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 if (e4b
->alloc_semp
)
1056 up_read(e4b
->alloc_semp
);
1060 static int mb_find_order_for_block(struct ext4_buddy
*e4b
, int block
)
1065 BUG_ON(EXT4_MB_BITMAP(e4b
) == EXT4_MB_BUDDY(e4b
));
1066 BUG_ON(block
>= (1 << (e4b
->bd_blkbits
+ 3)));
1068 bb
= EXT4_MB_BUDDY(e4b
);
1069 while (order
<= e4b
->bd_blkbits
+ 1) {
1071 if (!mb_test_bit(block
, bb
)) {
1072 /* this block is part of buddy of order 'order' */
1075 bb
+= 1 << (e4b
->bd_blkbits
- order
);
1081 static void mb_clear_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1087 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1088 /* fast path: clear whole word at once */
1089 addr
= bm
+ (cur
>> 3);
1095 mb_clear_bit_atomic(lock
, cur
, bm
);
1097 mb_clear_bit(cur
, bm
);
1102 static void mb_set_bits(spinlock_t
*lock
, void *bm
, int cur
, int len
)
1108 if ((cur
& 31) == 0 && (len
- cur
) >= 32) {
1109 /* fast path: set whole word at once */
1110 addr
= bm
+ (cur
>> 3);
1116 mb_set_bit_atomic(lock
, cur
, bm
);
1118 mb_set_bit(cur
, bm
);
1123 static void mb_free_blocks(struct inode
*inode
, struct ext4_buddy
*e4b
,
1124 int first
, int count
)
1131 struct super_block
*sb
= e4b
->bd_sb
;
1133 BUG_ON(first
+ count
> (sb
->s_blocksize
<< 3));
1134 BUG_ON(!ext4_is_group_locked(sb
, e4b
->bd_group
));
1135 mb_check_buddy(e4b
);
1136 mb_free_blocks_double(inode
, e4b
, first
, count
);
1138 e4b
->bd_info
->bb_free
+= count
;
1139 if (first
< e4b
->bd_info
->bb_first_free
)
1140 e4b
->bd_info
->bb_first_free
= first
;
1142 /* let's maintain fragments counter */
1144 block
= !mb_test_bit(first
- 1, EXT4_MB_BITMAP(e4b
));
1145 if (first
+ count
< EXT4_SB(sb
)->s_mb_maxs
[0])
1146 max
= !mb_test_bit(first
+ count
, EXT4_MB_BITMAP(e4b
));
1148 e4b
->bd_info
->bb_fragments
--;
1149 else if (!block
&& !max
)
1150 e4b
->bd_info
->bb_fragments
++;
1152 /* let's maintain buddy itself */
1153 while (count
-- > 0) {
1157 if (!mb_test_bit(block
, EXT4_MB_BITMAP(e4b
))) {
1158 ext4_fsblk_t blocknr
;
1159 blocknr
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
);
1162 le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
1163 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1164 __func__
, "double-free of inode"
1165 " %lu's block %llu(bit %u in group %u)",
1166 inode
? inode
->i_ino
: 0, blocknr
, block
,
1169 mb_clear_bit(block
, EXT4_MB_BITMAP(e4b
));
1170 e4b
->bd_info
->bb_counters
[order
]++;
1172 /* start of the buddy */
1173 buddy
= mb_find_buddy(e4b
, order
, &max
);
1177 if (mb_test_bit(block
, buddy
) ||
1178 mb_test_bit(block
+ 1, buddy
))
1181 /* both the buddies are free, try to coalesce them */
1182 buddy2
= mb_find_buddy(e4b
, order
+ 1, &max
);
1188 /* for special purposes, we don't set
1189 * free bits in bitmap */
1190 mb_set_bit(block
, buddy
);
1191 mb_set_bit(block
+ 1, buddy
);
1193 e4b
->bd_info
->bb_counters
[order
]--;
1194 e4b
->bd_info
->bb_counters
[order
]--;
1198 e4b
->bd_info
->bb_counters
[order
]++;
1200 mb_clear_bit(block
, buddy2
);
1204 mb_check_buddy(e4b
);
1207 static int mb_find_extent(struct ext4_buddy
*e4b
, int order
, int block
,
1208 int needed
, struct ext4_free_extent
*ex
)
1215 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1218 buddy
= mb_find_buddy(e4b
, order
, &max
);
1219 BUG_ON(buddy
== NULL
);
1220 BUG_ON(block
>= max
);
1221 if (mb_test_bit(block
, buddy
)) {
1228 /* FIXME dorp order completely ? */
1229 if (likely(order
== 0)) {
1230 /* find actual order */
1231 order
= mb_find_order_for_block(e4b
, block
);
1232 block
= block
>> order
;
1235 ex
->fe_len
= 1 << order
;
1236 ex
->fe_start
= block
<< order
;
1237 ex
->fe_group
= e4b
->bd_group
;
1239 /* calc difference from given start */
1240 next
= next
- ex
->fe_start
;
1242 ex
->fe_start
+= next
;
1244 while (needed
> ex
->fe_len
&&
1245 (buddy
= mb_find_buddy(e4b
, order
, &max
))) {
1247 if (block
+ 1 >= max
)
1250 next
= (block
+ 1) * (1 << order
);
1251 if (mb_test_bit(next
, EXT4_MB_BITMAP(e4b
)))
1254 ord
= mb_find_order_for_block(e4b
, next
);
1257 block
= next
>> order
;
1258 ex
->fe_len
+= 1 << order
;
1261 BUG_ON(ex
->fe_start
+ ex
->fe_len
> (1 << (e4b
->bd_blkbits
+ 3)));
1265 static int mb_mark_used(struct ext4_buddy
*e4b
, struct ext4_free_extent
*ex
)
1271 int start
= ex
->fe_start
;
1272 int len
= ex
->fe_len
;
1277 BUG_ON(start
+ len
> (e4b
->bd_sb
->s_blocksize
<< 3));
1278 BUG_ON(e4b
->bd_group
!= ex
->fe_group
);
1279 BUG_ON(!ext4_is_group_locked(e4b
->bd_sb
, e4b
->bd_group
));
1280 mb_check_buddy(e4b
);
1281 mb_mark_used_double(e4b
, start
, len
);
1283 e4b
->bd_info
->bb_free
-= len
;
1284 if (e4b
->bd_info
->bb_first_free
== start
)
1285 e4b
->bd_info
->bb_first_free
+= len
;
1287 /* let's maintain fragments counter */
1289 mlen
= !mb_test_bit(start
- 1, EXT4_MB_BITMAP(e4b
));
1290 if (start
+ len
< EXT4_SB(e4b
->bd_sb
)->s_mb_maxs
[0])
1291 max
= !mb_test_bit(start
+ len
, EXT4_MB_BITMAP(e4b
));
1293 e4b
->bd_info
->bb_fragments
++;
1294 else if (!mlen
&& !max
)
1295 e4b
->bd_info
->bb_fragments
--;
1297 /* let's maintain buddy itself */
1299 ord
= mb_find_order_for_block(e4b
, start
);
1301 if (((start
>> ord
) << ord
) == start
&& len
>= (1 << ord
)) {
1302 /* the whole chunk may be allocated at once! */
1304 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1305 BUG_ON((start
>> ord
) >= max
);
1306 mb_set_bit(start
>> ord
, buddy
);
1307 e4b
->bd_info
->bb_counters
[ord
]--;
1314 /* store for history */
1316 ret
= len
| (ord
<< 16);
1318 /* we have to split large buddy */
1320 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1321 mb_set_bit(start
>> ord
, buddy
);
1322 e4b
->bd_info
->bb_counters
[ord
]--;
1325 cur
= (start
>> ord
) & ~1U;
1326 buddy
= mb_find_buddy(e4b
, ord
, &max
);
1327 mb_clear_bit(cur
, buddy
);
1328 mb_clear_bit(cur
+ 1, buddy
);
1329 e4b
->bd_info
->bb_counters
[ord
]++;
1330 e4b
->bd_info
->bb_counters
[ord
]++;
1333 mb_set_bits(sb_bgl_lock(EXT4_SB(e4b
->bd_sb
), ex
->fe_group
),
1334 EXT4_MB_BITMAP(e4b
), ex
->fe_start
, len0
);
1335 mb_check_buddy(e4b
);
1341 * Must be called under group lock!
1343 static void ext4_mb_use_best_found(struct ext4_allocation_context
*ac
,
1344 struct ext4_buddy
*e4b
)
1346 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1349 BUG_ON(ac
->ac_b_ex
.fe_group
!= e4b
->bd_group
);
1350 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1352 ac
->ac_b_ex
.fe_len
= min(ac
->ac_b_ex
.fe_len
, ac
->ac_g_ex
.fe_len
);
1353 ac
->ac_b_ex
.fe_logical
= ac
->ac_g_ex
.fe_logical
;
1354 ret
= mb_mark_used(e4b
, &ac
->ac_b_ex
);
1356 /* preallocation can change ac_b_ex, thus we store actually
1357 * allocated blocks for history */
1358 ac
->ac_f_ex
= ac
->ac_b_ex
;
1360 ac
->ac_status
= AC_STATUS_FOUND
;
1361 ac
->ac_tail
= ret
& 0xffff;
1362 ac
->ac_buddy
= ret
>> 16;
1365 * take the page reference. We want the page to be pinned
1366 * so that we don't get a ext4_mb_init_cache_call for this
1367 * group until we update the bitmap. That would mean we
1368 * double allocate blocks. The reference is dropped
1369 * in ext4_mb_release_context
1371 ac
->ac_bitmap_page
= e4b
->bd_bitmap_page
;
1372 get_page(ac
->ac_bitmap_page
);
1373 ac
->ac_buddy_page
= e4b
->bd_buddy_page
;
1374 get_page(ac
->ac_buddy_page
);
1375 /* on allocation we use ac to track the held semaphore */
1376 ac
->alloc_semp
= e4b
->alloc_semp
;
1377 e4b
->alloc_semp
= NULL
;
1378 /* store last allocated for subsequent stream allocation */
1379 if ((ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1380 spin_lock(&sbi
->s_md_lock
);
1381 sbi
->s_mb_last_group
= ac
->ac_f_ex
.fe_group
;
1382 sbi
->s_mb_last_start
= ac
->ac_f_ex
.fe_start
;
1383 spin_unlock(&sbi
->s_md_lock
);
1388 * regular allocator, for general purposes allocation
1391 static void ext4_mb_check_limits(struct ext4_allocation_context
*ac
,
1392 struct ext4_buddy
*e4b
,
1395 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1396 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1397 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1398 struct ext4_free_extent ex
;
1401 if (ac
->ac_status
== AC_STATUS_FOUND
)
1404 * We don't want to scan for a whole year
1406 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
&&
1407 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1408 ac
->ac_status
= AC_STATUS_BREAK
;
1413 * Haven't found good chunk so far, let's continue
1415 if (bex
->fe_len
< gex
->fe_len
)
1418 if ((finish_group
|| ac
->ac_found
> sbi
->s_mb_min_to_scan
)
1419 && bex
->fe_group
== e4b
->bd_group
) {
1420 /* recheck chunk's availability - we don't know
1421 * when it was found (within this lock-unlock
1423 max
= mb_find_extent(e4b
, 0, bex
->fe_start
, gex
->fe_len
, &ex
);
1424 if (max
>= gex
->fe_len
) {
1425 ext4_mb_use_best_found(ac
, e4b
);
1432 * The routine checks whether found extent is good enough. If it is,
1433 * then the extent gets marked used and flag is set to the context
1434 * to stop scanning. Otherwise, the extent is compared with the
1435 * previous found extent and if new one is better, then it's stored
1436 * in the context. Later, the best found extent will be used, if
1437 * mballoc can't find good enough extent.
1439 * FIXME: real allocation policy is to be designed yet!
1441 static void ext4_mb_measure_extent(struct ext4_allocation_context
*ac
,
1442 struct ext4_free_extent
*ex
,
1443 struct ext4_buddy
*e4b
)
1445 struct ext4_free_extent
*bex
= &ac
->ac_b_ex
;
1446 struct ext4_free_extent
*gex
= &ac
->ac_g_ex
;
1448 BUG_ON(ex
->fe_len
<= 0);
1449 BUG_ON(ex
->fe_len
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1450 BUG_ON(ex
->fe_start
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
1451 BUG_ON(ac
->ac_status
!= AC_STATUS_CONTINUE
);
1456 * The special case - take what you catch first
1458 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
1460 ext4_mb_use_best_found(ac
, e4b
);
1465 * Let's check whether the chuck is good enough
1467 if (ex
->fe_len
== gex
->fe_len
) {
1469 ext4_mb_use_best_found(ac
, e4b
);
1474 * If this is first found extent, just store it in the context
1476 if (bex
->fe_len
== 0) {
1482 * If new found extent is better, store it in the context
1484 if (bex
->fe_len
< gex
->fe_len
) {
1485 /* if the request isn't satisfied, any found extent
1486 * larger than previous best one is better */
1487 if (ex
->fe_len
> bex
->fe_len
)
1489 } else if (ex
->fe_len
> gex
->fe_len
) {
1490 /* if the request is satisfied, then we try to find
1491 * an extent that still satisfy the request, but is
1492 * smaller than previous one */
1493 if (ex
->fe_len
< bex
->fe_len
)
1497 ext4_mb_check_limits(ac
, e4b
, 0);
1500 static int ext4_mb_try_best_found(struct ext4_allocation_context
*ac
,
1501 struct ext4_buddy
*e4b
)
1503 struct ext4_free_extent ex
= ac
->ac_b_ex
;
1504 ext4_group_t group
= ex
.fe_group
;
1508 BUG_ON(ex
.fe_len
<= 0);
1509 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1513 ext4_lock_group(ac
->ac_sb
, group
);
1514 max
= mb_find_extent(e4b
, 0, ex
.fe_start
, ex
.fe_len
, &ex
);
1518 ext4_mb_use_best_found(ac
, e4b
);
1521 ext4_unlock_group(ac
->ac_sb
, group
);
1522 ext4_mb_release_desc(e4b
);
1527 static int ext4_mb_find_by_goal(struct ext4_allocation_context
*ac
,
1528 struct ext4_buddy
*e4b
)
1530 ext4_group_t group
= ac
->ac_g_ex
.fe_group
;
1533 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
1534 struct ext4_super_block
*es
= sbi
->s_es
;
1535 struct ext4_free_extent ex
;
1537 if (!(ac
->ac_flags
& EXT4_MB_HINT_TRY_GOAL
))
1540 err
= ext4_mb_load_buddy(ac
->ac_sb
, group
, e4b
);
1544 ext4_lock_group(ac
->ac_sb
, group
);
1545 max
= mb_find_extent(e4b
, 0, ac
->ac_g_ex
.fe_start
,
1546 ac
->ac_g_ex
.fe_len
, &ex
);
1548 if (max
>= ac
->ac_g_ex
.fe_len
&& ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
) {
1551 start
= (e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
)) +
1552 ex
.fe_start
+ le32_to_cpu(es
->s_first_data_block
);
1553 /* use do_div to get remainder (would be 64-bit modulo) */
1554 if (do_div(start
, sbi
->s_stripe
) == 0) {
1557 ext4_mb_use_best_found(ac
, e4b
);
1559 } else if (max
>= ac
->ac_g_ex
.fe_len
) {
1560 BUG_ON(ex
.fe_len
<= 0);
1561 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1562 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1565 ext4_mb_use_best_found(ac
, e4b
);
1566 } else if (max
> 0 && (ac
->ac_flags
& EXT4_MB_HINT_MERGE
)) {
1567 /* Sometimes, caller may want to merge even small
1568 * number of blocks to an existing extent */
1569 BUG_ON(ex
.fe_len
<= 0);
1570 BUG_ON(ex
.fe_group
!= ac
->ac_g_ex
.fe_group
);
1571 BUG_ON(ex
.fe_start
!= ac
->ac_g_ex
.fe_start
);
1574 ext4_mb_use_best_found(ac
, e4b
);
1576 ext4_unlock_group(ac
->ac_sb
, group
);
1577 ext4_mb_release_desc(e4b
);
1583 * The routine scans buddy structures (not bitmap!) from given order
1584 * to max order and tries to find big enough chunk to satisfy the req
1586 static void ext4_mb_simple_scan_group(struct ext4_allocation_context
*ac
,
1587 struct ext4_buddy
*e4b
)
1589 struct super_block
*sb
= ac
->ac_sb
;
1590 struct ext4_group_info
*grp
= e4b
->bd_info
;
1596 BUG_ON(ac
->ac_2order
<= 0);
1597 for (i
= ac
->ac_2order
; i
<= sb
->s_blocksize_bits
+ 1; i
++) {
1598 if (grp
->bb_counters
[i
] == 0)
1601 buddy
= mb_find_buddy(e4b
, i
, &max
);
1602 BUG_ON(buddy
== NULL
);
1604 k
= mb_find_next_zero_bit(buddy
, max
, 0);
1609 ac
->ac_b_ex
.fe_len
= 1 << i
;
1610 ac
->ac_b_ex
.fe_start
= k
<< i
;
1611 ac
->ac_b_ex
.fe_group
= e4b
->bd_group
;
1613 ext4_mb_use_best_found(ac
, e4b
);
1615 BUG_ON(ac
->ac_b_ex
.fe_len
!= ac
->ac_g_ex
.fe_len
);
1617 if (EXT4_SB(sb
)->s_mb_stats
)
1618 atomic_inc(&EXT4_SB(sb
)->s_bal_2orders
);
1625 * The routine scans the group and measures all found extents.
1626 * In order to optimize scanning, caller must pass number of
1627 * free blocks in the group, so the routine can know upper limit.
1629 static void ext4_mb_complex_scan_group(struct ext4_allocation_context
*ac
,
1630 struct ext4_buddy
*e4b
)
1632 struct super_block
*sb
= ac
->ac_sb
;
1633 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1634 struct ext4_free_extent ex
;
1638 free
= e4b
->bd_info
->bb_free
;
1641 i
= e4b
->bd_info
->bb_first_free
;
1643 while (free
&& ac
->ac_status
== AC_STATUS_CONTINUE
) {
1644 i
= mb_find_next_zero_bit(bitmap
,
1645 EXT4_BLOCKS_PER_GROUP(sb
), i
);
1646 if (i
>= EXT4_BLOCKS_PER_GROUP(sb
)) {
1648 * IF we have corrupt bitmap, we won't find any
1649 * free blocks even though group info says we
1650 * we have free blocks
1652 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1653 __func__
, "%d free blocks as per "
1654 "group info. But bitmap says 0",
1659 mb_find_extent(e4b
, 0, i
, ac
->ac_g_ex
.fe_len
, &ex
);
1660 BUG_ON(ex
.fe_len
<= 0);
1661 if (free
< ex
.fe_len
) {
1662 ext4_grp_locked_error(sb
, e4b
->bd_group
,
1663 __func__
, "%d free blocks as per "
1664 "group info. But got %d blocks",
1667 * The number of free blocks differs. This mostly
1668 * indicate that the bitmap is corrupt. So exit
1669 * without claiming the space.
1674 ext4_mb_measure_extent(ac
, &ex
, e4b
);
1680 ext4_mb_check_limits(ac
, e4b
, 1);
1684 * This is a special case for storages like raid5
1685 * we try to find stripe-aligned chunks for stripe-size requests
1686 * XXX should do so at least for multiples of stripe size as well
1688 static void ext4_mb_scan_aligned(struct ext4_allocation_context
*ac
,
1689 struct ext4_buddy
*e4b
)
1691 struct super_block
*sb
= ac
->ac_sb
;
1692 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1693 void *bitmap
= EXT4_MB_BITMAP(e4b
);
1694 struct ext4_free_extent ex
;
1695 ext4_fsblk_t first_group_block
;
1700 BUG_ON(sbi
->s_stripe
== 0);
1702 /* find first stripe-aligned block in group */
1703 first_group_block
= e4b
->bd_group
* EXT4_BLOCKS_PER_GROUP(sb
)
1704 + le32_to_cpu(sbi
->s_es
->s_first_data_block
);
1705 a
= first_group_block
+ sbi
->s_stripe
- 1;
1706 do_div(a
, sbi
->s_stripe
);
1707 i
= (a
* sbi
->s_stripe
) - first_group_block
;
1709 while (i
< EXT4_BLOCKS_PER_GROUP(sb
)) {
1710 if (!mb_test_bit(i
, bitmap
)) {
1711 max
= mb_find_extent(e4b
, 0, i
, sbi
->s_stripe
, &ex
);
1712 if (max
>= sbi
->s_stripe
) {
1715 ext4_mb_use_best_found(ac
, e4b
);
1723 static int ext4_mb_good_group(struct ext4_allocation_context
*ac
,
1724 ext4_group_t group
, int cr
)
1726 unsigned free
, fragments
;
1728 struct ext4_group_desc
*desc
;
1729 struct ext4_group_info
*grp
= ext4_get_group_info(ac
->ac_sb
, group
);
1731 BUG_ON(cr
< 0 || cr
>= 4);
1732 BUG_ON(EXT4_MB_GRP_NEED_INIT(grp
));
1734 free
= grp
->bb_free
;
1735 fragments
= grp
->bb_fragments
;
1743 BUG_ON(ac
->ac_2order
== 0);
1744 /* If this group is uninitialized, skip it initially */
1745 desc
= ext4_get_group_desc(ac
->ac_sb
, group
, NULL
);
1746 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
))
1749 bits
= ac
->ac_sb
->s_blocksize_bits
+ 1;
1750 for (i
= ac
->ac_2order
; i
<= bits
; i
++)
1751 if (grp
->bb_counters
[i
] > 0)
1755 if ((free
/ fragments
) >= ac
->ac_g_ex
.fe_len
)
1759 if (free
>= ac
->ac_g_ex
.fe_len
)
1772 * lock the group_info alloc_sem of all the groups
1773 * belonging to the same buddy cache page. This
1774 * make sure other parallel operation on the buddy
1775 * cache doesn't happen whild holding the buddy cache
1778 int ext4_mb_get_buddy_cache_lock(struct super_block
*sb
, ext4_group_t group
)
1782 int blocks_per_page
;
1783 int groups_per_page
;
1784 ext4_group_t first_group
;
1785 struct ext4_group_info
*grp
;
1787 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1789 * the buddy cache inode stores the block bitmap
1790 * and buddy information in consecutive blocks.
1791 * So for each group we need two blocks.
1794 pnum
= block
/ blocks_per_page
;
1795 first_group
= pnum
* blocks_per_page
/ 2;
1797 groups_per_page
= blocks_per_page
>> 1;
1798 if (groups_per_page
== 0)
1799 groups_per_page
= 1;
1800 /* read all groups the page covers into the cache */
1801 for (i
= 0; i
< groups_per_page
; i
++) {
1803 if ((first_group
+ i
) >= EXT4_SB(sb
)->s_groups_count
)
1805 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1806 /* take all groups write allocation
1807 * semaphore. This make sure there is
1808 * no block allocation going on in any
1811 down_write_nested(&grp
->alloc_sem
, i
);
1816 void ext4_mb_put_buddy_cache_lock(struct super_block
*sb
,
1817 ext4_group_t group
, int locked_group
)
1821 int blocks_per_page
;
1822 ext4_group_t first_group
;
1823 struct ext4_group_info
*grp
;
1825 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1827 * the buddy cache inode stores the block bitmap
1828 * and buddy information in consecutive blocks.
1829 * So for each group we need two blocks.
1832 pnum
= block
/ blocks_per_page
;
1833 first_group
= pnum
* blocks_per_page
/ 2;
1834 /* release locks on all the groups */
1835 for (i
= 0; i
< locked_group
; i
++) {
1837 grp
= ext4_get_group_info(sb
, first_group
+ i
);
1838 /* take all groups write allocation
1839 * semaphore. This make sure there is
1840 * no block allocation going on in any
1843 up_write(&grp
->alloc_sem
);
1848 static int ext4_mb_init_group(struct super_block
*sb
, ext4_group_t group
)
1853 int blocks_per_page
;
1854 int block
, pnum
, poff
;
1855 int num_grp_locked
= 0;
1856 struct ext4_group_info
*this_grp
;
1857 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
1858 struct inode
*inode
= sbi
->s_buddy_cache
;
1859 struct page
*page
= NULL
, *bitmap_page
= NULL
;
1861 mb_debug("init group %lu\n", group
);
1862 blocks_per_page
= PAGE_CACHE_SIZE
/ sb
->s_blocksize
;
1863 this_grp
= ext4_get_group_info(sb
, group
);
1865 * This ensures we don't add group
1866 * to this buddy cache via resize
1868 num_grp_locked
= ext4_mb_get_buddy_cache_lock(sb
, group
);
1869 if (!EXT4_MB_GRP_NEED_INIT(this_grp
)) {
1871 * somebody initialized the group
1872 * return without doing anything
1878 * the buddy cache inode stores the block bitmap
1879 * and buddy information in consecutive blocks.
1880 * So for each group we need two blocks.
1883 pnum
= block
/ blocks_per_page
;
1884 poff
= block
% blocks_per_page
;
1885 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1887 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1888 ret
= ext4_mb_init_cache(page
, NULL
);
1895 if (page
== NULL
|| !PageUptodate(page
)) {
1899 mark_page_accessed(page
);
1901 bitmap
= page_address(page
) + (poff
* sb
->s_blocksize
);
1903 /* init buddy cache */
1905 pnum
= block
/ blocks_per_page
;
1906 poff
= block
% blocks_per_page
;
1907 page
= find_or_create_page(inode
->i_mapping
, pnum
, GFP_NOFS
);
1908 if (page
== bitmap_page
) {
1910 * If both the bitmap and buddy are in
1911 * the same page we don't need to force
1916 BUG_ON(page
->mapping
!= inode
->i_mapping
);
1917 ret
= ext4_mb_init_cache(page
, bitmap
);
1924 if (page
== NULL
|| !PageUptodate(page
)) {
1928 mark_page_accessed(page
);
1930 ext4_mb_put_buddy_cache_lock(sb
, group
, num_grp_locked
);
1932 page_cache_release(bitmap_page
);
1934 page_cache_release(page
);
1938 static noinline_for_stack
int
1939 ext4_mb_regular_allocator(struct ext4_allocation_context
*ac
)
1946 struct ext4_sb_info
*sbi
;
1947 struct super_block
*sb
;
1948 struct ext4_buddy e4b
;
1953 BUG_ON(ac
->ac_status
== AC_STATUS_FOUND
);
1955 /* first, try the goal */
1956 err
= ext4_mb_find_by_goal(ac
, &e4b
);
1957 if (err
|| ac
->ac_status
== AC_STATUS_FOUND
)
1960 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
1964 * ac->ac2_order is set only if the fe_len is a power of 2
1965 * if ac2_order is set we also set criteria to 0 so that we
1966 * try exact allocation using buddy.
1968 i
= fls(ac
->ac_g_ex
.fe_len
);
1971 * We search using buddy data only if the order of the request
1972 * is greater than equal to the sbi_s_mb_order2_reqs
1973 * You can tune it via /proc/fs/ext4/<partition>/order2_req
1975 if (i
>= sbi
->s_mb_order2_reqs
) {
1977 * This should tell if fe_len is exactly power of 2
1979 if ((ac
->ac_g_ex
.fe_len
& (~(1 << (i
- 1)))) == 0)
1980 ac
->ac_2order
= i
- 1;
1983 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
1984 /* if stream allocation is enabled, use global goal */
1985 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
1986 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
1990 if (size
< sbi
->s_mb_stream_request
&&
1991 (ac
->ac_flags
& EXT4_MB_HINT_DATA
)) {
1992 /* TBD: may be hot point */
1993 spin_lock(&sbi
->s_md_lock
);
1994 ac
->ac_g_ex
.fe_group
= sbi
->s_mb_last_group
;
1995 ac
->ac_g_ex
.fe_start
= sbi
->s_mb_last_start
;
1996 spin_unlock(&sbi
->s_md_lock
);
1998 /* Let's just scan groups to find more-less suitable blocks */
1999 cr
= ac
->ac_2order
? 0 : 1;
2001 * cr == 0 try to get exact allocation,
2002 * cr == 3 try to get anything
2005 for (; cr
< 4 && ac
->ac_status
== AC_STATUS_CONTINUE
; cr
++) {
2006 ac
->ac_criteria
= cr
;
2008 * searching for the right group start
2009 * from the goal value specified
2011 group
= ac
->ac_g_ex
.fe_group
;
2013 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; group
++, i
++) {
2014 struct ext4_group_info
*grp
;
2015 struct ext4_group_desc
*desc
;
2017 if (group
== EXT4_SB(sb
)->s_groups_count
)
2020 /* quick check to skip empty groups */
2021 grp
= ext4_get_group_info(sb
, group
);
2022 if (grp
->bb_free
== 0)
2026 * if the group is already init we check whether it is
2027 * a good group and if not we don't load the buddy
2029 if (EXT4_MB_GRP_NEED_INIT(grp
)) {
2031 * we need full data about the group
2032 * to make a good selection
2034 err
= ext4_mb_init_group(sb
, group
);
2040 * If the particular group doesn't satisfy our
2041 * criteria we continue with the next group
2043 if (!ext4_mb_good_group(ac
, group
, cr
))
2046 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2050 ext4_lock_group(sb
, group
);
2051 if (!ext4_mb_good_group(ac
, group
, cr
)) {
2052 /* someone did allocation from this group */
2053 ext4_unlock_group(sb
, group
);
2054 ext4_mb_release_desc(&e4b
);
2058 ac
->ac_groups_scanned
++;
2059 desc
= ext4_get_group_desc(sb
, group
, NULL
);
2060 if (cr
== 0 || (desc
->bg_flags
&
2061 cpu_to_le16(EXT4_BG_BLOCK_UNINIT
) &&
2062 ac
->ac_2order
!= 0))
2063 ext4_mb_simple_scan_group(ac
, &e4b
);
2065 ac
->ac_g_ex
.fe_len
== sbi
->s_stripe
)
2066 ext4_mb_scan_aligned(ac
, &e4b
);
2068 ext4_mb_complex_scan_group(ac
, &e4b
);
2070 ext4_unlock_group(sb
, group
);
2071 ext4_mb_release_desc(&e4b
);
2073 if (ac
->ac_status
!= AC_STATUS_CONTINUE
)
2078 if (ac
->ac_b_ex
.fe_len
> 0 && ac
->ac_status
!= AC_STATUS_FOUND
&&
2079 !(ac
->ac_flags
& EXT4_MB_HINT_FIRST
)) {
2081 * We've been searching too long. Let's try to allocate
2082 * the best chunk we've found so far
2085 ext4_mb_try_best_found(ac
, &e4b
);
2086 if (ac
->ac_status
!= AC_STATUS_FOUND
) {
2088 * Someone more lucky has already allocated it.
2089 * The only thing we can do is just take first
2091 printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2093 ac
->ac_b_ex
.fe_group
= 0;
2094 ac
->ac_b_ex
.fe_start
= 0;
2095 ac
->ac_b_ex
.fe_len
= 0;
2096 ac
->ac_status
= AC_STATUS_CONTINUE
;
2097 ac
->ac_flags
|= EXT4_MB_HINT_FIRST
;
2099 atomic_inc(&sbi
->s_mb_lost_chunks
);
2107 #ifdef EXT4_MB_HISTORY
2108 struct ext4_mb_proc_session
{
2109 struct ext4_mb_history
*history
;
2110 struct super_block
*sb
;
2115 static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session
*s
,
2116 struct ext4_mb_history
*hs
,
2119 if (hs
== s
->history
+ s
->max
)
2121 if (!first
&& hs
== s
->history
+ s
->start
)
2123 while (hs
->orig
.fe_len
== 0) {
2125 if (hs
== s
->history
+ s
->max
)
2127 if (hs
== s
->history
+ s
->start
)
2133 static void *ext4_mb_seq_history_start(struct seq_file
*seq
, loff_t
*pos
)
2135 struct ext4_mb_proc_session
*s
= seq
->private;
2136 struct ext4_mb_history
*hs
;
2140 return SEQ_START_TOKEN
;
2141 hs
= ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2144 while (--l
&& (hs
= ext4_mb_history_skip_empty(s
, ++hs
, 0)) != NULL
);
2148 static void *ext4_mb_seq_history_next(struct seq_file
*seq
, void *v
,
2151 struct ext4_mb_proc_session
*s
= seq
->private;
2152 struct ext4_mb_history
*hs
= v
;
2155 if (v
== SEQ_START_TOKEN
)
2156 return ext4_mb_history_skip_empty(s
, s
->history
+ s
->start
, 1);
2158 return ext4_mb_history_skip_empty(s
, ++hs
, 0);
2161 static int ext4_mb_seq_history_show(struct seq_file
*seq
, void *v
)
2163 char buf
[25], buf2
[25], buf3
[25], *fmt
;
2164 struct ext4_mb_history
*hs
= v
;
2166 if (v
== SEQ_START_TOKEN
) {
2167 seq_printf(seq
, "%-5s %-8s %-23s %-23s %-23s %-5s "
2168 "%-5s %-2s %-5s %-5s %-5s %-6s\n",
2169 "pid", "inode", "original", "goal", "result", "found",
2170 "grps", "cr", "flags", "merge", "tail", "broken");
2174 if (hs
->op
== EXT4_MB_HISTORY_ALLOC
) {
2175 fmt
= "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
2176 "%-5u %-5s %-5u %-6u\n";
2177 sprintf(buf2
, "%u/%d/%u@%u", hs
->result
.fe_group
,
2178 hs
->result
.fe_start
, hs
->result
.fe_len
,
2179 hs
->result
.fe_logical
);
2180 sprintf(buf
, "%u/%d/%u@%u", hs
->orig
.fe_group
,
2181 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2182 hs
->orig
.fe_logical
);
2183 sprintf(buf3
, "%u/%d/%u@%u", hs
->goal
.fe_group
,
2184 hs
->goal
.fe_start
, hs
->goal
.fe_len
,
2185 hs
->goal
.fe_logical
);
2186 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, buf3
, buf2
,
2187 hs
->found
, hs
->groups
, hs
->cr
, hs
->flags
,
2188 hs
->merged
? "M" : "", hs
->tail
,
2189 hs
->buddy
? 1 << hs
->buddy
: 0);
2190 } else if (hs
->op
== EXT4_MB_HISTORY_PREALLOC
) {
2191 fmt
= "%-5u %-8u %-23s %-23s %-23s\n";
2192 sprintf(buf2
, "%u/%d/%u@%u", hs
->result
.fe_group
,
2193 hs
->result
.fe_start
, hs
->result
.fe_len
,
2194 hs
->result
.fe_logical
);
2195 sprintf(buf
, "%u/%d/%u@%u", hs
->orig
.fe_group
,
2196 hs
->orig
.fe_start
, hs
->orig
.fe_len
,
2197 hs
->orig
.fe_logical
);
2198 seq_printf(seq
, fmt
, hs
->pid
, hs
->ino
, buf
, "", buf2
);
2199 } else if (hs
->op
== EXT4_MB_HISTORY_DISCARD
) {
2200 sprintf(buf2
, "%u/%d/%u", hs
->result
.fe_group
,
2201 hs
->result
.fe_start
, hs
->result
.fe_len
);
2202 seq_printf(seq
, "%-5u %-8u %-23s discard\n",
2203 hs
->pid
, hs
->ino
, buf2
);
2204 } else if (hs
->op
== EXT4_MB_HISTORY_FREE
) {
2205 sprintf(buf2
, "%u/%d/%u", hs
->result
.fe_group
,
2206 hs
->result
.fe_start
, hs
->result
.fe_len
);
2207 seq_printf(seq
, "%-5u %-8u %-23s free\n",
2208 hs
->pid
, hs
->ino
, buf2
);
2213 static void ext4_mb_seq_history_stop(struct seq_file
*seq
, void *v
)
2217 static struct seq_operations ext4_mb_seq_history_ops
= {
2218 .start
= ext4_mb_seq_history_start
,
2219 .next
= ext4_mb_seq_history_next
,
2220 .stop
= ext4_mb_seq_history_stop
,
2221 .show
= ext4_mb_seq_history_show
,
2224 static int ext4_mb_seq_history_open(struct inode
*inode
, struct file
*file
)
2226 struct super_block
*sb
= PDE(inode
)->data
;
2227 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2228 struct ext4_mb_proc_session
*s
;
2232 if (unlikely(sbi
->s_mb_history
== NULL
))
2234 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
2238 size
= sizeof(struct ext4_mb_history
) * sbi
->s_mb_history_max
;
2239 s
->history
= kmalloc(size
, GFP_KERNEL
);
2240 if (s
->history
== NULL
) {
2245 spin_lock(&sbi
->s_mb_history_lock
);
2246 memcpy(s
->history
, sbi
->s_mb_history
, size
);
2247 s
->max
= sbi
->s_mb_history_max
;
2248 s
->start
= sbi
->s_mb_history_cur
% s
->max
;
2249 spin_unlock(&sbi
->s_mb_history_lock
);
2251 rc
= seq_open(file
, &ext4_mb_seq_history_ops
);
2253 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2263 static int ext4_mb_seq_history_release(struct inode
*inode
, struct file
*file
)
2265 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2266 struct ext4_mb_proc_session
*s
= seq
->private;
2269 return seq_release(inode
, file
);
2272 static ssize_t
ext4_mb_seq_history_write(struct file
*file
,
2273 const char __user
*buffer
,
2274 size_t count
, loff_t
*ppos
)
2276 struct seq_file
*seq
= (struct seq_file
*)file
->private_data
;
2277 struct ext4_mb_proc_session
*s
= seq
->private;
2278 struct super_block
*sb
= s
->sb
;
2282 if (count
>= sizeof(str
)) {
2283 printk(KERN_ERR
"EXT4-fs: %s string too long, max %u bytes\n",
2284 "mb_history", (int)sizeof(str
));
2288 if (copy_from_user(str
, buffer
, count
))
2291 value
= simple_strtol(str
, NULL
, 0);
2294 EXT4_SB(sb
)->s_mb_history_filter
= value
;
2299 static struct file_operations ext4_mb_seq_history_fops
= {
2300 .owner
= THIS_MODULE
,
2301 .open
= ext4_mb_seq_history_open
,
2303 .write
= ext4_mb_seq_history_write
,
2304 .llseek
= seq_lseek
,
2305 .release
= ext4_mb_seq_history_release
,
2308 static void *ext4_mb_seq_groups_start(struct seq_file
*seq
, loff_t
*pos
)
2310 struct super_block
*sb
= seq
->private;
2311 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2314 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2318 return (void *) ((unsigned long) group
);
2321 static void *ext4_mb_seq_groups_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2323 struct super_block
*sb
= seq
->private;
2324 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2328 if (*pos
< 0 || *pos
>= sbi
->s_groups_count
)
2331 return (void *) ((unsigned long) group
);
2334 static int ext4_mb_seq_groups_show(struct seq_file
*seq
, void *v
)
2336 struct super_block
*sb
= seq
->private;
2337 ext4_group_t group
= (ext4_group_t
) ((unsigned long) v
);
2340 struct ext4_buddy e4b
;
2342 struct ext4_group_info info
;
2343 unsigned short counters
[16];
2348 seq_printf(seq
, "#%-5s: %-5s %-5s %-5s "
2349 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2350 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2351 "group", "free", "frags", "first",
2352 "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2353 "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2355 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(sg
.info
.bb_counters
[0]) +
2356 sizeof(struct ext4_group_info
);
2357 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
2359 seq_printf(seq
, "#%-5u: I/O error\n", group
);
2362 ext4_lock_group(sb
, group
);
2363 memcpy(&sg
, ext4_get_group_info(sb
, group
), i
);
2364 ext4_unlock_group(sb
, group
);
2365 ext4_mb_release_desc(&e4b
);
2367 seq_printf(seq
, "#%-5u: %-5u %-5u %-5u [", group
, sg
.info
.bb_free
,
2368 sg
.info
.bb_fragments
, sg
.info
.bb_first_free
);
2369 for (i
= 0; i
<= 13; i
++)
2370 seq_printf(seq
, " %-5u", i
<= sb
->s_blocksize_bits
+ 1 ?
2371 sg
.info
.bb_counters
[i
] : 0);
2372 seq_printf(seq
, " ]\n");
2377 static void ext4_mb_seq_groups_stop(struct seq_file
*seq
, void *v
)
2381 static struct seq_operations ext4_mb_seq_groups_ops
= {
2382 .start
= ext4_mb_seq_groups_start
,
2383 .next
= ext4_mb_seq_groups_next
,
2384 .stop
= ext4_mb_seq_groups_stop
,
2385 .show
= ext4_mb_seq_groups_show
,
2388 static int ext4_mb_seq_groups_open(struct inode
*inode
, struct file
*file
)
2390 struct super_block
*sb
= PDE(inode
)->data
;
2393 rc
= seq_open(file
, &ext4_mb_seq_groups_ops
);
2395 struct seq_file
*m
= (struct seq_file
*)file
->private_data
;
2402 static struct file_operations ext4_mb_seq_groups_fops
= {
2403 .owner
= THIS_MODULE
,
2404 .open
= ext4_mb_seq_groups_open
,
2406 .llseek
= seq_lseek
,
2407 .release
= seq_release
,
2410 static void ext4_mb_history_release(struct super_block
*sb
)
2412 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2414 if (sbi
->s_proc
!= NULL
) {
2415 remove_proc_entry("mb_groups", sbi
->s_proc
);
2416 remove_proc_entry("mb_history", sbi
->s_proc
);
2418 kfree(sbi
->s_mb_history
);
2421 static void ext4_mb_history_init(struct super_block
*sb
)
2423 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2426 if (sbi
->s_proc
!= NULL
) {
2427 proc_create_data("mb_history", S_IRUGO
, sbi
->s_proc
,
2428 &ext4_mb_seq_history_fops
, sb
);
2429 proc_create_data("mb_groups", S_IRUGO
, sbi
->s_proc
,
2430 &ext4_mb_seq_groups_fops
, sb
);
2433 sbi
->s_mb_history_max
= 1000;
2434 sbi
->s_mb_history_cur
= 0;
2435 spin_lock_init(&sbi
->s_mb_history_lock
);
2436 i
= sbi
->s_mb_history_max
* sizeof(struct ext4_mb_history
);
2437 sbi
->s_mb_history
= kzalloc(i
, GFP_KERNEL
);
2438 /* if we can't allocate history, then we simple won't use it */
2441 static noinline_for_stack
void
2442 ext4_mb_store_history(struct ext4_allocation_context
*ac
)
2444 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
2445 struct ext4_mb_history h
;
2447 if (unlikely(sbi
->s_mb_history
== NULL
))
2450 if (!(ac
->ac_op
& sbi
->s_mb_history_filter
))
2454 h
.pid
= current
->pid
;
2455 h
.ino
= ac
->ac_inode
? ac
->ac_inode
->i_ino
: 0;
2456 h
.orig
= ac
->ac_o_ex
;
2457 h
.result
= ac
->ac_b_ex
;
2458 h
.flags
= ac
->ac_flags
;
2459 h
.found
= ac
->ac_found
;
2460 h
.groups
= ac
->ac_groups_scanned
;
2461 h
.cr
= ac
->ac_criteria
;
2462 h
.tail
= ac
->ac_tail
;
2463 h
.buddy
= ac
->ac_buddy
;
2465 if (ac
->ac_op
== EXT4_MB_HISTORY_ALLOC
) {
2466 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
2467 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
2469 h
.goal
= ac
->ac_g_ex
;
2470 h
.result
= ac
->ac_f_ex
;
2473 spin_lock(&sbi
->s_mb_history_lock
);
2474 memcpy(sbi
->s_mb_history
+ sbi
->s_mb_history_cur
, &h
, sizeof(h
));
2475 if (++sbi
->s_mb_history_cur
>= sbi
->s_mb_history_max
)
2476 sbi
->s_mb_history_cur
= 0;
2477 spin_unlock(&sbi
->s_mb_history_lock
);
2481 #define ext4_mb_history_release(sb)
2482 #define ext4_mb_history_init(sb)
2486 /* Create and initialize ext4_group_info data for the given group. */
2487 int ext4_mb_add_groupinfo(struct super_block
*sb
, ext4_group_t group
,
2488 struct ext4_group_desc
*desc
)
2492 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2493 struct ext4_group_info
**meta_group_info
;
2496 * First check if this group is the first of a reserved block.
2497 * If it's true, we have to allocate a new table of pointers
2498 * to ext4_group_info structures
2500 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0) {
2501 metalen
= sizeof(*meta_group_info
) <<
2502 EXT4_DESC_PER_BLOCK_BITS(sb
);
2503 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2504 if (meta_group_info
== NULL
) {
2505 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2507 goto exit_meta_group_info
;
2509 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)] =
2514 * calculate needed size. if change bb_counters size,
2515 * don't forget about ext4_mb_generate_buddy()
2517 len
= offsetof(typeof(**meta_group_info
),
2518 bb_counters
[sb
->s_blocksize_bits
+ 2]);
2521 sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)];
2522 i
= group
& (EXT4_DESC_PER_BLOCK(sb
) - 1);
2524 meta_group_info
[i
] = kzalloc(len
, GFP_KERNEL
);
2525 if (meta_group_info
[i
] == NULL
) {
2526 printk(KERN_ERR
"EXT4-fs: can't allocate buddy mem\n");
2527 goto exit_group_info
;
2529 set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT
,
2530 &(meta_group_info
[i
]->bb_state
));
2533 * initialize bb_free to be able to skip
2534 * empty groups without initialization
2536 if (desc
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
2537 meta_group_info
[i
]->bb_free
=
2538 ext4_free_blocks_after_init(sb
, group
, desc
);
2540 meta_group_info
[i
]->bb_free
=
2541 ext4_free_blks_count(sb
, desc
);
2544 INIT_LIST_HEAD(&meta_group_info
[i
]->bb_prealloc_list
);
2545 init_rwsem(&meta_group_info
[i
]->alloc_sem
);
2546 meta_group_info
[i
]->bb_free_root
.rb_node
= NULL
;;
2550 struct buffer_head
*bh
;
2551 meta_group_info
[i
]->bb_bitmap
=
2552 kmalloc(sb
->s_blocksize
, GFP_KERNEL
);
2553 BUG_ON(meta_group_info
[i
]->bb_bitmap
== NULL
);
2554 bh
= ext4_read_block_bitmap(sb
, group
);
2556 memcpy(meta_group_info
[i
]->bb_bitmap
, bh
->b_data
,
2565 /* If a meta_group_info table has been allocated, release it now */
2566 if (group
% EXT4_DESC_PER_BLOCK(sb
) == 0)
2567 kfree(sbi
->s_group_info
[group
>> EXT4_DESC_PER_BLOCK_BITS(sb
)]);
2568 exit_meta_group_info
:
2570 } /* ext4_mb_add_groupinfo */
2573 * Update an existing group.
2574 * This function is used for online resize
2576 void ext4_mb_update_group_info(struct ext4_group_info
*grp
, ext4_grpblk_t add
)
2578 grp
->bb_free
+= add
;
2581 static int ext4_mb_init_backend(struct super_block
*sb
)
2585 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2586 struct ext4_super_block
*es
= sbi
->s_es
;
2587 int num_meta_group_infos
;
2588 int num_meta_group_infos_max
;
2590 struct ext4_group_info
**meta_group_info
;
2591 struct ext4_group_desc
*desc
;
2593 /* This is the number of blocks used by GDT */
2594 num_meta_group_infos
= (sbi
->s_groups_count
+ EXT4_DESC_PER_BLOCK(sb
) -
2595 1) >> EXT4_DESC_PER_BLOCK_BITS(sb
);
2598 * This is the total number of blocks used by GDT including
2599 * the number of reserved blocks for GDT.
2600 * The s_group_info array is allocated with this value
2601 * to allow a clean online resize without a complex
2602 * manipulation of pointer.
2603 * The drawback is the unused memory when no resize
2604 * occurs but it's very low in terms of pages
2605 * (see comments below)
2606 * Need to handle this properly when META_BG resizing is allowed
2608 num_meta_group_infos_max
= num_meta_group_infos
+
2609 le16_to_cpu(es
->s_reserved_gdt_blocks
);
2612 * array_size is the size of s_group_info array. We round it
2613 * to the next power of two because this approximation is done
2614 * internally by kmalloc so we can have some more memory
2615 * for free here (e.g. may be used for META_BG resize).
2618 while (array_size
< sizeof(*sbi
->s_group_info
) *
2619 num_meta_group_infos_max
)
2620 array_size
= array_size
<< 1;
2621 /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2622 * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2623 * So a two level scheme suffices for now. */
2624 sbi
->s_group_info
= kmalloc(array_size
, GFP_KERNEL
);
2625 if (sbi
->s_group_info
== NULL
) {
2626 printk(KERN_ERR
"EXT4-fs: can't allocate buddy meta group\n");
2629 sbi
->s_buddy_cache
= new_inode(sb
);
2630 if (sbi
->s_buddy_cache
== NULL
) {
2631 printk(KERN_ERR
"EXT4-fs: can't get new inode\n");
2634 EXT4_I(sbi
->s_buddy_cache
)->i_disksize
= 0;
2636 metalen
= sizeof(*meta_group_info
) << EXT4_DESC_PER_BLOCK_BITS(sb
);
2637 for (i
= 0; i
< num_meta_group_infos
; i
++) {
2638 if ((i
+ 1) == num_meta_group_infos
)
2639 metalen
= sizeof(*meta_group_info
) *
2640 (sbi
->s_groups_count
-
2641 (i
<< EXT4_DESC_PER_BLOCK_BITS(sb
)));
2642 meta_group_info
= kmalloc(metalen
, GFP_KERNEL
);
2643 if (meta_group_info
== NULL
) {
2644 printk(KERN_ERR
"EXT4-fs: can't allocate mem for a "
2648 sbi
->s_group_info
[i
] = meta_group_info
;
2651 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2652 desc
= ext4_get_group_desc(sb
, i
, NULL
);
2655 "EXT4-fs: can't read descriptor %u\n", i
);
2658 if (ext4_mb_add_groupinfo(sb
, i
, desc
) != 0)
2666 kfree(ext4_get_group_info(sb
, i
));
2667 i
= num_meta_group_infos
;
2670 kfree(sbi
->s_group_info
[i
]);
2671 iput(sbi
->s_buddy_cache
);
2673 kfree(sbi
->s_group_info
);
2677 int ext4_mb_init(struct super_block
*sb
, int needs_recovery
)
2679 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2685 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned short);
2687 sbi
->s_mb_offsets
= kmalloc(i
, GFP_KERNEL
);
2688 if (sbi
->s_mb_offsets
== NULL
) {
2692 i
= (sb
->s_blocksize_bits
+ 2) * sizeof(unsigned int);
2693 sbi
->s_mb_maxs
= kmalloc(i
, GFP_KERNEL
);
2694 if (sbi
->s_mb_maxs
== NULL
) {
2695 kfree(sbi
->s_mb_maxs
);
2699 /* order 0 is regular bitmap */
2700 sbi
->s_mb_maxs
[0] = sb
->s_blocksize
<< 3;
2701 sbi
->s_mb_offsets
[0] = 0;
2705 max
= sb
->s_blocksize
<< 2;
2707 sbi
->s_mb_offsets
[i
] = offset
;
2708 sbi
->s_mb_maxs
[i
] = max
;
2709 offset
+= 1 << (sb
->s_blocksize_bits
- i
);
2712 } while (i
<= sb
->s_blocksize_bits
+ 1);
2714 /* init file for buddy data */
2715 ret
= ext4_mb_init_backend(sb
);
2717 kfree(sbi
->s_mb_offsets
);
2718 kfree(sbi
->s_mb_maxs
);
2722 spin_lock_init(&sbi
->s_md_lock
);
2723 spin_lock_init(&sbi
->s_bal_lock
);
2725 sbi
->s_mb_max_to_scan
= MB_DEFAULT_MAX_TO_SCAN
;
2726 sbi
->s_mb_min_to_scan
= MB_DEFAULT_MIN_TO_SCAN
;
2727 sbi
->s_mb_stats
= MB_DEFAULT_STATS
;
2728 sbi
->s_mb_stream_request
= MB_DEFAULT_STREAM_THRESHOLD
;
2729 sbi
->s_mb_order2_reqs
= MB_DEFAULT_ORDER2_REQS
;
2730 sbi
->s_mb_history_filter
= EXT4_MB_HISTORY_DEFAULT
;
2731 sbi
->s_mb_group_prealloc
= MB_DEFAULT_GROUP_PREALLOC
;
2733 sbi
->s_locality_groups
= alloc_percpu(struct ext4_locality_group
);
2734 if (sbi
->s_locality_groups
== NULL
) {
2735 kfree(sbi
->s_mb_offsets
);
2736 kfree(sbi
->s_mb_maxs
);
2739 for_each_possible_cpu(i
) {
2740 struct ext4_locality_group
*lg
;
2741 lg
= per_cpu_ptr(sbi
->s_locality_groups
, i
);
2742 mutex_init(&lg
->lg_mutex
);
2743 for (j
= 0; j
< PREALLOC_TB_SIZE
; j
++)
2744 INIT_LIST_HEAD(&lg
->lg_prealloc_list
[j
]);
2745 spin_lock_init(&lg
->lg_prealloc_lock
);
2748 ext4_mb_init_per_dev_proc(sb
);
2749 ext4_mb_history_init(sb
);
2752 sbi
->s_journal
->j_commit_callback
= release_blocks_on_commit
;
2754 printk(KERN_INFO
"EXT4-fs: mballoc enabled\n");
2758 /* need to called with ext4 group lock (ext4_lock_group) */
2759 static void ext4_mb_cleanup_pa(struct ext4_group_info
*grp
)
2761 struct ext4_prealloc_space
*pa
;
2762 struct list_head
*cur
, *tmp
;
2765 list_for_each_safe(cur
, tmp
, &grp
->bb_prealloc_list
) {
2766 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
2767 list_del(&pa
->pa_group_list
);
2769 kmem_cache_free(ext4_pspace_cachep
, pa
);
2772 mb_debug("mballoc: %u PAs left\n", count
);
2776 int ext4_mb_release(struct super_block
*sb
)
2779 int num_meta_group_infos
;
2780 struct ext4_group_info
*grinfo
;
2781 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2783 if (sbi
->s_group_info
) {
2784 for (i
= 0; i
< sbi
->s_groups_count
; i
++) {
2785 grinfo
= ext4_get_group_info(sb
, i
);
2787 kfree(grinfo
->bb_bitmap
);
2789 ext4_lock_group(sb
, i
);
2790 ext4_mb_cleanup_pa(grinfo
);
2791 ext4_unlock_group(sb
, i
);
2794 num_meta_group_infos
= (sbi
->s_groups_count
+
2795 EXT4_DESC_PER_BLOCK(sb
) - 1) >>
2796 EXT4_DESC_PER_BLOCK_BITS(sb
);
2797 for (i
= 0; i
< num_meta_group_infos
; i
++)
2798 kfree(sbi
->s_group_info
[i
]);
2799 kfree(sbi
->s_group_info
);
2801 kfree(sbi
->s_mb_offsets
);
2802 kfree(sbi
->s_mb_maxs
);
2803 if (sbi
->s_buddy_cache
)
2804 iput(sbi
->s_buddy_cache
);
2805 if (sbi
->s_mb_stats
) {
2807 "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
2808 atomic_read(&sbi
->s_bal_allocated
),
2809 atomic_read(&sbi
->s_bal_reqs
),
2810 atomic_read(&sbi
->s_bal_success
));
2812 "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
2813 "%u 2^N hits, %u breaks, %u lost\n",
2814 atomic_read(&sbi
->s_bal_ex_scanned
),
2815 atomic_read(&sbi
->s_bal_goals
),
2816 atomic_read(&sbi
->s_bal_2orders
),
2817 atomic_read(&sbi
->s_bal_breaks
),
2818 atomic_read(&sbi
->s_mb_lost_chunks
));
2820 "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
2821 sbi
->s_mb_buddies_generated
++,
2822 sbi
->s_mb_generation_time
);
2824 "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
2825 atomic_read(&sbi
->s_mb_preallocated
),
2826 atomic_read(&sbi
->s_mb_discarded
));
2829 free_percpu(sbi
->s_locality_groups
);
2830 ext4_mb_history_release(sb
);
2831 ext4_mb_destroy_per_dev_proc(sb
);
2837 * This function is called by the jbd2 layer once the commit has finished,
2838 * so we know we can free the blocks that were released with that commit.
2840 static void release_blocks_on_commit(journal_t
*journal
, transaction_t
*txn
)
2842 struct super_block
*sb
= journal
->j_private
;
2843 struct ext4_buddy e4b
;
2844 struct ext4_group_info
*db
;
2845 int err
, count
= 0, count2
= 0;
2846 struct ext4_free_data
*entry
;
2847 ext4_fsblk_t discard_block
;
2848 struct list_head
*l
, *ltmp
;
2850 list_for_each_safe(l
, ltmp
, &txn
->t_private_list
) {
2851 entry
= list_entry(l
, struct ext4_free_data
, list
);
2853 mb_debug("gonna free %u blocks in group %u (0x%p):",
2854 entry
->count
, entry
->group
, entry
);
2856 err
= ext4_mb_load_buddy(sb
, entry
->group
, &e4b
);
2857 /* we expect to find existing buddy because it's pinned */
2861 /* there are blocks to put in buddy to make them really free */
2862 count
+= entry
->count
;
2864 ext4_lock_group(sb
, entry
->group
);
2865 /* Take it out of per group rb tree */
2866 rb_erase(&entry
->node
, &(db
->bb_free_root
));
2867 mb_free_blocks(NULL
, &e4b
, entry
->start_blk
, entry
->count
);
2869 if (!db
->bb_free_root
.rb_node
) {
2870 /* No more items in the per group rb tree
2871 * balance refcounts from ext4_mb_free_metadata()
2873 page_cache_release(e4b
.bd_buddy_page
);
2874 page_cache_release(e4b
.bd_bitmap_page
);
2876 ext4_unlock_group(sb
, entry
->group
);
2877 discard_block
= (ext4_fsblk_t
) entry
->group
* EXT4_BLOCKS_PER_GROUP(sb
)
2879 + le32_to_cpu(EXT4_SB(sb
)->s_es
->s_first_data_block
);
2880 trace_mark(ext4_discard_blocks
, "dev %s blk %llu count %u", sb
->s_id
,
2881 (unsigned long long) discard_block
, entry
->count
);
2882 sb_issue_discard(sb
, discard_block
, entry
->count
);
2884 kmem_cache_free(ext4_free_ext_cachep
, entry
);
2885 ext4_mb_release_desc(&e4b
);
2888 mb_debug("freed %u blocks in %u structures\n", count
, count2
);
2891 #define EXT4_MB_STATS_NAME "stats"
2892 #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan"
2893 #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan"
2894 #define EXT4_MB_ORDER2_REQ "order2_req"
2895 #define EXT4_MB_STREAM_REQ "stream_req"
2896 #define EXT4_MB_GROUP_PREALLOC "group_prealloc"
2898 static int ext4_mb_init_per_dev_proc(struct super_block
*sb
)
2900 #ifdef CONFIG_PROC_FS
2901 mode_t mode
= S_IFREG
| S_IRUGO
| S_IWUSR
;
2902 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2903 struct proc_dir_entry
*proc
;
2905 if (sbi
->s_proc
== NULL
)
2908 EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME
, mb_stats
);
2909 EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME
, mb_max_to_scan
);
2910 EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME
, mb_min_to_scan
);
2911 EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ
, mb_order2_reqs
);
2912 EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ
, mb_stream_request
);
2913 EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC
, mb_group_prealloc
);
2917 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2918 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2919 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2920 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2921 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2922 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2929 static int ext4_mb_destroy_per_dev_proc(struct super_block
*sb
)
2931 #ifdef CONFIG_PROC_FS
2932 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
2934 if (sbi
->s_proc
== NULL
)
2937 remove_proc_entry(EXT4_MB_GROUP_PREALLOC
, sbi
->s_proc
);
2938 remove_proc_entry(EXT4_MB_STREAM_REQ
, sbi
->s_proc
);
2939 remove_proc_entry(EXT4_MB_ORDER2_REQ
, sbi
->s_proc
);
2940 remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME
, sbi
->s_proc
);
2941 remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME
, sbi
->s_proc
);
2942 remove_proc_entry(EXT4_MB_STATS_NAME
, sbi
->s_proc
);
2947 int __init
init_ext4_mballoc(void)
2949 ext4_pspace_cachep
=
2950 kmem_cache_create("ext4_prealloc_space",
2951 sizeof(struct ext4_prealloc_space
),
2952 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2953 if (ext4_pspace_cachep
== NULL
)
2957 kmem_cache_create("ext4_alloc_context",
2958 sizeof(struct ext4_allocation_context
),
2959 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2960 if (ext4_ac_cachep
== NULL
) {
2961 kmem_cache_destroy(ext4_pspace_cachep
);
2965 ext4_free_ext_cachep
=
2966 kmem_cache_create("ext4_free_block_extents",
2967 sizeof(struct ext4_free_data
),
2968 0, SLAB_RECLAIM_ACCOUNT
, NULL
);
2969 if (ext4_free_ext_cachep
== NULL
) {
2970 kmem_cache_destroy(ext4_pspace_cachep
);
2971 kmem_cache_destroy(ext4_ac_cachep
);
2977 void exit_ext4_mballoc(void)
2979 /* XXX: synchronize_rcu(); */
2980 kmem_cache_destroy(ext4_pspace_cachep
);
2981 kmem_cache_destroy(ext4_ac_cachep
);
2982 kmem_cache_destroy(ext4_free_ext_cachep
);
2987 * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
2988 * Returns 0 if success or error code
2990 static noinline_for_stack
int
2991 ext4_mb_mark_diskspace_used(struct ext4_allocation_context
*ac
,
2992 handle_t
*handle
, unsigned int reserv_blks
)
2994 struct buffer_head
*bitmap_bh
= NULL
;
2995 struct ext4_super_block
*es
;
2996 struct ext4_group_desc
*gdp
;
2997 struct buffer_head
*gdp_bh
;
2998 struct ext4_sb_info
*sbi
;
2999 struct super_block
*sb
;
3003 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3004 BUG_ON(ac
->ac_b_ex
.fe_len
<= 0);
3012 bitmap_bh
= ext4_read_block_bitmap(sb
, ac
->ac_b_ex
.fe_group
);
3016 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
3021 gdp
= ext4_get_group_desc(sb
, ac
->ac_b_ex
.fe_group
, &gdp_bh
);
3025 ext4_debug("using block group %u(%d)\n", ac
->ac_b_ex
.fe_group
,
3026 gdp
->bg_free_blocks_count
);
3028 err
= ext4_journal_get_write_access(handle
, gdp_bh
);
3032 block
= ac
->ac_b_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(sb
)
3033 + ac
->ac_b_ex
.fe_start
3034 + le32_to_cpu(es
->s_first_data_block
);
3036 len
= ac
->ac_b_ex
.fe_len
;
3037 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, len
) ||
3038 in_range(ext4_inode_bitmap(sb
, gdp
), block
, len
) ||
3039 in_range(block
, ext4_inode_table(sb
, gdp
),
3040 EXT4_SB(sb
)->s_itb_per_group
) ||
3041 in_range(block
+ len
- 1, ext4_inode_table(sb
, gdp
),
3042 EXT4_SB(sb
)->s_itb_per_group
)) {
3043 ext4_error(sb
, __func__
,
3044 "Allocating block %llu in system zone of %d group\n",
3045 block
, ac
->ac_b_ex
.fe_group
);
3046 /* File system mounted not to panic on error
3047 * Fix the bitmap and repeat the block allocation
3048 * We leak some of the blocks here.
3050 mb_set_bits(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
),
3051 bitmap_bh
->b_data
, ac
->ac_b_ex
.fe_start
,
3052 ac
->ac_b_ex
.fe_len
);
3053 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3058 #ifdef AGGRESSIVE_CHECK
3061 for (i
= 0; i
< ac
->ac_b_ex
.fe_len
; i
++) {
3062 BUG_ON(mb_test_bit(ac
->ac_b_ex
.fe_start
+ i
,
3063 bitmap_bh
->b_data
));
3067 spin_lock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3068 mb_set_bits(NULL
, bitmap_bh
->b_data
,
3069 ac
->ac_b_ex
.fe_start
, ac
->ac_b_ex
.fe_len
);
3070 if (gdp
->bg_flags
& cpu_to_le16(EXT4_BG_BLOCK_UNINIT
)) {
3071 gdp
->bg_flags
&= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT
);
3072 ext4_free_blks_set(sb
, gdp
,
3073 ext4_free_blocks_after_init(sb
,
3074 ac
->ac_b_ex
.fe_group
, gdp
));
3076 len
= ext4_free_blks_count(sb
, gdp
) - ac
->ac_b_ex
.fe_len
;
3077 ext4_free_blks_set(sb
, gdp
, len
);
3078 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, ac
->ac_b_ex
.fe_group
, gdp
);
3079 spin_unlock(sb_bgl_lock(sbi
, ac
->ac_b_ex
.fe_group
));
3080 percpu_counter_sub(&sbi
->s_freeblocks_counter
, ac
->ac_b_ex
.fe_len
);
3082 * Now reduce the dirty block count also. Should not go negative
3084 if (!(ac
->ac_flags
& EXT4_MB_DELALLOC_RESERVED
))
3085 /* release all the reserved blocks if non delalloc */
3086 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
, reserv_blks
);
3088 percpu_counter_sub(&sbi
->s_dirtyblocks_counter
,
3089 ac
->ac_b_ex
.fe_len
);
3091 if (sbi
->s_log_groups_per_flex
) {
3092 ext4_group_t flex_group
= ext4_flex_group(sbi
,
3093 ac
->ac_b_ex
.fe_group
);
3094 spin_lock(sb_bgl_lock(sbi
, flex_group
));
3095 sbi
->s_flex_groups
[flex_group
].free_blocks
-= ac
->ac_b_ex
.fe_len
;
3096 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
3099 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
3102 err
= ext4_handle_dirty_metadata(handle
, NULL
, gdp_bh
);
3111 * here we normalize request for locality group
3112 * Group request are normalized to s_strip size if we set the same via mount
3113 * option. If not we set it to s_mb_group_prealloc which can be configured via
3114 * /proc/fs/ext4/<partition>/group_prealloc
3116 * XXX: should we try to preallocate more than the group has now?
3118 static void ext4_mb_normalize_group_request(struct ext4_allocation_context
*ac
)
3120 struct super_block
*sb
= ac
->ac_sb
;
3121 struct ext4_locality_group
*lg
= ac
->ac_lg
;
3124 if (EXT4_SB(sb
)->s_stripe
)
3125 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_stripe
;
3127 ac
->ac_g_ex
.fe_len
= EXT4_SB(sb
)->s_mb_group_prealloc
;
3128 mb_debug("#%u: goal %u blocks for locality group\n",
3129 current
->pid
, ac
->ac_g_ex
.fe_len
);
3133 * Normalization means making request better in terms of
3134 * size and alignment
3136 static noinline_for_stack
void
3137 ext4_mb_normalize_request(struct ext4_allocation_context
*ac
,
3138 struct ext4_allocation_request
*ar
)
3142 loff_t size
, orig_size
, start_off
;
3143 ext4_lblk_t start
, orig_start
;
3144 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3145 struct ext4_prealloc_space
*pa
;
3147 /* do normalize only data requests, metadata requests
3148 do not need preallocation */
3149 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3152 /* sometime caller may want exact blocks */
3153 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
3156 /* caller may indicate that preallocation isn't
3157 * required (it's a tail, for example) */
3158 if (ac
->ac_flags
& EXT4_MB_HINT_NOPREALLOC
)
3161 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
) {
3162 ext4_mb_normalize_group_request(ac
);
3166 bsbits
= ac
->ac_sb
->s_blocksize_bits
;
3168 /* first, let's learn actual file size
3169 * given current request is allocated */
3170 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
3171 size
= size
<< bsbits
;
3172 if (size
< i_size_read(ac
->ac_inode
))
3173 size
= i_size_read(ac
->ac_inode
);
3175 /* max size of free chunks */
3178 #define NRL_CHECK_SIZE(req, size, max, chunk_size) \
3179 (req <= (size) || max <= (chunk_size))
3181 /* first, try to predict filesize */
3182 /* XXX: should this table be tunable? */
3184 if (size
<= 16 * 1024) {
3186 } else if (size
<= 32 * 1024) {
3188 } else if (size
<= 64 * 1024) {
3190 } else if (size
<= 128 * 1024) {
3192 } else if (size
<= 256 * 1024) {
3194 } else if (size
<= 512 * 1024) {
3196 } else if (size
<= 1024 * 1024) {
3198 } else if (NRL_CHECK_SIZE(size
, 4 * 1024 * 1024, max
, 2 * 1024)) {
3199 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3200 (21 - bsbits
)) << 21;
3201 size
= 2 * 1024 * 1024;
3202 } else if (NRL_CHECK_SIZE(size
, 8 * 1024 * 1024, max
, 4 * 1024)) {
3203 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3204 (22 - bsbits
)) << 22;
3205 size
= 4 * 1024 * 1024;
3206 } else if (NRL_CHECK_SIZE(ac
->ac_o_ex
.fe_len
,
3207 (8<<20)>>bsbits
, max
, 8 * 1024)) {
3208 start_off
= ((loff_t
)ac
->ac_o_ex
.fe_logical
>>
3209 (23 - bsbits
)) << 23;
3210 size
= 8 * 1024 * 1024;
3212 start_off
= (loff_t
)ac
->ac_o_ex
.fe_logical
<< bsbits
;
3213 size
= ac
->ac_o_ex
.fe_len
<< bsbits
;
3215 orig_size
= size
= size
>> bsbits
;
3216 orig_start
= start
= start_off
>> bsbits
;
3218 /* don't cover already allocated blocks in selected range */
3219 if (ar
->pleft
&& start
<= ar
->lleft
) {
3220 size
-= ar
->lleft
+ 1 - start
;
3221 start
= ar
->lleft
+ 1;
3223 if (ar
->pright
&& start
+ size
- 1 >= ar
->lright
)
3224 size
-= start
+ size
- ar
->lright
;
3228 /* check we don't cross already preallocated blocks */
3230 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3235 spin_lock(&pa
->pa_lock
);
3236 if (pa
->pa_deleted
) {
3237 spin_unlock(&pa
->pa_lock
);
3241 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3243 /* PA must not overlap original request */
3244 BUG_ON(!(ac
->ac_o_ex
.fe_logical
>= pa_end
||
3245 ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
));
3247 /* skip PA normalized request doesn't overlap with */
3248 if (pa
->pa_lstart
>= end
) {
3249 spin_unlock(&pa
->pa_lock
);
3252 if (pa_end
<= start
) {
3253 spin_unlock(&pa
->pa_lock
);
3256 BUG_ON(pa
->pa_lstart
<= start
&& pa_end
>= end
);
3258 if (pa_end
<= ac
->ac_o_ex
.fe_logical
) {
3259 BUG_ON(pa_end
< start
);
3263 if (pa
->pa_lstart
> ac
->ac_o_ex
.fe_logical
) {
3264 BUG_ON(pa
->pa_lstart
> end
);
3265 end
= pa
->pa_lstart
;
3267 spin_unlock(&pa
->pa_lock
);
3272 /* XXX: extra loop to check we really don't overlap preallocations */
3274 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3276 spin_lock(&pa
->pa_lock
);
3277 if (pa
->pa_deleted
== 0) {
3278 pa_end
= pa
->pa_lstart
+ pa
->pa_len
;
3279 BUG_ON(!(start
>= pa_end
|| end
<= pa
->pa_lstart
));
3281 spin_unlock(&pa
->pa_lock
);
3285 if (start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3286 start
> ac
->ac_o_ex
.fe_logical
) {
3287 printk(KERN_ERR
"start %lu, size %lu, fe_logical %lu\n",
3288 (unsigned long) start
, (unsigned long) size
,
3289 (unsigned long) ac
->ac_o_ex
.fe_logical
);
3291 BUG_ON(start
+ size
<= ac
->ac_o_ex
.fe_logical
&&
3292 start
> ac
->ac_o_ex
.fe_logical
);
3293 BUG_ON(size
<= 0 || size
>= EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
));
3295 /* now prepare goal request */
3297 /* XXX: is it better to align blocks WRT to logical
3298 * placement or satisfy big request as is */
3299 ac
->ac_g_ex
.fe_logical
= start
;
3300 ac
->ac_g_ex
.fe_len
= size
;
3302 /* define goal start in order to merge */
3303 if (ar
->pright
&& (ar
->lright
== (start
+ size
))) {
3304 /* merge to the right */
3305 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pright
- size
,
3306 &ac
->ac_f_ex
.fe_group
,
3307 &ac
->ac_f_ex
.fe_start
);
3308 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3310 if (ar
->pleft
&& (ar
->lleft
+ 1 == start
)) {
3311 /* merge to the left */
3312 ext4_get_group_no_and_offset(ac
->ac_sb
, ar
->pleft
+ 1,
3313 &ac
->ac_f_ex
.fe_group
,
3314 &ac
->ac_f_ex
.fe_start
);
3315 ac
->ac_flags
|= EXT4_MB_HINT_TRY_GOAL
;
3318 mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size
,
3319 (unsigned) orig_size
, (unsigned) start
);
3322 static void ext4_mb_collect_stats(struct ext4_allocation_context
*ac
)
3324 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
3326 if (sbi
->s_mb_stats
&& ac
->ac_g_ex
.fe_len
> 1) {
3327 atomic_inc(&sbi
->s_bal_reqs
);
3328 atomic_add(ac
->ac_b_ex
.fe_len
, &sbi
->s_bal_allocated
);
3329 if (ac
->ac_o_ex
.fe_len
>= ac
->ac_g_ex
.fe_len
)
3330 atomic_inc(&sbi
->s_bal_success
);
3331 atomic_add(ac
->ac_found
, &sbi
->s_bal_ex_scanned
);
3332 if (ac
->ac_g_ex
.fe_start
== ac
->ac_b_ex
.fe_start
&&
3333 ac
->ac_g_ex
.fe_group
== ac
->ac_b_ex
.fe_group
)
3334 atomic_inc(&sbi
->s_bal_goals
);
3335 if (ac
->ac_found
> sbi
->s_mb_max_to_scan
)
3336 atomic_inc(&sbi
->s_bal_breaks
);
3339 ext4_mb_store_history(ac
);
3343 * use blocks preallocated to inode
3345 static void ext4_mb_use_inode_pa(struct ext4_allocation_context
*ac
,
3346 struct ext4_prealloc_space
*pa
)
3352 /* found preallocated blocks, use them */
3353 start
= pa
->pa_pstart
+ (ac
->ac_o_ex
.fe_logical
- pa
->pa_lstart
);
3354 end
= min(pa
->pa_pstart
+ pa
->pa_len
, start
+ ac
->ac_o_ex
.fe_len
);
3356 ext4_get_group_no_and_offset(ac
->ac_sb
, start
, &ac
->ac_b_ex
.fe_group
,
3357 &ac
->ac_b_ex
.fe_start
);
3358 ac
->ac_b_ex
.fe_len
= len
;
3359 ac
->ac_status
= AC_STATUS_FOUND
;
3362 BUG_ON(start
< pa
->pa_pstart
);
3363 BUG_ON(start
+ len
> pa
->pa_pstart
+ pa
->pa_len
);
3364 BUG_ON(pa
->pa_free
< len
);
3367 mb_debug("use %llu/%u from inode pa %p\n", start
, len
, pa
);
3371 * use blocks preallocated to locality group
3373 static void ext4_mb_use_group_pa(struct ext4_allocation_context
*ac
,
3374 struct ext4_prealloc_space
*pa
)
3376 unsigned int len
= ac
->ac_o_ex
.fe_len
;
3378 ext4_get_group_no_and_offset(ac
->ac_sb
, pa
->pa_pstart
,
3379 &ac
->ac_b_ex
.fe_group
,
3380 &ac
->ac_b_ex
.fe_start
);
3381 ac
->ac_b_ex
.fe_len
= len
;
3382 ac
->ac_status
= AC_STATUS_FOUND
;
3385 /* we don't correct pa_pstart or pa_plen here to avoid
3386 * possible race when the group is being loaded concurrently
3387 * instead we correct pa later, after blocks are marked
3388 * in on-disk bitmap -- see ext4_mb_release_context()
3389 * Other CPUs are prevented from allocating from this pa by lg_mutex
3391 mb_debug("use %u/%u from group pa %p\n", pa
->pa_lstart
-len
, len
, pa
);
3395 * Return the prealloc space that have minimal distance
3396 * from the goal block. @cpa is the prealloc
3397 * space that is having currently known minimal distance
3398 * from the goal block.
3400 static struct ext4_prealloc_space
*
3401 ext4_mb_check_group_pa(ext4_fsblk_t goal_block
,
3402 struct ext4_prealloc_space
*pa
,
3403 struct ext4_prealloc_space
*cpa
)
3405 ext4_fsblk_t cur_distance
, new_distance
;
3408 atomic_inc(&pa
->pa_count
);
3411 cur_distance
= abs(goal_block
- cpa
->pa_pstart
);
3412 new_distance
= abs(goal_block
- pa
->pa_pstart
);
3414 if (cur_distance
< new_distance
)
3417 /* drop the previous reference */
3418 atomic_dec(&cpa
->pa_count
);
3419 atomic_inc(&pa
->pa_count
);
3424 * search goal blocks in preallocated space
3426 static noinline_for_stack
int
3427 ext4_mb_use_preallocated(struct ext4_allocation_context
*ac
)
3430 struct ext4_inode_info
*ei
= EXT4_I(ac
->ac_inode
);
3431 struct ext4_locality_group
*lg
;
3432 struct ext4_prealloc_space
*pa
, *cpa
= NULL
;
3433 ext4_fsblk_t goal_block
;
3435 /* only data can be preallocated */
3436 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
3439 /* first, try per-file preallocation */
3441 list_for_each_entry_rcu(pa
, &ei
->i_prealloc_list
, pa_inode_list
) {
3443 /* all fields in this condition don't change,
3444 * so we can skip locking for them */
3445 if (ac
->ac_o_ex
.fe_logical
< pa
->pa_lstart
||
3446 ac
->ac_o_ex
.fe_logical
>= pa
->pa_lstart
+ pa
->pa_len
)
3449 /* found preallocated blocks, use them */
3450 spin_lock(&pa
->pa_lock
);
3451 if (pa
->pa_deleted
== 0 && pa
->pa_free
) {
3452 atomic_inc(&pa
->pa_count
);
3453 ext4_mb_use_inode_pa(ac
, pa
);
3454 spin_unlock(&pa
->pa_lock
);
3455 ac
->ac_criteria
= 10;
3459 spin_unlock(&pa
->pa_lock
);
3463 /* can we use group allocation? */
3464 if (!(ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
))
3467 /* inode may have no locality group for some reason */
3471 order
= fls(ac
->ac_o_ex
.fe_len
) - 1;
3472 if (order
> PREALLOC_TB_SIZE
- 1)
3473 /* The max size of hash table is PREALLOC_TB_SIZE */
3474 order
= PREALLOC_TB_SIZE
- 1;
3476 goal_block
= ac
->ac_g_ex
.fe_group
* EXT4_BLOCKS_PER_GROUP(ac
->ac_sb
) +
3477 ac
->ac_g_ex
.fe_start
+
3478 le32_to_cpu(EXT4_SB(ac
->ac_sb
)->s_es
->s_first_data_block
);
3480 * search for the prealloc space that is having
3481 * minimal distance from the goal block.
3483 for (i
= order
; i
< PREALLOC_TB_SIZE
; i
++) {
3485 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[i
],
3487 spin_lock(&pa
->pa_lock
);
3488 if (pa
->pa_deleted
== 0 &&
3489 pa
->pa_free
>= ac
->ac_o_ex
.fe_len
) {
3491 cpa
= ext4_mb_check_group_pa(goal_block
,
3494 spin_unlock(&pa
->pa_lock
);
3499 ext4_mb_use_group_pa(ac
, cpa
);
3500 ac
->ac_criteria
= 20;
3507 * the function goes through all block freed in the group
3508 * but not yet committed and marks them used in in-core bitmap.
3509 * buddy must be generated from this bitmap
3510 * Need to be called with ext4 group lock (ext4_lock_group)
3512 static void ext4_mb_generate_from_freelist(struct super_block
*sb
, void *bitmap
,
3516 struct ext4_group_info
*grp
;
3517 struct ext4_free_data
*entry
;
3519 grp
= ext4_get_group_info(sb
, group
);
3520 n
= rb_first(&(grp
->bb_free_root
));
3523 entry
= rb_entry(n
, struct ext4_free_data
, node
);
3524 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3525 bitmap
, entry
->start_blk
,
3533 * the function goes through all preallocation in this group and marks them
3534 * used in in-core bitmap. buddy must be generated from this bitmap
3535 * Need to be called with ext4 group lock (ext4_lock_group)
3537 static void ext4_mb_generate_from_pa(struct super_block
*sb
, void *bitmap
,
3540 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3541 struct ext4_prealloc_space
*pa
;
3542 struct list_head
*cur
;
3543 ext4_group_t groupnr
;
3544 ext4_grpblk_t start
;
3545 int preallocated
= 0;
3549 /* all form of preallocation discards first load group,
3550 * so the only competing code is preallocation use.
3551 * we don't need any locking here
3552 * notice we do NOT ignore preallocations with pa_deleted
3553 * otherwise we could leave used blocks available for
3554 * allocation in buddy when concurrent ext4_mb_put_pa()
3555 * is dropping preallocation
3557 list_for_each(cur
, &grp
->bb_prealloc_list
) {
3558 pa
= list_entry(cur
, struct ext4_prealloc_space
, pa_group_list
);
3559 spin_lock(&pa
->pa_lock
);
3560 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
3563 spin_unlock(&pa
->pa_lock
);
3564 if (unlikely(len
== 0))
3566 BUG_ON(groupnr
!= group
);
3567 mb_set_bits(sb_bgl_lock(EXT4_SB(sb
), group
),
3568 bitmap
, start
, len
);
3569 preallocated
+= len
;
3572 mb_debug("prellocated %u for group %u\n", preallocated
, group
);
3575 static void ext4_mb_pa_callback(struct rcu_head
*head
)
3577 struct ext4_prealloc_space
*pa
;
3578 pa
= container_of(head
, struct ext4_prealloc_space
, u
.pa_rcu
);
3579 kmem_cache_free(ext4_pspace_cachep
, pa
);
3583 * drops a reference to preallocated space descriptor
3584 * if this was the last reference and the space is consumed
3586 static void ext4_mb_put_pa(struct ext4_allocation_context
*ac
,
3587 struct super_block
*sb
, struct ext4_prealloc_space
*pa
)
3591 if (!atomic_dec_and_test(&pa
->pa_count
) || pa
->pa_free
!= 0)
3594 /* in this short window concurrent discard can set pa_deleted */
3595 spin_lock(&pa
->pa_lock
);
3596 if (pa
->pa_deleted
== 1) {
3597 spin_unlock(&pa
->pa_lock
);
3602 spin_unlock(&pa
->pa_lock
);
3604 /* -1 is to protect from crossing allocation group */
3605 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
- 1, &grp
, NULL
);
3610 * P1 (buddy init) P2 (regular allocation)
3611 * find block B in PA
3612 * copy on-disk bitmap to buddy
3613 * mark B in on-disk bitmap
3614 * drop PA from group
3615 * mark all PAs in buddy
3617 * thus, P1 initializes buddy with B available. to prevent this
3618 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3621 ext4_lock_group(sb
, grp
);
3622 list_del(&pa
->pa_group_list
);
3623 ext4_unlock_group(sb
, grp
);
3625 spin_lock(pa
->pa_obj_lock
);
3626 list_del_rcu(&pa
->pa_inode_list
);
3627 spin_unlock(pa
->pa_obj_lock
);
3629 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
3633 * creates new preallocated space for given inode
3635 static noinline_for_stack
int
3636 ext4_mb_new_inode_pa(struct ext4_allocation_context
*ac
)
3638 struct super_block
*sb
= ac
->ac_sb
;
3639 struct ext4_prealloc_space
*pa
;
3640 struct ext4_group_info
*grp
;
3641 struct ext4_inode_info
*ei
;
3643 /* preallocate only when found space is larger then requested */
3644 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3645 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3646 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3648 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3652 if (ac
->ac_b_ex
.fe_len
< ac
->ac_g_ex
.fe_len
) {
3658 /* we can't allocate as much as normalizer wants.
3659 * so, found space must get proper lstart
3660 * to cover original request */
3661 BUG_ON(ac
->ac_g_ex
.fe_logical
> ac
->ac_o_ex
.fe_logical
);
3662 BUG_ON(ac
->ac_g_ex
.fe_len
< ac
->ac_o_ex
.fe_len
);
3664 /* we're limited by original request in that
3665 * logical block must be covered any way
3666 * winl is window we can move our chunk within */
3667 winl
= ac
->ac_o_ex
.fe_logical
- ac
->ac_g_ex
.fe_logical
;
3669 /* also, we should cover whole original request */
3670 wins
= ac
->ac_b_ex
.fe_len
- ac
->ac_o_ex
.fe_len
;
3672 /* the smallest one defines real window */
3673 win
= min(winl
, wins
);
3675 offs
= ac
->ac_o_ex
.fe_logical
% ac
->ac_b_ex
.fe_len
;
3676 if (offs
&& offs
< win
)
3679 ac
->ac_b_ex
.fe_logical
= ac
->ac_o_ex
.fe_logical
- win
;
3680 BUG_ON(ac
->ac_o_ex
.fe_logical
< ac
->ac_b_ex
.fe_logical
);
3681 BUG_ON(ac
->ac_o_ex
.fe_len
> ac
->ac_b_ex
.fe_len
);
3684 /* preallocation can change ac_b_ex, thus we store actually
3685 * allocated blocks for history */
3686 ac
->ac_f_ex
= ac
->ac_b_ex
;
3688 pa
->pa_lstart
= ac
->ac_b_ex
.fe_logical
;
3689 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3690 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3691 pa
->pa_free
= pa
->pa_len
;
3692 atomic_set(&pa
->pa_count
, 1);
3693 spin_lock_init(&pa
->pa_lock
);
3697 mb_debug("new inode pa %p: %llu/%u for %u\n", pa
,
3698 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3700 ext4_mb_use_inode_pa(ac
, pa
);
3701 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3703 ei
= EXT4_I(ac
->ac_inode
);
3704 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3706 pa
->pa_obj_lock
= &ei
->i_prealloc_lock
;
3707 pa
->pa_inode
= ac
->ac_inode
;
3709 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3710 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3711 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3713 spin_lock(pa
->pa_obj_lock
);
3714 list_add_rcu(&pa
->pa_inode_list
, &ei
->i_prealloc_list
);
3715 spin_unlock(pa
->pa_obj_lock
);
3721 * creates new preallocated space for locality group inodes belongs to
3723 static noinline_for_stack
int
3724 ext4_mb_new_group_pa(struct ext4_allocation_context
*ac
)
3726 struct super_block
*sb
= ac
->ac_sb
;
3727 struct ext4_locality_group
*lg
;
3728 struct ext4_prealloc_space
*pa
;
3729 struct ext4_group_info
*grp
;
3731 /* preallocate only when found space is larger then requested */
3732 BUG_ON(ac
->ac_o_ex
.fe_len
>= ac
->ac_b_ex
.fe_len
);
3733 BUG_ON(ac
->ac_status
!= AC_STATUS_FOUND
);
3734 BUG_ON(!S_ISREG(ac
->ac_inode
->i_mode
));
3736 BUG_ON(ext4_pspace_cachep
== NULL
);
3737 pa
= kmem_cache_alloc(ext4_pspace_cachep
, GFP_NOFS
);
3741 /* preallocation can change ac_b_ex, thus we store actually
3742 * allocated blocks for history */
3743 ac
->ac_f_ex
= ac
->ac_b_ex
;
3745 pa
->pa_pstart
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
3746 pa
->pa_lstart
= pa
->pa_pstart
;
3747 pa
->pa_len
= ac
->ac_b_ex
.fe_len
;
3748 pa
->pa_free
= pa
->pa_len
;
3749 atomic_set(&pa
->pa_count
, 1);
3750 spin_lock_init(&pa
->pa_lock
);
3751 INIT_LIST_HEAD(&pa
->pa_inode_list
);
3755 mb_debug("new group pa %p: %llu/%u for %u\n", pa
,
3756 pa
->pa_pstart
, pa
->pa_len
, pa
->pa_lstart
);
3758 ext4_mb_use_group_pa(ac
, pa
);
3759 atomic_add(pa
->pa_free
, &EXT4_SB(sb
)->s_mb_preallocated
);
3761 grp
= ext4_get_group_info(sb
, ac
->ac_b_ex
.fe_group
);
3765 pa
->pa_obj_lock
= &lg
->lg_prealloc_lock
;
3766 pa
->pa_inode
= NULL
;
3768 ext4_lock_group(sb
, ac
->ac_b_ex
.fe_group
);
3769 list_add(&pa
->pa_group_list
, &grp
->bb_prealloc_list
);
3770 ext4_unlock_group(sb
, ac
->ac_b_ex
.fe_group
);
3773 * We will later add the new pa to the right bucket
3774 * after updating the pa_free in ext4_mb_release_context
3779 static int ext4_mb_new_preallocation(struct ext4_allocation_context
*ac
)
3783 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
3784 err
= ext4_mb_new_group_pa(ac
);
3786 err
= ext4_mb_new_inode_pa(ac
);
3791 * finds all unused blocks in on-disk bitmap, frees them in
3792 * in-core bitmap and buddy.
3793 * @pa must be unlinked from inode and group lists, so that
3794 * nobody else can find/use it.
3795 * the caller MUST hold group/inode locks.
3796 * TODO: optimize the case when there are no in-core structures yet
3798 static noinline_for_stack
int
3799 ext4_mb_release_inode_pa(struct ext4_buddy
*e4b
, struct buffer_head
*bitmap_bh
,
3800 struct ext4_prealloc_space
*pa
,
3801 struct ext4_allocation_context
*ac
)
3803 struct super_block
*sb
= e4b
->bd_sb
;
3804 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
3813 BUG_ON(pa
->pa_deleted
== 0);
3814 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3815 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3816 end
= bit
+ pa
->pa_len
;
3820 ac
->ac_inode
= pa
->pa_inode
;
3821 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3825 bit
= mb_find_next_zero_bit(bitmap_bh
->b_data
, end
, bit
);
3828 next
= mb_find_next_bit(bitmap_bh
->b_data
, end
, bit
);
3829 start
= group
* EXT4_BLOCKS_PER_GROUP(sb
) + bit
+
3830 le32_to_cpu(sbi
->s_es
->s_first_data_block
);
3831 mb_debug(" free preallocated %u/%u in group %u\n",
3832 (unsigned) start
, (unsigned) next
- bit
,
3837 ac
->ac_b_ex
.fe_group
= group
;
3838 ac
->ac_b_ex
.fe_start
= bit
;
3839 ac
->ac_b_ex
.fe_len
= next
- bit
;
3840 ac
->ac_b_ex
.fe_logical
= 0;
3841 ext4_mb_store_history(ac
);
3844 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, next
- bit
);
3847 if (free
!= pa
->pa_free
) {
3848 printk(KERN_CRIT
"pa %p: logic %lu, phys. %lu, len %lu\n",
3849 pa
, (unsigned long) pa
->pa_lstart
,
3850 (unsigned long) pa
->pa_pstart
,
3851 (unsigned long) pa
->pa_len
);
3852 ext4_grp_locked_error(sb
, group
,
3853 __func__
, "free %u, pa_free %u",
3856 * pa is already deleted so we use the value obtained
3857 * from the bitmap and continue.
3860 atomic_add(free
, &sbi
->s_mb_discarded
);
3865 static noinline_for_stack
int
3866 ext4_mb_release_group_pa(struct ext4_buddy
*e4b
,
3867 struct ext4_prealloc_space
*pa
,
3868 struct ext4_allocation_context
*ac
)
3870 struct super_block
*sb
= e4b
->bd_sb
;
3875 ac
->ac_op
= EXT4_MB_HISTORY_DISCARD
;
3877 BUG_ON(pa
->pa_deleted
== 0);
3878 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, &bit
);
3879 BUG_ON(group
!= e4b
->bd_group
&& pa
->pa_len
!= 0);
3880 mb_free_blocks(pa
->pa_inode
, e4b
, bit
, pa
->pa_len
);
3881 atomic_add(pa
->pa_len
, &EXT4_SB(sb
)->s_mb_discarded
);
3885 ac
->ac_inode
= NULL
;
3886 ac
->ac_b_ex
.fe_group
= group
;
3887 ac
->ac_b_ex
.fe_start
= bit
;
3888 ac
->ac_b_ex
.fe_len
= pa
->pa_len
;
3889 ac
->ac_b_ex
.fe_logical
= 0;
3890 ext4_mb_store_history(ac
);
3897 * releases all preallocations in given group
3899 * first, we need to decide discard policy:
3900 * - when do we discard
3902 * - how many do we discard
3903 * 1) how many requested
3905 static noinline_for_stack
int
3906 ext4_mb_discard_group_preallocations(struct super_block
*sb
,
3907 ext4_group_t group
, int needed
)
3909 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, group
);
3910 struct buffer_head
*bitmap_bh
= NULL
;
3911 struct ext4_prealloc_space
*pa
, *tmp
;
3912 struct ext4_allocation_context
*ac
;
3913 struct list_head list
;
3914 struct ext4_buddy e4b
;
3919 mb_debug("discard preallocation for group %u\n", group
);
3921 if (list_empty(&grp
->bb_prealloc_list
))
3924 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
3925 if (bitmap_bh
== NULL
) {
3926 ext4_error(sb
, __func__
, "Error in reading block "
3927 "bitmap for %u", group
);
3931 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
3933 ext4_error(sb
, __func__
, "Error in loading buddy "
3934 "information for %u", group
);
3940 needed
= EXT4_BLOCKS_PER_GROUP(sb
) + 1;
3942 INIT_LIST_HEAD(&list
);
3943 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
3945 ext4_lock_group(sb
, group
);
3946 list_for_each_entry_safe(pa
, tmp
,
3947 &grp
->bb_prealloc_list
, pa_group_list
) {
3948 spin_lock(&pa
->pa_lock
);
3949 if (atomic_read(&pa
->pa_count
)) {
3950 spin_unlock(&pa
->pa_lock
);
3954 if (pa
->pa_deleted
) {
3955 spin_unlock(&pa
->pa_lock
);
3959 /* seems this one can be freed ... */
3962 /* we can trust pa_free ... */
3963 free
+= pa
->pa_free
;
3965 spin_unlock(&pa
->pa_lock
);
3967 list_del(&pa
->pa_group_list
);
3968 list_add(&pa
->u
.pa_tmp_list
, &list
);
3971 /* if we still need more blocks and some PAs were used, try again */
3972 if (free
< needed
&& busy
) {
3974 ext4_unlock_group(sb
, group
);
3976 * Yield the CPU here so that we don't get soft lockup
3977 * in non preempt case.
3983 /* found anything to free? */
3984 if (list_empty(&list
)) {
3989 /* now free all selected PAs */
3990 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
3992 /* remove from object (inode or locality group) */
3993 spin_lock(pa
->pa_obj_lock
);
3994 list_del_rcu(&pa
->pa_inode_list
);
3995 spin_unlock(pa
->pa_obj_lock
);
3998 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4000 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
4002 list_del(&pa
->u
.pa_tmp_list
);
4003 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4007 ext4_unlock_group(sb
, group
);
4009 kmem_cache_free(ext4_ac_cachep
, ac
);
4010 ext4_mb_release_desc(&e4b
);
4016 * releases all non-used preallocated blocks for given inode
4018 * It's important to discard preallocations under i_data_sem
4019 * We don't want another block to be served from the prealloc
4020 * space when we are discarding the inode prealloc space.
4022 * FIXME!! Make sure it is valid at all the call sites
4024 void ext4_discard_preallocations(struct inode
*inode
)
4026 struct ext4_inode_info
*ei
= EXT4_I(inode
);
4027 struct super_block
*sb
= inode
->i_sb
;
4028 struct buffer_head
*bitmap_bh
= NULL
;
4029 struct ext4_prealloc_space
*pa
, *tmp
;
4030 struct ext4_allocation_context
*ac
;
4031 ext4_group_t group
= 0;
4032 struct list_head list
;
4033 struct ext4_buddy e4b
;
4036 if (!S_ISREG(inode
->i_mode
)) {
4037 /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
4041 mb_debug("discard preallocation for inode %lu\n", inode
->i_ino
);
4043 INIT_LIST_HEAD(&list
);
4045 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4047 /* first, collect all pa's in the inode */
4048 spin_lock(&ei
->i_prealloc_lock
);
4049 while (!list_empty(&ei
->i_prealloc_list
)) {
4050 pa
= list_entry(ei
->i_prealloc_list
.next
,
4051 struct ext4_prealloc_space
, pa_inode_list
);
4052 BUG_ON(pa
->pa_obj_lock
!= &ei
->i_prealloc_lock
);
4053 spin_lock(&pa
->pa_lock
);
4054 if (atomic_read(&pa
->pa_count
)) {
4055 /* this shouldn't happen often - nobody should
4056 * use preallocation while we're discarding it */
4057 spin_unlock(&pa
->pa_lock
);
4058 spin_unlock(&ei
->i_prealloc_lock
);
4059 printk(KERN_ERR
"uh-oh! used pa while discarding\n");
4061 schedule_timeout_uninterruptible(HZ
);
4065 if (pa
->pa_deleted
== 0) {
4067 spin_unlock(&pa
->pa_lock
);
4068 list_del_rcu(&pa
->pa_inode_list
);
4069 list_add(&pa
->u
.pa_tmp_list
, &list
);
4073 /* someone is deleting pa right now */
4074 spin_unlock(&pa
->pa_lock
);
4075 spin_unlock(&ei
->i_prealloc_lock
);
4077 /* we have to wait here because pa_deleted
4078 * doesn't mean pa is already unlinked from
4079 * the list. as we might be called from
4080 * ->clear_inode() the inode will get freed
4081 * and concurrent thread which is unlinking
4082 * pa from inode's list may access already
4083 * freed memory, bad-bad-bad */
4085 /* XXX: if this happens too often, we can
4086 * add a flag to force wait only in case
4087 * of ->clear_inode(), but not in case of
4088 * regular truncate */
4089 schedule_timeout_uninterruptible(HZ
);
4092 spin_unlock(&ei
->i_prealloc_lock
);
4094 list_for_each_entry_safe(pa
, tmp
, &list
, u
.pa_tmp_list
) {
4095 BUG_ON(pa
->pa_linear
!= 0);
4096 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4098 err
= ext4_mb_load_buddy(sb
, group
, &e4b
);
4100 ext4_error(sb
, __func__
, "Error in loading buddy "
4101 "information for %u", group
);
4105 bitmap_bh
= ext4_read_block_bitmap(sb
, group
);
4106 if (bitmap_bh
== NULL
) {
4107 ext4_error(sb
, __func__
, "Error in reading block "
4108 "bitmap for %u", group
);
4109 ext4_mb_release_desc(&e4b
);
4113 ext4_lock_group(sb
, group
);
4114 list_del(&pa
->pa_group_list
);
4115 ext4_mb_release_inode_pa(&e4b
, bitmap_bh
, pa
, ac
);
4116 ext4_unlock_group(sb
, group
);
4118 ext4_mb_release_desc(&e4b
);
4121 list_del(&pa
->u
.pa_tmp_list
);
4122 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4125 kmem_cache_free(ext4_ac_cachep
, ac
);
4129 * finds all preallocated spaces and return blocks being freed to them
4130 * if preallocated space becomes full (no block is used from the space)
4131 * then the function frees space in buddy
4132 * XXX: at the moment, truncate (which is the only way to free blocks)
4133 * discards all preallocations
4135 static void ext4_mb_return_to_preallocation(struct inode
*inode
,
4136 struct ext4_buddy
*e4b
,
4137 sector_t block
, int count
)
4139 BUG_ON(!list_empty(&EXT4_I(inode
)->i_prealloc_list
));
4142 static void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4144 struct super_block
*sb
= ac
->ac_sb
;
4147 printk(KERN_ERR
"EXT4-fs: Can't allocate:"
4148 " Allocation context details:\n");
4149 printk(KERN_ERR
"EXT4-fs: status %d flags %d\n",
4150 ac
->ac_status
, ac
->ac_flags
);
4151 printk(KERN_ERR
"EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
4152 "best %lu/%lu/%lu@%lu cr %d\n",
4153 (unsigned long)ac
->ac_o_ex
.fe_group
,
4154 (unsigned long)ac
->ac_o_ex
.fe_start
,
4155 (unsigned long)ac
->ac_o_ex
.fe_len
,
4156 (unsigned long)ac
->ac_o_ex
.fe_logical
,
4157 (unsigned long)ac
->ac_g_ex
.fe_group
,
4158 (unsigned long)ac
->ac_g_ex
.fe_start
,
4159 (unsigned long)ac
->ac_g_ex
.fe_len
,
4160 (unsigned long)ac
->ac_g_ex
.fe_logical
,
4161 (unsigned long)ac
->ac_b_ex
.fe_group
,
4162 (unsigned long)ac
->ac_b_ex
.fe_start
,
4163 (unsigned long)ac
->ac_b_ex
.fe_len
,
4164 (unsigned long)ac
->ac_b_ex
.fe_logical
,
4165 (int)ac
->ac_criteria
);
4166 printk(KERN_ERR
"EXT4-fs: %lu scanned, %d found\n", ac
->ac_ex_scanned
,
4168 printk(KERN_ERR
"EXT4-fs: groups: \n");
4169 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
; i
++) {
4170 struct ext4_group_info
*grp
= ext4_get_group_info(sb
, i
);
4171 struct ext4_prealloc_space
*pa
;
4172 ext4_grpblk_t start
;
4173 struct list_head
*cur
;
4174 ext4_lock_group(sb
, i
);
4175 list_for_each(cur
, &grp
->bb_prealloc_list
) {
4176 pa
= list_entry(cur
, struct ext4_prealloc_space
,
4178 spin_lock(&pa
->pa_lock
);
4179 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
,
4181 spin_unlock(&pa
->pa_lock
);
4182 printk(KERN_ERR
"PA:%lu:%d:%u \n", i
,
4185 ext4_unlock_group(sb
, i
);
4187 if (grp
->bb_free
== 0)
4189 printk(KERN_ERR
"%lu: %d/%d \n",
4190 i
, grp
->bb_free
, grp
->bb_fragments
);
4192 printk(KERN_ERR
"\n");
4195 static inline void ext4_mb_show_ac(struct ext4_allocation_context
*ac
)
4202 * We use locality group preallocation for small size file. The size of the
4203 * file is determined by the current size or the resulting size after
4204 * allocation which ever is larger
4206 * One can tune this size via /proc/fs/ext4/<partition>/stream_req
4208 static void ext4_mb_group_or_file(struct ext4_allocation_context
*ac
)
4210 struct ext4_sb_info
*sbi
= EXT4_SB(ac
->ac_sb
);
4211 int bsbits
= ac
->ac_sb
->s_blocksize_bits
;
4214 if (!(ac
->ac_flags
& EXT4_MB_HINT_DATA
))
4217 size
= ac
->ac_o_ex
.fe_logical
+ ac
->ac_o_ex
.fe_len
;
4218 isize
= i_size_read(ac
->ac_inode
) >> bsbits
;
4219 size
= max(size
, isize
);
4221 /* don't use group allocation for large files */
4222 if (size
>= sbi
->s_mb_stream_request
)
4225 if (unlikely(ac
->ac_flags
& EXT4_MB_HINT_GOAL_ONLY
))
4228 BUG_ON(ac
->ac_lg
!= NULL
);
4230 * locality group prealloc space are per cpu. The reason for having
4231 * per cpu locality group is to reduce the contention between block
4232 * request from multiple CPUs.
4234 ac
->ac_lg
= per_cpu_ptr(sbi
->s_locality_groups
, raw_smp_processor_id());
4236 /* we're going to use group allocation */
4237 ac
->ac_flags
|= EXT4_MB_HINT_GROUP_ALLOC
;
4239 /* serialize all allocations in the group */
4240 mutex_lock(&ac
->ac_lg
->lg_mutex
);
4243 static noinline_for_stack
int
4244 ext4_mb_initialize_context(struct ext4_allocation_context
*ac
,
4245 struct ext4_allocation_request
*ar
)
4247 struct super_block
*sb
= ar
->inode
->i_sb
;
4248 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4249 struct ext4_super_block
*es
= sbi
->s_es
;
4253 ext4_grpblk_t block
;
4255 /* we can't allocate > group size */
4258 /* just a dirty hack to filter too big requests */
4259 if (len
>= EXT4_BLOCKS_PER_GROUP(sb
) - 10)
4260 len
= EXT4_BLOCKS_PER_GROUP(sb
) - 10;
4262 /* start searching from the goal */
4264 if (goal
< le32_to_cpu(es
->s_first_data_block
) ||
4265 goal
>= ext4_blocks_count(es
))
4266 goal
= le32_to_cpu(es
->s_first_data_block
);
4267 ext4_get_group_no_and_offset(sb
, goal
, &group
, &block
);
4269 /* set up allocation goals */
4270 ac
->ac_b_ex
.fe_logical
= ar
->logical
;
4271 ac
->ac_b_ex
.fe_group
= 0;
4272 ac
->ac_b_ex
.fe_start
= 0;
4273 ac
->ac_b_ex
.fe_len
= 0;
4274 ac
->ac_status
= AC_STATUS_CONTINUE
;
4275 ac
->ac_groups_scanned
= 0;
4276 ac
->ac_ex_scanned
= 0;
4279 ac
->ac_inode
= ar
->inode
;
4280 ac
->ac_o_ex
.fe_logical
= ar
->logical
;
4281 ac
->ac_o_ex
.fe_group
= group
;
4282 ac
->ac_o_ex
.fe_start
= block
;
4283 ac
->ac_o_ex
.fe_len
= len
;
4284 ac
->ac_g_ex
.fe_logical
= ar
->logical
;
4285 ac
->ac_g_ex
.fe_group
= group
;
4286 ac
->ac_g_ex
.fe_start
= block
;
4287 ac
->ac_g_ex
.fe_len
= len
;
4288 ac
->ac_f_ex
.fe_len
= 0;
4289 ac
->ac_flags
= ar
->flags
;
4291 ac
->ac_criteria
= 0;
4293 ac
->ac_bitmap_page
= NULL
;
4294 ac
->ac_buddy_page
= NULL
;
4295 ac
->alloc_semp
= NULL
;
4298 /* we have to define context: we'll we work with a file or
4299 * locality group. this is a policy, actually */
4300 ext4_mb_group_or_file(ac
);
4302 mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4303 "left: %u/%u, right %u/%u to %swritable\n",
4304 (unsigned) ar
->len
, (unsigned) ar
->logical
,
4305 (unsigned) ar
->goal
, ac
->ac_flags
, ac
->ac_2order
,
4306 (unsigned) ar
->lleft
, (unsigned) ar
->pleft
,
4307 (unsigned) ar
->lright
, (unsigned) ar
->pright
,
4308 atomic_read(&ar
->inode
->i_writecount
) ? "" : "non-");
4313 static noinline_for_stack
void
4314 ext4_mb_discard_lg_preallocations(struct super_block
*sb
,
4315 struct ext4_locality_group
*lg
,
4316 int order
, int total_entries
)
4318 ext4_group_t group
= 0;
4319 struct ext4_buddy e4b
;
4320 struct list_head discard_list
;
4321 struct ext4_prealloc_space
*pa
, *tmp
;
4322 struct ext4_allocation_context
*ac
;
4324 mb_debug("discard locality group preallocation\n");
4326 INIT_LIST_HEAD(&discard_list
);
4327 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4329 spin_lock(&lg
->lg_prealloc_lock
);
4330 list_for_each_entry_rcu(pa
, &lg
->lg_prealloc_list
[order
],
4332 spin_lock(&pa
->pa_lock
);
4333 if (atomic_read(&pa
->pa_count
)) {
4335 * This is the pa that we just used
4336 * for block allocation. So don't
4339 spin_unlock(&pa
->pa_lock
);
4342 if (pa
->pa_deleted
) {
4343 spin_unlock(&pa
->pa_lock
);
4346 /* only lg prealloc space */
4347 BUG_ON(!pa
->pa_linear
);
4349 /* seems this one can be freed ... */
4351 spin_unlock(&pa
->pa_lock
);
4353 list_del_rcu(&pa
->pa_inode_list
);
4354 list_add(&pa
->u
.pa_tmp_list
, &discard_list
);
4357 if (total_entries
<= 5) {
4359 * we want to keep only 5 entries
4360 * allowing it to grow to 8. This
4361 * mak sure we don't call discard
4362 * soon for this list.
4367 spin_unlock(&lg
->lg_prealloc_lock
);
4369 list_for_each_entry_safe(pa
, tmp
, &discard_list
, u
.pa_tmp_list
) {
4371 ext4_get_group_no_and_offset(sb
, pa
->pa_pstart
, &group
, NULL
);
4372 if (ext4_mb_load_buddy(sb
, group
, &e4b
)) {
4373 ext4_error(sb
, __func__
, "Error in loading buddy "
4374 "information for %u", group
);
4377 ext4_lock_group(sb
, group
);
4378 list_del(&pa
->pa_group_list
);
4379 ext4_mb_release_group_pa(&e4b
, pa
, ac
);
4380 ext4_unlock_group(sb
, group
);
4382 ext4_mb_release_desc(&e4b
);
4383 list_del(&pa
->u
.pa_tmp_list
);
4384 call_rcu(&(pa
)->u
.pa_rcu
, ext4_mb_pa_callback
);
4387 kmem_cache_free(ext4_ac_cachep
, ac
);
4391 * We have incremented pa_count. So it cannot be freed at this
4392 * point. Also we hold lg_mutex. So no parallel allocation is
4393 * possible from this lg. That means pa_free cannot be updated.
4395 * A parallel ext4_mb_discard_group_preallocations is possible.
4396 * which can cause the lg_prealloc_list to be updated.
4399 static void ext4_mb_add_n_trim(struct ext4_allocation_context
*ac
)
4401 int order
, added
= 0, lg_prealloc_count
= 1;
4402 struct super_block
*sb
= ac
->ac_sb
;
4403 struct ext4_locality_group
*lg
= ac
->ac_lg
;
4404 struct ext4_prealloc_space
*tmp_pa
, *pa
= ac
->ac_pa
;
4406 order
= fls(pa
->pa_free
) - 1;
4407 if (order
> PREALLOC_TB_SIZE
- 1)
4408 /* The max size of hash table is PREALLOC_TB_SIZE */
4409 order
= PREALLOC_TB_SIZE
- 1;
4410 /* Add the prealloc space to lg */
4412 list_for_each_entry_rcu(tmp_pa
, &lg
->lg_prealloc_list
[order
],
4414 spin_lock(&tmp_pa
->pa_lock
);
4415 if (tmp_pa
->pa_deleted
) {
4416 spin_unlock(&pa
->pa_lock
);
4419 if (!added
&& pa
->pa_free
< tmp_pa
->pa_free
) {
4420 /* Add to the tail of the previous entry */
4421 list_add_tail_rcu(&pa
->pa_inode_list
,
4422 &tmp_pa
->pa_inode_list
);
4425 * we want to count the total
4426 * number of entries in the list
4429 spin_unlock(&tmp_pa
->pa_lock
);
4430 lg_prealloc_count
++;
4433 list_add_tail_rcu(&pa
->pa_inode_list
,
4434 &lg
->lg_prealloc_list
[order
]);
4437 /* Now trim the list to be not more than 8 elements */
4438 if (lg_prealloc_count
> 8) {
4439 ext4_mb_discard_lg_preallocations(sb
, lg
,
4440 order
, lg_prealloc_count
);
4447 * release all resource we used in allocation
4449 static int ext4_mb_release_context(struct ext4_allocation_context
*ac
)
4451 struct ext4_prealloc_space
*pa
= ac
->ac_pa
;
4453 if (pa
->pa_linear
) {
4454 /* see comment in ext4_mb_use_group_pa() */
4455 spin_lock(&pa
->pa_lock
);
4456 pa
->pa_pstart
+= ac
->ac_b_ex
.fe_len
;
4457 pa
->pa_lstart
+= ac
->ac_b_ex
.fe_len
;
4458 pa
->pa_free
-= ac
->ac_b_ex
.fe_len
;
4459 pa
->pa_len
-= ac
->ac_b_ex
.fe_len
;
4460 spin_unlock(&pa
->pa_lock
);
4462 * We want to add the pa to the right bucket.
4463 * Remove it from the list and while adding
4464 * make sure the list to which we are adding
4467 if (likely(pa
->pa_free
)) {
4468 spin_lock(pa
->pa_obj_lock
);
4469 list_del_rcu(&pa
->pa_inode_list
);
4470 spin_unlock(pa
->pa_obj_lock
);
4471 ext4_mb_add_n_trim(ac
);
4474 ext4_mb_put_pa(ac
, ac
->ac_sb
, pa
);
4477 up_read(ac
->alloc_semp
);
4478 if (ac
->ac_bitmap_page
)
4479 page_cache_release(ac
->ac_bitmap_page
);
4480 if (ac
->ac_buddy_page
)
4481 page_cache_release(ac
->ac_buddy_page
);
4482 if (ac
->ac_flags
& EXT4_MB_HINT_GROUP_ALLOC
)
4483 mutex_unlock(&ac
->ac_lg
->lg_mutex
);
4484 ext4_mb_collect_stats(ac
);
4488 static int ext4_mb_discard_preallocations(struct super_block
*sb
, int needed
)
4494 for (i
= 0; i
< EXT4_SB(sb
)->s_groups_count
&& needed
> 0; i
++) {
4495 ret
= ext4_mb_discard_group_preallocations(sb
, i
, needed
);
4504 * Main entry point into mballoc to allocate blocks
4505 * it tries to use preallocation first, then falls back
4506 * to usual allocation
4508 ext4_fsblk_t
ext4_mb_new_blocks(handle_t
*handle
,
4509 struct ext4_allocation_request
*ar
, int *errp
)
4512 struct ext4_allocation_context
*ac
= NULL
;
4513 struct ext4_sb_info
*sbi
;
4514 struct super_block
*sb
;
4515 ext4_fsblk_t block
= 0;
4516 unsigned int inquota
;
4517 unsigned int reserv_blks
= 0;
4519 sb
= ar
->inode
->i_sb
;
4522 if (!EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
) {
4524 * With delalloc we already reserved the blocks
4526 while (ar
->len
&& ext4_claim_free_blocks(sbi
, ar
->len
)) {
4527 /* let others to free the space */
4529 ar
->len
= ar
->len
>> 1;
4535 reserv_blks
= ar
->len
;
4537 while (ar
->len
&& DQUOT_ALLOC_BLOCK(ar
->inode
, ar
->len
)) {
4538 ar
->flags
|= EXT4_MB_HINT_NOPREALLOC
;
4547 if (EXT4_I(ar
->inode
)->i_delalloc_reserved_flag
)
4548 ar
->flags
|= EXT4_MB_DELALLOC_RESERVED
;
4550 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4557 *errp
= ext4_mb_initialize_context(ac
, ar
);
4563 ac
->ac_op
= EXT4_MB_HISTORY_PREALLOC
;
4564 if (!ext4_mb_use_preallocated(ac
)) {
4565 ac
->ac_op
= EXT4_MB_HISTORY_ALLOC
;
4566 ext4_mb_normalize_request(ac
, ar
);
4568 /* allocate space in core */
4569 ext4_mb_regular_allocator(ac
);
4571 /* as we've just preallocated more space than
4572 * user requested orinally, we store allocated
4573 * space in a special descriptor */
4574 if (ac
->ac_status
== AC_STATUS_FOUND
&&
4575 ac
->ac_o_ex
.fe_len
< ac
->ac_b_ex
.fe_len
)
4576 ext4_mb_new_preallocation(ac
);
4578 if (likely(ac
->ac_status
== AC_STATUS_FOUND
)) {
4579 *errp
= ext4_mb_mark_diskspace_used(ac
, handle
, reserv_blks
);
4580 if (*errp
== -EAGAIN
) {
4582 * drop the reference that we took
4583 * in ext4_mb_use_best_found
4585 ext4_mb_release_context(ac
);
4586 ac
->ac_b_ex
.fe_group
= 0;
4587 ac
->ac_b_ex
.fe_start
= 0;
4588 ac
->ac_b_ex
.fe_len
= 0;
4589 ac
->ac_status
= AC_STATUS_CONTINUE
;
4592 ac
->ac_b_ex
.fe_len
= 0;
4594 ext4_mb_show_ac(ac
);
4596 block
= ext4_grp_offs_to_block(sb
, &ac
->ac_b_ex
);
4597 ar
->len
= ac
->ac_b_ex
.fe_len
;
4600 freed
= ext4_mb_discard_preallocations(sb
, ac
->ac_o_ex
.fe_len
);
4604 ac
->ac_b_ex
.fe_len
= 0;
4606 ext4_mb_show_ac(ac
);
4609 ext4_mb_release_context(ac
);
4612 kmem_cache_free(ext4_ac_cachep
, ac
);
4614 if (ar
->len
< inquota
)
4615 DQUOT_FREE_BLOCK(ar
->inode
, inquota
- ar
->len
);
4621 * We can merge two free data extents only if the physical blocks
4622 * are contiguous, AND the extents were freed by the same transaction,
4623 * AND the blocks are associated with the same group.
4625 static int can_merge(struct ext4_free_data
*entry1
,
4626 struct ext4_free_data
*entry2
)
4628 if ((entry1
->t_tid
== entry2
->t_tid
) &&
4629 (entry1
->group
== entry2
->group
) &&
4630 ((entry1
->start_blk
+ entry1
->count
) == entry2
->start_blk
))
4635 static noinline_for_stack
int
4636 ext4_mb_free_metadata(handle_t
*handle
, struct ext4_buddy
*e4b
,
4637 struct ext4_free_data
*new_entry
)
4639 ext4_grpblk_t block
;
4640 struct ext4_free_data
*entry
;
4641 struct ext4_group_info
*db
= e4b
->bd_info
;
4642 struct super_block
*sb
= e4b
->bd_sb
;
4643 struct ext4_sb_info
*sbi
= EXT4_SB(sb
);
4644 struct rb_node
**n
= &db
->bb_free_root
.rb_node
, *node
;
4645 struct rb_node
*parent
= NULL
, *new_node
;
4647 BUG_ON(!ext4_handle_valid(handle
));
4648 BUG_ON(e4b
->bd_bitmap_page
== NULL
);
4649 BUG_ON(e4b
->bd_buddy_page
== NULL
);
4651 new_node
= &new_entry
->node
;
4652 block
= new_entry
->start_blk
;
4655 /* first free block exent. We need to
4656 protect buddy cache from being freed,
4657 * otherwise we'll refresh it from
4658 * on-disk bitmap and lose not-yet-available
4660 page_cache_get(e4b
->bd_buddy_page
);
4661 page_cache_get(e4b
->bd_bitmap_page
);
4665 entry
= rb_entry(parent
, struct ext4_free_data
, node
);
4666 if (block
< entry
->start_blk
)
4668 else if (block
>= (entry
->start_blk
+ entry
->count
))
4669 n
= &(*n
)->rb_right
;
4671 ext4_grp_locked_error(sb
, e4b
->bd_group
, __func__
,
4672 "Double free of blocks %d (%d %d)",
4673 block
, entry
->start_blk
, entry
->count
);
4678 rb_link_node(new_node
, parent
, n
);
4679 rb_insert_color(new_node
, &db
->bb_free_root
);
4681 /* Now try to see the extent can be merged to left and right */
4682 node
= rb_prev(new_node
);
4684 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4685 if (can_merge(entry
, new_entry
)) {
4686 new_entry
->start_blk
= entry
->start_blk
;
4687 new_entry
->count
+= entry
->count
;
4688 rb_erase(node
, &(db
->bb_free_root
));
4689 spin_lock(&sbi
->s_md_lock
);
4690 list_del(&entry
->list
);
4691 spin_unlock(&sbi
->s_md_lock
);
4692 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4696 node
= rb_next(new_node
);
4698 entry
= rb_entry(node
, struct ext4_free_data
, node
);
4699 if (can_merge(new_entry
, entry
)) {
4700 new_entry
->count
+= entry
->count
;
4701 rb_erase(node
, &(db
->bb_free_root
));
4702 spin_lock(&sbi
->s_md_lock
);
4703 list_del(&entry
->list
);
4704 spin_unlock(&sbi
->s_md_lock
);
4705 kmem_cache_free(ext4_free_ext_cachep
, entry
);
4708 /* Add the extent to transaction's private list */
4709 spin_lock(&sbi
->s_md_lock
);
4710 list_add(&new_entry
->list
, &handle
->h_transaction
->t_private_list
);
4711 spin_unlock(&sbi
->s_md_lock
);
4716 * Main entry point into mballoc to free blocks
4718 void ext4_mb_free_blocks(handle_t
*handle
, struct inode
*inode
,
4719 unsigned long block
, unsigned long count
,
4720 int metadata
, unsigned long *freed
)
4722 struct buffer_head
*bitmap_bh
= NULL
;
4723 struct super_block
*sb
= inode
->i_sb
;
4724 struct ext4_allocation_context
*ac
= NULL
;
4725 struct ext4_group_desc
*gdp
;
4726 struct ext4_super_block
*es
;
4727 unsigned int overflow
;
4729 struct buffer_head
*gd_bh
;
4730 ext4_group_t block_group
;
4731 struct ext4_sb_info
*sbi
;
4732 struct ext4_buddy e4b
;
4739 es
= EXT4_SB(sb
)->s_es
;
4740 if (block
< le32_to_cpu(es
->s_first_data_block
) ||
4741 block
+ count
< block
||
4742 block
+ count
> ext4_blocks_count(es
)) {
4743 ext4_error(sb
, __func__
,
4744 "Freeing blocks not in datazone - "
4745 "block = %lu, count = %lu", block
, count
);
4749 ext4_debug("freeing block %lu\n", block
);
4751 ac
= kmem_cache_alloc(ext4_ac_cachep
, GFP_NOFS
);
4753 ac
->ac_op
= EXT4_MB_HISTORY_FREE
;
4754 ac
->ac_inode
= inode
;
4760 ext4_get_group_no_and_offset(sb
, block
, &block_group
, &bit
);
4763 * Check to see if we are freeing blocks across a group
4766 if (bit
+ count
> EXT4_BLOCKS_PER_GROUP(sb
)) {
4767 overflow
= bit
+ count
- EXT4_BLOCKS_PER_GROUP(sb
);
4770 bitmap_bh
= ext4_read_block_bitmap(sb
, block_group
);
4775 gdp
= ext4_get_group_desc(sb
, block_group
, &gd_bh
);
4781 if (in_range(ext4_block_bitmap(sb
, gdp
), block
, count
) ||
4782 in_range(ext4_inode_bitmap(sb
, gdp
), block
, count
) ||
4783 in_range(block
, ext4_inode_table(sb
, gdp
),
4784 EXT4_SB(sb
)->s_itb_per_group
) ||
4785 in_range(block
+ count
- 1, ext4_inode_table(sb
, gdp
),
4786 EXT4_SB(sb
)->s_itb_per_group
)) {
4788 ext4_error(sb
, __func__
,
4789 "Freeing blocks in system zone - "
4790 "Block = %lu, count = %lu", block
, count
);
4791 /* err = 0. ext4_std_error should be a no op */
4795 BUFFER_TRACE(bitmap_bh
, "getting write access");
4796 err
= ext4_journal_get_write_access(handle
, bitmap_bh
);
4801 * We are about to modify some metadata. Call the journal APIs
4802 * to unshare ->b_data if a currently-committing transaction is
4805 BUFFER_TRACE(gd_bh
, "get_write_access");
4806 err
= ext4_journal_get_write_access(handle
, gd_bh
);
4809 #ifdef AGGRESSIVE_CHECK
4812 for (i
= 0; i
< count
; i
++)
4813 BUG_ON(!mb_test_bit(bit
+ i
, bitmap_bh
->b_data
));
4817 ac
->ac_b_ex
.fe_group
= block_group
;
4818 ac
->ac_b_ex
.fe_start
= bit
;
4819 ac
->ac_b_ex
.fe_len
= count
;
4820 ext4_mb_store_history(ac
);
4823 err
= ext4_mb_load_buddy(sb
, block_group
, &e4b
);
4826 if (metadata
&& ext4_handle_valid(handle
)) {
4827 struct ext4_free_data
*new_entry
;
4829 * blocks being freed are metadata. these blocks shouldn't
4830 * be used until this transaction is committed
4832 new_entry
= kmem_cache_alloc(ext4_free_ext_cachep
, GFP_NOFS
);
4833 new_entry
->start_blk
= bit
;
4834 new_entry
->group
= block_group
;
4835 new_entry
->count
= count
;
4836 new_entry
->t_tid
= handle
->h_transaction
->t_tid
;
4837 ext4_lock_group(sb
, block_group
);
4838 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4840 ext4_mb_free_metadata(handle
, &e4b
, new_entry
);
4841 ext4_unlock_group(sb
, block_group
);
4843 ext4_lock_group(sb
, block_group
);
4844 /* need to update group_info->bb_free and bitmap
4845 * with group lock held. generate_buddy look at
4846 * them with group lock_held
4848 mb_clear_bits(sb_bgl_lock(sbi
, block_group
), bitmap_bh
->b_data
,
4850 mb_free_blocks(inode
, &e4b
, bit
, count
);
4851 ext4_mb_return_to_preallocation(inode
, &e4b
, block
, count
);
4852 ext4_unlock_group(sb
, block_group
);
4855 spin_lock(sb_bgl_lock(sbi
, block_group
));
4856 ret
= ext4_free_blks_count(sb
, gdp
) + count
;
4857 ext4_free_blks_set(sb
, gdp
, ret
);
4858 gdp
->bg_checksum
= ext4_group_desc_csum(sbi
, block_group
, gdp
);
4859 spin_unlock(sb_bgl_lock(sbi
, block_group
));
4860 percpu_counter_add(&sbi
->s_freeblocks_counter
, count
);
4862 if (sbi
->s_log_groups_per_flex
) {
4863 ext4_group_t flex_group
= ext4_flex_group(sbi
, block_group
);
4864 spin_lock(sb_bgl_lock(sbi
, flex_group
));
4865 sbi
->s_flex_groups
[flex_group
].free_blocks
+= count
;
4866 spin_unlock(sb_bgl_lock(sbi
, flex_group
));
4869 ext4_mb_release_desc(&e4b
);
4873 /* We dirtied the bitmap block */
4874 BUFFER_TRACE(bitmap_bh
, "dirtied bitmap block");
4875 err
= ext4_handle_dirty_metadata(handle
, NULL
, bitmap_bh
);
4877 /* And the group descriptor block */
4878 BUFFER_TRACE(gd_bh
, "dirtied group descriptor block");
4879 ret
= ext4_handle_dirty_metadata(handle
, NULL
, gd_bh
);
4883 if (overflow
&& !err
) {
4892 ext4_std_error(sb
, err
);
4894 kmem_cache_free(ext4_ac_cachep
, ac
);