them really free */
2859 count
+= entry
->count
;
2861 ext4_lock_group(sb
, entry
->group
);
2862 /* Take it out of per group rb tree */
2863 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2864 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2866 if (!db
->bb_free_root
.rb_node
) {
2867 /* No more items in the per group rb tree
2868 * balance refcounts from ext4_mb_free_metadata()
2870 page_cache_release(e4b
.bd_buddy_page
);
2871 page_cache_release(e4b
.bd_bitmap_page
);
2873 ext4_unlock_group(sb
, entry
->group
);
2874 discard_block
= (ext4_fsblk_t
) entry
->group
* EXT4_BLOCKS_PER_GROUP(sb
)
2876 + le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
2877 trace_mark(ext4_discard_blocks
, "dev %s blk %llu count %u", sb
->s_id
,
2878 (unsigned long long) discard_block
, entry
->count
);
2879 sb_issue_discard(sb
, discard_block
, entry
->count
);
2881 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2882 ext4_mb_release_desc(&e4b
);
2885 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2888 #define EXT4_MB_STATS_NAME "stats"
2889 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2890 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2891 #define EXT4_MB_ORDER2_REQ "order2_req"
2892 #define EXT4_MB_STREAM_REQ "stream_req"
2893 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2895 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2897 #ifdef CONFIG_PROC_FS
2898 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2899 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2900 struct proc_dir_entry
*proc
;
2902 if (sbi
->s_proc
== NULL
)
2905 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME
, mb_stats
);
2906 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, mb_max_to_scan
);
2907 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, mb_min_to_scan
);
2908 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, mb_order2_reqs
);
2909 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ
, mb_stream_request
);
2910 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, mb_group_prealloc
);
2914 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2915 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2916 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2917 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2918 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2919 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2926 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2928 #ifdef CONFIG_PROC_FS
2929 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2931 if (sbi
->s_proc
== NULL
)
2934 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2935 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2936 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2937 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2938 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2939 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2944 int __init
init_ext4_mballoc(void)
2946 ext4_pspace_cachep
=
2947 kmem_cache_create("ext4_prealloc_space",
2948 sizeof(struct ext4_prealloc_space
),
2949 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2950 if (ext4_pspace_cachep
== NULL
)
2954 kmem_cache_create("ext4_alloc_context",
2955 sizeof(struct ext4_allocation_context
),
2956 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2957 if (ext4_ac_cachep
== NULL
) {
2958 kmem_cache_destroy(ext4_pspace_cachep
);
2962 ext4_free_ext_cachep
=
2963 kmem_cache_create("ext4_free_block_extents",
2964 sizeof(struct ext4_free_data
),
2965 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2966 if (ext4_free_ext_cachep
== NULL
) {
2967 kmem_cache_destroy(ext4_pspace_cachep
);
2968 kmem_cache_destroy(ext4_ac_cachep
);
2974 void exit_ext4_mballoc(void)
2976 /* XXX: synchronize_rcu(); */
2977 kmem_cache_destroy(ext4_pspace_cachep
);
2978 kmem_cache_destroy(ext4_ac_cachep
);
2979 kmem_cache_destroy(ext4_free_ext_cachep
);
2984 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2985 * Returns 0 if success or error code
2987 static noinline_for_stack
int
2988 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2989 handle_t
*handle
, unsigned int reserv_blks
)
2991 struct buffer_head
*bitmap_bh
= NULL
;
2992 struct ext4_super_block
*es
;
2993 struct ext4_group_desc
*gdp
;
2994 struct buffer_head
*gdp_bh
;
2995 struct ext4_sb_info
*sbi
;
2996 struct super_block
*sb
;
3000 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3001 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
3009 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
3013 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
3018 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
3022 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
3023 gdp
->bg_free_blocks_count
);
3025 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
3029 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
3030 + ac
->ac_b_ex
.fe_start
3031 + le32_to_cpu(es
->s_first_data_block
);
3033 len
= ac
->ac_b_ex
.fe_len
;
3034 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, len
) ||
3035 in_range(ext4_inode_bitmap(sb
, gdp
), block
, len
) ||
3036 in_range(block
, ext4_inode_table(sb
, gdp
),
3037 EXT4_SB(sb
)->s_itb_per_group
) ||
3038 in_range(block
+ len
- 1, ext4_inode_table(sb
, gdp
),
3039 EXT4_SB(sb
)->s_itb_per_group
)) {
3040 ext4_error(sb
, __func__
,
3041 "Allocating block in system zone - block = %llu",
3043 /* File system mounted not to panic on error
3044 * Fix the bitmap and repeat the block allocation
3045 * We leak some of the blocks here.
3047 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
),
3048 bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3049 ac
->ac_b_ex
.fe_len
);
3050 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3055 #ifdef AGGRESSIVE_CHECK
3058 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3059 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3060 bitmap_bh
->b_data
));
3064 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3065 mb_set_bits(NULL
, bitmap_bh
->b_data
,
3066 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
3067 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
3068 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3069 ext4_free_blks_set(sb
, gdp
,
3070 ext4_free_blocks_after_init(sb
,
3071 ac
->ac_b_ex
.fe_group
, gdp
));
3073 len
= ext4_free_blks_count(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
3074 ext4_free_blks_set(sb
, gdp
, len
);
3075 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
3076 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3077 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
3079 * Now reduce the dirty block count also. Should not go negative
3081 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
3082 /* release all the reserved blocks if non delalloc */
3083 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
3085 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
3086 ac
->ac_b_ex
.fe_len
);
3088 if (sbi
->s_log_groups_per_flex
) {
3089 ext4_group_t flex_group
= ext4_flex_group(sbi
,
3090 ac
->ac_b_ex
.fe_group
);
3091 spin_lock(sb_bgl_lock(sbi
, flex_group
));
3092 sbi
->s_flex_groups
[flex_group
].free_blocks
-= ac
->ac_b_ex
.fe_len
;
3093 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
3096 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3099 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3108 * here we normalize request for locality group
3109 * Group request are normalized to s_strip size if we set the same via mount
3110 * option. If not we set it to s_mb_group_prealloc which can be configured via
3111 * /proc/fs/ext4/<partition>/group_prealloc
3113 * XXX: should we try to preallocate more than the group has now?
3115 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3117 struct super_block
*sb
= ac
->ac_sb
;
3118 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3121 if (EXT4_SB(sb
)->s_stripe
)
3122 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
3124 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3125 mb_debug("#%u: goal %u blocks for locality group\n",
3126 current
->pid
, ac
->ac_g_ex
.fe_len
);
3130 * Normalization means making request better in terms of
3131 * size and alignment
3133 static noinline_for_stack
void
3134 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3135 struct ext4_allocation_request
*ar
)
3139 loff_t size
, orig_size
, start_off
;
3140 ext4_lblk_t start
, orig_start
;
3141 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3142 struct ext4_prealloc_space
*pa
;
3144 /* do normalize only data requests, metadata requests
3145 do not need preallocation */
3146 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3149 /* sometime caller may want exact blocks */
3150 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3153 /* caller may indicate that preallocation isn't
3154 * required (it's a tail, for example) */
3155 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3158 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3159 ext4_mb_normalize_group_request(ac
);
3163 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3165 /* first, let's learn actual file size
3166 * given current request is allocated */
3167 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3168 size
= size
<< bsbits
;
3169 if (size
< i_size_read(ac
->ac_inode
))
3170 size
= i_size_read(ac
->ac_inode
);
3172 /* max size of free chunks */
3175 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3176 (req <= (size) || max <= (chunk_size))
3178 /* first, try to predict filesize */
3179 /* XXX: should this table be tunable? */
3181 if (size
<= 16 * 1024) {
3183 } else if (size
<= 32 * 1024) {
3185 } else if (size
<= 64 * 1024) {
3187 } else if (size
<= 128 * 1024) {
3189 } else if (size
<= 256 * 1024) {
3191 } else if (size
<= 512 * 1024) {
3193 } else if (size
<= 1024 * 1024) {
3195 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3196 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3197 (21 - bsbits
)) << 21;
3198 size
= 2 * 1024 * 1024;
3199 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3200 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3201 (22 - bsbits
)) << 22;
3202 size
= 4 * 1024 * 1024;
3203 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3204 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3205 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3206 (23 - bsbits
)) << 23;
3207 size
= 8 * 1024 * 1024;
3209 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3210 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3212 orig_size
= size
= size
>> bsbits
;
3213 orig_start
= start
= start_off
>> bsbits
;
3215 /* don't cover already allocated blocks in selected range */
3216 if (ar
->pleft
&& start
<= ar
->lleft
) {
3217 size
-= ar
->lleft
+ 1 - start
;
3218 start
= ar
->lleft
+ 1;
3220 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3221 size
-= start
+ size
- ar
->lright
;
3225 /* check we don't cross already preallocated blocks */
3227 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3232 spin_lock(&pa
->pa_lock
);
3233 if (pa
->pa_deleted
) {
3234 spin_unlock(&pa
->pa_lock
);
3238 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3240 /* PA must not overlap original request */
3241 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3242 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3244 /* skip PA normalized request doesn't overlap with */
3245 if (pa
->pa_lstart
>= end
) {
3246 spin_unlock(&pa
->pa_lock
);
3249 if (pa_end
<= start
) {
3250 spin_unlock(&pa
->pa_lock
);
3253 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3255 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3256 BUG_ON(pa_end
< start
);
3260 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3261 BUG_ON(pa
->pa_lstart
> end
);
3262 end
= pa
->pa_lstart
;
3264 spin_unlock(&pa
->pa_lock
);
3269 /* XXX: extra loop to check we really don't overlap preallocations */
3271 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3273 spin_lock(&pa
->pa_lock
);
3274 if (pa
->pa_deleted
== 0) {
3275 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3276 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3278 spin_unlock(&pa
->pa_lock
);
3282 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3283 start
> ac
->ac_o_ex
.fe_logical
) {
3284 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3285 (unsigned long) start
, (unsigned long) size
,
3286 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3288 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3289 start
> ac
->ac_o_ex
.fe_logical
);
3290 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3292 /* now prepare goal request */
3294 /* XXX: is it better to align blocks WRT to logical
3295 * placement or satisfy big request as is */
3296 ac
->ac_g_ex
.fe_logical
= start
;
3297 ac
->ac_g_ex
.fe_len
= size
;
3299 /* define goal start in order to merge */
3300 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3301 /* merge to the right */
3302 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3303 &ac
->ac_f_ex
.fe_group
,
3304 &ac
->ac_f_ex
.fe_start
);
3305 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3307 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3308 /* merge to the left */
3309 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3310 &ac
->ac_f_ex
.fe_group
,
3311 &ac
->ac_f_ex
.fe_start
);
3312 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3315 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3316 (unsigned) orig_size
, (unsigned) start
);
3319 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3321 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3323 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3324 atomic_inc(&sbi
->s_bal_reqs
);
3325 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3326 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3327 atomic_inc(&sbi
->s_bal_success
);
3328 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3329 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3330 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3331 atomic_inc(&sbi
->s_bal_goals
);
3332 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3333 atomic_inc(&sbi
->s_bal_breaks
);
3336 ext4_mb_store_history(ac
);
3340 * use blocks preallocated to inode
3342 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3343 struct ext4_prealloc_space
*pa
)
3349 /* found preallocated blocks, use them */
3350 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3351 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3353 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3354 &ac
->ac_b_ex
.fe_start
);
3355 ac
->ac_b_ex
.fe_len
= len
;
3356 ac
->ac_status
= AC_STATUS_FOUND
;
3359 BUG_ON(start
< pa
->pa_pstart
);
3360 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3361 BUG_ON(pa
->pa_free
< len
);
3364 mb_debug("use %llu/%u from inode pa %p\n", start
, len
, pa
);
3368 * use blocks preallocated to locality group
3370 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3371 struct ext4_prealloc_space
*pa
)
3373 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3375 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3376 &ac
->ac_b_ex
.fe_group
,
3377 &ac
->ac_b_ex
.fe_start
);
3378 ac
->ac_b_ex
.fe_len
= len
;
3379 ac
->ac_status
= AC_STATUS_FOUND
;
3382 /* we don't correct pa_pstart or pa_plen here to avoid
3383 * possible race when the group is being loaded concurrently
3384 * instead we correct pa later, after blocks are marked
3385 * in on-disk bitmap -- see ext4_mb_release_context()
3386 * Other CPUs are prevented from allocating from this pa by lg_mutex
3388 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3392 * Return the prealloc space that have minimal distance
3393 * from the goal block. @cpa is the prealloc
3394 * space that is having currently known minimal distance
3395 * from the goal block.
3397 static struct ext4_prealloc_space
*
3398 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3399 struct ext4_prealloc_space
*pa
,
3400 struct ext4_prealloc_space
*cpa
)
3402 ext4_fsblk_t cur_distance
, new_distance
;
3405 atomic_inc(&pa
->pa_count
);
3408 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3409 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3411 if (cur_distance
< new_distance
)
3414 /* drop the previous reference */
3415 atomic_dec(&cpa
->pa_count
);
3416 atomic_inc(&pa
->pa_count
);
3421 * search goal blocks in preallocated space
3423 static noinline_for_stack
int
3424 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3427 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3428 struct ext4_locality_group
*lg
;
3429 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3430 ext4_fsblk_t goal_block
;
3432 /* only data can be preallocated */
3433 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3436 /* first, try per-file preallocation */
3438 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3440 /* all fields in this condition don't change,
3441 * so we can skip locking for them */
3442 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3443 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3446 /* found preallocated blocks, use them */
3447 spin_lock(&pa
->pa_lock
);
3448 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3449 atomic_inc(&pa
->pa_count
);
3450 ext4_mb_use_inode_pa(ac
, pa
);
3451 spin_unlock(&pa
->pa_lock
);
3452 ac
->ac_criteria
= 10;
3456 spin_unlock(&pa
->pa_lock
);
3460 /* can we use group allocation? */
3461 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3464 /* inode may have no locality group for some reason */
3468 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3469 if (order
> PREALLOC_TB_SIZE
- 1)
3470 /* The max size of hash table is PREALLOC_TB_SIZE */
3471 order
= PREALLOC_TB_SIZE
- 1;
3473 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3474 ac
->ac_g_ex
.fe_start
+
3475 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3477 * search for the prealloc space that is having
3478 * minimal distance from the goal block.
3480 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3482 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3484 spin_lock(&pa
->pa_lock
);
3485 if (pa
->pa_deleted
== 0 &&
3486 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3488 cpa
= ext4_mb_check_group_pa(goal_block
,
3491 spin_unlock(&pa
->pa_lock
);
3496 ext4_mb_use_group_pa(ac
, cpa
);
3497 ac
->ac_criteria
= 20;
3504 * the function goes through all block freed in the group
3505 * but not yet committed and marks them used in in-core bitmap.
3506 * buddy must be generated from this bitmap
3507 * Need to be called with ext4 group lock (ext4_lock_group)
3509 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3513 struct ext4_group_info
*grp
;
3514 struct ext4_free_data
*entry
;
3516 grp
= ext4_get_group_info(sb
, group
);
3517 n
= rb_first(&(grp
->bb_free_root
));
3520 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3521 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3522 bitmap
, entry
->start_blk
,
3530 * the function goes through all preallocation in this group and marks them
3531 * used in in-core bitmap. buddy must be generated from this bitmap
3532 * Need to be called with ext4 group lock (ext4_lock_group)
3534 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3537 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3538 struct ext4_prealloc_space
*pa
;
3539 struct list_head
*cur
;
3540 ext4_group_t groupnr
;
3541 ext4_grpblk_t start
;
3542 int preallocated
= 0;
3546 /* all form of preallocation discards first load group,
3547 * so the only competing code is preallocation use.
3548 * we don't need any locking here
3549 * notice we do NOT ignore preallocations with pa_deleted
3550 * otherwise we could leave used blocks available for
3551 * allocation in buddy when concurrent ext4_mb_put_pa()
3552 * is dropping preallocation
3554 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3555 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3556 spin_lock(&pa
->pa_lock
);
3557 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3560 spin_unlock(&pa
->pa_lock
);
3561 if (unlikely(len
== 0))
3563 BUG_ON(groupnr
!= group
);
3564 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3565 bitmap
, start
, len
);
3566 preallocated
+= len
;
3569 mb_debug("prellocated %u for group %u\n", preallocated
, group
);
3572 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3574 struct ext4_prealloc_space
*pa
;
3575 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3576 kmem_cache_free(ext4_pspace_cachep
, pa
);
3580 * drops a reference to preallocated space descriptor
3581 * if this was the last reference and the space is consumed
3583 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3584 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3588 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3591 /* in this short window concurrent discard can set pa_deleted */
3592 spin_lock(&pa
->pa_lock
);
3593 if (pa
->pa_deleted
== 1) {
3594 spin_unlock(&pa
->pa_lock
);
3599 spin_unlock(&pa
->pa_lock
);
3601 /* -1 is to protect from crossing allocation group */
3602 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3607 * P1 (buddy init) P2 (regular allocation)
3608 * find block B in PA
3609 * copy on-disk bitmap to buddy
3610 * mark B in on-disk bitmap
3611 * drop PA from group
3612 * mark all PAs in buddy
3614 * thus, P1 initializes buddy with B available. to prevent this
3615 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3618 ext4_lock_group(sb
, grp
);
3619 list_del(&pa
->pa_group_list
);
3620 ext4_unlock_group(sb
, grp
);
3622 spin_lock(pa
->pa_obj_lock
);
3623 list_del_rcu(&pa
->pa_inode_list
);
3624 spin_unlock(pa
->pa_obj_lock
);
3626 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3630 * creates new preallocated space for given inode
3632 static noinline_for_stack
int
3633 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3635 struct super_block
*sb
= ac
->ac_sb
;
3636 struct ext4_prealloc_space
*pa
;
3637 struct ext4_group_info
*grp
;
3638 struct ext4_inode_info
*ei
;
3640 /* preallocate only when found space is larger then requested */
3641 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3642 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3643 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3645 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3649 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3655 /* we can't allocate as much as normalizer wants.
3656 * so, found space must get proper lstart
3657 * to cover original request */
3658 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3659 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3661 /* we're limited by original request in that
3662 * logical block must be covered any way
3663 * winl is window we can move our chunk within */
3664 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3666 /* also, we should cover whole original request */
3667 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3669 /* the smallest one defines real window */
3670 win
= min(winl
, wins
);
3672 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3673 if (offs
&& offs
< win
)
3676 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3677 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3678 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3681 /* preallocation can change ac_b_ex, thus we store actually
3682 * allocated blocks for history */
3683 ac
->ac_f_ex
= ac
->ac_b_ex
;
3685 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3686 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3687 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3688 pa
->pa_free
= pa
->pa_len
;
3689 atomic_set(&pa
->pa_count
, 1);
3690 spin_lock_init(&pa
->pa_lock
);
3694 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3695 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3697 ext4_mb_use_inode_pa(ac
, pa
);
3698 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3700 ei
= EXT4_I(ac
->ac_inode
);
3701 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3703 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3704 pa
->pa_inode
= ac
->ac_inode
;
3706 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3707 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3708 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3710 spin_lock(pa
->pa_obj_lock
);
3711 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3712 spin_unlock(pa
->pa_obj_lock
);
3718 * creates new preallocated space for locality group inodes belongs to
3720 static noinline_for_stack
int
3721 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3723 struct super_block
*sb
= ac
->ac_sb
;
3724 struct ext4_locality_group
*lg
;
3725 struct ext4_prealloc_space
*pa
;
3726 struct ext4_group_info
*grp
;
3728 /* preallocate only when found space is larger then requested */
3729 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3730 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3731 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3733 BUG_ON(ext4_pspace_cachep
== NULL
);
3734 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3738 /* preallocation can change ac_b_ex, thus we store actually
3739 * allocated blocks for history */
3740 ac
->ac_f_ex
= ac
->ac_b_ex
;
3742 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3743 pa
->pa_lstart
= pa
->pa_pstart
;
3744 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3745 pa
->pa_free
= pa
->pa_len
;
3746 atomic_set(&pa
->pa_count
, 1);
3747 spin_lock_init(&pa
->pa_lock
);
3748 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3752 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3753 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3755 ext4_mb_use_group_pa(ac
, pa
);
3756 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3758 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3762 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3763 pa
->pa_inode
= NULL
;
3765 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3766 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3767 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3770 * We will later add the new pa to the right bucket
3771 * after updating the pa_free in ext4_mb_release_context
3776 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3780 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3781 err
= ext4_mb_new_group_pa(ac
);
3783 err
= ext4_mb_new_inode_pa(ac
);
3788 * finds all unused blocks in on-disk bitmap, frees them in
3789 * in-core bitmap and buddy.
3790 * @pa must be unlinked from inode and group lists, so that
3791 * nobody else can find/use it.
3792 * the caller MUST hold group/inode locks.
3793 * TODO: optimize the case when there are no in-core structures yet
3795 static noinline_for_stack
int
3796 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3797 struct ext4_prealloc_space
*pa
,
3798 struct ext4_allocation_context
*ac
)
3800 struct super_block
*sb
= e4b
->bd_sb
;
3801 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3810 BUG_ON(pa
->pa_deleted
== 0);
3811 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3812 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3813 end
= bit
+ pa
->pa_len
;
3817 ac
->ac_inode
= pa
->pa_inode
;
3818 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3822 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3825 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3826 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3827 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3828 mb_debug(" free preallocated %u/%u in group %u\n",
3829 (unsigned) start
, (unsigned) next
- bit
,
3834 ac
->ac_b_ex
.fe_group
= group
;
3835 ac
->ac_b_ex
.fe_start
= bit
;
3836 ac
->ac_b_ex
.fe_len
= next
- bit
;
3837 ac
->ac_b_ex
.fe_logical
= 0;
3838 ext4_mb_store_history(ac
);
3841 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3844 if (free
!= pa
->pa_free
) {
3845 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3846 pa
, (unsigned long) pa
->pa_lstart
,
3847 (unsigned long) pa
->pa_pstart
,
3848 (unsigned long) pa
->pa_len
);
3849 ext4_grp_locked_error(sb
, group
,
3850 __func__
, "free %u, pa_free %u",
3853 * pa is already deleted so we use the value obtained
3854 * from the bitmap and continue.
3857 atomic_add(free
, &sbi
->s_mb_discarded
);
3862 static noinline_for_stack
int
3863 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3864 struct ext4_prealloc_space
*pa
,
3865 struct ext4_allocation_context
*ac
)
3867 struct super_block
*sb
= e4b
->bd_sb
;
3872 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3874 BUG_ON(pa
->pa_deleted
== 0);
3875 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3876 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3877 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3878 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3882 ac
->ac_inode
= NULL
;
3883 ac
->ac_b_ex
.fe_group
= group
;
3884 ac
->ac_b_ex
.fe_start
= bit
;
3885 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3886 ac
->ac_b_ex
.fe_logical
= 0;
3887 ext4_mb_store_history(ac
);
3894 * releases all preallocations in given group
3896 * first, we need to decide discard policy:
3897 * - when do we discard
3899 * - how many do we discard
3900 * 1) how many requested
3902 static noinline_for_stack
int
3903 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3904 ext4_group_t group
, int needed
)
3906 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3907 struct buffer_head
*bitmap_bh
= NULL
;
3908 struct ext4_prealloc_space
*pa
, *tmp
;
3909 struct ext4_allocation_context
*ac
;
3910 struct list_head list
;
3911 struct ext4_buddy e4b
;
3916 mb_debug("discard preallocation for group %u\n", group
);
3918 if (list_empty(&grp
->bb_prealloc_list
))
3921 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3922 if (bitmap_bh
== NULL
) {
3923 ext4_error(sb
, __func__
, "Error in reading block "
3924 "bitmap for %u", group
);
3928 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3930 ext4_error(sb
, __func__
, "Error in loading buddy "
3931 "information for %u", group
);
3937 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3939 INIT_LIST_HEAD(&list
);
3940 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3942 ext4_lock_group(sb
, group
);
3943 list_for_each_entry_safe(pa
, tmp
,
3944 &grp
->bb_prealloc_list
, pa_group_list
) {
3945 spin_lock(&pa
->pa_lock
);
3946 if (atomic_read(&pa
->pa_count
)) {
3947 spin_unlock(&pa
->pa_lock
);
3951 if (pa
->pa_deleted
) {
3952 spin_unlock(&pa
->pa_lock
);
3956 /* seems this one can be freed ... */
3959 /* we can trust pa_free ... */
3960 free
+= pa
->pa_free
;
3962 spin_unlock(&pa
->pa_lock
);
3964 list_del(&pa
->pa_group_list
);
3965 list_add(&pa
->u
.pa_tmp_list
, &list
);
3968 /* if we still need more blocks and some PAs were used, try again */
3969 if (free
< needed
&& busy
) {
3971 ext4_unlock_group(sb
, group
);
3973 * Yield the CPU here so that we don't get soft lockup
3974 * in non preempt case.
3980 /* found anything to free? */
3981 if (list_empty(&list
)) {
3986 /* now free all selected PAs */
3987 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3989 /* remove from object (inode or locality group) */
3990 spin_lock(pa
->pa_obj_lock
);
3991 list_del_rcu(&pa
->pa_inode_list
);
3992 spin_unlock(pa
->pa_obj_lock
);
3995 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
3997 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
3999 list_del(&pa
->u
.pa_tmp_list
);
4000 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4004 ext4_unlock_group(sb
, group
);
4006 kmem_cache_free(ext4_ac_cachep
, ac
);
4007 ext4_mb_release_desc(&e4b
);
4013 * releases all non-used preallocated blocks for given inode
4015 * It's important to discard preallocations under i_data_sem
4016 * We don't want another block to be served from the prealloc
4017 * space when we are discarding the inode prealloc space.
4019 * FIXME!! Make sure it is valid at all the call sites
4021 void ext4_discard_preallocations(struct inode
*inode
)
4023 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4024 struct super_block
*sb
= inode
->i_sb
;
4025 struct buffer_head
*bitmap_bh
= NULL
;
4026 struct ext4_prealloc_space
*pa
, *tmp
;
4027 struct ext4_allocation_context
*ac
;
4028 ext4_group_t group
= 0;
4029 struct list_head list
;
4030 struct ext4_buddy e4b
;
4033 if (!S_ISREG(inode
->i_mode
)) {
4034 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4038 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
4040 INIT_LIST_HEAD(&list
);
4042 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4044 /* first, collect all pa's in the inode */
4045 spin_lock(&ei
->i_prealloc_lock
);
4046 while (!list_empty(&ei
->i_prealloc_list
)) {
4047 pa
= list_entry(ei
->i_prealloc_list
.next
,
4048 struct ext4_prealloc_space
, pa_inode_list
);
4049 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
4050 spin_lock(&pa
->pa_lock
);
4051 if (atomic_read(&pa
->pa_count
)) {
4052 /* this shouldn't happen often - nobody should
4053 * use preallocation while we're discarding it */
4054 spin_unlock(&pa
->pa_lock
);
4055 spin_unlock(&ei
->i_prealloc_lock
);
4056 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
4058 schedule_timeout_uninterruptible(HZ
);
4062 if (pa
->pa_deleted
== 0) {
4064 spin_unlock(&pa
->pa_lock
);
4065 list_del_rcu(&pa
->pa_inode_list
);
4066 list_add(&pa
->u
.pa_tmp_list
, &list
);
4070 /* someone is deleting pa right now */
4071 spin_unlock(&pa
->pa_lock
);
4072 spin_unlock(&ei
->i_prealloc_lock
);
4074 /* we have to wait here because pa_deleted
4075 * doesn't mean pa is already unlinked from
4076 * the list. as we might be called from
4077 * ->clear_inode() the inode will get freed
4078 * and concurrent thread which is unlinking
4079 * pa from inode's list may access already
4080 * freed memory, bad-bad-bad */
4082 /* XXX: if this happens too often, we can
4083 * add a flag to force wait only in case
4084 * of ->clear_inode(), but not in case of
4085 * regular truncate */
4086 schedule_timeout_uninterruptible(HZ
);
4089 spin_unlock(&ei
->i_prealloc_lock
);
4091 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4092 BUG_ON(pa
->pa_linear
!= 0);
4093 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4095 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4097 ext4_error(sb
, __func__
, "Error in loading buddy "
4098 "information for %u", group
);
4102 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4103 if (bitmap_bh
== NULL
) {
4104 ext4_error(sb
, __func__
, "Error in reading block "
4105 "bitmap for %u", group
);
4106 ext4_mb_release_desc(&e4b
);
4110 ext4_lock_group(sb
, group
);
4111 list_del(&pa
->pa_group_list
);
4112 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
4113 ext4_unlock_group(sb
, group
);
4115 ext4_mb_release_desc(&e4b
);
4118 list_del(&pa
->u
.pa_tmp_list
);
4119 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4122 kmem_cache_free(ext4_ac_cachep
, ac
);
4126 * finds all preallocated spaces and return blocks being freed to them
4127 * if preallocated space becomes full (no block is used from the space)
4128 * then the function frees space in buddy
4129 * XXX: at the moment, truncate (which is the only way to free blocks)
4130 * discards all preallocations
4132 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
4133 struct ext4_buddy
*e4b
,
4134 sector_t block
, int count
)
4136 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
4139 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4141 struct super_block
*sb
= ac
->ac_sb
;
4144 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
4145 " Allocation context details:\n");
4146 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
4147 ac
->ac_status
, ac
->ac_flags
);
4148 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4149 "best %lu/%lu/%lu@%lu cr %d\n",
4150 (unsigned long)ac
->ac_o_ex
.fe_group
,
4151 (unsigned long)ac
->ac_o_ex
.fe_start
,
4152 (unsigned long)ac
->ac_o_ex
.fe_len
,
4153 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4154 (unsigned long)ac
->ac_g_ex
.fe_group
,
4155 (unsigned long)ac
->ac_g_ex
.fe_start
,
4156 (unsigned long)ac
->ac_g_ex
.fe_len
,
4157 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4158 (unsigned long)ac
->ac_b_ex
.fe_group
,
4159 (unsigned long)ac
->ac_b_ex
.fe_start
,
4160 (unsigned long)ac
->ac_b_ex
.fe_len
,
4161 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4162 (int)ac
->ac_criteria
);
4163 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
4165 printk(KERN_ERR
"EXT4-fs: groups: \n");
4166 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
4167 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4168 struct ext4_prealloc_space
*pa
;
4169 ext4_grpblk_t start
;
4170 struct list_head
*cur
;
4171 ext4_lock_group(sb
, i
);
4172 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4173 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4175 spin_lock(&pa
->pa_lock
);
4176 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4178 spin_unlock(&pa
->pa_lock
);
4179 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
4182 ext4_unlock_group(sb
, i
);
4184 if (grp
->bb_free
== 0)
4186 printk(KERN_ERR
"%lu: %d/%d \n",
4187 i
, grp
->bb_free
, grp
->bb_fragments
);
4189 printk(KERN_ERR
"\n");
4192 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4199 * We use locality group preallocation for small size file. The size of the
4200 * file is determined by the current size or the resulting size after
4201 * allocation which ever is larger
4203 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4205 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4207 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4208 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4211 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4214 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
4215 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
4216 size
= max(size
, isize
);
4218 /* don't use group allocation for large files */
4219 if (size
>= sbi
->s_mb_stream_request
)
4222 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4225 BUG_ON(ac
->ac_lg
!= NULL
);
4227 * locality group prealloc space are per cpu. The reason for having
4228 * per cpu locality group is to reduce the contention between block
4229 * request from multiple CPUs.
4231 ac
->ac_lg
= per_cpu_ptr(sbi
->s_locality_groups
, raw_smp_processor_id());
4233 /* we're going to use group allocation */
4234 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4236 /* serialize all allocations in the group */
4237 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4240 static noinline_for_stack
int
4241 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4242 struct ext4_allocation_request
*ar
)
4244 struct super_block
*sb
= ar
->inode
->i_sb
;
4245 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4246 struct ext4_super_block
*es
= sbi
->s_es
;
4250 ext4_grpblk_t block
;
4252 /* we can't allocate > group size */
4255 /* just a dirty hack to filter too big requests */
4256 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4257 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4259 /* start searching from the goal */
4261 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4262 goal
>= ext4_blocks_count(es
))
4263 goal
= le32_to_cpu(es
->s_first_data_block
);
4264 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4266 /* set up allocation goals */
4267 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4268 ac
->ac_b_ex
.fe_group
= 0;
4269 ac
->ac_b_ex
.fe_start
= 0;
4270 ac
->ac_b_ex
.fe_len
= 0;
4271 ac
->ac_status
= AC_STATUS_CONTINUE
;
4272 ac
->ac_groups_scanned
= 0;
4273 ac
->ac_ex_scanned
= 0;
4276 ac
->ac_inode
= ar
->inode
;
4277 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4278 ac
->ac_o_ex
.fe_group
= group
;
4279 ac
->ac_o_ex
.fe_start
= block
;
4280 ac
->ac_o_ex
.fe_len
= len
;
4281 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4282 ac
->ac_g_ex
.fe_group
= group
;
4283 ac
->ac_g_ex
.fe_start
= block
;
4284 ac
->ac_g_ex
.fe_len
= len
;
4285 ac
->ac_f_ex
.fe_len
= 0;
4286 ac
->ac_flags
= ar
->flags
;
4288 ac
->ac_criteria
= 0;
4290 ac
->ac_bitmap_page
= NULL
;
4291 ac
->ac_buddy_page
= NULL
;
4294 /* we have to define context: we'll we work with a file or
4295 * locality group. this is a policy, actually */
4296 ext4_mb_group_or_file(ac
);
4298 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4299 "left: %u/%u, right %u/%u to %swritable\n",
4300 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4301 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4302 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4303 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4304 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4309 static noinline_for_stack
void
4310 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4311 struct ext4_locality_group
*lg
,
4312 int order
, int total_entries
)
4314 ext4_group_t group
= 0;
4315 struct ext4_buddy e4b
;
4316 struct list_head discard_list
;
4317 struct ext4_prealloc_space
*pa
, *tmp
;
4318 struct ext4_allocation_context
*ac
;
4320 mb_debug("discard locality group preallocation\n");
4322 INIT_LIST_HEAD(&discard_list
);
4323 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4325 spin_lock(&lg
->lg_prealloc_lock
);
4326 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4328 spin_lock(&pa
->pa_lock
);
4329 if (atomic_read(&pa
->pa_count
)) {
4331 * This is the pa that we just used
4332 * for block allocation. So don't
4335 spin_unlock(&pa
->pa_lock
);
4338 if (pa
->pa_deleted
) {
4339 spin_unlock(&pa
->pa_lock
);
4342 /* only lg prealloc space */
4343 BUG_ON(!pa
->pa_linear
);
4345 /* seems this one can be freed ... */
4347 spin_unlock(&pa
->pa_lock
);
4349 list_del_rcu(&pa
->pa_inode_list
);
4350 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4353 if (total_entries
<= 5) {
4355 * we want to keep only 5 entries
4356 * allowing it to grow to 8. This
4357 * mak sure we don't call discard
4358 * soon for this list.
4363 spin_unlock(&lg
->lg_prealloc_lock
);
4365 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4367 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4368 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4369 ext4_error(sb
, __func__
, "Error in loading buddy "
4370 "information for %u", group
);
4373 ext4_lock_group(sb
, group
);
4374 list_del(&pa
->pa_group_list
);
4375 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4376 ext4_unlock_group(sb
, group
);
4378 ext4_mb_release_desc(&e4b
);
4379 list_del(&pa
->u
.pa_tmp_list
);
4380 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4383 kmem_cache_free(ext4_ac_cachep
, ac
);
4387 * We have incremented pa_count. So it cannot be freed at this
4388 * point. Also we hold lg_mutex. So no parallel allocation is
4389 * possible from this lg. That means pa_free cannot be updated.
4391 * A parallel ext4_mb_discard_group_preallocations is possible.
4392 * which can cause the lg_prealloc_list to be updated.
4395 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4397 int order
, added
= 0, lg_prealloc_count
= 1;
4398 struct super_block
*sb
= ac
->ac_sb
;
4399 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4400 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4402 order
= fls(pa
->pa_free
) - 1;
4403 if (order
> PREALLOC_TB_SIZE
- 1)
4404 /* The max size of hash table is PREALLOC_TB_SIZE */
4405 order
= PREALLOC_TB_SIZE
- 1;
4406 /* Add the prealloc space to lg */
4408 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4410 spin_lock(&tmp_pa
->pa_lock
);
4411 if (tmp_pa
->pa_deleted
) {
4412 spin_unlock(&pa
->pa_lock
);
4415 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4416 /* Add to the tail of the previous entry */
4417 list_add_tail_rcu(&pa
->pa_inode_list
,
4418 &tmp_pa
->pa_inode_list
);
4421 * we want to count the total
4422 * number of entries in the list
4425 spin_unlock(&tmp_pa
->pa_lock
);
4426 lg_prealloc_count
++;
4429 list_add_tail_rcu(&pa
->pa_inode_list
,
4430 &lg
->lg_prealloc_list
[order
]);
4433 /* Now trim the list to be not more than 8 elements */
4434 if (lg_prealloc_count
> 8) {
4435 ext4_mb_discard_lg_preallocations(sb
, lg
,
4436 order
, lg_prealloc_count
);
4443 * release all resource we used in allocation
4445 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4447 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4449 if (pa
->pa_linear
) {
4450 /* see comment in ext4_mb_use_group_pa() */
4451 spin_lock(&pa
->pa_lock
);
4452 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4453 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4454 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4455 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4456 spin_unlock(&pa
->pa_lock
);
4458 * We want to add the pa to the right bucket.
4459 * Remove it from the list and while adding
4460 * make sure the list to which we are adding
4463 if (likely(pa
->pa_free
)) {
4464 spin_lock(pa
->pa_obj_lock
);
4465 list_del_rcu(&pa
->pa_inode_list
);
4466 spin_unlock(pa
->pa_obj_lock
);
4467 ext4_mb_add_n_trim(ac
);
4470 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4472 if (ac
->ac_bitmap_page
)
4473 page_cache_release(ac
->ac_bitmap_page
);
4474 if (ac
->ac_buddy_page
)
4475 page_cache_release(ac
->ac_buddy_page
);
4476 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4477 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4478 ext4_mb_collect_stats(ac
);
4482 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4488 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4489 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4498 * Main entry point into mballoc to allocate blocks
4499 * it tries to use preallocation first, then falls back
4500 * to usual allocation
4502 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4503 struct ext4_allocation_request
*ar
, int *errp
)
4506 struct ext4_allocation_context
*ac
= NULL
;
4507 struct ext4_sb_info
*sbi
;
4508 struct super_block
*sb
;
4509 ext4_fsblk_t block
= 0;
4510 unsigned int inquota
;
4511 unsigned int reserv_blks
= 0;
4513 sb
= ar
->inode
->i_sb
;
4516 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
) {
4518 * With delalloc we already reserved the blocks
4520 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4521 /* let others to free the space */
4523 ar
->len
= ar
->len
>> 1;
4529 reserv_blks
= ar
->len
;
4531 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4532 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4541 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4542 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4544 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4551 *errp
= ext4_mb_initialize_context(ac
, ar
);
4557 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4558 if (!ext4_mb_use_preallocated(ac
)) {
4559 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4560 ext4_mb_normalize_request(ac
, ar
);
4562 /* allocate space in core */
4563 ext4_mb_regular_allocator(ac
);
4565 /* as we've just preallocated more space than
4566 * user requested orinally, we store allocated
4567 * space in a special descriptor */
4568 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4569 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4570 ext4_mb_new_preallocation(ac
);
4573 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4574 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4575 if (*errp
== -EAGAIN
) {
4576 ac
->ac_b_ex
.fe_group
= 0;
4577 ac
->ac_b_ex
.fe_start
= 0;
4578 ac
->ac_b_ex
.fe_len
= 0;
4579 ac
->ac_status
= AC_STATUS_CONTINUE
;
4582 ac
->ac_b_ex
.fe_len
= 0;
4584 ext4_mb_show_ac(ac
);
4586 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4587 ar
->len
= ac
->ac_b_ex
.fe_len
;
4590 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4594 ac
->ac_b_ex
.fe_len
= 0;
4596 ext4_mb_show_ac(ac
);
4599 ext4_mb_release_context(ac
);
4602 kmem_cache_free(ext4_ac_cachep
, ac
);
4604 if (ar
->len
< inquota
)
4605 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4611 * We can merge two free data extents only if the physical blocks
4612 * are contiguous, AND the extents were freed by the same transaction,
4613 * AND the blocks are associated with the same group.
4615 static int can_merge(struct ext4_free_data
*entry1
,
4616 struct ext4_free_data
*entry2
)
4618 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4619 (entry1
->group
== entry2
->group
) &&
4620 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4625 static noinline_for_stack
int
4626 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4627 struct ext4_free_data
*new_entry
)
4629 ext4_grpblk_t block
;
4630 struct ext4_free_data
*entry
;
4631 struct ext4_group_info
*db
= e4b
->bd_info
;
4632 struct super_block
*sb
= e4b
->bd_sb
;
4633 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4634 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4635 struct rb_node
*parent
= NULL
, *new_node
;
4637 BUG_ON(!ext4_handle_valid(handle
));
4638 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4639 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4641 new_node
= &new_entry
->node
;
4642 block
= new_entry
->start_blk
;
4645 /* first free block exent. We need to
4646 protect buddy cache from being freed,
4647 * otherwise we'll refresh it from
4648 * on-disk bitmap and lose not-yet-available
4650 page_cache_get(e4b
->bd_buddy_page
);
4651 page_cache_get(e4b
->bd_bitmap_page
);
4655 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4656 if (block
< entry
->start_blk
)
4658 else if (block
>= (entry
->start_blk
+ entry
->count
))
4659 n
= &(*n
)->rb_right
;
4661 ext4_grp_locked_error(sb
, e4b
->bd_group
, __func__
,
4662 "Double free of blocks %d (%d %d)",
4663 block
, entry
->start_blk
, entry
->count
);
4668 rb_link_node(new_node
, parent
, n
);
4669 rb_insert_color(new_node
, &db
->bb_free_root
);
4671 /* Now try to see the extent can be merged to left and right */
4672 node
= rb_prev(new_node
);
4674 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4675 if (can_merge(entry
, new_entry
)) {
4676 new_entry
->start_blk
= entry
->start_blk
;
4677 new_entry
->count
+= entry
->count
;
4678 rb_erase(node
, &(db
->bb_free_root
));
4679 spin_lock(&sbi
->s_md_lock
);
4680 list_del(&entry
->list
);
4681 spin_unlock(&sbi
->s_md_lock
);
4682 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4686 node
= rb_next(new_node
);
4688 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4689 if (can_merge(new_entry
, entry
)) {
4690 new_entry
->count
+= entry
->count
;
4691 rb_erase(node
, &(db
->bb_free_root
));
4692 spin_lock(&sbi
->s_md_lock
);
4693 list_del(&entry
->list
);
4694 spin_unlock(&sbi
->s_md_lock
);
4695 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4698 /* Add the extent to transaction's private list */
4699 spin_lock(&sbi
->s_md_lock
);
4700 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4701 spin_unlock(&sbi
->s_md_lock
);
4706 * Main entry point into mballoc to free blocks
4708 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4709 unsigned long block
, unsigned long count
,
4710 int metadata
, unsigned long *freed
)
4712 struct buffer_head
*bitmap_bh
= NULL
;
4713 struct super_block
*sb
= inode
->i_sb
;
4714 struct ext4_allocation_context
*ac
= NULL
;
4715 struct ext4_group_desc
*gdp
;
4716 struct ext4_super_block
*es
;
4717 unsigned int overflow
;
4719 struct buffer_head
*gd_bh
;
4720 ext4_group_t block_group
;
4721 struct ext4_sb_info
*sbi
;
4722 struct ext4_buddy e4b
;
4729 es
= EXT4_SB(sb
)->s_es
;
4730 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4731 block
+ count
< block
||
4732 block
+ count
> ext4_blocks_count(es
)) {
4733 ext4_error(sb
, __func__
,
4734 "Freeing blocks not in datazone - "
4735 "block = %lu, count = %lu", block
, count
);
4739 ext4_debug("freeing block %lu\n", block
);
4741 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4743 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4744 ac
->ac_inode
= inode
;
4750 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4753 * Check to see if we are freeing blocks across a group
4756 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4757 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4760 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4765 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4771 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4772 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4773 in_range(block
, ext4_inode_table(sb
, gdp
),
4774 EXT4_SB(sb
)->s_itb_per_group
) ||
4775 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4776 EXT4_SB(sb
)->s_itb_per_group
)) {
4778 ext4_error(sb
, __func__
,
4779 "Freeing blocks in system zone - "
4780 "Block = %lu, count = %lu", block
, count
);
4781 /* err = 0. ext4_std_error should be a no op */
4785 BUFFER_TRACE(bitmap_bh
, "getting write access");
4786 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4791 * We are about to modify some metadata. Call the journal APIs
4792 * to unshare ->b_data if a currently-committing transaction is
4795 BUFFER_TRACE(gd_bh
, "get_write_access");
4796 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4799 #ifdef AGGRESSIVE_CHECK
4802 for (i
= 0; i
< count
; i
++)
4803 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4807 ac
->ac_b_ex
.fe_group
= block_group
;
4808 ac
->ac_b_ex
.fe_start
= bit
;
4809 ac
->ac_b_ex
.fe_len
= count
;
4810 ext4_mb_store_history(ac
);
4813 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4816 if (metadata
&& ext4_handle_valid(handle
)) {
4817 struct ext4_free_data
*new_entry
;
4819 * blocks being freed are metadata. these blocks shouldn't
4820 * be used until this transaction is committed
4822 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4823 new_entry
->start_blk
= bit
;
4824 new_entry
->group
= block_group
;
4825 new_entry
->count
= count
;
4826 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4827 ext4_lock_group(sb
, block_group
);
4828 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4830 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4831 ext4_unlock_group(sb
, block_group
);
4833 ext4_lock_group(sb
, block_group
);
4834 /* need to update group_info->bb_free and bitmap
4835 * with group lock held. generate_buddy look at
4836 * them with group lock_held
4838 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4840 mb_free_blocks(inode
, &e4b
, bit
, count
);
4841 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4842 ext4_unlock_group(sb
, block_group
);
4845 spin_lock(sb_bgl_lock(sbi
, block_group
));
4846 ret
= ext4_free_blks_count(sb
, gdp
) + count
;
4847 ext4_free_blks_set(sb
, gdp
, ret
);
4848 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4849 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4850 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4852 if (sbi
->s_log_groups_per_flex
) {
4853 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4854 spin_lock(sb_bgl_lock(sbi
, flex_group
));
4855 sbi
->s_flex_groups
[flex_group
].free_blocks
+= count
;
4856 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
4859 ext4_mb_release_desc(&e4b
);
4863 /* We dirtied the bitmap block */
4864 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4865 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4867 /* And the group descriptor block */
4868 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4869 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4873 if (overflow
&& !err
) {
4882 ext4_std_error(sb
, err
);
4884 kmem_cache_free(ext4_ac_cachep
, ac
);